谈谈fork/join实现原理
害，又是一个炒冷饭的时间。fork/join是在jdk1.7中出现的一个并发工作包，其特点是可以将一个大的任务拆分成多个子任务进行并行处理，最后将子任务结果合并成最后的计算结果，并进行输出。从而达到多线程分发任务，达到高效处理的目的。
1. 关于fork/join的一点想法
以上说法，也许大家没什么感觉。但换个说法可能会更让人体会深切。总体上，相当于一个map阶段数据拆分，一个reduce阶段数据收集。即一个mapreduce过程，是不是有大数据的思想在了。只不过这fork/join的拆分难度可见性更大（自己手动拆，mapreduce由shuffle组件自动拆），另外fork/join是在一个机器上运行，而大数据的框架，则是在分布式系统中运行的。
从这个点说来，好像研究fork/join就显得有些意义了。
只是，按照fork/join的语义解释，是将任务拆分，然后处理，然后再合并结果。如果没有了合并结果这一步，那么，它就等同于线程池了，这也就是有人说它与线程池有啥差别的疑惑所在了。再说有需要收集结果的这一语义，其实我们也是可以通过线程池去执行任务，然后再用get()得到结果，然后在外部做合并，也是一样咯。
2. fork/join的几个核心类
fork/join被称作执行框架，自然不会是一个单一组件问题了。
首先，它会有一个 ForkJoinPool, 相当于线程池, 所有的任务都要通过它来进行提交，然后由其进行统一调度。
然后，每个任务都会有许多相同的代码，只有业务实现是不一样的，所以它会有一个基类： RecursiveTask . 实现上还有一个无返回结果的类：RecursiveAction, 只是没有返回结果时，往往又可能可以使用普通线程池执行替代了。（没有绝对）
ForkJoinWorkerThreadFactory, 是fork/join框架的线程工厂类，原本含义与普通的线程工厂类一致，只是它的入参不再是一个个 Runnable 任务，而是 ForkJoinPool, 因为它们所处的上下文是不一样的。
ForkJoinWorkerThread, 执行fork/join的具体线程，它可能在执行过程中，再去主动添加task。而它自身拥有一个队列，它的主要任务就是获取队列任务，然后执行。但当其自身的队列完成时，它可以通过work-steal算法窃取其他线程的队列任务。这也是fork/join的核心所在。
sun.misc.Unsafe, 之所以要提到这个jdk类，是因为在fork/join框架中，对于队列的管理，不是通过普通的list或数组来实现，而是通过 U.putOrderedObject(a, j, task); 来存放，虽然效果与数组是一样的，但它会更简单地实现线程安全的操作。只是，其中有许多的位操作，值得学习的同时，也显得有些麻烦了。
3. fork/join使用样例
我们通过对一个数组的排序过程，使用fork/join来实现看看如何使用这框架。尤其对于大数组的排序，显得还是有用的。这种大数组的排序，一般都会使用快速排序或者归并排序来处理。此处使用fork/join框架来处理，也是暗合了归并排序的道理了。
import java.util.Arrays; import java.util.Random; import java.util.concurrent.ExecutionException; import java.util.concurrent.ForkJoinPool; import java.util.concurrent.ForkJoinTask; import java.util.concurrent.RecursiveTask; /** * Fork/join框架测试 */ public class TestForkJoinFramework { public static void main(String[] args) { long beginTime = System.currentTimeMillis(); ForkJoinPool pool = new ForkJoinPool(); int mockArrLen = 1000_0000; int[] arr = new int[mockArrLen]; Random r = new Random(); for (int index = 1; index <= mockArrLen; index++) { arr[index - 1] = r.nextInt(1000_0000); } FJOrderTask task = new FJOrderTask(arr); ForkJoinTask<int[]> taskResult = pool.submit(task); try { // 等待结果完成 taskResult.get(); } catch (InterruptedException | ExecutionException e) { e.printStackTrace(); } long endTime = System.currentTimeMillis(); System.out.println("耗时=" + (endTime - beginTime)); } /** * 单个排序的子任务 */ private static class FJOrderTask extends RecursiveTask<int[]> { /** * 当前排序的数组值 */ private final int[] source; public FJOrderTask(int[] source) { this.source = source; } /** * 真正的业务计算逻辑 * * @see java.util.concurrent.RecursiveTask#compute() */ @Override protected int[] compute() { int sourceLen = source.length; // 如果条件成立，说明任务中要进行排序的集合还不够小 System.out.println(Thread.currentThread()); if (sourceLen > 2) { int midIndex = sourceLen / 2; // 拆分成两个子任务, 0 -> mid - 1, mid -> len FJOrderTask task1 = new FJOrderTask( Arrays.copyOf(source, midIndex)); task1.fork(); FJOrderTask task2 = new FJOrderTask( Arrays.copyOfRange(source, midIndex, sourceLen)); task2.fork(); // 将两个有序的数组，合并成一个有序的数组 int[] result1 = task1.join(); int[] result2 = task2.join(); return insertMerge(result1, result2); } // 否则说明集合中只有一个或者两个元素，可以进行这两个元素的比较排序了 else { // 如果条件成立，说明数组中只有一个元素，或者是数组中的元素都已经排列好位置了 if (sourceLen == 1 || source[0] <= source[1]) { return source; } else { int[] orderedArr = new int[sourceLen]; orderedArr[0] = source[1]; orderedArr[1] = source[0]; return orderedArr; } } } /** * 使用插入排序，将两个有序数组合并起来 * * @param arr1 有序数组1 * @param arr2 有序数组2 * @return 合并后的有序数组 */ private int[] insertMerge(int[] arr1, int[] arr2) { int[] result = new int[arr1.length + arr2.length]; int arr1Len = arr1.length; int arr2Len = arr2.length; int destLen = result.length; // 简单插入排序 for (int i = 0, array1Index = 0, array2Index = 0; i < destLen; i++) { int value1 = array1Index >= arr1Len ? Integer.MAX_VALUE : arr1[array1Index]; int value2 = array2Index >= arr2Len ? Integer.MAX_VALUE : arr2[array2Index]; if (value1 < value2) { array1Index++; result[i] = value1; } else { array2Index++; result[i] = value2; } } return result; } } }
思路很简单，就是将数组一直拆分，直到最后一个或者两个时，从最下面来开始排序，然后依次往上回溯，使用插入排序归并结果集，最终返回排好序的值。如果除去任务拆分的过程，则时间复杂度还是非常好的 O(nlog(n)), 只是这任务拆分的过程，需要大量的空间复杂度，也不见得是什么好事。且不管它。
4. fork/join框架的实现原理
我们以上面的demo为出发点，观察fork/join的工作过程，不知道100%，也八九不离十了。上面主要有几个动作，一ForkJoinPool实例化，submit一个Task, get()等待最终结果完成。这三个看得见的动作好办，只是其核心也许还在背后。
4.1. ForkJoinPool构造器
每个要调用框架的应用，必先初始化一个pool实例，这是自然。如上使用无参构造器，实际上是使用了框架的各种默认值而已, 这种默认值往往是能够满足大部分的场景的，从而体现其易用性。
// java.util.concurrent.ForkJoinPool#ForkJoinPool() /** * Creates a {@code ForkJoinPool} with parallelism equal to {@link * java.lang.Runtime#availableProcessors}, using the {@linkplain * #defaultForkJoinWorkerThreadFactory default thread factory}, * no UncaughtExceptionHandler, and non-async LIFO processing mode. * * @throws SecurityException if a security manager exists and * the caller is not permitted to modify threads * because it does not hold {@link * java.lang.RuntimePermission}{@code ("modifyThread")} */ public ForkJoinPool() { // 并行度默认是cpu的核数 this(Math.min(MAX_CAP, Runtime.getRuntime().availableProcessors()), defaultForkJoinWorkerThreadFactory, null, false); } /** * Creates a {@code ForkJoinPool} with the given parameters. * * @param parallelism the parallelism level. For default value, * use {@link java.lang.Runtime#availableProcessors}. * @param factory the factory for creating new threads. For default value, * use {@link #defaultForkJoinWorkerThreadFactory}. * @param handler the handler for internal worker threads that * terminate due to unrecoverable errors encountered while executing * tasks. For default value, use {@code null}. * @param asyncMode if true, * establishes local first-in-first-out scheduling mode for forked * tasks that are never joined. This mode may be more appropriate * than default locally stack-based mode in applications in which * worker threads only process event-style asynchronous tasks. * For default value, use {@code false}. * @throws IllegalArgumentException if parallelism less than or * equal to zero, or greater than implementation limit * @throws NullPointerException if the factory is null * @throws SecurityException if a security manager exists and * the caller is not permitted to modify threads * because it does not hold {@link * java.lang.RuntimePermission}{@code ("modifyThread")} */ public ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory, UncaughtExceptionHandler handler, boolean asyncMode) { this(checkParallelism(parallelism), checkFactory(factory), handler, // FIFO_QUEUE = 1 << 16, LIFO_QUEUE = 0 asyncMode ? FIFO_QUEUE : LIFO_QUEUE, "ForkJoinPool-" + nextPoolId() + "-worker-"); checkPermission(); } /** * Creates a {@code ForkJoinPool} with the given parameters, without * any security checks or parameter validation. Invoked directly by * makeCommonPool. */ private ForkJoinPool(int parallelism, ForkJoinWorkerThreadFactory factory, UncaughtExceptionHandler handler, int mode, String workerNamePrefix) { this.workerNamePrefix = workerNamePrefix; this.factory = factory; this.ueh = handler; this.config = (parallelism & SMASK) | mode; long np = (long)(-parallelism); // offset ctl counts this.ctl = ((np << AC_SHIFT) & AC_MASK) | ((np << TC_SHIFT) & TC_MASK); }
构造器自然没啥好说的，就是设置几个并行度，初始化线程工厂，标识等等。为下文做准备。
4.2. 任务submit过程
上面的例子中，submit只有一次调用，而实际应用中则不一定。但即使如此，一次submit, 其实背后也是有许多的动作的。因为这一个task里，又会生出许多task来。
// java.util.concurrent.ForkJoinPool#submit /** * Submits a ForkJoinTask for execution. * * @param task the task to submit * @param <T> the type of the task's result * @return the task * @throws NullPointerException if the task is null * @throws RejectedExecutionException if the task cannot be * scheduled for execution */ public <T> ForkJoinTask<T> submit(ForkJoinTask<T> task) { if (task == null) throw new NullPointerException(); // submit主要是向pool中加入任务队列 externalPush(task); return task; } /** * Tries to add the given task to a submission queue at * submitter's current queue. Only the (vastly) most common path * is directly handled in this method, while screening for need * for externalSubmit. * * @param task the task. Caller must ensure non-null. */ final void externalPush(ForkJoinTask<?> task) { WorkQueue[] ws; WorkQueue q; int m; int r = ThreadLocalRandom.getProbe(); int rs = runState; // 如果线程不是第一次进入，且获得锁，则直接放队列即可 // 否则走普通加入队列逻辑 if ((ws = workQueues) != null && (m = (ws.length - 1)) >= 0 && (q = ws[m & r & SQMASK]) != null && r != 0 && rs > 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { ForkJoinTask<?>[] a; int am, n, s; if ((a = q.array) != null && (am = a.length - 1) > (n = (s = q.top) - q.base)) { int j = ((am & s) << ASHIFT) + ABASE; // 通过 putOrderedObject 添加任务到队列中 U.putOrderedObject(a, j, task); U.putOrderedInt(q, QTOP, s + 1); U.putIntVolatile(q, QLOCK, 0); if (n <= 1) signalWork(ws, q); return; } U.compareAndSwapInt(q, QLOCK, 1, 0); } // 初始化时的submit或者通用 submit externalSubmit(task); } /** * Full version of externalPush, handling uncommon cases, as well * as performing secondary initialization upon the first * submission of the first task to the pool. It also detects * first submission by an external thread and creates a new shared * queue if the one at index if empty or contended. * * @param task the task. Caller must ensure non-null. */ private void externalSubmit(ForkJoinTask<?> task) { int r; // initialize caller's probe if ((r = ThreadLocalRandom.getProbe()) == 0) { ThreadLocalRandom.localInit(); r = ThreadLocalRandom.getProbe(); } for (;;) { WorkQueue[] ws; WorkQueue q; int rs, m, k; boolean move = false; // 停止运行 if ((rs = runState) < 0) { tryTerminate(false, false); // help terminate throw new RejectedExecutionException(); } // 未被初始化，先执行初始化 else if ((rs & STARTED) == 0 || // initialize ((ws = workQueues) == null || (m = ws.length - 1) < 0)) { int ns = 0; // 上锁初始化 rs = lockRunState(); try { if ((rs & STARTED) == 0) { U.compareAndSwapObject(this, STEALCOUNTER, null, new AtomicLong()); // create workQueues array with size a power of two int p = config & SMASK; // ensure at least 2 slots int n = (p > 1) ? p - 1 : 1; n |= n >>> 1; n |= n >>> 2; n |= n >>> 4; n |= n >>> 8; n |= n >>> 16; n = (n + 1) << 1; // 队列数量初始化 workQueues = new WorkQueue[n]; ns = STARTED; } } finally { unlockRunState(rs, (rs & ~RSLOCK) | ns); } } // 当前线程已添加过队列 else if ((q = ws[k = r & m & SQMASK]) != null) { // 上锁添加到队列中 if (q.qlock == 0 && U.compareAndSwapInt(q, QLOCK, 0, 1)) { ForkJoinTask<?>[] a = q.array; // 取出栈顶指针，向其中放入任务 int s = q.top; boolean submitted = false; // initial submission or resizing try { // locked version of push if ((a != null && a.length > s + 1 - q.base) || (a = q.growArray()) != null) { int j = (((a.length - 1) & s) << ASHIFT) + ABASE; U.putOrderedObject(a, j, task); U.putOrderedInt(q, QTOP, s + 1); submitted = true; } } finally { U.compareAndSwapInt(q, QLOCK, 1, 0); } // 如果队列添加成功，则唤醒一个 worker, 返回 // 否则进入下一次尝试添加过程 if (submitted) { signalWork(ws, q); return; } } move = true; // move on failure } else if (((rs = runState) & RSLOCK) == 0) { // create new queue q = new WorkQueue(this, null); q.hint = r; q.config = k | SHARED_QUEUE; q.scanState = INACTIVE; rs = lockRunState(); // publish index if (rs > 0 && (ws = workQueues) != null && k < ws.length && ws[k] == null) ws[k] = q; // else terminated unlockRunState(rs, rs & ~RSLOCK); } else move = true; // move if busy // 如有必要，为当前线程生成新的标识 if (move) r = ThreadLocalRandom.advanceProbe(r); } }
由上可知，submit主要初始化队列以及向队列中添加任务，并在唤醒worker处理任务。但实际上，worker Thread 我们还没有看到被激活，只是看到有队workQueue的初始化。那么，worker又是在哪进行初始化的呢？只可能是在 signal 的时候了。
4.3. worker的初始化
worker是真正执行任务的线程，前面光看到添加队列，以及唤醒worker了。只是这时还未见worker被初始化，实际上它是在被唤醒的逻辑中进行初始化的。
// java.util.concurrent.ForkJoinPool#signalWork /** * Tries to create or activate a worker if too few are active. * * @param ws the worker array to use to find signallees * @param q a WorkQueue --if non-null, don't retry if now empty */ final void signalWork(WorkQueue[] ws, WorkQueue q) { long c; int sp, i; WorkQueue v; Thread p; while ((c = ctl) < 0L) { // too few active，一个标识，分两段使用，低位为0代表worker还可以添加 if ((sp = (int)c) == 0) { // no idle workers if ((c & ADD_WORKER) != 0L) // too few workers tryAddWorker(c); break; } if (ws == null) // unstarted/terminated break; if (ws.length <= (i = sp & SMASK)) // terminated break; if ((v = ws[i]) == null) // terminating break; int vs = (sp + SS_SEQ) & ~INACTIVE; // next scanState int d = sp - v.scanState; // screen CAS long nc = (UC_MASK & (c + AC_UNIT)) | (SP_MASK & v.stackPred); if (d == 0 && U.compareAndSwapLong(this, CTL, c, nc)) { v.scanState = vs; // activate v if ((p = v.parker) != null) U.unpark(p); break; } if (q != null && q.base == q.top) // no more work break; } } /** * Tries to add one worker, incrementing ctl counts before doing * so, relying on createWorker to back out on failure. * * @param c incoming ctl value, with total count negative and no * idle workers. On CAS failure, c is refreshed and retried if * this holds (otherwise, a new worker is not needed). */ private void tryAddWorker(long c) { boolean add = false; do { long nc = ((AC_MASK & (c + AC_UNIT)) | (TC_MASK & (c + TC_UNIT))); if (ctl == c) { int rs, stop; // check if terminating if ((stop = (rs = lockRunState()) & STOP) == 0) add = U.compareAndSwapLong(this, CTL, c, nc); unlockRunState(rs, rs & ~RSLOCK); if (stop != 0) break; // 添加标识成功，再创建worker if (add) { createWorker(); break; } } } while (((c = ctl) & ADD_WORKER) != 0L && (int)c == 0); } /** * Tries to construct and start one worker. Assumes that total * count has already been incremented as a reservation. Invokes * deregisterWorker on any failure. * * @return true if successful */ private boolean createWorker() { ForkJoinWorkerThreadFactory fac = factory; Throwable ex = null; ForkJoinWorkerThread wt = null; try { // 调用线程工厂创建新的worker, 并立即启动worker if (fac != null && (wt = fac.newThread(this)) != null) { wt.start(); return true; } } catch (Throwable rex) { ex = rex; } // 创建失败，处理异常 deregisterWorker(wt, ex); return false; } /** * Default ForkJoinWorkerThreadFactory implementation; creates a * new ForkJoinWorkerThread. */ static final class DefaultForkJoinWorkerThreadFactory implements ForkJoinWorkerThreadFactory { public final ForkJoinWorkerThread newThread(ForkJoinPool pool) { return new ForkJoinWorkerThread(pool); } }
果然在signal时，创建worker。值得一提的，为了实现安全地添加worker，它会先更新成功ctl，然后再执行真正的create操作。避免多创建出worker来。
4.4. worker的工作原理
前面看到worker创建过程，传入了pool的实例，即当前上下文都是被worker可见的。所以，它能很好地复用当前的配置信息，而它自身是一个异步线程，在创建之后，立即被启动起来了。那它后续则必然尝试从队列获取任务，进行执行了。具体如何？
1. WorkerThread 构造方法
// java.util.concurrent.ForkJoinWorkerThread#ForkJoinWorkerThread /** * Creates a ForkJoinWorkerThread operating in the given pool. * * @param pool the pool this thread works in * @throws NullPointerException if pool is null */ protected ForkJoinWorkerThread(ForkJoinPool pool) { // Use a placeholder until a useful name can be set in registerWorker super("aForkJoinWorkerThread"); this.pool = pool; // workQueue 临时向 pool 中进行注册所得 this.workQueue = pool.registerWorker(this); } /** * Callback from ForkJoinWorkerThread constructor to establish and * record its WorkQueue. * * @param wt the worker thread * @return the worker's queue */ final WorkQueue registerWorker(ForkJoinWorkerThread wt) { UncaughtExceptionHandler handler; wt.setDaemon(true); // configure thread if ((handler = ueh) != null) wt.setUncaughtExceptionHandler(handler); WorkQueue w = new WorkQueue(this, wt); int i = 0; // assign a pool index int mode = config & MODE_MASK; int rs = lockRunState(); try { WorkQueue[] ws; int n; // skip if no array if ((ws = workQueues) != null && (n = ws.length) > 0) { int s = indexSeed += SEED_INCREMENT; // unlikely to collide int m = n - 1; i = ((s << 1) | 1) & m; // odd-numbered indices if (ws[i] != null) { // collision int probes = 0; // step by approx half n int step = (n <= 4) ? 2 : ((n >>> 1) & EVENMASK) + 2; while (ws[i = (i + step) & m] != null) { if (++probes >= n) { workQueues = ws = Arrays.copyOf(ws, n <<= 1); m = n - 1; probes = 0; } } } w.hint = s; // use as random seed w.config = i | mode; w.scanState = i; // publication fence ws[i] = w; } } finally { unlockRunState(rs, rs & ~RSLOCK); } wt.setName(workerNamePrefix.concat(Integer.toString(i >>> 1))); return w; }
重点则是在 pool 中注册自身，得到一个 workQueue. 而其具体业务，则是在run方法中实现。
// java.util.concurrent.ForkJoinWorkerThread#run /** * This method is required to be public, but should never be * called explicitly. It performs the main run loop to execute * {@link ForkJoinTask}s. */ public void run() { if (workQueue.array == null) { // only run once Throwable exception = null; try { onStart(); pool.runWorker(workQueue); } catch (Throwable ex) { exception = ex; } finally { try { onTermination(exception); } catch (Throwable ex) { if (exception == null) exception = ex; } finally { pool.deregisterWorker(this, exception); } } } } // java.util.concurrent.ForkJoinPool#runWorker /** * Top-level runloop for workers, called by ForkJoinWorkerThread.run. */ final void runWorker(WorkQueue w) { w.growArray(); // allocate queue int seed = w.hint; // initially holds randomization hint int r = (seed == 0) ? 1 : seed; // avoid 0 for xorShift for (ForkJoinTask<?> t;;) { // 取任务，执行 if ((t = scan(w, r)) != null) w.runTask(t); else if (!awaitWork(w, r)) break; r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // xorshift } } /** * Executes the given task and any remaining local tasks. */ final void runTask(ForkJoinTask<?> task) { if (task != null) { scanState &= ~SCANNING; // mark as busy (currentSteal = task).doExec(); U.putOrderedObject(this, QCURRENTSTEAL, null); // release for GC execLocalTasks(); ForkJoinWorkerThread thread = owner; if (++nsteals < 0) // collect on overflow transferStealCount(pool); scanState |= SCANNING; if (thread != null) thread.afterTopLevelExec(); } } // java.util.concurrent.ForkJoinTask#doExec /** * Primary execution method for stolen tasks. Unless done, calls * exec and records status if completed, but doesn't wait for * completion otherwise. * * @return status on exit from this method */ final int doExec() { int s; boolean completed; if ((s = status) >= 0) { try { completed = exec(); } catch (Throwable rex) { return setExceptionalCompletion(rex); } if (completed) s = setCompletion(NORMAL); } return s; } // java.util.concurrent.RecursiveTask#exec /** * Implements execution conventions for RecursiveTask. */ protected final boolean exec() { // 即调用具体业务类的 compute 方法 result = compute(); return true; }
咱们草草看了 worker 如何运行任务。这和线程池没多少差别，大致仍是从队列获取任务，然后执行业务方法compute . 我们暂时略去了如何获取任务，以及如何执行work-steal了。且看下节。
4.5. 任务获取实现
主要是通过scan处理。
// java.util.concurrent.ForkJoinPool#scan /** * Scans for and tries to steal a top-level task. Scans start at a * random location, randomly moving on apparent contention, * otherwise continuing linearly until reaching two consecutive * empty passes over all queues with the same checksum (summing * each base index of each queue, that moves on each steal), at * which point the worker tries to inactivate and then re-scans, * attempting to re-activate (itself or some other worker) if * finding a task; otherwise returning null to await work. Scans * otherwise touch as little memory as possible, to reduce * disruption on other scanning threads. * * @param w the worker (via its WorkQueue) * @param r a random seed * @return a task, or null if none found */ private ForkJoinTask<?> scan(WorkQueue w, int r) { WorkQueue[] ws; int m; if ((ws = workQueues) != null && (m = ws.length - 1) > 0 && w != null) { int ss = w.scanState; // initially non-negative for (int origin = r & m, k = origin, oldSum = 0, checkSum = 0;;) { WorkQueue q; ForkJoinTask<?>[] a; ForkJoinTask<?> t; int b, n; long c; // 首次获取时，是从自身队列中获取 if ((q = ws[k]) != null) { if ((n = (b = q.base) - q.top) < 0 && (a = q.array) != null) { // non-empty long i = (((a.length - 1) & b) << ASHIFT) + ABASE; if ((t = ((ForkJoinTask<?>) U.getObjectVolatile(a, i))) != null && q.base == b) { if (ss >= 0) { if (U.compareAndSwapObject(a, i, t, null)) { q.base = b + 1; if (n < -1) // signal others signalWork(ws, q); return t; } } else if (oldSum == 0 && // try to activate w.scanState < 0) tryRelease(c = ctl, ws[m & (int)c], AC_UNIT); } if (ss < 0) // refresh ss = w.scanState; r ^= r << 1; r ^= r >>> 3; r ^= r << 10; origin = k = r & m; // move and rescan oldSum = checkSum = 0; continue; } checkSum += b; } if ((k = (k + 1) & m) == origin) { // continue until stable if ((ss >= 0 || (ss == (ss = w.scanState))) && oldSum == (oldSum = checkSum)) { if (ss < 0 || w.qlock < 0) // already inactive break; int ns = ss | INACTIVE; // try to inactivate long nc = ((SP_MASK & ns) | (UC_MASK & ((c = ctl) - AC_UNIT))); w.stackPred = (int)c; // hold prev stack top U.putInt(w, QSCANSTATE, ns); if (U.compareAndSwapLong(this, CTL, c, nc)) ss = ns; else w.scanState = ss; // back out } checkSum = 0; } } } return null; }
要安全高效地实现一个获取队列还是不易啊。
4.6. task.fork 实现
一般地，能用上fork一词的场景，一般是对于当前环境的一个copy. 难道这里的fork也是这样吗？新开一个线程？不然又是如何找到需要处理的队列的呢？
// java.util.concurrent.ForkJoinTask#fork /** * Arranges to asynchronously execute this task in the pool the * current task is running in, if applicable, or using the {@link * ForkJoinPool#commonPool()} if not {@link #inForkJoinPool}. While * it is not necessarily enforced, it is a usage error to fork a * task more than once unless it has completed and been * reinitialized. Subsequent modifications to the state of this * task or any data it operates on are not necessarily * consistently observable by any thread other than the one * executing it unless preceded by a call to {@link #join} or * related methods, or a call to {@link #isDone} returning {@code * true}. * * @return {@code this}, to simplify usage */ public final ForkJoinTask<V> fork() { Thread t; // ForkJoinWorkerThread 中持有workQueue实例，可直接向其添加任务 if ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) ((ForkJoinWorkerThread)t).workQueue.push(this); else // 如果是外部线程，则添加到一共享pool中即可，后续将其各空闲线程处理 ForkJoinPool.common.externalPush(this); return this; } // java.util.concurrent.ForkJoinPool.WorkQueue#push /** * Pushes a task. Call only by owner in unshared queues. (The * shared-queue version is embedded in method externalPush.) * * @param task the task. Caller must ensure non-null. * @throws RejectedExecutionException if array cannot be resized */ final void push(ForkJoinTask<?> task) { ForkJoinTask<?>[] a; ForkJoinPool p; int b = base, s = top, n; if ((a = array) != null) { // ignore if queue removed int m = a.length - 1; // fenced write for task visibility U.putOrderedObject(a, ((m & s) << ASHIFT) + ABASE, task); U.putOrderedInt(this, QTOP, s + 1); if ((n = s - b) <= 1) { if ((p = pool) != null) p.signalWork(p.workQueues, this); } else if (n >= m) growArray(); } } /** * A thread managed by a {@link ForkJoinPool}, which executes * {@link ForkJoinTask}s. * This class is subclassable solely for the sake of adding * functionality -- there are no overridable methods dealing with * scheduling or execution. However, you can override initialization * and termination methods surrounding the main task processing loop. * If you do create such a subclass, you will also need to supply a * custom {@link ForkJoinPool.ForkJoinWorkerThreadFactory} to * {@linkplain ForkJoinPool#ForkJoinPool use it} in a {@code ForkJoinPool}. * * @since 1.7 * @author Doug Lea */ public class ForkJoinWorkerThread extends Thread { /* * ForkJoinWorkerThreads are managed by ForkJoinPools and perform * ForkJoinTasks. For explanation, see the internal documentation * of class ForkJoinPool. * * This class just maintains links to its pool and WorkQueue. The * pool field is set immediately upon construction, but the * workQueue field is not set until a call to registerWorker * completes. This leads to a visibility race, that is tolerated * by requiring that the workQueue field is only accessed by the * owning thread. * * Support for (non-public) subclass InnocuousForkJoinWorkerThread * requires that we break quite a lot of encapsulation (via Unsafe) * both here and in the subclass to access and set Thread fields. */ final ForkJoinPool pool; // the pool this thread works in final ForkJoinPool.WorkQueue workQueue; // work-stealing mechanics ... }
可见，fork的过程，即是向当前线程中添加当前任务而已，并没有所谓的上下文copy过程。
4.7. task.join 实现
join的语义是，等待任务完成后返回。与 Thread.join()一致。只是有一个问题，即如果某个线程阻塞等待结果去了，那当前线程自然就相当于无法再被利用了。那后续的任务又何从谈起呢？想来只有递归能够解决这个问题了。但是递归往往又是在单线程中完成的，这岂不无法利用并发特性了？
实际上，之所以被分作fork/join两个步骤，意义就是在这。上一节我们看到，fork的过程是向队列中添加了任务，随后就返回了。这时，如果当前worker比较繁忙（在做任务拆分），则这些任务就会被其他worker窃取过去处理了。而其他任务在处理时，又会遇到自己的递归，从而将一个单线程的递归变为多线程的递归了。
下面我们主要看一个线程的递归过程。join的本义只是等待当前任务完成，但是当前任务完成又要依赖于其子任务完成join, 子任务又要等待其子任务join, 因此形成递归。而join()返回的表象是compute()完成，所以这过程其实是伴随着compute的运算的。
// java.util.concurrent.ForkJoinTask#join /** * Returns the result of the computation when it {@link #isDone is * done}. This method differs from {@link #get()} in that * abnormal completion results in {@code RuntimeException} or * {@code Error}, not {@code ExecutionException}, and that * interrupts of the calling thread do <em>not</em> cause the * method to abruptly return by throwing {@code * InterruptedException}. * * @return the computed result */ public final V join() { int s; if ((s = doJoin() & DONE_MASK) != NORMAL) reportException(s); // 任务完成后，主动获取结果 return getRawResult(); } /** * Throws exception, if any, associated with the given status. */ private void reportException(int s) { if (s == CANCELLED) throw new CancellationException(); if (s == EXCEPTIONAL) rethrow(getThrowableException()); } // java.util.concurrent.RecursiveTask#getRawResult public final V getRawResult() { return result; } /** * Implementation for join, get, quietlyJoin. Directly handles * only cases of already-completed, external wait, and * unfork+exec. Others are relayed to ForkJoinPool.awaitJoin. * * @return status upon completion */ private int doJoin() { int s; Thread t; ForkJoinWorkerThread wt; ForkJoinPool.WorkQueue w; return (s = status) < 0 ? s : ((t = Thread.currentThread()) instanceof ForkJoinWorkerThread) ? // 取当前任务执行， doExec 执行任务，awaitJoin 等待执行完成 (w = (wt = (ForkJoinWorkerThread)t).workQueue). tryUnpush(this) && (s = doExec()) < 0 ? s : wt.pool.awaitJoin(w, this, 0L) : externalAwaitDone(); } // java.util.concurrent.ForkJoinPool#awaitJoin /** * Helps and/or blocks until the given task is done or timeout. * * @param w caller * @param task the task * @param deadline for timed waits, if nonzero * @return task status on exit */ final int awaitJoin(WorkQueue w, ForkJoinTask<?> task, long deadline) { int s = 0; if (task != null && w != null) { ForkJoinTask<?> prevJoin = w.currentJoin; U.putOrderedObject(w, QCURRENTJOIN, task); CountedCompleter<?> cc = (task instanceof CountedCompleter) ? (CountedCompleter<?>)task : null; for (;;) { if ((s = task.status) < 0) break; if (cc != null) helpComplete(w, cc, 0); // 递归添加任务等待完成 else if (w.base == w.top || w.tryRemoveAndExec(task)) helpStealer(w, task); if ((s = task.status) < 0) break; long ms, ns; if (deadline == 0L) ms = 0L; else if ((ns = deadline - System.nanoTime()) <= 0L) break; else if ((ms = TimeUnit.NANOSECONDS.toMillis(ns)) <= 0L) ms = 1L; if (tryCompensate(w)) { task.internalWait(ms); U.getAndAddLong(this, CTL, AC_UNIT); } } U.putOrderedObject(w, QCURRENTJOIN, prevJoin); } return s; } // java.util.concurrent.ForkJoinPool.WorkQueue#tryRemoveAndExec /** * If present, removes from queue and executes the given task, * or any other cancelled task. Used only by awaitJoin. * * @return true if queue empty and task not known to be done */ final boolean tryRemoveAndExec(ForkJoinTask<?> task) { ForkJoinTask<?>[] a; int m, s, b, n; if ((a = array) != null && (m = a.length - 1) >= 0 && task != null) { while ((n = (s = top) - (b = base)) > 0) { for (ForkJoinTask<?> t;;) { // traverse from s to b long j = ((--s & m) << ASHIFT) + ABASE; if ((t = (ForkJoinTask<?>)U.getObject(a, j)) == null) return s + 1 == top; // shorter than expected else if (t == task) { boolean removed = false; if (s + 1 == top) { // pop if (U.compareAndSwapObject(a, j, task, null)) { U.putOrderedInt(this, QTOP, s); removed = true; } } else if (base == b) // replace with proxy removed = U.compareAndSwapObject( a, j, task, new EmptyTask()); // 执行子任务 if (removed) task.doExec(); break; } else if (t.status < 0 && s + 1 == top) { if (U.compareAndSwapObject(a, j, t, null)) U.putOrderedInt(this, QTOP, s); break; // was cancelled } if (--n == 0) return false; } if (task.status < 0) return false; } } return true; }
可见，最终fork/join还是使用递归完成join任务等待。差别在于其利用了多线程的优势，同时执行多个任务。这有两个好处，一是减轻了单线程的任务处理压力，二是让递归的深度也分担到了多个点上。避免了栈早早溢出的可能。
只是每个线程被分配的任务数是多少，join需要等待的结果有多少，就不太好说了。比如最上层的线程如果任务被别的线程抢走，则它就只需一直在等结果就行了。而最下面的线程，则需要承担最深的递归深度，以保证程序的最终出口。其实从这个点，我们自己可以做个猜想，如果没有做好控制，让线程之间任意执行任务，是否会造成死锁呢？这恐怕是个问题。
不要害怕今日的苦，你要相信明天，更苦！
