Tue, 17 Oct 2017 12:58:25 +0800
merge
1 /*
2 * Copyright (c) 2001, 2013, 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 */
25 #ifndef SHARE_VM_UTILITIES_TASKQUEUE_HPP
26 #define SHARE_VM_UTILITIES_TASKQUEUE_HPP
28 #include "memory/allocation.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "runtime/mutex.hpp"
31 #include "runtime/orderAccess.inline.hpp"
32 #include "utilities/stack.hpp"
34 // Simple TaskQueue stats that are collected by default in debug builds.
36 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
37 #define TASKQUEUE_STATS 1
38 #elif !defined(TASKQUEUE_STATS)
39 #define TASKQUEUE_STATS 0
40 #endif
42 #if TASKQUEUE_STATS
43 #define TASKQUEUE_STATS_ONLY(code) code
44 #else
45 #define TASKQUEUE_STATS_ONLY(code)
46 #endif // TASKQUEUE_STATS
48 #if TASKQUEUE_STATS
49 class TaskQueueStats {
50 public:
51 enum StatId {
52 push, // number of taskqueue pushes
53 pop, // number of taskqueue pops
54 pop_slow, // subset of taskqueue pops that were done slow-path
55 steal_attempt, // number of taskqueue steal attempts
56 steal, // number of taskqueue steals
57 overflow, // number of overflow pushes
58 overflow_max_len, // max length of overflow stack
59 last_stat_id
60 };
62 public:
63 inline TaskQueueStats() { reset(); }
65 inline void record_push() { ++_stats[push]; }
66 inline void record_pop() { ++_stats[pop]; }
67 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
68 inline void record_steal(bool success);
69 inline void record_overflow(size_t new_length);
71 TaskQueueStats & operator +=(const TaskQueueStats & addend);
73 inline size_t get(StatId id) const { return _stats[id]; }
74 inline const size_t* get() const { return _stats; }
76 inline void reset();
78 // Print the specified line of the header (does not include a line separator).
79 static void print_header(unsigned int line, outputStream* const stream = tty,
80 unsigned int width = 10);
81 // Print the statistics (does not include a line separator).
82 void print(outputStream* const stream = tty, unsigned int width = 10) const;
84 DEBUG_ONLY(void verify() const;)
86 private:
87 size_t _stats[last_stat_id];
88 static const char * const _names[last_stat_id];
89 };
91 void TaskQueueStats::record_steal(bool success) {
92 ++_stats[steal_attempt];
93 if (success) ++_stats[steal];
94 }
96 void TaskQueueStats::record_overflow(size_t new_len) {
97 ++_stats[overflow];
98 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
99 }
101 void TaskQueueStats::reset() {
102 memset(_stats, 0, sizeof(_stats));
103 }
104 #endif // TASKQUEUE_STATS
106 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances.
108 template <unsigned int N, MEMFLAGS F>
109 class TaskQueueSuper: public CHeapObj<F> {
110 protected:
111 // Internal type for indexing the queue; also used for the tag.
112 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
114 #ifdef MIPS64
115 private:
116 #endif
117 // The first free element after the last one pushed (mod N).
118 volatile uint _bottom;
120 #ifdef MIPS64
121 protected:
122 inline uint get_bottom() const {
123 return OrderAccess::load_acquire((volatile juint*)&_bottom);
124 }
126 inline void set_bottom(uint new_bottom) {
127 OrderAccess::release_store(&_bottom, new_bottom);
128 }
129 #endif
131 enum { MOD_N_MASK = N - 1 };
133 class Age {
134 public:
135 Age(size_t data = 0) { _data = data; }
136 Age(const Age& age) { _data = age._data; }
137 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; }
139 #ifndef MIPS64
140 Age get() const volatile { return _data; }
141 void set(Age age) volatile { _data = age._data; }
142 idx_t top() const volatile { return _fields._top; }
143 idx_t tag() const volatile { return _fields._tag; }
144 #else
145 Age get() const volatile {
146 size_t res = OrderAccess::load_ptr_acquire((volatile intptr_t*) &_data);
147 return *(Age*)(&res);
148 }
150 void set(Age age) volatile { OrderAccess::release_store_ptr((volatile intptr_t*) &_data, *(size_t*)(&age._data)); }
151 idx_t top() const volatile { return OrderAccess::load_acquire((volatile idx_t*) &(_fields._top)); }
152 idx_t tag() const volatile { return OrderAccess::load_acquire((volatile idx_t*) &(_fields._tag)); }
153 #endif
155 // Increment top; if it wraps, increment tag also.
