序
本文主要研究一下PowerJob的HashedWheelTimer
Timer
tech/powerjob/server/common/timewheel/Timer.java
public interface Timer {
/**
* 调度定时任务
*/
TimerFuture schedule(TimerTask task, long delay, TimeUnit unit);
/**
* 停止所有调度任务
*/
Set<TimerTask> stop();
}Timer接口定义了schedule方法，用于在指定时间之后调度TimerTask，它返回TimerFuture；stop方法返回未处理的TimerTask
TimerTask
@FunctionalInterface
public interface TimerTask extends Runnable {
}TimerTask继承了Runnable
TimerFuture
tech/powerjob/server/common/timewheel/TimerFuture.java
public interface TimerFuture {
TimerTask getTask();
boolean cancel();
boolean isCancelled();
boolean isDone();
} TimerFuture定于了getTask、cancel、isCancelled、isDone方法
HashedWheelTimer
tech/powerjob/server/common/timewheel/HashedWheelTimer.java
@Slf4j
public class HashedWheelTimer implements Timer {
private final long tickDuration;
private final HashedWheelBucket[] wheel;
private final int mask;
private final Indicator indicator;
private final long startTime;
private final Queue<HashedWheelTimerFuture> waitingTasks = Queues.newLinkedBlockingQueue();
private final Queue<HashedWheelTimerFuture> canceledTasks = Queues.newLinkedBlockingQueue();
private final ExecutorService taskProcessPool;
public HashedWheelTimer(long tickDuration, int ticksPerWheel) {
this(tickDuration, ticksPerWheel, 0);
}
/**
* 新建时间轮定时器
* @param tickDuration 时间间隔，单位毫秒（ms）
* @param ticksPerWheel 轮盘个数
* @param processThreadNum 处理任务的线程个数，0代表不启用新线程（如果定时任务需要耗时操作，请启用线程池）
*/
public HashedWheelTimer(long tickDuration, int ticksPerWheel, int processThreadNum) {
this.tickDuration = tickDuration;
// 初始化轮盘，大小格式化为2的N次，可以使用 & 代替取余
int ticksNum = CommonUtils.formatSize(ticksPerWheel);
wheel = new HashedWheelBucket[ticksNum];
for (int i = 0; i < ticksNum; i++) {
wheel[i] = new HashedWheelBucket();
}
mask = wheel.length - 1;
// 初始化执行线程池
if (processThreadNum <= 0) {
taskProcessPool = null;
}else {
ThreadFactory threadFactory = new ThreadFactoryBuilder().setNameFormat("HashedWheelTimer-Executor-%d").build();
// 这里需要调整一下队列大小
BlockingQueue<Runnable> queue = Queues.newLinkedBlockingQueue(8192);
int core = Math.max(Runtime.getRuntime().availableProcessors(), processThreadNum);
// 基本都是 io 密集型任务
taskProcessPool = new ThreadPoolExecutor(core, 2 * core,
60, TimeUnit.SECONDS,
queue, threadFactory, RejectedExecutionHandlerFactory.newCallerRun("PowerJobTimeWheelPool"));
}
startTime = System.currentTimeMillis();
// 启动后台线程
indicator = new Indicator();
new Thread(indicator, "HashedWheelTimer-Indicator").start();
}
//......
} HashedWheelTimer实现了Timer接口，其构造器要求输入tickDuration、ticksPerWheel，它会将ticksPerWheel转换为2的N次，然后创建对应的HashedWheelBucket，若processThreadNum大于0则同时创建ThreadPoolExecutor用于处理任务，最后启动异步线程执行Indicator
schedule
@Override
public TimerFuture schedule(TimerTask task, long delay, TimeUnit unit) {
long targetTime = System.currentTimeMillis() + unit.toMillis(delay);
HashedWheelTimerFuture timerFuture = new HashedWheelTimerFuture(task, targetTime);
// 直接运行到期、过期任务
if (delay <= 0) {
runTask(timerFuture);
return timerFuture;
}
// 写入阻塞队列，保证并发安全（性能进一步优化可以考虑 Netty 的 Multi-Producer-Single-Consumer队列）
waitingTasks.add(timerFuture);
return timerFuture;
}schedule方法先计算目标时间，然后创建对应的HashedWheelTimerFuture，若delay小于等于0则执行runTask，否则添加到waitingTasks
stop
@Override
public Set<TimerTask> stop() {
indicator.stop.set(true);
taskProcessPool.shutdown();
while (!taskProcessPool.isTerminated()) {
try {
Thread.sleep(100);
}catch (Exception ignore) {
}
}
return indicator.getUnprocessedTasks();
}stop方法先设置indicator的stop为true，然后执行taskProcessPool.shutdown()，等待关闭，最后返回indicator.getUnprocessedTasks()
HashedWheelBucket
private final class HashedWheelBucket extends LinkedList<HashedWheelTimerFuture> {
public void expireTimerTasks(long currentTick) {
removeIf(timerFuture -> {
// processCanceledTasks 后外部操作取消任务会导致 BUCKET 中仍存在 CANCELED 任务的情况
if (timerFuture.status == HashedWheelTimerFuture.CANCELED) {
return true;
}
if (timerFuture.status != HashedWheelTimerFuture.WAITING) {
log.warn("[HashedWheelTimer] impossible, please fix the bug");
return true;
}
// 本轮直接调度
if (timerFuture.totalTicks <= currentTick) {
if (timerFuture.totalTicks < currentTick) {
log.warn("[HashedWheelTimer] timerFuture.totalTicks < currentTick, please fix the bug");
}
try {
// 提交执行
runTask(timerFuture);
}catch (Exception ignore) {
} finally {
timerFuture.status = HashedWheelTimerFuture.FINISHED;
}
return true;
}
return false;
});
}
}
private void runTask(HashedWheelTimerFuture timerFuture) {
timerFuture.status = HashedWheelTimerFuture.RUNNING;
if (taskProcessPool == null) {
timerFuture.timerTask.run();
}else {
taskProcessPool.submit(timerFuture.timerTask);
}
} HashedWheelBucket继承了LinkedList，其泛型为HashedWheelTimerFuture，它提供了expireTimerTasks方法，通过removeIf删除status为CANCELED、status不为WAITING，以及执行runTask(注意这里忽略了异常)之后标记status为FINISHED的元素；runTask先标记为RUNNING，对于taskProcessPool为null则直接执行，否则提交到taskProcessPoolHashedWheelTimerFuture
tech/powerjob/server/common/timewheel/HashedWheelTimer.java
private final class HashedWheelTimerFuture implements TimerFuture {
// 预期执行时间
private final long targetTime;
private final TimerTask timerTask;
// 所属的时间格，用于快速删除该任务
private HashedWheelBucket bucket;
// 总圈数
private long totalTicks;
// 当前状态 0 - 初始化等待中，1 - 运行中，2 - 完成，3 - 已取消
private int status;
// 状态枚举值
private static final int WAITING = 0;
private static final int RUNNING = 1;
private static final int FINISHED = 2;
private static final int CANCELED = 3;
public HashedWheelTimerFuture(TimerTask timerTask, long targetTime) {
this.targetTime = targetTime;
this.timerTask = timerTask;
this.status = WAITING;
}
@Override
public TimerTask getTask() {
return timerTask;
}
@Override
public boolean cancel() {
if (status == WAITING) {
status = CANCELED;
canceledTasks.add(this);
return true;
}
return false;
}
@Override
public boolean isCancelled() {
return status == CANCELED;
}
@Override
public boolean isDone() {
return status == FINISHED;
}
}HashedWheelTimerFuture实现了TimerFuture接口，它定义了WAITING、RUNNING、FINISHED、CANCELED状态；初始状态为WAITING，对于WAITING状态的可以设置为CANCELED，并添加到canceledTasks；isCancelled判断状态是不是CANCELED，isDone判断状态是不是FINISHED
getUnprocessedTasks
public Set<TimerTask> getUnprocessedTasks() {
try {
latch.await();
}catch (Exception ignore) {
}
Set<TimerTask> tasks = Sets.newHashSet();
Consumer<HashedWheelTimerFuture> consumer = timerFuture -> {
if (timerFuture.status == HashedWheelTimerFuture.WAITING) {
tasks.add(timerFuture.timerTask);
}
};
waitingTasks.forEach(consumer);
for (HashedWheelBucket bucket : wheel) {
bucket.forEach(consumer);
}
return tasks;
}getUnprocessedTasks会等待Indicator的while循环结束，然后遍历所有的HashedWheelBucket找出状态还是WAITING的任务
Indicator
private class Indicator implements Runnable {
private long tick = 0;
private final AtomicBoolean stop = new AtomicBoolean(false);
private final CountDownLatch latch = new CountDownLatch(1);
@Override
public void run() {
while (!stop.get()) {
// 1. 将任务从队列推入时间轮
pushTaskToBucket();
// 2. 处理取消的任务
processCanceledTasks();
// 3. 等待指针跳向下一刻
tickTack();
// 4. 执行定时任务
int currentIndex = (int) (tick & mask);
HashedWheelBucket bucket = wheel[currentIndex];
bucket.expireTimerTasks(tick);
tick ++;
}
latch.countDown();
}
/**
* 模拟指针转动，当返回时指针已经转到了下一个刻度
*/
private void tickTack() {
// 下一次调度的绝对时间
long nextTime = startTime + (tick + 1) * tickDuration;
long sleepTime = nextTime - System.currentTimeMillis();
if (sleepTime > 0) {
try {
Thread.sleep(sleepTime);
}catch (Exception ignore) {
}
}
}
/**
* 处理被取消的任务
*/
private void processCanceledTasks() {
while (true) {
HashedWheelTimerFuture canceledTask = canceledTasks.poll();
if (canceledTask == null) {
return;
}
// 从链表中删除该任务（bucket为null说明还没被正式推入时间格中，不需要处理）
if (canceledTask.bucket != null) {
canceledTask.bucket.remove(canceledTask);
}
}
}
/**
* 将队列中的任务推入时间轮中
*/
private void pushTaskToBucket() {
while (true) {
HashedWheelTimerFuture timerTask = waitingTasks.poll();
if (timerTask == null) {
return;
}
// 总共的偏移量
long offset = timerTask.targetTime - startTime;
// 总共需要走的指针步数
timerTask.totalTicks = offset / tickDuration;
// 取余计算 bucket index
int index = (int) (timerTask.totalTicks & mask);
HashedWheelBucket bucket = wheel[index];
// TimerTask 维护 Bucket 引用，用于删除该任务
timerTask.bucket = bucket;
if (timerTask.status == HashedWheelTimerFuture.WAITING) {
bucket.add(timerTask);
}
}
}
public Set<TimerTask> getUnprocessedTasks() {
try {
latch.await();
}catch (Exception ignore) {
}
Set<TimerTask> tasks = Sets.newHashSet();
Consumer<HashedWheelTimerFuture> consumer = timerFuture -> {
if (timerFuture.status == HashedWheelTimerFuture.WAITING) {
tasks.add(timerFuture.timerTask);
}
};
waitingTasks.forEach(consumer);
for (HashedWheelBucket bucket : wheel) {
bucket.forEach(consumer);
}
return tasks;
}
}Indicator实现了Runnable接口，其run方法在stop为false的时候循环执行，pushTaskToBucket、processCanceledTasks、tickTack、expireTimerTasks
pushTaskToBucket
private void pushTaskToBucket() {
while (true) {
HashedWheelTimerFuture timerTask = waitingTasks.poll();
if (timerTask == null) {
return;
}
// 总共的偏移量
long offset = timerTask.targetTime - startTime;
// 总共需要走的指针步数
timerTask.totalTicks = offset / tickDuration;
// 取余计算 bucket index
int index = (int) (timerTask.totalTicks & mask);
HashedWheelBucket bucket = wheel[index];
// TimerTask 维护 Bucket 引用，用于删除该任务
timerTask.bucket = bucket;
if (timerTask.status == HashedWheelTimerFuture.WAITING) {
bucket.add(timerTask);
}
}
}pushTaskToBucket通过waitingTasks.poll()拉取任务，若为null直接返回，否则通过timerTask.targetTime与startTime计算offset，再根据tickDuration计算需要走的步数，然后计算并获取目标HashedWheelBucket，然后将timerTask添加到bucket中
processCanceledTasks
private void processCanceledTasks() {
while (true) {
HashedWheelTimerFuture canceledTask = canceledTasks.poll();
if (canceledTask == null) {
return;
}
// 从链表中删除该任务（bucket为null说明还没被正式推入时间格中，不需要处理）
if (canceledTask.bucket != null) {
canceledTask.bucket.remove(canceledTask);
}
}
}processCanceledTasks会执行canceledTasks.poll()拉取canceledTask，若canceledTask.bucket不为null则将canceledTask从该bucket中移除
tickTack
private void tickTack() {
// 下一次调度的绝对时间
long nextTime = startTime + (tick + 1) * tickDuration;
long sleepTime = nextTime - System.currentTimeMillis();
if (sleepTime > 0) {
try {
Thread.sleep(sleepTime);
}catch (Exception ignore) {
}
}
}tickTack模拟指针移动，它先计算nextTime，再计算需要sleep多久，然后执行Thread.sleep(sleepTime)
小结
PowerJob定义了Timer接口，并提供了HashedWheelTimer的实现，它定义了waitingTasks、canceledTasks两个LinkedBlockingQueue(无界队列)，同时还支持定义任务处理线程池的core线程数；它通过Indicator线程来处理时间轮的转动及任务处理，Indicator循环将waitingTasks的任务放入到对应的bucket，然后模拟时间轮等待，然后通过bucket.expireTimerTasks(tick)处理到期任务，最后再递增tick。
**粗体** _斜体_ [链接](http://example.com) `代码` - 列表 > 引用。你还可以使用@来通知其他用户。