LightArrayRevolverScheduler

2020/02/09 Akka源码 共 11082 字,约 32 分钟
梦境迷离
/*
 * Copyright (C) 2009-2019 Lightbend Inc. <https://www.lightbend.com>
 */

package akka.actor

import java.io.Closeable
import java.util.concurrent.ThreadFactory
import java.util.concurrent.atomic.{ AtomicLong, AtomicReference }
import scala.annotation.tailrec
import scala.collection.immutable
import scala.concurrent.{ Await, ExecutionContext, Future, Promise }
import scala.concurrent.duration._
import scala.util.control.NonFatal
import com.typesafe.config.Config
import akka.event.LoggingAdapter
import akka.util.Helpers
import akka.util.Unsafe.{ instance => unsafe }
import akka.dispatch.AbstractNodeQueue

//本类是Akka actor 2.5.23 默认定时器实现,感兴趣可以看看整理的一个PPT https://github.com/jxnu-liguobin/scala-examples/blob/master/scala-other/src/main/scala/io/github/dreamylost/timer/TimingWheels.ppt
//此调度程序的实现基于诸如Netty的HashedWheelTimer之类的由桶组成的旋转轮,它以固定的滴答速度前进,将在当前存储桶中找到的任务分派到各自的任务ExecutionContexts(线程池)
//任务保存在TaskHolders中,取消后,它们将取消对实际任务的引用,仅在下次轮子到达该存储桶时才清理该外壳
//这样就可以使用一个简单的链接列表来将TaskHolders链接到轮子
//值得注意的是,该调度程序在调度单发任务时不会获得当前时间戳,而是始终将任务延迟四舍五入为TickDuration的整数倍
//这意味着任务安排的时间可能比计划的要晚一滴答(轮子的一个桶的位置),如果 now() + delay <= nextTick,表示已经完成
class LightArrayRevolverScheduler(config: Config, log: LoggingAdapter, threadFactory: ThreadFactory)
    extends Scheduler
    with Closeable {

  import Helpers.Requiring
  import Helpers.ConfigOps

  val WheelSize =
    config
      .getInt("akka.scheduler.ticks-per-wheel")
      //默认512,必须是2的幂次
      //理解为轮子(类似时钟的刻度盘)上有512个刻度
      .requiring(ticks => (ticks & (ticks - 1)) == 0, "ticks-per-wheel must be a power of 2")
  val TickDuration =
    config
    //轮上每次移动1个刻度的延迟,这个值影响精度。windows上默认10ms,且最小是1ms
    //停止定时器可能需要1个滴答,这个值可能影响到关闭actor的时间
      .getMillisDuration("akka.scheduler.tick-duration")
      .requiring(
        _ >= 10.millis || !Helpers.isWindows,
        "minimum supported akka.scheduler.tick-duration on Windows is 10ms")
      .requiring(_ >= 1.millis, "minimum supported akka.scheduler.tick-duration is 1ms")
      //当关闭调度程序时,通常会有一个线程需要停止,而此超时决定了等待多长时间
      //在超时的情况下,actor系统的关闭将继续进行,而不运行可能仍在排队的任务
      //这个时间默认是5ms
  val ShutdownTimeout = config.getMillisDuration("akka.scheduler.shutdown-timeout")

  import LightArrayRevolverScheduler._
 
  //远离零值,向上舍入数字
  private def roundUp(d: FiniteDuration): FiniteDuration = {
    val dn = d.toNanos
    val r = ((dn - 1) / tickNanos + 1) * tickNanos
    if (r != dn && r > 0 && dn > 0) r.nanos else d
  }

  //当前纳米时间戳,时钟实现是可替换的(用于测试),其实现必须返回单调递增的纳秒序列
  protected def clock(): Long = System.nanoTime

  //可替换用于测试
  protected def startTick: Int = 0

  //可重写用于测试
  protected def getShutdownTimeout: FiniteDuration = ShutdownTimeout

  //可重写用于测试
  protected def waitNanos(nanos: Long): Unit = {
    // see http://www.javamex.com/tutorials/threads/sleep_issues.shtml 解释了这里为什么需要这么做
    val sleepMs = if (Helpers.isWindows) (nanos + 4999999) / 10000000 * 10 else (nanos + 999999) / 1000000
    try Thread.sleep(sleepMs)
    catch {
      case _: InterruptedException => Thread.currentThread().interrupt() // we got woken up
    }
  }

  //固定延迟的调度任务
  override def scheduleWithFixedDelay(initialDelay: FiniteDuration, delay: FiniteDuration)(runnable: Runnable)(
      implicit executor: ExecutionContext): Cancellable = {
    //验证延迟的最大值
    checkMaxDelay(roundUp(delay).toNanos)
    super.scheduleWithFixedDelay(initialDelay, delay)(runnable)
  }

  override def schedule(initialDelay: FiniteDuration, delay: FiniteDuration, runnable: Runnable)(
      implicit executor: ExecutionContext): Cancellable = {
    checkMaxDelay(roundUp(delay).toNanos)
    try new AtomicReference[Cancellable](InitialRepeatMarker) with Cancellable { self =>
    //原子性的初始化操作
      compareAndSet(
        InitialRepeatMarker,
        schedule(
          executor,
          //原子性的计算任务的开始时间
          new AtomicLong(clock() + initialDelay.toNanos) with Runnable {
            override def run(): Unit = {
              try {
                runnable.run()
                //原子性的给当前AtomicLong增加delay值,并计算与当前时间的间隔
                val driftNanos = clock() - getAndAdd(delay.toNanos)
                if (self.get != null)
                //还没结束,继续调度,更新延迟时间,本实现默认使用的是纳秒的时间戳,最小精度是1纳米
                  swap(schedule(executor, this, Duration.fromNanos(Math.max(delay.toNanos - driftNanos, 1))))
              } catch {
                case _: SchedulerException => //忽略排队失败或终止的目标actor
              }
            }
          },
          roundUp(initialDelay)))//远离0,向数字舍入,值可能会变大

      @tailrec private def swap(c: Cancellable): Unit = {
        get match {
            //更新,直到c都被关闭了
            //cancel方法会取消此可取消项,如果成功,则返回true。如果此可取消的已被(同时)取消,则此方法将返回false,尽管isCancelled将返回true。
          case null => if (c != null) c.cancel()
          //cas预期更新失败,交换c
          case old  => if (!compareAndSet(old, c)) swap(c)
        }
      }

      @tailrec final def cancel(): Boolean = {
        get match {
          case null => false
          case c =>
            //可取消的,使用cas更新
            if (c.cancel()) compareAndSet(c, null)
            else compareAndSet(c, null) || cancel()//不可取消的,尝试调用cancel取消
        }
      }

      override def isCancelled: Boolean = get == null//AtomicReference get为空时表示可以取消
    } catch {
      case cause @ SchedulerException(msg) => throw new IllegalStateException(msg, cause)
    }
  }

  //将Runnable延迟执行一次,属于单发任务,是直接调用内部方法schedule
  override def scheduleOnce(delay: FiniteDuration, runnable: Runnable)(
      implicit executor: ExecutionContext): Cancellable =
    try schedule(executor, runnable, roundUp(delay))
    catch {
      case cause @ SchedulerException(msg) => throw new IllegalStateException(msg, cause)
    }

  override def close(): Unit = {

    def runTask(task: Runnable): Unit = {
      try task.run()
      catch {
        case e: InterruptedException => throw e
        case _: SchedulerException   => //忽略终止的actor
        //执行定时器任务时发生非致命的异常
        case NonFatal(e)             => log.error(e, "exception while executing timer task")
      }
    }
    Await.result(stop(), getShutdownTimeout).foreach {
      case task: Scheduler.TaskRunOnClose =>
        runTask(task)
      case holder: TaskHolder => // don't run
        holder.task match {
          case task: Scheduler.TaskRunOnClose =>
            runTask(task)
          case _ => // don't run
        }
      case _ => // don't run
    }
  }

  //这个调度器的最大支持的任务频率,即在Hz中重复任务执行之间的最小时间间隔的倒数
  override val maxFrequency: Double = 1.second / TickDuration

  /*
   * 下面是实际的定时器实现
   */
  private val start = clock()
  private val tickNanos = TickDuration.toNanos
  private val wheelMask = WheelSize - 1
  private val queue = new TaskQueue

  private def schedule(ec: ExecutionContext, r: Runnable, delay: FiniteDuration): TimerTask =
    if (delay <= Duration.Zero) {
     //计时器关闭后无法排队
      if (stopped.get != null) throw SchedulerException("cannot enqueue after timer shutdown")
      ec.execute(r)
      NotCancellable
    } else if (stopped.get != null) {
        //计时器关闭后无法排队
      throw SchedulerException("cannot enqueue after timer shutdown")
    } else {
      val delayNanos = delay.toNanos
      checkMaxDelay(delayNanos)
      //计算这个任务需要的滴答数量(指针需要走多少步)
      val ticks = (delayNanos / tickNanos).toInt
      //任务入队
      val task = new TaskHolder(r, ticks, ec)
      queue.add(task)
      if (stopped.get != null && task.cancel())
      //计时器关闭后无法排队
        throw SchedulerException("cannot enqueue after timer shutdown")
      task
    }

  private def checkMaxDelay(delayNanos: Long): Unit =
    //延迟/滴答的值不能大于MaxValue
    if (delayNanos / tickNanos > Int.MaxValue)
      //由于舍入,错误消息中有1秒的误差
      throw new IllegalArgumentException(
        s"Task scheduled with [${delayNanos.nanos.toSeconds}] seconds delay, " +
        s"which is too far in future, maximum delay is [${(tickNanos * Int.MaxValue).nanos.toSeconds - 1}] seconds")

  private val stopped = new AtomicReference[Promise[immutable.Seq[TimerTask]]]
  //关闭,使用超时
  private def stop(): Future[immutable.Seq[TimerTask]] = {
    val p = Promise[immutable.Seq[TimerTask]]()
    //http://hg.openjdk.java.net/jdk7/jdk7/hotspot/file/9b0ca45cd756/src/share/vm/prims/unsafe.cpp
    if (stopped.compareAndSet(null, p)) {//期望为null,失败为false
    //中断定时器线程以使其更快地关闭是不好的,因为它可能正在执行计划的任务,这可能对被中断的响应不好。相反,我们只需再等待一个滴答就可以完成。
      p.future
    } else Future.successful(Nil)//Future.successful(Nil)
  }

  //定时任务的主线程
  @volatile private var timerThread: Thread = threadFactory.newThread(new Runnable {
    
    var tick = startTick
    var totalTick: Long = tick //不环绕(不会因为轮子满一周而被重置)的滴答数,用于计算睡眠时间
    val wheel = Array.fill(WheelSize)(new TaskQueue)//将轮子中的所有桶都初始化为TaskQueue队列

    private def clearAll(): immutable.Seq[TimerTask] = {
      @tailrec def collect(q: TaskQueue, acc: Vector[TimerTask]): Vector[TimerTask] = {
        q.poll() match {
          case null => acc
          case x    => collect(q, acc :+ x)
        }
      }
      (0 until WheelSize).flatMap(i => collect(wheel(i), Vector.empty)) ++ collect(queue, Vector.empty)
    }

    @tailrec
    private def checkQueue(time: Long): Unit = queue.pollNode() match {
      case null => ()
      case node =>
        node.value.ticks match {
          //到时间了,开始执行,value是TaskHolder类型的
          case 0 => node.value.executeTask()
          case ticks =>
            //ticks是延迟时间,除以每个滴答的延迟,得到总的步数&后得到桶中的位置,与总共开始的步数tick相减得到偏移步数
            val futureTick = ((
              time - start + //计算自定时器启动以来的纳秒数
              (ticks * tickNanos) + // 添加所需延迟
              tickNanos - 1 // 四舍五入
            ) / tickNanos).toInt //并转换为槽号(此处计算出的是步数,&wheelMask后得到是槽号或桶的位置)
            //桶中队列里面的任务的偏移量
            val offset = futureTick - tick
            //得到在桶上的刻度
            val bucket = futureTick & wheelMask
            //还有offset步才能开始执行
            node.value.ticks = offset
            //没有能执行,重新入队
            wheel(bucket).addNode(node)
        }
        checkQueue(time)
    }

    override final def run(): Unit =
      try nextTick()
      catch {
        case t: Throwable =>
          log.error(t, "exception on LARS’ timer thread")
          stopped.get match {
            case null =>
              val thread = threadFactory.newThread(this)
              log.info("starting new LARS thread")
              try thread.start()
              catch {
                case e: Throwable =>
                  log.error(e, "LARS cannot start new thread, ship’s going down!")
                  stopped.set(Promise.successful(Nil))
                  clearAll()
              }
              //当工作线程发生异常退出时重新根据当前任务创建新的工作线程
              timerThread = thread
            case p =>
              //不为空时表面已经stop过,填充promise success,"取消"所有任务
              //p为Promise.successful(Nil))时,比较成功,来自stop方法返回的promise
              assert(stopped.compareAndSet(p, Promise.successful(Nil)), "Stop signal violated in LARS")
              p.success(clearAll())
          }
          throw t
      }

    @tailrec final def nextTick(): Unit = {
      val time = clock()
      //开始时间+总的步数*每步延迟 - 实例化定时器时的时钟时间 = 需要睡眠的时间
      val sleepTime = start + (totalTick * tickNanos) - time
      if (sleepTime > 0) {
        //小睡前检查队列
        checkQueue(time)
        waitNanos(sleepTime)
      } else {
        //求得开始的刻度在轮子上桶中的位置
        val bucket = tick & wheelMask
        //获得本刻度上对应的任务队列
        val tasks = wheel(bucket)
        val putBack = new TaskQueue
        //执行轮子上的桶
        @tailrec def executeBucket(): Unit = tasks.pollNode() match {
          case null => ()
          case node =>
            val task = node.value
            if (!task.isCancelled) {
              if (task.ticks >= WheelSize) {
                task.ticks -= WheelSize//如果需要滴答的次数大于轮子的大小,说明还需要转圈,暂不执行该桶中的队列中的任务
                putBack.addNode(node)//需要重新入任务队列的任务
              } else task.executeTask()//执行队列中的任务
            }
            executeBucket()//递归调用自己判断桶上是否有任务达到执行时间
        }
        executeBucket()
        //更新本刻度对应的任务队列(队列中的TaskTimer会随着时间的流逝变少)
        wheel(bucket) = putBack
        //刻度加1(轮子上的指针走了一步)
        tick += 1
        totalTick += 1//总的刻度数加1
      }
      stopped.get match {
         //没有被关闭,指针继续走
        case null => nextTick()
        case p =>
        //已经关闭了,清空任务
        //p为Promise.successful(Nil))时,比较成功
          assert(stopped.compareAndSet(p, Promise.successful(Nil)), "Stop signal violated in LARS")
          p.success(clearAll())
      }
    }
  })
  //启动工作线程
  timerThread.start()
}

//内部api
object LightArrayRevolverScheduler {
 
  //使用TaskHolder包装了task
  private[this] val taskOffset = unsafe.objectFieldOffset(classOf[TaskHolder].getDeclaredField("task"))
  
  //基于Dmitriy Vyukov的非侵入式MPSC队列的无锁MPSC链接队列实现
  //http://www.1024cores.net/home/lock-free-algorithms/queues/non-intrusive-mpsc-node-based-queue
  //如果在队列不再很空但入队的元素尚不可见的情况下允许此队列返回null,则该队列可以是无等待的(即在peekNode和pollNode中没有旋转循环)。但是,这会破坏actor schedule
  private class TaskQueue extends AbstractNodeQueue[TaskHolder]

  protected[actor] trait TimerTask extends Runnable with Cancellable

  protected[actor] class TaskHolder(@volatile var task: Runnable, var ticks: Int, executionContext: ExecutionContext)
      extends TimerTask {
 
    //提取任务
    @tailrec
    private final def extractTask(replaceWith: Runnable): Runnable =
      task match {
        case t @ (ExecutedTask | CancelledTask) => t
        case x                                  => if (unsafe.compareAndSwapObject(this, taskOffset, x, replaceWith)) x else extractTask(replaceWith)
      }

    private[akka] final def executeTask(): Boolean = extractTask(ExecutedTask) match {
      case ExecutedTask | CancelledTask => false
      case other =>
        try {
          executionContext.execute(other)
          true
        } catch {
          case _: InterruptedException => Thread.currentThread().interrupt(); false
          case NonFatal(e)             => executionContext.reportFailure(e); false
        }
    }

    //仅应在直接执行中调用,入close方法调用时
    override def run(): Unit = extractTask(ExecutedTask).run()

    override def cancel(): Boolean = extractTask(CancelledTask) match {
      case ExecutedTask | CancelledTask => false
      case _                            => true
    }

    override def isCancelled: Boolean = task eq CancelledTask
  }

  private[this] val CancelledTask = new Runnable { def run = () }
  private[this] val ExecutedTask = new Runnable { def run = () }

  private val NotCancellable: TimerTask = new TimerTask {
    def cancel(): Boolean = false
    def isCancelled: Boolean = false
    def run(): Unit = ()
  }
 
  //初始重复标记
  private val InitialRepeatMarker: Cancellable = new Cancellable {
    def cancel(): Boolean = false
    def isCancelled: Boolean = false
  }
}

文档信息

Search

    Table of Contents