156 void increment() {
157 _fields._top = increment_index(_fields._top);
158 if (_fields._top == 0) ++_fields._tag;
159 }
161 Age cmpxchg(const Age new_age, const Age old_age) volatile {
162 return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
163 (volatile intptr_t *)&_data,
164 (intptr_t)old_age._data);
165 }
167 bool operator ==(const Age& other) const { return _data == other._data; }
169 private:
170 struct fields {
171 idx_t _top;
172 idx_t _tag;
173 };
174 union {
175 size_t _data;
176 fields _fields;
177 };
178 };
180 volatile Age _age;
182 // These both operate mod N.
183 static uint increment_index(uint ind) {
184 return (ind + 1) & MOD_N_MASK;
185 }
186 static uint decrement_index(uint ind) {
187 return (ind - 1) & MOD_N_MASK;
188 }
190 // Returns a number in the range [0..N). If the result is "N-1", it should be
191 // interpreted as 0.
192 uint dirty_size(uint bot, uint top) const {
193 return (bot - top) & MOD_N_MASK;
194 }
196 // Returns the size corresponding to the given "bot" and "top".
197 uint size(uint bot, uint top) const {
198 uint sz = dirty_size(bot, top);
199 // Has the queue "wrapped", so that bottom is less than top? There's a
200 // complicated special case here. A pair of threads could perform pop_local
201 // and pop_global operations concurrently, starting from a state in which
202 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
203 // and the pop_global in incrementing _top (in which case the pop_global
204 // will be awarded the contested queue element.) The resulting state must
205 // be interpreted as an empty queue. (We only need to worry about one such
206 // event: only the queue owner performs pop_local's, and several concurrent
207 // threads attempting to perform the pop_global will all perform the same
208 // CAS, and only one can succeed.) Any stealing thread that reads after
209 // either the increment or decrement will see an empty queue, and will not
210 // join the competitors. The "sz == -1 || sz == N-1" state will not be
211 // modified by concurrent queues, so the owner thread can reset the state to
212 // _bottom == top so subsequent pushes will be performed normally.
213 return (sz == N - 1) ? 0 : sz;
214 }
216 public:
217 TaskQueueSuper() : _bottom(0), _age() {}
219 // Return true if the TaskQueue contains/does not contain any tasks.
220 bool peek() const {
221 #ifdef MIPS64
222 return get_bottom() != _age.top();
223 #else
224 return _bottom != _age.top();
225 #endif
226 }
228 bool is_empty() const { return size() == 0; }
230 // Return an estimate of the number of elements in the queue.
231 // The "careful" version admits the possibility of pop_local/pop_global
232 // races.
233 uint size() const {
234 #ifdef MIPS64
235 return size(get_bottom(), _age.top());
236 #else
237 return size(_bottom, _age.top());
238 #endif
239 }
241 uint dirty_size() const {
242 #ifdef MIPS64
243 return dirty_size(get_bottom(), _age.top());
244 #else
245 return dirty_size(_bottom, _age.top());
246 #endif
247 }
249 void set_empty() {
250 #ifdef MIPS64
251 set_bottom(0);
252 #else
253 _bottom = 0;
254 #endif
255 _age.set(0);
256 }
258 // Maximum number of elements allowed in the queue. This is two less
259 // than the actual queue size, for somewhat complicated reasons.
260 uint max_elems() const { return N - 2; }
262 // Total size of queue.
263 static const uint total_size() { return N; }
265 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
266 };
268 //
269 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
270 // ended-queue (deque), intended for use in work stealing. Queue operations
271 // are non-blocking.
272 //
273 // A queue owner thread performs push() and pop_local() operations on one end
274 // of the queue, while other threads may steal work using the pop_global()
275 // method.
276 //
277 // The main difference to the original algorithm is that this
278 // implementation allows wrap-around at the end of its allocated
279 // storage, which is an array.
280 //
281 // The original paper is:
282 //
283 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
284 // Thread scheduling for multiprogrammed multiprocessors.
285 // Theory of Computing Systems 34, 2 (2001), 115-144.
286 //
287 // The following paper provides an correctness proof and an
288 // implementation for weakly ordered memory models including (pseudo-)
289 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
290 // similar to ABP, with the main difference that it allows resizing of the
291 // underlying storage:
292 //
293 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
294 // Correct and efficient work-stealing for weak memory models
295 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
296 // practice of parallel programming (PPoPP 2013), 69-80
297 //
299 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
300 class GenericTaskQueue: public TaskQueueSuper<N, F> {
301 ArrayAllocator<E, F> _array_allocator;
302 protected:
303 typedef typename TaskQueueSuper<N, F>::Age Age;
304 typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
306 #ifndef MIPS64
307 using TaskQueueSuper<N, F>::_bottom;
308 #endif
309 using TaskQueueSuper<N, F>::_age;
310 using TaskQueueSuper<N, F>::increment_index;
311 using TaskQueueSuper<N, F>::decrement_index;
312 using TaskQueueSuper<N, F>::dirty_size;
314 public:
315 using TaskQueueSuper<N, F>::max_elems;
316 using TaskQueueSuper<N, F>::size;
318 #if TASKQUEUE_STATS
319 using TaskQueueSuper<N, F>::stats;
320 #endif
322 private:
323 // Slow paths for push, pop_local. (pop_global has no fast path.)
324 bool push_slow(E t, uint dirty_n_elems);
325 bool pop_local_slow(uint localBot, Age oldAge);
327 public:
328 typedef E element_type;
330 // Initializes the queue to empty.
331 GenericTaskQueue();
333 void initialize();
335 // Push the task "t" on the queue. Returns "false" iff the queue is full.
336 inline bool push(E t);
338 // Attempts to claim a task from the "local" end of the queue (the most
339 // recently pushed). If successful, returns true and sets t to the task;
340 // otherwise, returns false (the queue is empty).
341 inline bool pop_local(volatile E& t);
343 // Like pop_local(), but uses the "global" end of the queue (the least
344 // recently pushed).
345 bool pop_global(volatile E& t);
347 // Delete any resource associated with the queue.
348 ~GenericTaskQueue();
350 // apply the closure to all elements in the task queue
351 void oops_do(OopClosure* f);
353 private:
354 // Element array.
355 volatile E* _elems;
356 };
358 template<class E, MEMFLAGS F, unsigned int N>
359 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
360 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
361 }
363 template<class E, MEMFLAGS F, unsigned int N>
364 void GenericTaskQueue<E, F, N>::initialize() {
365 _elems = _array_allocator.allocate(N);
366 }
368 template<class E, MEMFLAGS F, unsigned int N>
369 void GenericTaskQueue<E, F, N>::oops_do(OopClosure* f) {
370 // tty->print_cr("START OopTaskQueue::oops_do");
371 uint iters = size();
372 #ifdef MIPS64
373 uint index = this->get_bottom();
374 #else
375 uint index = _bottom;
376 #endif
377 for (uint i = 0; i < iters; ++i) {
378 index = decrement_index(index);
379 // tty->print_cr(" doing entry %d," INTPTR_T " -> " INTPTR_T,
380 // index, &_elems[index], _elems[index]);
381 E* t = (E*)&_elems[index]; // cast away volatility
382 oop* p = (oop*)t;
383 assert((*t)->is_oop_or_null(), "Not an oop or null");
384 f->do_oop(p);
385 }
386 // tty->print_cr("END OopTaskQueue::oops_do");
387 }
389 template<class E, MEMFLAGS F, unsigned int N>
390 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) {
391 if (dirty_n_elems == N - 1) {
392 // Actually means 0, so do the push.
393 #ifdef MIPS64
394 uint localBot = this->get_bottom();
395 #else
396 uint localBot = _bottom;
397 #endif
398 // g++ complains if the volatile result of the assignment is
399 // unused, so we cast the volatile away. We cannot cast directly
400 // to void, because gcc treats that as not using the result of the
401 // assignment. However, casting to E& means that we trigger an
402 // unused-value warning. So, we cast the E& to void.
403 (void)const_cast<E&>(_elems[localBot] = t);
404 #ifdef MIPS64
405 this->set_bottom(increment_index(localBot));
406 #else
407 OrderAccess::release_store(&_bottom, increment_index(localBot));
408 #endif
409 TASKQUEUE_STATS_ONLY(stats.record_push());
410 return true;
411 }
412 return false;
413 }
415 // pop_local_slow() is done by the owning thread and is trying to
416 // get the last task in the queue. It will compete with pop_global()
417 // that will be used by other threads. The tag age is incremented
418 // whenever the queue goes empty which it will do here if this thread
419 // gets the last task or in pop_global() if the queue wraps (top == 0
420 // and pop_global() succeeds, see pop_global()).
421 template<class E, MEMFLAGS F, unsigned int N>
422 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) {
423 // This queue was observed to contain exactly one element; either this
424 // thread will claim it, or a competing "pop_global". In either case,
425 // the queue will be logically empty afterwards. Create a new Age value
426 // that represents the empty queue for the given value of "_bottom". (We
427 // must also increment "tag" because of the case where "bottom == 1",
428 // "top == 0". A pop_global could read the queue element in that case,
429 // then have the owner thread do a pop followed by another push. Without
430 // the incrementing of "tag", the pop_global's CAS could succeed,
431 // allowing it to believe it has claimed the stale element.)
432 Age newAge((idx_t)localBot, oldAge.tag() + 1);
433 // Perhaps a competing pop_global has already incremented "top", in which
434 // case it wins the element.
435 if (localBot == oldAge.top()) {
436 // No competing pop_global has yet incremented "top"; we'll try to
437 // install new_age, thus claiming the element.
438 Age tempAge = _age.cmpxchg(newAge, oldAge);
439 if (tempAge == oldAge) {
440 // We win.
441 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
442 TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
443 return true;
444 }
445 }
446 // We lose; a completing pop_global gets the element. But the queue is empty
447 // and top is greater than bottom. Fix this representation of the empty queue
448 // to become the canonical one.
449 _age.set(newAge);
450 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
451 return false;
452 }
454 template<class E, MEMFLAGS F, unsigned int N>
455 bool GenericTaskQueue<E, F, N>::pop_global(volatile E& t) {
456 Age oldAge = _age.get();
457 // Architectures with weak memory model require a barrier here
458 // to guarantee that bottom is not older than age,
459 // which is crucial for the correctness of the algorithm.
460 #if !(defined SPARC || defined IA32 || defined AMD64)
461 OrderAccess::fence();
462 #endif
463 #ifdef MIPS64
464 uint localBot = this->get_bottom();
465 #else
466 uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom);
467 #endif
468 uint n_elems = size(localBot, oldAge.top());
469 if (n_elems == 0) {
470 return false;
471 }
473 // g++ complains if the volatile result of the assignment is
474 // unused, so we cast the volatile away. We cannot cast directly
475 // to void, because gcc treats that as not using the result of the
476 // assignment. However, casting to E& means that we trigger an
477 // unused-value warning. So, we cast the E& to void.
478 (void) const_cast<E&>(t = _elems[oldAge.top()]);
479 Age newAge(oldAge);
480 newAge.increment();
481 Age resAge = _age.cmpxchg(newAge, oldAge);
483 // Note that using "_bottom" here might fail, since a pop_local might
484 // have decremented it.
485 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
486 return resAge == oldAge;
487 }
489 template<class E, MEMFLAGS F, unsigned int N>
490 GenericTaskQueue<E, F, N>::~GenericTaskQueue() {
491 FREE_C_HEAP_ARRAY(E, _elems, F);
492 }
494 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
495 // elements that do not fit in the TaskQueue.
496 //
497 // This class hides two methods from super classes:
498 //
499 // push() - push onto the task queue or, if that fails, onto the overflow stack
500 // is_empty() - return true if both the TaskQueue and overflow stack are empty
501 //
502 // Note that size() is not hidden--it returns the number of elements in the
503 // TaskQueue, and does not include the size of the overflow stack. This
504 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
505 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
506 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
507 {
508 public:
509 typedef Stack<E, F> overflow_t;
510 typedef GenericTaskQueue<E, F, N> taskqueue_t;
512 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
514 // Push task t onto the queue or onto the overflow stack. Return true.
515 inline bool push(E t);
517 // Attempt to pop from the overflow stack; return true if anything was popped.
518 inline bool pop_overflow(E& t);
520 inline overflow_t* overflow_stack() { return &_overflow_stack; }
522 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
523 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
524 inline bool is_empty() const {
525 return taskqueue_empty() && overflow_empty();
526 }
528 private:
529 overflow_t _overflow_stack;
530 };
532 template <class E, MEMFLAGS F, unsigned int N>
533 bool OverflowTaskQueue<E, F, N>::push(E t)
534 {
535 if (!taskqueue_t::push(t)) {
536 overflow_stack()->push(t);
537 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
538 }
539 return true;
540 }
542 template <class E, MEMFLAGS F, unsigned int N>
543 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t)
544 {
545 if (overflow_empty()) return false;
546 t = overflow_stack()->pop();
547 return true;
548 }
550 class TaskQueueSetSuper {
551 protected:
552 static int randomParkAndMiller(int* seed0);
553 public:
554 // Returns "true" if some TaskQueue in the set contains a task.
555 virtual bool peek() = 0;
556 };
558 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
559 };
561 template<class T, MEMFLAGS F>
562 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
563 private:
564 uint _n;
565 T** _queues;
567 public:
568 typedef typename T::element_type E;
570 GenericTaskQueueSet(int n) : _n(n) {
571 typedef T* GenericTaskQueuePtr;
572 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F);
573 for (int i = 0; i < n; i++) {
574 _queues[i] = NULL;
575 }
576 }
578 bool steal_best_of_2(uint queue_num, int* seed, E& t);
580 void register_queue(uint i, T* q);
582 T* queue(uint n);
584 // The thread with queue number "queue_num" (and whose random number seed is
585 // at "seed") is trying to steal a task from some other queue. (It may try
586 // several queues, according to some configuration parameter.) If some steal
587 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
588 // false.
589 bool steal(uint queue_num, int* seed, E& t);
591 bool peek();
592 };
594 template<class T, MEMFLAGS F> void
595 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
596 assert(i < _n, "index out of range.");
597 _queues[i] = q;
598 }
600 template<class T, MEMFLAGS F> T*
601 GenericTaskQueueSet<T, F>::queue(uint i) {
602 return _queues[i];
603 }
605 template<class T, MEMFLAGS F> bool
606 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
607 for (uint i = 0; i < 2 * _n; i++) {
608 if (steal_best_of_2(queue_num, seed, t)) {
609 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
610 return true;
611 }
612 }
613 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
614 return false;
615 }
617 template<class T, MEMFLAGS F> bool
618 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) {
619 if (_n > 2) {
620 uint k1 = queue_num;
621 while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
622 uint k2 = queue_num;
623 while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
624 // Sample both and try the larger.
625 uint sz1 = _queues[k1]->size();
626 uint sz2 = _queues[k2]->size();
627 if (sz2 > sz1) return _queues[k2]->pop_global(t);
628 else return _queues[k1]->pop_global(t);
629 } else if (_n == 2) {
630 // Just try the other one.
631 uint k = (queue_num + 1) % 2;
632 return _queues[k]->pop_global(t);
633 } else {
634 assert(_n == 1, "can't be zero.");
635 return false;
636 }
637 }
639 template<class T, MEMFLAGS F>
640 bool GenericTaskQueueSet<T, F>::peek() {
641 // Try all the queues.
642 for (uint j = 0; j < _n; j++) {
643 if (_queues[j]->peek())
644 return true;
645 }
646 return false;
647 }
649 // When to terminate from the termination protocol.
650 class TerminatorTerminator: public CHeapObj<mtInternal> {
651 public:
652 virtual bool should_exit_termination() = 0;
653 };
655 // A class to aid in the termination of a set of parallel tasks using
656 // TaskQueueSet's for work stealing.
658 #undef TRACESPINNING
660 class ParallelTaskTerminator: public StackObj {
661 private:
662 int _n_threads;
663 TaskQueueSetSuper* _queue_set;
664 int _offered_termination;
666 #ifdef TRACESPINNING
667 static uint _total_yields;
668 static uint _total_spins;
669 static uint _total_peeks;
670 #endif
672 bool peek_in_queue_set();
673 protected:
674 virtual void yield();
675 void sleep(uint millis);
677 public:
679 // "n_threads" is the number of threads to be terminated. "queue_set" is a
680 // queue sets of work queues of other threads.
681 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
683 // The current thread has no work, and is ready to terminate if everyone
684 // else is. If returns "true", all threads are terminated. If returns
685 // "false", available work has been observed in one of the task queues,
686 // so the global task is not complete.
687 bool offer_termination() {
688 return offer_termination(NULL);
689 }
691 // As above, but it also terminates if the should_exit_termination()
692 // method of the terminator parameter returns true. If terminator is
693 // NULL, then it is ignored.
694 bool offer_termination(TerminatorTerminator* terminator);
696 // Reset the terminator, so that it may be reused again.
697 // The caller is responsible for ensuring that this is done
698 // in an MT-safe manner, once the previous round of use of
699 // the terminator is finished.
700 void reset_for_reuse();
701 // Same as above but the number of parallel threads is set to the
702 // given number.
703 void reset_for_reuse(int n_threads);
705 #ifdef TRACESPINNING
706 static uint total_yields() { return _total_yields; }
707 static uint total_spins() { return _total_spins; }
708 static uint total_peeks() { return _total_peeks; }
709 static void print_termination_counts();
710 #endif
711 };
713 template<class E, MEMFLAGS F, unsigned int N> inline bool
714 GenericTaskQueue<E, F, N>::push(E t) {
715 #ifdef MIPS64
716 uint localBot = this->get_bottom();
717 #else
718 uint localBot = _bottom;
719 #endif
720 assert(localBot < N, "_bottom out of range.");
721 idx_t top = _age.top();
722 uint dirty_n_elems = dirty_size(localBot, top);
723 assert(dirty_n_elems < N, "n_elems out of range.");
724 if (dirty_n_elems < max_elems()) {
725 // g++ complains if the volatile result of the assignment is
726 // unused, so we cast the volatile away. We cannot cast directly
727 // to void, because gcc treats that as not using the result of the
728 // assignment. However, casting to E& means that we trigger an
729 // unused-value warning. So, we cast the E& to void.
730 (void) const_cast<E&>(_elems[localBot] = t);
731 #ifdef MIPS64
732 this->set_bottom(increment_index(localBot));
733 #else
734 OrderAccess::release_store(&_bottom, increment_index(localBot));
735 #endif
736 TASKQUEUE_STATS_ONLY(stats.record_push());
737 return true;
738 } else {
739 return push_slow(t, dirty_n_elems);
740 }
741 }
743 template<class E, MEMFLAGS F, unsigned int N> inline bool
744 GenericTaskQueue<E, F, N>::pop_local(volatile E& t) {
745 #ifdef MIPS64
746 uint localBot = this->get_bottom();
747 #else
748 uint localBot = _bottom;
749 #endif
750 // This value cannot be N-1. That can only occur as a result of
751 // the assignment to bottom in this method. If it does, this method
752 // resets the size to 0 before the next call (which is sequential,
753 // since this is pop_local.)
754 uint dirty_n_elems = dirty_size(localBot, _age.top());
755 assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
756 if (dirty_n_elems == 0) return false;
757 localBot = decrement_index(localBot);
758 #ifdef MIPS64
759 this->set_bottom(localBot);
760 #else
761 _bottom = localBot;
762 #endif
763 // This is necessary to prevent any read below from being reordered
764 // before the store just above.
765 OrderAccess::fence();
766 // g++ complains if the volatile result of the assignment is
767 // unused, so we cast the volatile away. We cannot cast directly
768 // to void, because gcc treats that as not using the result of the
769 // assignment. However, casting to E& means that we trigger an
770 // unused-value warning. So, we cast the E& to void.
771 (void) const_cast<E&>(t = _elems[localBot]);
772 // This is a second read of "age"; the "size()" above is the first.
773 // If there's still at least one element in the queue, based on the
774 // "_bottom" and "age" we've read, then there can be no interference with
775 // a "pop_global" operation, and we're done.
776 idx_t tp = _age.top(); // XXX
777 if (size(localBot, tp) > 0) {
778 assert(dirty_size(localBot, tp) != N - 1, "sanity");
779 TASKQUEUE_STATS_ONLY(stats.record_pop());
780 return true;
781 } else {
782 // Otherwise, the queue contained exactly one element; we take the slow
783 // path.
784 return pop_local_slow(localBot, _age.get());
785 }
786 }
788 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue;
789 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
791 #ifdef _MSC_VER
792 #pragma warning(push)
793 // warning C4522: multiple assignment operators specified
794 #pragma warning(disable:4522)
795 #endif
797 // This is a container class for either an oop* or a narrowOop*.
798 // Both are pushed onto a task queue and the consumer will test is_narrow()
799 // to determine which should be processed.
800 class StarTask {
801 void* _holder; // either union oop* or narrowOop*
803 enum { COMPRESSED_OOP_MASK = 1 };
805 public:
806 StarTask(narrowOop* p) {
807 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
808 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
809 }
810 StarTask(oop* p) {
811 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
812 _holder = (void*)p;
813 }
814 StarTask() { _holder = NULL; }
815 operator oop*() { return (oop*)_holder; }
816 operator narrowOop*() {
817 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
818 }
820 StarTask& operator=(const StarTask& t) {
821 _holder = t._holder;
822 return *this;
823 }
824 volatile StarTask& operator=(const volatile StarTask& t) volatile {
825 _holder = t._holder;
826 return *this;
827 }
829 bool is_narrow() const {
830 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
831 }
832 };
834 class ObjArrayTask
835 {
836 public:
837 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
838 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
839 assert(idx <= size_t(max_jint), "too big");
840 }
841 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
843 ObjArrayTask& operator =(const ObjArrayTask& t) {
844 _obj = t._obj;
845 _index = t._index;
846 return *this;
847 }
848 volatile ObjArrayTask&
849 operator =(const volatile ObjArrayTask& t) volatile {
850 (void)const_cast<oop&>(_obj = t._obj);
851 _index = t._index;
852 return *this;
853 }
855 inline oop obj() const { return _obj; }
856 inline int index() const { return _index; }
858 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
860 private:
861 oop _obj;
862 int _index;
863 };
865 #ifdef _MSC_VER
866 #pragma warning(pop)
867 #endif
869 typedef OverflowTaskQueue<StarTask, mtClass> OopStarTaskQueue;
870 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet;
872 typedef OverflowTaskQueue<size_t, mtInternal> RegionTaskQueue;
873 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass> RegionTaskQueueSet;
876 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP