ViewVC Help
View File | Revision Log | Show Annotations | Download File | Root Listing
root/jsr166/jsr166/src/jsr166y/ForkJoinWorkerThread.java
(Generate patch)

Comparing jsr166/src/jsr166y/ForkJoinWorkerThread.java (file contents):
Revision 1.39 by dl, Sat Jul 24 20:28:18 2010 UTC vs.
Revision 1.72 by dl, Mon Nov 19 18:12:42 2012 UTC

# Line 1 | Line 1
1   /*
2   * Written by Doug Lea with assistance from members of JCP JSR-166
3   * Expert Group and released to the public domain, as explained at
4 < * http://creativecommons.org/licenses/publicdomain
4 > * http://creativecommons.org/publicdomain/zero/1.0/
5   */
6  
7   package jsr166y;
8  
9 < import java.util.concurrent.*;
10 <
11 < import java.util.Random;
12 < import java.util.Collection;
13 < import java.util.concurrent.locks.LockSupport;
9 > import java.util.concurrent.atomic.AtomicInteger;
10  
11   /**
12 < * A thread managed by a {@link ForkJoinPool}.  This class is
13 < * subclassable solely for the sake of adding functionality -- there
14 < * are no overridable methods dealing with scheduling or execution.
15 < * However, you can override initialization and termination methods
16 < * surrounding the main task processing loop.  If you do create such a
17 < * subclass, you will also need to supply a custom {@link
18 < * ForkJoinPool.ForkJoinWorkerThreadFactory} to use it in a {@code
19 < * ForkJoinPool}.
12 > * A thread managed by a {@link ForkJoinPool}, which executes
13 > * {@link ForkJoinTask}s.
14 > * This class is subclassable solely for the sake of adding
15 > * functionality -- there are no overridable methods dealing with
16 > * scheduling or execution.  However, you can override initialization
17 > * and termination methods surrounding the main task processing loop.
18 > * If you do create such a subclass, you will also need to supply a
19 > * custom {@link ForkJoinPool.ForkJoinWorkerThreadFactory} to use it
20 > * in a {@code ForkJoinPool}.
21   *
22   * @since 1.7
23   * @author Doug Lea
24   */
25   public class ForkJoinWorkerThread extends Thread {
26      /*
30     * Overview:
31     *
27       * ForkJoinWorkerThreads are managed by ForkJoinPools and perform
28 <     * ForkJoinTasks. This class includes bookkeeping in support of
29 <     * worker activation, suspension, and lifecycle control described
35 <     * in more detail in the internal documentation of class
36 <     * ForkJoinPool. And as described further below, this class also
37 <     * includes special-cased support for some ForkJoinTask
38 <     * methods. But the main mechanics involve work-stealing:
39 <     *
40 <     * Work-stealing queues are special forms of Deques that support
41 <     * only three of the four possible end-operations -- push, pop,
42 <     * and deq (aka steal), under the further constraints that push
43 <     * and pop are called only from the owning thread, while deq may
44 <     * be called from other threads.  (If you are unfamiliar with
45 <     * them, you probably want to read Herlihy and Shavit's book "The
46 <     * Art of Multiprocessor programming", chapter 16 describing these
47 <     * in more detail before proceeding.)  The main work-stealing
48 <     * queue design is roughly similar to those in the papers "Dynamic
49 <     * Circular Work-Stealing Deque" by Chase and Lev, SPAA 2005
50 <     * (http://research.sun.com/scalable/pubs/index.html) and
51 <     * "Idempotent work stealing" by Michael, Saraswat, and Vechev,
52 <     * PPoPP 2009 (http://portal.acm.org/citation.cfm?id=1504186).
53 <     * The main differences ultimately stem from gc requirements that
54 <     * we null out taken slots as soon as we can, to maintain as small
55 <     * a footprint as possible even in programs generating huge
56 <     * numbers of tasks. To accomplish this, we shift the CAS
57 <     * arbitrating pop vs deq (steal) from being on the indices
58 <     * ("base" and "sp") to the slots themselves (mainly via method
59 <     * "casSlotNull()"). So, both a successful pop and deq mainly
60 <     * entail a CAS of a slot from non-null to null.  Because we rely
61 <     * on CASes of references, we do not need tag bits on base or sp.
62 <     * They are simple ints as used in any circular array-based queue
63 <     * (see for example ArrayDeque).  Updates to the indices must
64 <     * still be ordered in a way that guarantees that sp == base means
65 <     * the queue is empty, but otherwise may err on the side of
66 <     * possibly making the queue appear nonempty when a push, pop, or
67 <     * deq have not fully committed. Note that this means that the deq
68 <     * operation, considered individually, is not wait-free. One thief
69 <     * cannot successfully continue until another in-progress one (or,
70 <     * if previously empty, a push) completes.  However, in the
71 <     * aggregate, we ensure at least probabilistic non-blockingness.
72 <     * If an attempted steal fails, a thief always chooses a different
73 <     * random victim target to try next. So, in order for one thief to
74 <     * progress, it suffices for any in-progress deq or new push on
75 <     * any empty queue to complete. One reason this works well here is
76 <     * that apparently-nonempty often means soon-to-be-stealable,
77 <     * which gives threads a chance to set activation status if
78 <     * necessary before stealing.
79 <     *
80 <     * This approach also enables support for "async mode" where local
81 <     * task processing is in FIFO, not LIFO order; simply by using a
82 <     * version of deq rather than pop when locallyFifo is true (as set
83 <     * by the ForkJoinPool).  This allows use in message-passing
84 <     * frameworks in which tasks are never joined.
85 <     *
86 <     * When a worker would otherwise be blocked waiting to join a
87 <     * task, it first tries a form of linear helping: Each worker
88 <     * records (in field currentSteal) the most recent task it stole
89 <     * from some other worker. Plus, it records (in field currentJoin)
90 <     * the task it is currently actively joining. Method joinTask uses
91 <     * these markers to try to find a worker to help (i.e., steal back
92 <     * a task from and execute it) that could hasten completion of the
93 <     * actively joined task. In essence, the joiner executes a task
94 <     * that would be on its own local deque had the to-be-joined task
95 <     * not been stolen. This may be seen as a conservative variant of
96 <     * the approach in Wagner & Calder "Leapfrogging: a portable
97 <     * technique for implementing efficient futures" SIGPLAN Notices,
98 <     * 1993 (http://portal.acm.org/citation.cfm?id=155354). It differs
99 <     * in that: (1) We only maintain dependency links across workers
100 <     * upon steals, rather than use per-task bookkeeping.  This may
101 <     * require a linear scan of workers array to locate stealers, but
102 <     * usually doesn't because stealers leave hints (that may become
103 <     * stale/wrong) of where to locate them. This isolates cost to
104 <     * when it is needed, rather than adding to per-task overhead.
105 <     * (2) It is "shallow", ignoring nesting and potentially cyclic
106 <     * mutual steals.  (3) It is intentionally racy: field currentJoin
107 <     * is updated only while actively joining, which means that we
108 <     * miss links in the chain during long-lived tasks, GC stalls etc
109 <     * (which is OK since blocking in such cases is usually a good
110 <     * idea).  (4) We bound the number of attempts to find work (see
111 <     * MAX_HELP_DEPTH) and fall back to suspending the worker and if
112 <     * necessary replacing it with a spare (see
113 <     * ForkJoinPool.tryAwaitJoin).
28 >     * ForkJoinTasks. For explanation, see the internal documentation
29 >     * of class ForkJoinPool.
30       *
31 <     * Efficient implementation of these algorithms currently relies
32 <     * on an uncomfortable amount of "Unsafe" mechanics. To maintain
33 <     * correct orderings, reads and writes of variable base require
34 <     * volatile ordering.  Variable sp does not require volatile
35 <     * writes but still needs store-ordering, which we accomplish by
36 <     * pre-incrementing sp before filling the slot with an ordered
37 <     * store.  (Pre-incrementing also enables backouts used in
122 <     * joinTask.)  Because they are protected by volatile base reads,
123 <     * reads of the queue array and its slots by other threads do not
124 <     * need volatile load semantics, but writes (in push) require
125 <     * store order and CASes (in pop and deq) require (volatile) CAS
126 <     * semantics.  (Michael, Saraswat, and Vechev's algorithm has
127 <     * similar properties, but without support for nulling slots.)
128 <     * Since these combinations aren't supported using ordinary
129 <     * volatiles, the only way to accomplish these efficiently is to
130 <     * use direct Unsafe calls. (Using external AtomicIntegers and
131 <     * AtomicReferenceArrays for the indices and array is
132 <     * significantly slower because of memory locality and indirection
133 <     * effects.)
134 <     *
135 <     * Further, performance on most platforms is very sensitive to
136 <     * placement and sizing of the (resizable) queue array.  Even
137 <     * though these queues don't usually become all that big, the
138 <     * initial size must be large enough to counteract cache
139 <     * contention effects across multiple queues (especially in the
140 <     * presence of GC cardmarking). Also, to improve thread-locality,
141 <     * queues are initialized after starting.  All together, these
142 <     * low-level implementation choices produce as much as a factor of
143 <     * 4 performance improvement compared to naive implementations,
144 <     * and enable the processing of billions of tasks per second,
145 <     * sometimes at the expense of ugliness.
146 <     */
147 <
148 <    /**
149 <     * Generator for initial random seeds for random victim
150 <     * selection. This is used only to create initial seeds. Random
151 <     * steals use a cheaper xorshift generator per steal attempt. We
152 <     * expect only rare contention on seedGenerator, so just use a
153 <     * plain Random.
154 <     */
155 <    private static final Random seedGenerator = new Random();
156 <
157 <    /**
158 <     * The timeout value for suspending spares. Spare workers that
159 <     * remain unsignalled for more than this time may be trimmed
160 <     * (killed and removed from pool).  Since our goal is to avoid
161 <     * long-term thread buildup, the exact value of timeout does not
162 <     * matter too much so long as it avoids most false-alarm timeouts
163 <     * under GC stalls or momentarily high system load.
164 <     */
165 <    private static final long SPARE_KEEPALIVE_NANOS =
166 <        5L * 1000L * 1000L * 1000L; // 5 secs
167 <
168 <    /**
169 <     * The maximum stolen->joining link depth allowed in helpJoinTask.
170 <     * Depths for legitimate chains are unbounded, but we use a fixed
171 <     * constant to avoid (otherwise unchecked) cycles and bound
172 <     * staleness of traversal parameters at the expense of sometimes
173 <     * blocking when we could be helping.
174 <     */
175 <    private static final int MAX_HELP_DEPTH = 8;
176 <
177 <    /**
178 <     * Capacity of work-stealing queue array upon initialization.
179 <     * Must be a power of two. Initial size must be at least 4, but is
180 <     * padded to minimize cache effects.
181 <     */
182 <    private static final int INITIAL_QUEUE_CAPACITY = 1 << 13;
183 <
184 <    /**
185 <     * Maximum work-stealing queue array size.  Must be less than or
186 <     * equal to 1 << 28 to ensure lack of index wraparound. (This
187 <     * is less than usual bounds, because we need leftshift by 3
188 <     * to be in int range).
189 <     */
190 <    private static final int MAXIMUM_QUEUE_CAPACITY = 1 << 28;
191 <
192 <    /**
193 <     * The pool this thread works in. Accessed directly by ForkJoinTask.
194 <     */
195 <    final ForkJoinPool pool;
196 <
197 <    /**
198 <     * The work-stealing queue array. Size must be a power of two.
199 <     * Initialized in onStart, to improve memory locality.
200 <     */
201 <    private ForkJoinTask<?>[] queue;
202 <
203 <    /**
204 <     * Index (mod queue.length) of least valid queue slot, which is
205 <     * always the next position to steal from if nonempty.
206 <     */
207 <    private volatile int base;
208 <
209 <    /**
210 <     * Index (mod queue.length) of next queue slot to push to or pop
211 <     * from. It is written only by owner thread, and accessed by other
212 <     * threads only after reading (volatile) base.  Both sp and base
213 <     * are allowed to wrap around on overflow, but (sp - base) still
214 <     * estimates size.
215 <     */
216 <    private int sp;
217 <
218 <    /**
219 <     * The index of most recent stealer, used as a hint to avoid
220 <     * traversal in method helpJoinTask. This is only a hint because a
221 <     * worker might have had multiple steals and this only holds one
222 <     * of them (usually the most current). Declared non-volatile,
223 <     * relying on other prevailing sync to keep reasonably current.
224 <     */
225 <    private int stealHint;
226 <
227 <    /**
228 <     * Run state of this worker. In addition to the usual run levels,
229 <     * tracks if this worker is suspended as a spare, and if it was
230 <     * killed (trimmed) while suspended. However, "active" status is
231 <     * maintained separately.
232 <     */
233 <    private volatile int runState;
234 <
235 <    private static final int TERMINATING = 0x01;
236 <    private static final int TERMINATED  = 0x02;
237 <    private static final int SUSPENDED   = 0x04; // inactive spare
238 <    private static final int TRIMMED     = 0x08; // killed while suspended
239 <
240 <    /**
241 <     * Number of LockSupport.park calls to block this thread for
242 <     * suspension or event waits. Used for internal instrumention;
243 <     * currently not exported but included because volatile write upon
244 <     * park also provides a workaround for a JVM bug.
245 <     */
246 <    volatile int parkCount;
247 <
248 <    /**
249 <     * Number of steals, transferred and reset in pool callbacks pool
250 <     * when idle Accessed directly by pool.
31 >     * This class just maintains links to its pool and WorkQueue.  The
32 >     * pool field is set upon construction, but the workQueue field is
33 >     * not set until the thread has started (unless forced early by a
34 >     * subclass constructor call to poolIndex()).  This provides
35 >     * better memory placement (because this thread allocates queue
36 >     * and bookkeeping fields) but because the field is non-final, we
37 >     * require that it never be accessed except by the owning thread.
38       */
252    int stealCount;
39  
40 <    /**
41 <     * Seed for random number generator for choosing steal victims.
256 <     * Uses Marsaglia xorshift. Must be initialized as nonzero.
257 <     */
258 <    private int seed;
40 >    final ForkJoinPool pool;                // the pool this thread works in
41 >    ForkJoinPool.WorkQueue workQueue;       // Work-stealing mechanics
42  
43      /**
44 <     * Activity status. When true, this worker is considered active.
262 <     * Accessed directly by pool.  Must be false upon construction.
44 >     * Sequence number for creating worker Names
45       */
46 <    boolean active;
265 <
266 <    /**
267 <     * True if use local fifo, not default lifo, for local polling.
268 <     * Shadows value from ForkJoinPool.
269 <     */
270 <    private final boolean locallyFifo;
271 <
272 <    /**
273 <     * Index of this worker in pool array. Set once by pool before
274 <     * running, and accessed directly by pool to locate this worker in
275 <     * its workers array.
276 <     */
277 <    int poolIndex;
278 <
279 <    /**
280 <     * The last pool event waited for. Accessed only by pool in
281 <     * callback methods invoked within this thread.
282 <     */
283 <    int lastEventCount;
284 <
285 <    /**
286 <     * Encoded index and event count of next event waiter. Used only
287 <     * by ForkJoinPool for managing event waiters.
288 <     */
289 <    volatile long nextWaiter;
290 <
291 <    /**
292 <     * The task currently being joined, set only when actively trying
293 <     * to helpStealer. Written only by current thread, but read by
294 <     * others.
295 <     */
296 <    private volatile ForkJoinTask<?> currentJoin;
297 <
298 <    /**
299 <     * The task most recently stolen from another worker (or
300 <     * submission queue).  Not volatile because always read/written in
301 <     * presence of related volatiles in those cases where it matters.
302 <     */
303 <    private ForkJoinTask<?> currentSteal;
46 >    private static final AtomicInteger threadNumber = new AtomicInteger();
47  
48      /**
49       * Creates a ForkJoinWorkerThread operating in the given pool.
# Line 309 | Line 52 | public class ForkJoinWorkerThread extend
52       * @throws NullPointerException if pool is null
53       */
54      protected ForkJoinWorkerThread(ForkJoinPool pool) {
55 <        this.pool = pool;
313 <        this.locallyFifo = pool.locallyFifo;
314 <        // To avoid exposing construction details to subclasses,
315 <        // remaining initialization is in start() and onStart()
316 <    }
317 <
318 <    /**
319 <     * Performs additional initialization and starts this thread
320 <     */
321 <    final void start(int poolIndex, UncaughtExceptionHandler ueh) {
322 <        this.poolIndex = poolIndex;
55 >        super(pool.workerNamePrefix.concat(Integer.toString(threadNumber.incrementAndGet())));
56          setDaemon(true);
57 +        this.pool = pool;
58 +        Thread.UncaughtExceptionHandler ueh = pool.ueh;
59          if (ueh != null)
60              setUncaughtExceptionHandler(ueh);
326        start();
61      }
62  
329    // Public/protected methods
330
63      /**
64       * Returns the pool hosting this thread.
65       *
# Line 347 | Line 79 | public class ForkJoinWorkerThread extend
79       * @return the index number
80       */
81      public int getPoolIndex() {
82 <        return poolIndex;
82 >        // force early registration if called before started
83 >        ForkJoinPool.WorkQueue q;
84 >        if ((q = workQueue) == null) {
85 >            pool.registerWorker(this);
86 >            q = workQueue;
87 >        }
88 >        return q.poolIndex;
89      }
90  
91      /**
92       * Initializes internal state after construction but before
93       * processing any tasks. If you override this method, you must
94 <     * invoke super.onStart() at the beginning of the method.
94 >     * invoke {@code super.onStart()} at the beginning of the method.
95       * Initialization requires care: Most fields must have legal
96       * default values, to ensure that attempted accesses from other
97       * threads work correctly even before this thread starts
98       * processing tasks.
99       */
100      protected void onStart() {
363        int rs = seedGenerator.nextInt();
364        seed = rs == 0? 1 : rs; // seed must be nonzero
365
366        // Allocate name string and arrays in this thread
367        String pid = Integer.toString(pool.getPoolNumber());
368        String wid = Integer.toString(poolIndex);
369        setName("ForkJoinPool-" + pid + "-worker-" + wid);
370
371        queue = new ForkJoinTask<?>[INITIAL_QUEUE_CAPACITY];
101      }
102  
103      /**
# Line 380 | Line 109 | public class ForkJoinWorkerThread extend
109       * to an unrecoverable error, or {@code null} if completed normally
110       */
111      protected void onTermination(Throwable exception) {
383        try {
384            cancelTasks();
385            setTerminated();
386            pool.workerTerminated(this);
387        } catch (Throwable ex) {        // Shouldn't ever happen
388            if (exception == null)      // but if so, at least rethrown
389                exception = ex;
390        } finally {
391            if (exception != null)
392                UNSAFE.throwException(exception);
393        }
112      }
113  
114      /**
115       * This method is required to be public, but should never be
116       * called explicitly. It performs the main run loop to execute
117 <     * ForkJoinTasks.
117 >     * {@link ForkJoinTask}s.
118       */
119      public void run() {
120          Throwable exception = null;
121          try {
122 +            pool.registerWorker(this);
123              onStart();
124 <            mainLoop();
124 >            pool.runWorker(workQueue);
125          } catch (Throwable ex) {
126              exception = ex;
127          } finally {
409            onTermination(exception);
410        }
411    }
412
413    // helpers for run()
414
415    /**
416     * Find and execute tasks and check status while running
417     */
418    private void mainLoop() {
419        int emptyScans = 0; // consecutive times failed to find work
420        ForkJoinPool p = pool;
421        for (;;) {
422            p.preStep(this, emptyScans);
423            if (runState != 0)
424                return;
425            ForkJoinTask<?> t; // try to get and run stolen or submitted task
426            if ((t = scan()) != null || (t = pollSubmission()) != null) {
427                t.tryExec();
428                if (base != sp)
429                    runLocalTasks();
430                currentSteal = null;
431                emptyScans = 0;
432            }
433            else
434                ++emptyScans;
435        }
436    }
437
438    /**
439     * Runs local tasks until queue is empty or shut down.  Call only
440     * while active.
441     */
442    private void runLocalTasks() {
443        while (runState == 0) {
444            ForkJoinTask<?> t = locallyFifo? locallyDeqTask() : popTask();
445            if (t != null)
446                t.tryExec();
447            else if (base == sp)
448                break;
449        }
450    }
451
452    /**
453     * If a submission exists, try to activate and take it
454     *
455     * @return a task, if available
456     */
457    private ForkJoinTask<?> pollSubmission() {
458        ForkJoinPool p = pool;
459        while (p.hasQueuedSubmissions()) {
460            if (active || (active = p.tryIncrementActiveCount())) {
461                ForkJoinTask<?> t = p.pollSubmission();
462                if (t != null) {
463                    currentSteal = t;
464                    return t;
465                }
466                return scan(); // if missed, rescan
467            }
468        }
469        return null;
470    }
471
472    /*
473     * Intrinsics-based atomic writes for queue slots. These are
474     * basically the same as methods in AtomicObjectArray, but
475     * specialized for (1) ForkJoinTask elements (2) requirement that
476     * nullness and bounds checks have already been performed by
477     * callers and (3) effective offsets are known not to overflow
478     * from int to long (because of MAXIMUM_QUEUE_CAPACITY). We don't
479     * need corresponding version for reads: plain array reads are OK
480     * because they protected by other volatile reads and are
481     * confirmed by CASes.
482     *
483     * Most uses don't actually call these methods, but instead contain
484     * inlined forms that enable more predictable optimization.  We
485     * don't define the version of write used in pushTask at all, but
486     * instead inline there a store-fenced array slot write.
487     */
488
489    /**
490     * CASes slot i of array q from t to null. Caller must ensure q is
491     * non-null and index is in range.
492     */
493    private static final boolean casSlotNull(ForkJoinTask<?>[] q, int i,
494                                             ForkJoinTask<?> t) {
495        return UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null);
496    }
497
498    /**
499     * Performs a volatile write of the given task at given slot of
500     * array q.  Caller must ensure q is non-null and index is in
501     * range. This method is used only during resets and backouts.
502     */
503    private static final void writeSlot(ForkJoinTask<?>[] q, int i,
504                                              ForkJoinTask<?> t) {
505        UNSAFE.putObjectVolatile(q, (i << qShift) + qBase, t);
506    }
507
508    // queue methods
509
510    /**
511     * Pushes a task. Call only from this thread.
512     *
513     * @param t the task. Caller must ensure non-null.
514     */
515    final void pushTask(ForkJoinTask<?> t) {
516        ForkJoinTask<?>[] q = queue;
517        int mask = q.length - 1; // implicit assert q != null
518        int s = sp++;            // ok to increment sp before slot write
519        UNSAFE.putOrderedObject(q, ((s & mask) << qShift) + qBase, t);
520        if ((s -= base) == 0)
521            pool.signalWork();   // was empty
522        else if (s == mask)
523            growQueue();         // is full
524    }
525
526    /**
527     * Tries to take a task from the base of the queue, failing if
528     * empty or contended. Note: Specializations of this code appear
529     * in locallyDeqTask and elsewhere.
530     *
531     * @return a task, or null if none or contended
532     */
533    final ForkJoinTask<?> deqTask() {
534        ForkJoinTask<?> t;
535        ForkJoinTask<?>[] q;
536        int b, i;
537        if ((b = base) != sp &&
538            (q = queue) != null && // must read q after b
539            (t = q[i = (q.length - 1) & b]) != null && base == b &&
540            UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase, t, null)) {
541            base = b + 1;
542            return t;
543        }
544        return null;
545    }
546
547    /**
548     * Tries to take a task from the base of own queue. Assumes active
549     * status.  Called only by current thread.
550     *
551     * @return a task, or null if none
552     */
553    final ForkJoinTask<?> locallyDeqTask() {
554        ForkJoinTask<?>[] q = queue;
555        if (q != null) {
556            ForkJoinTask<?> t;
557            int b, i;
558            while (sp != (b = base)) {
559                if ((t = q[i = (q.length - 1) & b]) != null && base == b &&
560                    UNSAFE.compareAndSwapObject(q, (i << qShift) + qBase,
561                                                t, null)) {
562                    base = b + 1;
563                    return t;
564                }
565            }
566        }
567        return null;
568    }
569
570    /**
571     * Returns a popped task, or null if empty. Assumes active status.
572     * Called only by current thread.
573     */
574    final ForkJoinTask<?> popTask() {
575        int s;
576        ForkJoinTask<?>[] q;
577        if (base != (s = sp) && (q = queue) != null) {
578            int i = (q.length - 1) & --s;
579            ForkJoinTask<?> t = q[i];
580            if (t != null && UNSAFE.compareAndSwapObject
581                (q, (i << qShift) + qBase, t, null)) {
582                sp = s;
583                return t;
584            }
585        }
586        return null;
587    }
588
589    /**
590     * Specialized version of popTask to pop only if topmost element
591     * is the given task. Called only by current thread while
592     * active.
593     *
594     * @param t the task. Caller must ensure non-null.
595     */
596    final boolean unpushTask(ForkJoinTask<?> t) {
597        int s;
598        ForkJoinTask<?>[] q;
599        if (base != (s = sp) && (q = queue) != null &&
600            UNSAFE.compareAndSwapObject
601            (q, (((q.length - 1) & --s) << qShift) + qBase, t, null)) {
602            sp = s;
603            return true;
604        }
605        return false;
606    }
607
608    /**
609     * Returns next task or null if empty or contended
610     */
611    final ForkJoinTask<?> peekTask() {
612        ForkJoinTask<?>[] q = queue;
613        if (q == null)
614            return null;
615        int mask = q.length - 1;
616        int i = locallyFifo ? base : (sp - 1);
617        return q[i & mask];
618    }
619
620    /**
621     * Doubles queue array size. Transfers elements by emulating
622     * steals (deqs) from old array and placing, oldest first, into
623     * new array.
624     */
625    private void growQueue() {
626        ForkJoinTask<?>[] oldQ = queue;
627        int oldSize = oldQ.length;
628        int newSize = oldSize << 1;
629        if (newSize > MAXIMUM_QUEUE_CAPACITY)
630            throw new RejectedExecutionException("Queue capacity exceeded");
631        ForkJoinTask<?>[] newQ = queue = new ForkJoinTask<?>[newSize];
632
633        int b = base;
634        int bf = b + oldSize;
635        int oldMask = oldSize - 1;
636        int newMask = newSize - 1;
637        do {
638            int oldIndex = b & oldMask;
639            ForkJoinTask<?> t = oldQ[oldIndex];
640            if (t != null && !casSlotNull(oldQ, oldIndex, t))
641                t = null;
642            writeSlot(newQ, b & newMask, t);
643        } while (++b != bf);
644        pool.signalWork();
645    }
646
647    /**
648     * Computes next value for random victim probe in scan().  Scans
649     * don't require a very high quality generator, but also not a
650     * crummy one.  Marsaglia xor-shift is cheap and works well enough.
651     * Note: This is manually inlined in scan()
652     */
653    private static final int xorShift(int r) {
654        r ^= r << 13;
655        r ^= r >>> 17;
656        return r ^ (r << 5);
657    }
658
659    /**
660     * Tries to steal a task from another worker. Starts at a random
661     * index of workers array, and probes workers until finding one
662     * with non-empty queue or finding that all are empty.  It
663     * randomly selects the first n probes. If these are empty, it
664     * resorts to a circular sweep, which is necessary to accurately
665     * set active status. (The circular sweep uses steps of
666     * approximately half the array size plus 1, to avoid bias
667     * stemming from leftmost packing of the array in ForkJoinPool.)
668     *
669     * This method must be both fast and quiet -- usually avoiding
670     * memory accesses that could disrupt cache sharing etc other than
671     * those needed to check for and take tasks (or to activate if not
672     * already active). This accounts for, among other things,
673     * updating random seed in place without storing it until exit.
674     *
675     * @return a task, or null if none found
676     */
677    private ForkJoinTask<?> scan() {
678        ForkJoinPool p = pool;
679        ForkJoinWorkerThread[] ws;        // worker array
680        int n;                            // upper bound of #workers
681        if ((ws = p.workers) != null && (n = ws.length) > 1) {
682            boolean canSteal = active;    // shadow active status
683            int r = seed;                 // extract seed once
684            int mask = n - 1;
685            int j = -n;                   // loop counter
686            int k = r;                    // worker index, random if j < 0
687            for (;;) {
688                ForkJoinWorkerThread v = ws[k & mask];
689                r ^= r << 13; r ^= r >>> 17; r ^= r << 5; // inline xorshift
690                if (v != null && v.base != v.sp) {
691                    if (canSteal ||       // ensure active status
692                        (canSteal = active = p.tryIncrementActiveCount())) {
693                        int b = v.base;   // inline specialized deqTask
694                        ForkJoinTask<?>[] q;
695                        if (b != v.sp && (q = v.queue) != null) {
696                            ForkJoinTask<?> t;
697                            int i = (q.length - 1) & b;
698                            long u = (i << qShift) + qBase; // raw offset
699                            if ((t = q[i]) != null && v.base == b &&
700                                UNSAFE.compareAndSwapObject(q, u, t, null)) {
701                                currentSteal = t;
702                                v.stealHint = poolIndex;
703                                v.base = b + 1;
704                                seed = r;
705                                ++stealCount;
706                                return t;
707                            }
708                        }
709                    }
710                    j = -n;
711                    k = r;                // restart on contention
712                }
713                else if (++j <= 0)
714                    k = r;
715                else if (j <= n)
716                    k += (n >>> 1) | 1;
717                else
718                    break;
719            }
720        }
721        return null;
722    }
723
724    // Run State management
725
726    // status check methods used mainly by ForkJoinPool
727    final boolean isTerminating() { return (runState & TERMINATING) != 0; }
728    final boolean isTerminated()  { return (runState & TERMINATED) != 0; }
729    final boolean isSuspended()   { return (runState & SUSPENDED) != 0; }
730    final boolean isTrimmed()     { return (runState & TRIMMED) != 0; }
731
732    /**
733     * Sets state to TERMINATING, also resuming if suspended.
734     */
735    final void shutdown() {
736        for (;;) {
737            int s = runState;
738            if ((s & SUSPENDED) != 0) { // kill and wakeup if suspended
739                if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
740                                             (s & ~SUSPENDED) |
741                                             (TRIMMED|TERMINATING))) {
742                    LockSupport.unpark(this);
743                    break;
744                }
745            }
746            else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
747                                              s | TERMINATING))
748                break;
749        }
750    }
751
752    /**
753     * Sets state to TERMINATED. Called only by this thread.
754     */
755    private void setTerminated() {
756        int s;
757        do {} while (!UNSAFE.compareAndSwapInt(this, runStateOffset,
758                                               s = runState,
759                                               s | (TERMINATING|TERMINATED)));
760    }
761
762    /**
763     * Instrumented version of park used by ForkJoinPool.eventSync
764     */
765    final void doPark() {
766        ++parkCount;
767        LockSupport.park(this);
768    }
769
770    /**
771     * If suspended, tries to set status to unsuspended and unparks.
772     *
773     * @return true if successful
774     */
775    final boolean tryResumeSpare() {
776        int s = runState;
777        if ((s & SUSPENDED) != 0 &&
778            UNSAFE.compareAndSwapInt(this, runStateOffset, s,
779                                     s & ~SUSPENDED)) {
780            LockSupport.unpark(this);
781            return true;
782        }
783        return false;
784    }
785
786    /**
787     * Sets suspended status and blocks as spare until resumed,
788     * shutdown, or timed out.
789     *
790     * @return false if trimmed
791     */
792    final boolean suspendAsSpare() {
793        for (;;) {               // set suspended unless terminating
794            int s = runState;
795            if ((s & TERMINATING) != 0) { // must kill
796                if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
797                                             s | (TRIMMED | TERMINATING)))
798                    return false;
799            }
800            else if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
801                                              s | SUSPENDED))
802                break;
803        }
804        int pc = pool.parallelism;
805        pool.accumulateStealCount(this);
806        boolean timed;
807        long nanos;
808        long startTime;
809        if (poolIndex < pc) { // untimed wait for core threads
810            timed = false;
811            nanos = 0L;
812            startTime = 0L;
813        }
814        else {                // timed wait for added threads
815            timed = true;
816            nanos = SPARE_KEEPALIVE_NANOS;
817            startTime = System.nanoTime();
818        }
819        lastEventCount = 0;      // reset upon resume
820        interrupted();           // clear/ignore interrupts
821        while ((runState & SUSPENDED) != 0) {
822            ++parkCount;
823            if (!timed)
824                LockSupport.park(this);
825            else if ((nanos -= (System.nanoTime() - startTime)) > 0)
826                LockSupport.parkNanos(this, nanos);
827            else { // try to trim on timeout
828                int s = runState;
829                if (UNSAFE.compareAndSwapInt(this, runStateOffset, s,
830                                             (s & ~SUSPENDED) |
831                                             (TRIMMED|TERMINATING)))
832                    return false;
833            }
834        }
835        return true;
836    }
837
838    // Misc support methods for ForkJoinPool
839
840    /**
841     * Returns an estimate of the number of tasks in the queue.  Also
842     * used by ForkJoinTask.
843     */
844    final int getQueueSize() {
845        return -base + sp;
846    }
847
848    /**
849     * Removes and cancels all tasks in queue.  Can be called from any
850     * thread.
851     */
852    final void cancelTasks() {
853        ForkJoinTask<?> cj = currentJoin; // try to kill live tasks
854        if (cj != null) {
855            currentJoin = null;
856            cj.cancelIgnoringExceptions();
857        }
858        ForkJoinTask<?> cs = currentSteal;
859        if (cs != null) {
860            currentSteal = null;
861            cs.cancelIgnoringExceptions();
862        }
863        while (base != sp) {
864            ForkJoinTask<?> t = deqTask();
865            if (t != null)
866                t.cancelIgnoringExceptions();
867        }
868    }
869
870    /**
871     * Drains tasks to given collection c.
872     *
873     * @return the number of tasks drained
874     */
875    final int drainTasksTo(Collection<? super ForkJoinTask<?>> c) {
876        int n = 0;
877        while (base != sp) {
878            ForkJoinTask<?> t = deqTask();
879            if (t != null) {
880                c.add(t);
881                ++n;
882            }
883        }
884        return n;
885    }
886
887    // Support methods for ForkJoinTask
888
889    /**
890     * Gets and removes a local task.
891     *
892     * @return a task, if available
893     */
894    final ForkJoinTask<?> pollLocalTask() {
895        while (sp != base) {
896            if (active || (active = pool.tryIncrementActiveCount()))
897                return locallyFifo? locallyDeqTask() : popTask();
898        }
899        return null;
900    }
901
902    /**
903     * Gets and removes a local or stolen task.
904     *
905     * @return a task, if available
906     */
907    final ForkJoinTask<?> pollTask() {
908        ForkJoinTask<?> t = pollLocalTask();
909        if (t == null) {
910            t = scan();
911            currentSteal = null; // cannot retain/track
912        }
913        return t;
914    }
915
916    /**
917     * Possibly runs some tasks and/or blocks, until task is done.
918     * The main body is basically a big spinloop, alternating between
919     * calls to helpJoinTask and pool.tryAwaitJoin with increased
920     * patience parameters until either the task is done without
921     * waiting, or we have, if necessary, created or resumed a
922     * replacement for this thread while it blocks.
923     *
924     * @param joinMe the task to join
925     * @return task status on exit
926     */
927     final int joinTask(ForkJoinTask<?> joinMe) {
928        int stat;
929        ForkJoinTask<?> prevJoin = currentJoin;
930        // Only written by this thread; only need ordered store
931        UNSAFE.putOrderedObject(this, currentJoinOffset, joinMe);
932        if ((stat = joinMe.status) >= 0 &&
933            (sp == base || (stat = localHelpJoinTask(joinMe)) >= 0)) {
934            for (int retries = 0; ; ++retries) {
935                helpJoinTask(joinMe, retries);
936                if ((stat = joinMe.status) < 0)
937                    break;
938                pool.tryAwaitJoin(joinMe, retries);
939                if ((stat = joinMe.status) < 0)
940                    break;
941                Thread.yield(); // tame unbounded loop
942            }
943        }
944        UNSAFE.putOrderedObject(this, currentJoinOffset, prevJoin);
945        return stat;
946    }
947
948    /**
949     * Run tasks in local queue until given task is done.
950     *
951     * @param joinMe the task to join
952     * @return task status on exit
953     */
954    private int localHelpJoinTask(ForkJoinTask<?> joinMe) {
955        int stat, s;
956        ForkJoinTask<?>[] q;
957        while ((stat = joinMe.status) >= 0 &&
958               base != (s = sp) && (q = queue) != null) {
959            ForkJoinTask<?> t;
960            int i = (q.length - 1) & --s;
961            long u = (i << qShift) + qBase; // raw offset
962            if ((t = q[i]) != null &&
963                UNSAFE.compareAndSwapObject(q, u, t, null)) {
964                /*
965                 * This recheck (and similarly in helpJoinTask)
966                 * handles cases where joinMe is independently
967                 * cancelled or forced even though there is other work
968                 * available. Back out of the pop by putting t back
969                 * into slot before we commit by writing sp.
970                 */
971                if ((stat = joinMe.status) < 0) {
972                    UNSAFE.putObjectVolatile(q, u, t);
973                    break;
974                }
975                sp = s;
976                t.tryExec();
977            }
978        }
979        return stat;
980    }
981
982    /**
983     * Tries to locate and help perform tasks for a stealer of the
984     * given task, or in turn one of its stealers.  Traces
985     * currentSteal->currentJoin links looking for a thread working on
986     * a descendant of the given task and with a non-empty queue to
987     * steal back and execute tasks from. Restarts search upon
988     * encountering chains that are stale, unknown, or of length
989     * greater than MAX_HELP_DEPTH links, to avoid unbounded cycles.
990     *
991     * The implementation is very branchy to cope with the restart
992     * cases.  Returns void, not task status (which must be reread by
993     * caller anyway) to slightly simplify control paths.
994     *
995     * @param joinMe the task to join
996     * @param rescans the number of times to recheck for work
997     */
998    private void helpJoinTask(ForkJoinTask<?> joinMe, int rescans) {
999        ForkJoinWorkerThread[] ws = pool.workers;
1000        int n;
1001        if (ws == null || (n = ws.length) <= 1)
1002            return;                   // need at least 2 workers
1003        restart:while (rescans-- >= 0 && joinMe.status >= 0) {
1004            ForkJoinTask<?> task = joinMe;        // base of chain
1005            ForkJoinWorkerThread thread = this;   // thread with stolen task
1006            for (int depth = 0; depth < MAX_HELP_DEPTH; ++depth) {
1007                // Try to find v, the stealer of task, by first using hint
1008                ForkJoinWorkerThread v = ws[thread.stealHint & (n - 1)];
1009                if (v == null || v.currentSteal != task) {
1010                    for (int j = 0; ; ++j) {      // search array
1011                        if (task.status < 0 || j == n)
1012                            continue restart;     // stale or no stealer
1013                        if ((v = ws[j]) != null && v.currentSteal == task) {
1014                            thread.stealHint = j; // save for next time
1015                            break;
1016                        }
1017                    }
1018                }
1019                // Try to help v, using specialized form of deqTask
1020                int b;
1021                ForkJoinTask<?>[] q;
1022                while ((b = v.base) != v.sp && (q = v.queue) != null) {
1023                    int i = (q.length - 1) & b;
1024                    long u = (i << qShift) + qBase;
1025                    ForkJoinTask<?> t = q[i];
1026                    if (task.status < 0)          // stale
1027                        continue restart;
1028                    if (t != null) {
1029                        if (v.base == b &&
1030                            UNSAFE.compareAndSwapObject(q, u, t, null)) {
1031                            if (joinMe.status < 0) {
1032                                UNSAFE.putObjectVolatile(q, u, t);
1033                                return;           // back out on cancel
1034                            }
1035                            ForkJoinTask<?> prevSteal = currentSteal;
1036                            currentSteal = t;
1037                            v.stealHint = poolIndex;
1038                            v.base = b + 1;
1039                            t.tryExec();
1040                            currentSteal = prevSteal;
1041                        }
1042                    }
1043                    else if (v.base == b)          // producer stalled
1044                        continue restart;          // retry via restart
1045                    if (joinMe.status < 0)
1046                        return;
1047                }
1048                // Try to descend to find v's stealer
1049                ForkJoinTask<?> next = v.currentJoin;
1050                if (next == null || next == task || task.status < 0)
1051                    continue restart;             // no descendent or stale
1052                if (joinMe.status < 0)
1053                    return;
1054                task = next;
1055                thread = v;
1056            }
1057        }
1058    }
1059
1060    /**
1061     * Returns an estimate of the number of tasks, offset by a
1062     * function of number of idle workers.
1063     *
1064     * This method provides a cheap heuristic guide for task
1065     * partitioning when programmers, frameworks, tools, or languages
1066     * have little or no idea about task granularity.  In essence by
1067     * offering this method, we ask users only about tradeoffs in
1068     * overhead vs expected throughput and its variance, rather than
1069     * how finely to partition tasks.
1070     *
1071     * In a steady state strict (tree-structured) computation, each
1072     * thread makes available for stealing enough tasks for other
1073     * threads to remain active. Inductively, if all threads play by
1074     * the same rules, each thread should make available only a
1075     * constant number of tasks.
1076     *
1077     * The minimum useful constant is just 1. But using a value of 1
1078     * would require immediate replenishment upon each steal to
1079     * maintain enough tasks, which is infeasible.  Further,
1080     * partitionings/granularities of offered tasks should minimize
1081     * steal rates, which in general means that threads nearer the top
1082     * of computation tree should generate more than those nearer the
1083     * bottom. In perfect steady state, each thread is at
1084     * approximately the same level of computation tree. However,
1085     * producing extra tasks amortizes the uncertainty of progress and
1086     * diffusion assumptions.
1087     *
1088     * So, users will want to use values larger, but not much larger
1089     * than 1 to both smooth over transient shortages and hedge
1090     * against uneven progress; as traded off against the cost of
1091     * extra task overhead. We leave the user to pick a threshold
1092     * value to compare with the results of this call to guide
1093     * decisions, but recommend values such as 3.
1094     *
1095     * When all threads are active, it is on average OK to estimate
1096     * surplus strictly locally. In steady-state, if one thread is
1097     * maintaining say 2 surplus tasks, then so are others. So we can
1098     * just use estimated queue length (although note that (sp - base)
1099     * can be an overestimate because of stealers lagging increments
1100     * of base).  However, this strategy alone leads to serious
1101     * mis-estimates in some non-steady-state conditions (ramp-up,
1102     * ramp-down, other stalls). We can detect many of these by
1103     * further considering the number of "idle" threads, that are
1104     * known to have zero queued tasks, so compensate by a factor of
1105     * (#idle/#active) threads.
1106     */
1107    final int getEstimatedSurplusTaskCount() {
1108        return sp - base - pool.idlePerActive();
1109    }
1110
1111    /**
1112     * Runs tasks until {@code pool.isQuiescent()}.
1113     */
1114    final void helpQuiescePool() {
1115        for (;;) {
1116            ForkJoinTask<?> t = pollLocalTask();
1117            if (t != null || (t = scan()) != null) {
1118                t.tryExec();
1119                currentSteal = null;
1120            }
1121            else {
1122                ForkJoinPool p = pool;
1123                if (active) {
1124                    active = false; // inactivate
1125                    do {} while (!p.tryDecrementActiveCount());
1126                }
1127                if (p.isQuiescent()) {
1128                    active = true; // re-activate
1129                    do {} while (!p.tryIncrementActiveCount());
1130                    return;
1131                }
1132            }
1133        }
1134    }
1135
1136    // Unsafe mechanics
1137
1138    private static final sun.misc.Unsafe UNSAFE = getUnsafe();
1139    private static final long runStateOffset =
1140        objectFieldOffset("runState", ForkJoinWorkerThread.class);
1141    private static final long currentJoinOffset =
1142        objectFieldOffset("currentJoin", ForkJoinWorkerThread.class);
1143    private static final long qBase =
1144        UNSAFE.arrayBaseOffset(ForkJoinTask[].class);
1145    private static final int qShift;
1146
1147    static {
1148        int s = UNSAFE.arrayIndexScale(ForkJoinTask[].class);
1149        if ((s & (s-1)) != 0)
1150            throw new Error("data type scale not a power of two");
1151        qShift = 31 - Integer.numberOfLeadingZeros(s);
1152    }
1153
1154    private static long objectFieldOffset(String field, Class<?> klazz) {
1155        try {
1156            return UNSAFE.objectFieldOffset(klazz.getDeclaredField(field));
1157        } catch (NoSuchFieldException e) {
1158            // Convert Exception to corresponding Error
1159            NoSuchFieldError error = new NoSuchFieldError(field);
1160            error.initCause(e);
1161            throw error;
1162        }
1163    }
1164
1165    /**
1166     * Returns a sun.misc.Unsafe.  Suitable for use in a 3rd party package.
1167     * Replace with a simple call to Unsafe.getUnsafe when integrating
1168     * into a jdk.
1169     *
1170     * @return a sun.misc.Unsafe
1171     */
1172    private static sun.misc.Unsafe getUnsafe() {
1173        try {
1174            return sun.misc.Unsafe.getUnsafe();
1175        } catch (SecurityException se) {
128              try {
129 <                return java.security.AccessController.doPrivileged
130 <                    (new java.security
131 <                     .PrivilegedExceptionAction<sun.misc.Unsafe>() {
132 <                        public sun.misc.Unsafe run() throws Exception {
133 <                            java.lang.reflect.Field f = sun.misc
134 <                                .Unsafe.class.getDeclaredField("theUnsafe");
1183 <                            f.setAccessible(true);
1184 <                            return (sun.misc.Unsafe) f.get(null);
1185 <                        }});
1186 <            } catch (java.security.PrivilegedActionException e) {
1187 <                throw new RuntimeException("Could not initialize intrinsics",
1188 <                                           e.getCause());
129 >                onTermination(exception);
130 >            } catch (Throwable ex) {
131 >                if (exception == null)
132 >                    exception = ex;
133 >            } finally {
134 >                pool.deregisterWorker(this, exception);
135              }
136          }
137      }

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines