1 /*
2 * Copyright (c) 2001, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #ifndef SHARE_VM_UTILITIES_TASKQUEUE_HPP
26 #define SHARE_VM_UTILITIES_TASKQUEUE_HPP
27
28 #include "memory/allocation.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "memory/padded.hpp"
31 #include "runtime/mutex.hpp"
32 #include "runtime/orderAccess.inline.hpp"
33 #include "utilities/globalDefinitions.hpp"
34 #include "utilities/stack.hpp"
35
36 // Simple TaskQueue stats that are collected by default in debug builds.
37
38 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
39 #define TASKQUEUE_STATS 1
40 #elif !defined(TASKQUEUE_STATS)
41 #define TASKQUEUE_STATS 0
42 #endif
43
44 #if TASKQUEUE_STATS
45 #define TASKQUEUE_STATS_ONLY(code) code
46 #else
47 #define TASKQUEUE_STATS_ONLY(code)
48 #endif // TASKQUEUE_STATS
49
50 #if TASKQUEUE_STATS
51 class TaskQueueStats {
52 public:
53 enum StatId {
54 push, // number of taskqueue pushes
55 pop, // number of taskqueue pops
56 pop_slow, // subset of taskqueue pops that were done slow-path
57 steal_attempt, // number of taskqueue steal attempts
58 steal, // number of taskqueue steals
59 overflow, // number of overflow pushes
60 overflow_max_len, // max length of overflow stack
61 last_stat_id
62 };
63
64 public:
65 inline TaskQueueStats() { reset(); }
66
67 inline void record_push() { ++_stats[push]; }
68 inline void record_pop() { ++_stats[pop]; }
69 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
70 inline void record_steal(bool success);
71 inline void record_overflow(size_t new_length);
72
73 TaskQueueStats & operator +=(const TaskQueueStats & addend);
74
75 inline size_t get(StatId id) const { return _stats[id]; }
76 inline const size_t* get() const { return _stats; }
77
78 inline void reset();
79
80 // Print the specified line of the header (does not include a line separator).
81 static void print_header(unsigned int line, outputStream* const stream = tty,
82 unsigned int width = 10);
83 // Print the statistics (does not include a line separator).
84 void print(outputStream* const stream = tty, unsigned int width = 10) const;
85
86 DEBUG_ONLY(void verify() const;)
87
88 private:
89 size_t _stats[last_stat_id];
90 static const char * const _names[last_stat_id];
91 };
92
93 void TaskQueueStats::record_steal(bool success) {
94 ++_stats[steal_attempt];
95 if (success) ++_stats[steal];
96 }
97
98 void TaskQueueStats::record_overflow(size_t new_len) {
99 ++_stats[overflow];
100 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
101 }
102
103 void TaskQueueStats::reset() {
104 memset(_stats, 0, sizeof(_stats));
105 }
106 #endif // TASKQUEUE_STATS
107
108 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances.
109
110 template <unsigned int N, MEMFLAGS F>
111 class TaskQueueSuper: public CHeapObj<F> {
112 protected:
113 // Internal type for indexing the queue; also used for the tag.
114 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
115
116 enum { MOD_N_MASK = N - 1 };
117
118 class Age {
119 public:
120 Age(size_t data = 0) { _data = data; }
121 Age(const Age& age) { _data = age._data; }
122 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; }
123
124 Age get() const volatile { return _data; }
125 void set(Age age) volatile { _data = age._data; }
126
127 idx_t top() const volatile { return _fields._top; }
128 idx_t tag() const volatile { return _fields._tag; }
129
130 // Increment top; if it wraps, increment tag also.
131 void increment() {
132 _fields._top = increment_index(_fields._top);
133 if (_fields._top == 0) ++_fields._tag;
134 }
135
136 Age cmpxchg(const Age new_age, const Age old_age) volatile {
137 return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
138 (volatile intptr_t *)&_data,
139 (intptr_t)old_age._data);
140 }
141
142 bool operator ==(const Age& other) const { return _data == other._data; }
143
144 private:
145 struct fields {
146 idx_t _top;
147 idx_t _tag;
148 };
149 union {
150 size_t _data;
151 fields _fields;
152 };
153 };
154
155 // The first free element after the last one pushed (mod N).
156 volatile uint _bottom;
157 // Add paddings to reduce false-sharing cache contention between _bottom and _age
158 DEFINE_PAD_MINUS_SIZE(0, DEFAULT_CACHE_LINE_SIZE, sizeof(uint));
159 volatile Age _age;
160
161 // These both operate mod N.
162 static uint increment_index(uint ind) {
163 return (ind + 1) & MOD_N_MASK;
164 }
165 static uint decrement_index(uint ind) {
166 return (ind - 1) & MOD_N_MASK;
167 }
168
169 // Returns a number in the range [0..N). If the result is "N-1", it should be
170 // interpreted as 0.
171 uint dirty_size(uint bot, uint top) const {
172 return (bot - top) & MOD_N_MASK;
173 }
174
175 // Returns the size corresponding to the given "bot" and "top".
176 uint size(uint bot, uint top) const {
177 uint sz = dirty_size(bot, top);
178 // Has the queue "wrapped", so that bottom is less than top? There's a
179 // complicated special case here. A pair of threads could perform pop_local
180 // and pop_global operations concurrently, starting from a state in which
181 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
182 // and the pop_global in incrementing _top (in which case the pop_global
183 // will be awarded the contested queue element.) The resulting state must
184 // be interpreted as an empty queue. (We only need to worry about one such
185 // event: only the queue owner performs pop_local's, and several concurrent
186 // threads attempting to perform the pop_global will all perform the same
187 // CAS, and only one can succeed.) Any stealing thread that reads after
188 // either the increment or decrement will see an empty queue, and will not
189 // join the competitors. The "sz == -1 || sz == N-1" state will not be
190 // modified by concurrent queues, so the owner thread can reset the state to
191 // _bottom == top so subsequent pushes will be performed normally.
192 return (sz == N - 1) ? 0 : sz;
193 }
194
195 public:
196 TaskQueueSuper() : _bottom(0), _age() {}
197
198 // Return true if the TaskQueue contains/does not contain any tasks.
199 bool peek() const { return _bottom != _age.top(); }
200 bool is_empty() const { return size() == 0; }
201
202 // Return an estimate of the number of elements in the queue.
203 // The "careful" version admits the possibility of pop_local/pop_global
204 // races.
205 uint size() const {
206 return size(_bottom, _age.top());
207 }
208
209 uint dirty_size() const {
210 return dirty_size(_bottom, _age.top());
211 }
212
213 void set_empty() {
214 _bottom = 0;
215 _age.set(0);
216 }
217
218 // Maximum number of elements allowed in the queue. This is two less
219 // than the actual queue size, for somewhat complicated reasons.
220 uint max_elems() const { return N - 2; }
221
222 // Total size of queue.
223 static const uint total_size() { return N; }
224
225 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
226 };
227
228 //
229 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
230 // ended-queue (deque), intended for use in work stealing. Queue operations
231 // are non-blocking.
232 //
233 // A queue owner thread performs push() and pop_local() operations on one end
234 // of the queue, while other threads may steal work using the pop_global()
235 // method.
236 //
237 // The main difference to the original algorithm is that this
238 // implementation allows wrap-around at the end of its allocated
239 // storage, which is an array.
240 //
241 // The original paper is:
242 //
243 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
244 // Thread scheduling for multiprogrammed multiprocessors.
245 // Theory of Computing Systems 34, 2 (2001), 115-144.
246 //
247 // The following paper provides an correctness proof and an
248 // implementation for weakly ordered memory models including (pseudo-)
249 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
250 // similar to ABP, with the main difference that it allows resizing of the
251 // underlying storage:
252 //
253 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
254 // Correct and efficient work-stealing for weak memory models
255 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
256 // practice of parallel programming (PPoPP 2013), 69-80
257 //
258
259 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
260 class GenericTaskQueue: public TaskQueueSuper<N, F> {
261 ArrayAllocator<E, F> _array_allocator;
262 protected:
263 typedef typename TaskQueueSuper<N, F>::Age Age;
264 typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
265
266 using TaskQueueSuper<N, F>::_bottom;
267 using TaskQueueSuper<N, F>::_age;
268 using TaskQueueSuper<N, F>::increment_index;
269 using TaskQueueSuper<N, F>::decrement_index;
270 using TaskQueueSuper<N, F>::dirty_size;
271
272 public:
273 using TaskQueueSuper<N, F>::max_elems;
274 using TaskQueueSuper<N, F>::size;
275
276 #if TASKQUEUE_STATS
277 using TaskQueueSuper<N, F>::stats;
278 #endif
279
280 private:
281 // Slow paths for push, pop_local. (pop_global has no fast path.)
282 bool push_slow(E t, uint dirty_n_elems);
283 bool pop_local_slow(uint localBot, Age oldAge);
284
285 public:
286 typedef E element_type;
287
288 // Initializes the queue to empty.
289 GenericTaskQueue();
290
291 void initialize();
292
293 // Push the task "t" on the queue. Returns "false" iff the queue is full.
294 inline bool push(E t);
295
296 // Attempts to claim a task from the "local" end of the queue (the most
297 // recently pushed). If successful, returns true and sets t to the task;
298 // otherwise, returns false (the queue is empty).
299 inline bool pop_local(volatile E& t);
300
301 // Like pop_local(), but uses the "global" end of the queue (the least
302 // recently pushed).
303 bool pop_global(volatile E& t);
304
305 // Delete any resource associated with the queue.
306 ~GenericTaskQueue();
307
308 // apply the closure to all elements in the task queue
309 void oops_do(OopClosure* f);
310
311 private:
312 // Element array.
313 volatile E* _elems;
314 };
315
316 template<class E, MEMFLAGS F, unsigned int N>
317 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
318 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
319 }
320
321 template<class E, MEMFLAGS F, unsigned int N>
322 void GenericTaskQueue<E, F, N>::initialize() {
323 _elems = _array_allocator.allocate(N);
324 }
325
326 template<class E, MEMFLAGS F, unsigned int N>
327 void GenericTaskQueue<E, F, N>::oops_do(OopClosure* f) {
328 // tty->print_cr("START OopTaskQueue::oops_do");
329 uint iters = size();
330 uint index = _bottom;
331 for (uint i = 0; i < iters; ++i) {
332 index = decrement_index(index);
333 // tty->print_cr(" doing entry %d," INTPTR_T " -> " INTPTR_T,
334 // index, &_elems[index], _elems[index]);
335 E* t = (E*)&_elems[index]; // cast away volatility
336 oop* p = (oop*)t;
337 // G1 does its own checking
338 assert(UseG1GC || (*t)->is_oop_or_null(), "Not an oop or null");
339 f->do_oop(p);
340 }
341 // tty->print_cr("END OopTaskQueue::oops_do");
342 }
343
344 template<class E, MEMFLAGS F, unsigned int N>
345 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) {
346 if (dirty_n_elems == N - 1) {
347 // Actually means 0, so do the push.
348 uint localBot = _bottom;
349 // g++ complains if the volatile result of the assignment is
350 // unused, so we cast the volatile away. We cannot cast directly
351 // to void, because gcc treats that as not using the result of the
352 // assignment. However, casting to E& means that we trigger an
353 // unused-value warning. So, we cast the E& to void.
354 (void)const_cast<E&>(_elems[localBot] = t);
355 OrderAccess::release_store(&_bottom, increment_index(localBot));
356 TASKQUEUE_STATS_ONLY(stats.record_push());
357 return true;
358 }
359 return false;
360 }
361
362 // pop_local_slow() is done by the owning thread and is trying to
363 // get the last task in the queue. It will compete with pop_global()
364 // that will be used by other threads. The tag age is incremented
365 // whenever the queue goes empty which it will do here if this thread
366 // gets the last task or in pop_global() if the queue wraps (top == 0
367 // and pop_global() succeeds, see pop_global()).
368 template<class E, MEMFLAGS F, unsigned int N>
369 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) {
370 // This queue was observed to contain exactly one element; either this
371 // thread will claim it, or a competing "pop_global". In either case,
372 // the queue will be logically empty afterwards. Create a new Age value
373 // that represents the empty queue for the given value of "_bottom". (We
374 // must also increment "tag" because of the case where "bottom == 1",
375 // "top == 0". A pop_global could read the queue element in that case,
376 // then have the owner thread do a pop followed by another push. Without
377 // the incrementing of "tag", the pop_global's CAS could succeed,
378 // allowing it to believe it has claimed the stale element.)
379 Age newAge((idx_t)localBot, oldAge.tag() + 1);
380 // Perhaps a competing pop_global has already incremented "top", in which
381 // case it wins the element.
382 if (localBot == oldAge.top()) {
383 // No competing pop_global has yet incremented "top"; we'll try to
384 // install new_age, thus claiming the element.
385 Age tempAge = _age.cmpxchg(newAge, oldAge);
386 if (tempAge == oldAge) {
387 // We win.
388 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
389 TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
390 return true;
391 }
392 }
393 // We lose; a completing pop_global gets the element. But the queue is empty
394 // and top is greater than bottom. Fix this representation of the empty queue
395 // to become the canonical one.
396 _age.set(newAge);
397 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
398 return false;
399 }
400
401 template<class E, MEMFLAGS F, unsigned int N>
402 bool GenericTaskQueue<E, F, N>::pop_global(volatile E& t) {
403 Age oldAge = _age.get();
404 // Architectures with weak memory model require a barrier here
405 // to guarantee that bottom is not older than age,
406 // which is crucial for the correctness of the algorithm.
407 #if !(defined SPARC || defined IA32 || defined AMD64)
408 OrderAccess::fence();
409 #endif
410 uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom);
411 uint n_elems = size(localBot, oldAge.top());
412 if (n_elems == 0) {
413 return false;
414 }
415
416 // g++ complains if the volatile result of the assignment is
417 // unused, so we cast the volatile away. We cannot cast directly
418 // to void, because gcc treats that as not using the result of the
419 // assignment. However, casting to E& means that we trigger an
420 // unused-value warning. So, we cast the E& to void.
421 (void) const_cast<E&>(t = _elems[oldAge.top()]);
422 Age newAge(oldAge);
423 newAge.increment();
424 Age resAge = _age.cmpxchg(newAge, oldAge);
425
426 // Note that using "_bottom" here might fail, since a pop_local might
427 // have decremented it.
428 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
429 return resAge == oldAge;
430 }
431
432 template<class E, MEMFLAGS F, unsigned int N>
433 GenericTaskQueue<E, F, N>::~GenericTaskQueue() {}
434
435 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
436 // elements that do not fit in the TaskQueue.
437 //
438 // This class hides two methods from super classes:
439 //
440 // push() - push onto the task queue or, if that fails, onto the overflow stack
441 // is_empty() - return true if both the TaskQueue and overflow stack are empty
442 //
443 // Note that size() is not hidden--it returns the number of elements in the
444 // TaskQueue, and does not include the size of the overflow stack. This
445 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
446 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
447 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
448 {
449 public:
450 typedef Stack<E, F> overflow_t;
451 typedef GenericTaskQueue<E, F, N> taskqueue_t;
452
453 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
454
455 // Push task t onto the queue or onto the overflow stack. Return true.
456 inline bool push(E t);
457
458 // Try to push task t onto the queue only. Returns true if successful, false otherwise.
459 inline bool try_push_to_taskqueue(E t);
460
461 // Attempt to pop from the overflow stack; return true if anything was popped.
462 inline bool pop_overflow(E& t);
463
464 inline overflow_t* overflow_stack() { return &_overflow_stack; }
465
466 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
467 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
468 inline bool is_empty() const {
469 return taskqueue_empty() && overflow_empty();
470 }
471
472 private:
473 overflow_t _overflow_stack;
474 };
475
476 template <class E, MEMFLAGS F, unsigned int N>
477 bool OverflowTaskQueue<E, F, N>::push(E t)
478 {
479 if (!taskqueue_t::push(t)) {
480 overflow_stack()->push(t);
481 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
482 }
483 return true;
484 }
485
486 template <class E, MEMFLAGS F, unsigned int N>
487 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t)
488 {
489 if (overflow_empty()) return false;
490 t = overflow_stack()->pop();
491 return true;
492 }
493
494 template <class E, MEMFLAGS F, unsigned int N>
495 bool OverflowTaskQueue<E, F, N>::try_push_to_taskqueue(E t) {
496 return taskqueue_t::push(t);
497 }
498 class TaskQueueSetSuper {
499 protected:
500 static int randomParkAndMiller(int* seed0);
501 public:
502 // Returns "true" if some TaskQueue in the set contains a task.
503 virtual bool peek() = 0;
504 virtual size_t tasks() = 0;
505 };
506
507 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
508 };
509
510 template<class T, MEMFLAGS F>
511 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
512 private:
513 uint _n;
514 T** _queues;
515
516 public:
517 typedef typename T::element_type E;
518
519 GenericTaskQueueSet(int n) : _n(n) {
520 typedef T* GenericTaskQueuePtr;
521 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F);
522 for (int i = 0; i < n; i++) {
523 _queues[i] = NULL;
524 }
525 }
526
527 bool steal_best_of_2(uint queue_num, int* seed, E& t);
528
529 void register_queue(uint i, T* q);
530
531 T* queue(uint n);
532
533 // The thread with queue number "queue_num" (and whose random number seed is
534 // at "seed") is trying to steal a task from some other queue. (It may try
535 // several queues, according to some configuration parameter.) If some steal
536 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
537 // false.
538 bool steal(uint queue_num, int* seed, E& t);
539
540 bool peek();
541 size_t tasks();
542
543 uint size() const { return _n; }
544 };
545
546 template<class T, MEMFLAGS F> void
547 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
548 assert(i < _n, "index out of range.");
549 _queues[i] = q;
550 }
551
552 template<class T, MEMFLAGS F> T*
553 GenericTaskQueueSet<T, F>::queue(uint i) {
554 return _queues[i];
555 }
556
557 template<class T, MEMFLAGS F> bool
558 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
559 for (uint i = 0; i < 2 * _n; i++) {
560 if (steal_best_of_2(queue_num, seed, t)) {
561 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
562 return true;
563 }
564 }
565 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
566 return false;
567 }
568
569 template<class T, MEMFLAGS F> bool
570 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) {
571 if (_n > 2) {
572 uint k1 = queue_num;
573 while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
574 uint k2 = queue_num;
575 while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
576 // Sample both and try the larger.
577 uint sz1 = _queues[k1]->size();
578 uint sz2 = _queues[k2]->size();
579 if (sz2 > sz1) return _queues[k2]->pop_global(t);
580 else return _queues[k1]->pop_global(t);
581 } else if (_n == 2) {
582 // Just try the other one.
583 uint k = (queue_num + 1) % 2;
584 return _queues[k]->pop_global(t);
585 } else {
586 assert(_n == 1, "can't be zero.");
587 return false;
588 }
589 }
590
591 template<class T, MEMFLAGS F>
592 bool GenericTaskQueueSet<T, F>::peek() {
593 // Try all the queues.
594 for (uint j = 0; j < _n; j++) {
595 if (_queues[j]->peek())
596 return true;
597 }
598 return false;
599 }
600
601 template<class T, MEMFLAGS F>
602 size_t GenericTaskQueueSet<T, F>::tasks() {
603 size_t n = 0;
604 for (uint j = 0; j < _n; j++) {
605 n += _queues[j]->size();
606 }
607 return n;
608 }
609
610 // When to terminate from the termination protocol.
611 class TerminatorTerminator: public CHeapObj<mtInternal> {
612 public:
613 virtual bool should_exit_termination() = 0;
614 };
615
616 // A class to aid in the termination of a set of parallel tasks using
617 // TaskQueueSet's for work stealing.
618
619 #undef TRACESPINNING
620
621 class ParallelTaskTerminator: public StackObj {
622 protected:
623 int _n_threads;
624 TaskQueueSetSuper* _queue_set;
625 char _pad_before[DEFAULT_CACHE_LINE_SIZE];
626 int _offered_termination;
627 char _pad_after[DEFAULT_CACHE_LINE_SIZE];
628
629 #ifdef TRACESPINNING
630 static uint _total_yields;
631 static uint _total_spins;
632 static uint _total_peeks;
633 #endif
634
635 bool peek_in_queue_set();
636 protected:
637 virtual void yield();
638 void sleep(uint millis);
639
640 public:
641
642 // "n_threads" is the number of threads to be terminated. "queue_set" is a
643 // queue sets of work queues of other threads.
644 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
645
646 // The current thread has no work, and is ready to terminate if everyone
647 // else is. If returns "true", all threads are terminated. If returns
648 // "false", available work has been observed in one of the task queues,
649 // so the global task is not complete.
650 virtual bool offer_termination() {
651 return offer_termination(NULL);
652 }
653
654 // As above, but it also terminates if the should_exit_termination()
655 // method of the terminator parameter returns true. If terminator is
656 // NULL, then it is ignored.
657 bool offer_termination(TerminatorTerminator* terminator);
658
659 // Reset the terminator, so that it may be reused again.
660 // The caller is responsible for ensuring that this is done
661 // in an MT-safe manner, once the previous round of use of
662 // the terminator is finished.
663 void reset_for_reuse();
664 // Same as above but the number of parallel threads is set to the
665 // given number.
666 void reset_for_reuse(int n_threads);
667
668 #ifdef TRACESPINNING
669 static uint total_yields() { return _total_yields; }
670 static uint total_spins() { return _total_spins; }
671 static uint total_peeks() { return _total_peeks; }
672 static void print_termination_counts();
673 #endif
674 };
675
676 template<class E, MEMFLAGS F, unsigned int N> inline bool
677 GenericTaskQueue<E, F, N>::push(E t) {
678 uint localBot = _bottom;
679 assert(localBot < N, "_bottom out of range.");
680 idx_t top = _age.top();
681 uint dirty_n_elems = dirty_size(localBot, top);
682 assert(dirty_n_elems < N, "n_elems out of range.");
683 if (dirty_n_elems < max_elems()) {
684 // g++ complains if the volatile result of the assignment is
685 // unused, so we cast the volatile away. We cannot cast directly
686 // to void, because gcc treats that as not using the result of the
687 // assignment. However, casting to E& means that we trigger an
688 // unused-value warning. So, we cast the E& to void.
689 (void) const_cast<E&>(_elems[localBot] = t);
690 OrderAccess::release_store(&_bottom, increment_index(localBot));
691 TASKQUEUE_STATS_ONLY(stats.record_push());
692 return true;
693 } else {
694 return push_slow(t, dirty_n_elems);
695 }
696 }
697
698 template<class E, MEMFLAGS F, unsigned int N> inline bool
699 GenericTaskQueue<E, F, N>::pop_local(volatile E& t) {
700 uint localBot = _bottom;
701 // This value cannot be N-1. That can only occur as a result of
702 // the assignment to bottom in this method. If it does, this method
703 // resets the size to 0 before the next call (which is sequential,
704 // since this is pop_local.)
705 uint dirty_n_elems = dirty_size(localBot, _age.top());
706 assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
707 if (dirty_n_elems == 0) return false;
708 localBot = decrement_index(localBot);
709 _bottom = localBot;
710 // This is necessary to prevent any read below from being reordered
711 // before the store just above.
712 OrderAccess::fence();
713 // g++ complains if the volatile result of the assignment is
714 // unused, so we cast the volatile away. We cannot cast directly
715 // to void, because gcc treats that as not using the result of the
716 // assignment. However, casting to E& means that we trigger an
717 // unused-value warning. So, we cast the E& to void.
718 (void) const_cast<E&>(t = _elems[localBot]);
719 // This is a second read of "age"; the "size()" above is the first.
720 // If there's still at least one element in the queue, based on the
721 // "_bottom" and "age" we've read, then there can be no interference with
722 // a "pop_global" operation, and we're done.
723 idx_t tp = _age.top(); // XXX
724 if (size(localBot, tp) > 0) {
725 assert(dirty_size(localBot, tp) != N - 1, "sanity");
726 TASKQUEUE_STATS_ONLY(stats.record_pop());
727 return true;
728 } else {
729 // Otherwise, the queue contained exactly one element; we take the slow
730 // path.
731
732 // The barrier is required to prevent reordering the two reads of _age:
733 // one is the _age.get() below, and the other is _age.top() above the if-stmt.
734 // The algorithm may fail if _age.get() reads an older value than _age.top().
735 OrderAccess::loadload();
736 return pop_local_slow(localBot, _age.get());
737 }
738 }
739
740 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue;
741 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
742
743 #ifdef _MSC_VER
744 #pragma warning(push)
745 // warning C4522: multiple assignment operators specified
746 #pragma warning(disable:4522)
747 #endif
748
749 // This is a container class for either an oop* or a narrowOop*.
750 // Both are pushed onto a task queue and the consumer will test is_narrow()
751 // to determine which should be processed.
752 class StarTask {
753 void* _holder; // either union oop* or narrowOop*
754
755 enum { COMPRESSED_OOP_MASK = 1 };
756
757 public:
758 StarTask(narrowOop* p) {
759 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
760 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
761 }
762 StarTask(oop* p) {
763 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
764 _holder = (void*)p;
765 }
766 StarTask() { _holder = NULL; }
767 operator oop*() { return (oop*)_holder; }
768 operator narrowOop*() {
769 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
770 }
771
772 StarTask& operator=(const StarTask& t) {
773 _holder = t._holder;
774 return *this;
775 }
776 volatile StarTask& operator=(const volatile StarTask& t) volatile {
777 _holder = t._holder;
778 return *this;
779 }
780
781 bool is_narrow() const {
782 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
783 }
784 };
785
786 class ObjArrayTask
787 {
788 public:
789 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
790 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
791 assert(idx <= size_t(max_jint), "too big");
792 }
793 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
794
795 ObjArrayTask& operator =(const ObjArrayTask& t) {
796 _obj = t._obj;
797 _index = t._index;
798 return *this;
799 }
800 volatile ObjArrayTask&
801 operator =(const volatile ObjArrayTask& t) volatile {
802 (void)const_cast<oop&>(_obj = t._obj);
803 _index = t._index;
804 return *this;
805 }
806
807 inline oop obj() const { return _obj; }
808 inline int index() const { return _index; }
809
810 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
811
812 private:
813 oop _obj;
814 int _index;
815 };
816
817 #ifdef _MSC_VER
818 #pragma warning(pop)
819 #endif
820
821 typedef OverflowTaskQueue<StarTask, mtClass> OopStarTaskQueue;
822 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet;
823
824 typedef OverflowTaskQueue<size_t, mtInternal> RegionTaskQueue;
825 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass> RegionTaskQueueSet;
826
827
828 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP