Tue, 04 Feb 2020 18:13:14 +0800
Merge
1 /*
2 * Copyright (c) 2001, 2016, 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/globalDefinitions.hpp"
33 #include "utilities/stack.hpp"
35 // Simple TaskQueue stats that are collected by default in debug builds.
37 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
38 #define TASKQUEUE_STATS 1
39 #elif !defined(TASKQUEUE_STATS)
40 #define TASKQUEUE_STATS 0
41 #endif
43 #if TASKQUEUE_STATS
44 #define TASKQUEUE_STATS_ONLY(code) code
45 #else
46 #define TASKQUEUE_STATS_ONLY(code)
47 #endif // TASKQUEUE_STATS
49 #if TASKQUEUE_STATS
50 class TaskQueueStats {
51 public:
52 enum StatId {
53 push, // number of taskqueue pushes
54 pop, // number of taskqueue pops
55 pop_slow, // subset of taskqueue pops that were done slow-path
56 steal_attempt, // number of taskqueue steal attempts
57 steal, // number of taskqueue steals
58 overflow, // number of overflow pushes
59 overflow_max_len, // max length of overflow stack
60 last_stat_id
61 };
63 public:
64 inline TaskQueueStats() { reset(); }
66 inline void record_push() { ++_stats[push]; }
67 inline void record_pop() { ++_stats[pop]; }
68 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
69 inline void record_steal(bool success);
70 inline void record_overflow(size_t new_length);
72 TaskQueueStats & operator +=(const TaskQueueStats & addend);
74 inline size_t get(StatId id) const { return _stats[id]; }
75 inline const size_t* get() const { return _stats; }
77 inline void reset();
79 // Print the specified line of the header (does not include a line separator).
80 static void print_header(unsigned int line, outputStream* const stream = tty,
81 unsigned int width = 10);
82 // Print the statistics (does not include a line separator).
83 void print(outputStream* const stream = tty, unsigned int width = 10) const;
85 DEBUG_ONLY(void verify() const;)
87 private:
88 size_t _stats[last_stat_id];
89 static const char * const _names[last_stat_id];
90 };
92 void TaskQueueStats::record_steal(bool success) {
93 ++_stats[steal_attempt];
94 if (success) ++_stats[steal];
95 }
97 void TaskQueueStats::record_overflow(size_t new_len) {
98 ++_stats[overflow];
99 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
100 }
102 void TaskQueueStats::reset() {
103 memset(_stats, 0, sizeof(_stats));
104 }
105 #endif // TASKQUEUE_STATS
107 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances.
109 template <unsigned int N, MEMFLAGS F>
110 class TaskQueueSuper: public CHeapObj<F> {
111 protected:
112 // Internal type for indexing the queue; also used for the tag.
113 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
115 #ifdef MIPS
116 private:
117 #endif
118 // The first free element after the last one pushed (mod N).
119 volatile uint _bottom;
121 #ifdef MIPS
122 protected:
123 inline uint get_bottom() const {
124 return OrderAccess::load_acquire((volatile juint*)&_bottom);
125 }
127 inline void set_bottom(uint new_bottom) {
128 OrderAccess::release_store(&_bottom, new_bottom);
129 }
130 #endif
132 enum { MOD_N_MASK = N - 1 };
134 class Age {
135 public:
136 Age(size_t data = 0) { _data = data; }
137 Age(const Age& age) { _data = age._data; }
138 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; }
140 #ifndef MIPS
141 Age get() const volatile { return _data; }
142 void set(Age age) volatile { _data = age._data; }
144 idx_t top() const volatile { return _fields._top; }
145 idx_t tag() const volatile { return _fields._tag; }
146 #else
147 Age get() const volatile {
148 size_t res = OrderAccess::load_ptr_acquire((volatile intptr_t*) &_data);
149 return *(Age*)(&res);
150 }
151 void set(Age age) volatile { OrderAccess::release_store_ptr((volatile intptr_t*) &_data, *(size_t*)(&age._data)); }
153 idx_t top() const volatile { return OrderAccess::load_acquire((volatile idx_t*) &(_fields._top)); }
154 idx_t tag() const volatile { return OrderAccess::load_acquire((volatile idx_t*) &(_fields._tag)); }
155 #endif
157 // Increment top; if it wraps, increment tag also.
158 void increment() {
159 _fields._top = increment_index(_fields._top);
160 if (_fields._top == 0) ++_fields._tag;
161 }
163 Age cmpxchg(const Age new_age, const Age old_age) volatile {
164 return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
165 (volatile intptr_t *)&_data,
166 (intptr_t)old_age._data);
167 }
169 bool operator ==(const Age& other) const { return _data == other._data; }
171 private:
172 struct fields {
173 idx_t _top;
174 idx_t _tag;
175 };
176 union {
177 size_t _data;
178 fields _fields;
179 };
180 };
182 volatile Age _age;
184 // These both operate mod N.
185 static uint increment_index(uint ind) {
186 return (ind + 1) & MOD_N_MASK;
187 }
188 static uint decrement_index(uint ind) {
189 return (ind - 1) & MOD_N_MASK;
190 }
192 // Returns a number in the range [0..N). If the result is "N-1", it should be
193 // interpreted as 0.
194 uint dirty_size(uint bot, uint top) const {
195 return (bot - top) & MOD_N_MASK;
196 }
198 // Returns the size corresponding to the given "bot" and "top".
199 uint size(uint bot, uint top) const {
200 uint sz = dirty_size(bot, top);
201 // Has the queue "wrapped", so that bottom is less than top? There's a
202 // complicated special case here. A pair of threads could perform pop_local
203 // and pop_global operations concurrently, starting from a state in which
204 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
205 // and the pop_global in incrementing _top (in which case the pop_global
206 // will be awarded the contested queue element.) The resulting state must
207 // be interpreted as an empty queue. (We only need to worry about one such
208 // event: only the queue owner performs pop_local's, and several concurrent
209 // threads attempting to perform the pop_global will all perform the same
210 // CAS, and only one can succeed.) Any stealing thread that reads after
211 // either the increment or decrement will see an empty queue, and will not
212 // join the competitors. The "sz == -1 || sz == N-1" state will not be
213 // modified by concurrent queues, so the owner thread can reset the state to
214 // _bottom == top so subsequent pushes will be performed normally.
215 return (sz == N - 1) ? 0 : sz;
216 }
218 public:
219 TaskQueueSuper() : _bottom(0), _age() {}
221 // Return true if the TaskQueue contains/does not contain any tasks.
222 bool peek() const {
223 #ifdef MIPS
224 return get_bottom() != _age.top();
225 #else
226 return _bottom != _age.top();
227 #endif
228 }
229 bool is_empty() const { return size() == 0; }
231 // Return an estimate of the number of elements in the queue.
232 // The "careful" version admits the possibility of pop_local/pop_global
233 // races.
234 uint size() const {
235 #ifdef MIPS
236 return size(get_bottom(), _age.top());
237 #else
238 return size(_bottom, _age.top());
239 #endif
240 }
242 uint dirty_size() const {
243 #ifdef MIPS
244 return dirty_size(get_bottom(), _age.top());
245 #else
246 return dirty_size(_bottom, _age.top());
247 #endif
248 }
250 void set_empty() {
251 #ifdef MIPS
252 set_bottom(0);
253 #else
254 _bottom = 0;
255 #endif
256 _age.set(0);
257 }
259 // Maximum number of elements allowed in the queue. This is two less
260 // than the actual queue size, for somewhat complicated reasons.
261 uint max_elems() const { return N - 2; }
263 // Total size of queue.
264 static const uint total_size() { return N; }
266 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
267 };
269 //
270 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
271 // ended-queue (deque), intended for use in work stealing. Queue operations
272 // are non-blocking.
273 //
274 // A queue owner thread performs push() and pop_local() operations on one end
275 // of the queue, while other threads may steal work using the pop_global()
276 // method.
277 //
278 // The main difference to the original algorithm is that this
279 // implementation allows wrap-around at the end of its allocated
280 // storage, which is an array.
281 //
282 // The original paper is:
283 //
284 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
285 // Thread scheduling for multiprogrammed multiprocessors.
286 // Theory of Computing Systems 34, 2 (2001), 115-144.
287 //
288 // The following paper provides an correctness proof and an
289 // implementation for weakly ordered memory models including (pseudo-)
290 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
291 // similar to ABP, with the main difference that it allows resizing of the
292 // underlying storage:
293 //
294 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
295 // Correct and efficient work-stealing for weak memory models
296 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
297 // practice of parallel programming (PPoPP 2013), 69-80
298 //
300 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
301 class GenericTaskQueue: public TaskQueueSuper<N, F> {
302 ArrayAllocator<E, F> _array_allocator;
303 protected:
304 typedef typename TaskQueueSuper<N, F>::Age Age;
305 typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
307 #ifndef MIPS
308 using TaskQueueSuper<N, F>::_bottom;
309 #endif
310 using TaskQueueSuper<N, F>::_age;
311 using TaskQueueSuper<N, F>::increment_index;
312 using TaskQueueSuper<N, F>::decrement_index;
313 using TaskQueueSuper<N, F>::dirty_size;
315 public:
316 using TaskQueueSuper<N, F>::max_elems;
317 using TaskQueueSuper<N, F>::size;
319 #if TASKQUEUE_STATS
320 using TaskQueueSuper<N, F>::stats;
321 #endif
323 private:
324 // Slow paths for push, pop_local. (pop_global has no fast path.)
325 bool push_slow(E t, uint dirty_n_elems);
326 bool pop_local_slow(uint localBot, Age oldAge);
328 public:
329 typedef E element_type;
331 // Initializes the queue to empty.
332 GenericTaskQueue();
334 void initialize();
336 // Push the task "t" on the queue. Returns "false" iff the queue is full.
337 inline bool push(E t);
339 // Attempts to claim a task from the "local" end of the queue (the most
340 // recently pushed). If successful, returns true and sets t to the task;
341 // otherwise, returns false (the queue is empty).
342 inline bool pop_local(volatile E& t);
344 // Like pop_local(), but uses the "global" end of the queue (the least
345 // recently pushed).
346 bool pop_global(volatile E& t);
348 // Delete any resource associated with the queue.
349 ~GenericTaskQueue();
351 // apply the closure to all elements in the task queue
352 void oops_do(OopClosure* f);
354 private:
355 // Element array.
356 volatile E* _elems;
357 };
359 template<class E, MEMFLAGS F, unsigned int N>
360 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
361 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
362 }
364 template<class E, MEMFLAGS F, unsigned int N>
365 void GenericTaskQueue<E, F, N>::initialize() {
366 _elems = _array_allocator.allocate(N);
367 }
369 template<class E, MEMFLAGS F, unsigned int N>
370 void GenericTaskQueue<E, F, N>::oops_do(OopClosure* f) {
371 // tty->print_cr("START OopTaskQueue::oops_do");
372 uint iters = size();
373 #ifdef MIPS
374 uint index = this->get_bottom();
375 #else
376 uint index = _bottom;
377 #endif
378 for (uint i = 0; i < iters; ++i) {
379 index = decrement_index(index);
380 // tty->print_cr(" doing entry %d," INTPTR_T " -> " INTPTR_T,
381 // index, &_elems[index], _elems[index]);
382 E* t = (E*)&_elems[index]; // cast away volatility
383 oop* p = (oop*)t;
384 assert((*t)->is_oop_or_null(), "Not an oop or null");
385 f->do_oop(p);
386 }
387 // tty->print_cr("END OopTaskQueue::oops_do");
388 }
390 template<class E, MEMFLAGS F, unsigned int N>
391 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) {
392 if (dirty_n_elems == N - 1) {
393 // Actually means 0, so do the push.
394 #ifdef MIPS
395 uint localBot = this->get_bottom();
396 #else
397 uint localBot = _bottom;
398 #endif
399 // g++ complains if the volatile result of the assignment is
400 // unused, so we cast the volatile away. We cannot cast directly
401 // to void, because gcc treats that as not using the result of the
402 // assignment. However, casting to E& means that we trigger an
403 // unused-value warning. So, we cast the E& to void.
404 (void)const_cast<E&>(_elems[localBot] = t);
405 #ifdef MIPS
406 this->set_bottom(increment_index(localBot));
407 #else
408 OrderAccess::release_store(&_bottom, increment_index(localBot));
409 #endif
410 TASKQUEUE_STATS_ONLY(stats.record_push());
411 return true;
412 }
413 return false;
414 }
416 // pop_local_slow() is done by the owning thread and is trying to
417 // get the last task in the queue. It will compete with pop_global()
418 // that will be used by other threads. The tag age is incremented
419 // whenever the queue goes empty which it will do here if this thread
420 // gets the last task or in pop_global() if the queue wraps (top == 0
421 // and pop_global() succeeds, see pop_global()).
422 template<class E, MEMFLAGS F, unsigned int N>
423 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) {
424 // This queue was observed to contain exactly one element; either this
425 // thread will claim it, or a competing "pop_global". In either case,
426 // the queue will be logically empty afterwards. Create a new Age value
427 // that represents the empty queue for the given value of "_bottom". (We
428 // must also increment "tag" because of the case where "bottom == 1",
429 // "top == 0". A pop_global could read the queue element in that case,
430 // then have the owner thread do a pop followed by another push. Without
431 // the incrementing of "tag", the pop_global's CAS could succeed,
432 // allowing it to believe it has claimed the stale element.)
433 Age newAge((idx_t)localBot, oldAge.tag() + 1);
434 // Perhaps a competing pop_global has already incremented "top", in which
435 // case it wins the element.
436 if (localBot == oldAge.top()) {
437 // No competing pop_global has yet incremented "top"; we'll try to
438 // install new_age, thus claiming the element.
439 Age tempAge = _age.cmpxchg(newAge, oldAge);
440 if (tempAge == oldAge) {
441 // We win.
442 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
443 TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
444 return true;
445 }
446 }
447 // We lose; a completing pop_global gets the element. But the queue is empty
448 // and top is greater than bottom. Fix this representation of the empty queue
449 // to become the canonical one.
450 _age.set(newAge);
451 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
452 return false;
453 }
455 template<class E, MEMFLAGS F, unsigned int N>
456 bool GenericTaskQueue<E, F, N>::pop_global(volatile E& t) {
457 Age oldAge = _age.get();
458 // Architectures with weak memory model require a barrier here
459 // to guarantee that bottom is not older than age,
460 // which is crucial for the correctness of the algorithm.
461 #if !(defined SPARC || defined IA32 || defined AMD64)
462 OrderAccess::fence();
463 #endif
464 #ifdef MIPS
465 uint localBot = this->get_bottom();
466 #else
467 uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom);
468 #endif
469 uint n_elems = size(localBot, oldAge.top());
470 if (n_elems == 0) {
471 return false;
472 }
474 // g++ complains if the volatile result of the assignment is
475 // unused, so we cast the volatile away. We cannot cast directly
476 // to void, because gcc treats that as not using the result of the
477 // assignment. However, casting to E& means that we trigger an
478 // unused-value warning. So, we cast the E& to void.
479 (void) const_cast<E&>(t = _elems[oldAge.top()]);
480 Age newAge(oldAge);
481 newAge.increment();
482 Age resAge = _age.cmpxchg(newAge, oldAge);
484 // Note that using "_bottom" here might fail, since a pop_local might
485 // have decremented it.
486 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
487 return resAge == oldAge;
488 }
490 template<class E, MEMFLAGS F, unsigned int N>
491 GenericTaskQueue<E, F, N>::~GenericTaskQueue() {
492 FREE_C_HEAP_ARRAY(E, _elems, F);
493 }
495 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
496 // elements that do not fit in the TaskQueue.
497 //
498 // This class hides two methods from super classes:
499 //
500 // push() - push onto the task queue or, if that fails, onto the overflow stack
501 // is_empty() - return true if both the TaskQueue and overflow stack are empty
502 //
503 // Note that size() is not hidden--it returns the number of elements in the
504 // TaskQueue, and does not include the size of the overflow stack. This
505 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
506 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
507 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
508 {
509 public:
510 typedef Stack<E, F> overflow_t;
511 typedef GenericTaskQueue<E, F, N> taskqueue_t;
513 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
515 // Push task t onto the queue or onto the overflow stack. Return true.
516 inline bool push(E t);
518 // Try to push task t onto the queue only. Returns true if successful, false otherwise.
519 inline bool try_push_to_taskqueue(E t);
521 // Attempt to pop from the overflow stack; return true if anything was popped.
522 inline bool pop_overflow(E& t);
524 inline overflow_t* overflow_stack() { return &_overflow_stack; }
526 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
527 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
528 inline bool is_empty() const {
529 return taskqueue_empty() && overflow_empty();
530 }
532 private:
533 overflow_t _overflow_stack;
534 };
536 template <class E, MEMFLAGS F, unsigned int N>
537 bool OverflowTaskQueue<E, F, N>::push(E t)
538 {
539 if (!taskqueue_t::push(t)) {
540 overflow_stack()->push(t);
541 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
542 }
543 return true;
544 }
546 template <class E, MEMFLAGS F, unsigned int N>
547 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t)
548 {
549 if (overflow_empty()) return false;
550 t = overflow_stack()->pop();
551 return true;
552 }
554 template <class E, MEMFLAGS F, unsigned int N>
555 bool OverflowTaskQueue<E, F, N>::try_push_to_taskqueue(E t) {
556 return taskqueue_t::push(t);
557 }
558 class TaskQueueSetSuper {
559 protected:
560 static int randomParkAndMiller(int* seed0);
561 public:
562 // Returns "true" if some TaskQueue in the set contains a task.
563 virtual bool peek() = 0;
564 };
566 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
567 };
569 template<class T, MEMFLAGS F>
570 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
571 private:
572 uint _n;
573 T** _queues;
575 public:
576 typedef typename T::element_type E;
578 GenericTaskQueueSet(int n) : _n(n) {
579 typedef T* GenericTaskQueuePtr;
580 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F);
581 for (int i = 0; i < n; i++) {
582 _queues[i] = NULL;
583 }
584 }
586 bool steal_best_of_2(uint queue_num, int* seed, E& t);
588 void register_queue(uint i, T* q);
590 T* queue(uint n);
592 // The thread with queue number "queue_num" (and whose random number seed is
593 // at "seed") is trying to steal a task from some other queue. (It may try
594 // several queues, according to some configuration parameter.) If some steal
595 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
596 // false.
597 bool steal(uint queue_num, int* seed, E& t);
599 bool peek();
600 };
602 template<class T, MEMFLAGS F> void
603 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
604 assert(i < _n, "index out of range.");
605 _queues[i] = q;
606 }
608 template<class T, MEMFLAGS F> T*
609 GenericTaskQueueSet<T, F>::queue(uint i) {
610 return _queues[i];
611 }
613 template<class T, MEMFLAGS F> bool
614 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
615 for (uint i = 0; i < 2 * _n; i++) {
616 if (steal_best_of_2(queue_num, seed, t)) {
617 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
618 return true;
619 }
620 }
621 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
622 return false;
623 }
625 template<class T, MEMFLAGS F> bool
626 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) {
627 if (_n > 2) {
628 uint k1 = queue_num;
629 while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
630 uint k2 = queue_num;
631 while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
632 // Sample both and try the larger.
633 uint sz1 = _queues[k1]->size();
634 uint sz2 = _queues[k2]->size();
635 if (sz2 > sz1) return _queues[k2]->pop_global(t);
636 else return _queues[k1]->pop_global(t);
637 } else if (_n == 2) {
638 // Just try the other one.
639 uint k = (queue_num + 1) % 2;
640 return _queues[k]->pop_global(t);
641 } else {
642 assert(_n == 1, "can't be zero.");
643 return false;
644 }
645 }
647 template<class T, MEMFLAGS F>
648 bool GenericTaskQueueSet<T, F>::peek() {
649 // Try all the queues.
650 for (uint j = 0; j < _n; j++) {
651 if (_queues[j]->peek())
652 return true;
653 }
654 return false;
655 }
657 // When to terminate from the termination protocol.
658 class TerminatorTerminator: public CHeapObj<mtInternal> {
659 public:
660 virtual bool should_exit_termination() = 0;
661 };
663 // A class to aid in the termination of a set of parallel tasks using
664 // TaskQueueSet's for work stealing.
666 #undef TRACESPINNING
668 class ParallelTaskTerminator: public StackObj {
669 private:
670 int _n_threads;
671 TaskQueueSetSuper* _queue_set;
672 char _pad_before[DEFAULT_CACHE_LINE_SIZE];
673 int _offered_termination;
674 char _pad_after[DEFAULT_CACHE_LINE_SIZE];
676 #ifdef TRACESPINNING
677 static uint _total_yields;
678 static uint _total_spins;
679 static uint _total_peeks;
680 #endif
682 bool peek_in_queue_set();
683 protected:
684 virtual void yield();
685 void sleep(uint millis);
687 public:
689 // "n_threads" is the number of threads to be terminated. "queue_set" is a
690 // queue sets of work queues of other threads.
691 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
693 // The current thread has no work, and is ready to terminate if everyone
694 // else is. If returns "true", all threads are terminated. If returns
695 // "false", available work has been observed in one of the task queues,
696 // so the global task is not complete.
697 bool offer_termination() {
698 return offer_termination(NULL);
699 }
701 // As above, but it also terminates if the should_exit_termination()
702 // method of the terminator parameter returns true. If terminator is
703 // NULL, then it is ignored.
704 bool offer_termination(TerminatorTerminator* terminator);
706 // Reset the terminator, so that it may be reused again.
707 // The caller is responsible for ensuring that this is done
708 // in an MT-safe manner, once the previous round of use of
709 // the terminator is finished.
710 void reset_for_reuse();
711 // Same as above but the number of parallel threads is set to the
712 // given number.
713 void reset_for_reuse(int n_threads);
715 #ifdef TRACESPINNING
716 static uint total_yields() { return _total_yields; }
717 static uint total_spins() { return _total_spins; }
718 static uint total_peeks() { return _total_peeks; }
719 static void print_termination_counts();
720 #endif
721 };
723 template<class E, MEMFLAGS F, unsigned int N> inline bool
724 GenericTaskQueue<E, F, N>::push(E t) {
725 #ifdef MIPS
726 uint localBot = this->get_bottom();
727 #else
728 uint localBot = _bottom;
729 #endif
730 assert(localBot < N, "_bottom out of range.");
731 idx_t top = _age.top();
732 uint dirty_n_elems = dirty_size(localBot, top);
733 assert(dirty_n_elems < N, "n_elems out of range.");
734 if (dirty_n_elems < max_elems()) {
735 // g++ complains if the volatile result of the assignment is
736 // unused, so we cast the volatile away. We cannot cast directly
737 // to void, because gcc treats that as not using the result of the
738 // assignment. However, casting to E& means that we trigger an
739 // unused-value warning. So, we cast the E& to void.
740 (void) const_cast<E&>(_elems[localBot] = t);
741 #ifdef MIPS
742 this->set_bottom(increment_index(localBot));
743 #else
744 OrderAccess::release_store(&_bottom, increment_index(localBot));
745 #endif
746 TASKQUEUE_STATS_ONLY(stats.record_push());
747 return true;
748 } else {
749 return push_slow(t, dirty_n_elems);
750 }
751 }
753 template<class E, MEMFLAGS F, unsigned int N> inline bool
754 GenericTaskQueue<E, F, N>::pop_local(volatile E& t) {
755 #ifdef MIPS
756 uint localBot = this->get_bottom();
757 #else
758 uint localBot = _bottom;
759 #endif
760 // This value cannot be N-1. That can only occur as a result of
761 // the assignment to bottom in this method. If it does, this method
762 // resets the size to 0 before the next call (which is sequential,
763 // since this is pop_local.)
764 uint dirty_n_elems = dirty_size(localBot, _age.top());
765 assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
766 if (dirty_n_elems == 0) return false;
767 localBot = decrement_index(localBot);
768 #ifdef MIPS
769 this->set_bottom(localBot);
770 #else
771 _bottom = localBot;
772 #endif
773 // This is necessary to prevent any read below from being reordered
774 // before the store just above.
775 OrderAccess::fence();
776 // g++ complains if the volatile result of the assignment is
777 // unused, so we cast the volatile away. We cannot cast directly
778 // to void, because gcc treats that as not using the result of the
779 // assignment. However, casting to E& means that we trigger an
780 // unused-value warning. So, we cast the E& to void.
781 (void) const_cast<E&>(t = _elems[localBot]);
782 // This is a second read of "age"; the "size()" above is the first.
783 // If there's still at least one element in the queue, based on the
784 // "_bottom" and "age" we've read, then there can be no interference with
785 // a "pop_global" operation, and we're done.
786 idx_t tp = _age.top(); // XXX
787 if (size(localBot, tp) > 0) {
788 assert(dirty_size(localBot, tp) != N - 1, "sanity");
789 TASKQUEUE_STATS_ONLY(stats.record_pop());
790 return true;
791 } else {
792 // Otherwise, the queue contained exactly one element; we take the slow
793 // path.
795 // The barrier is required to prevent reordering the two reads of _age:
796 // one is the _age.get() below, and the other is _age.top() above the if-stmt.
797 // The algorithm may fail if _age.get() reads an older value than _age.top().
798 OrderAccess::loadload();
799 return pop_local_slow(localBot, _age.get());
800 }
801 }
803 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue;
804 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
806 #ifdef _MSC_VER
807 #pragma warning(push)
808 // warning C4522: multiple assignment operators specified
809 #pragma warning(disable:4522)
810 #endif
812 // This is a container class for either an oop* or a narrowOop*.
813 // Both are pushed onto a task queue and the consumer will test is_narrow()
814 // to determine which should be processed.
815 class StarTask {
816 void* _holder; // either union oop* or narrowOop*
818 enum { COMPRESSED_OOP_MASK = 1 };
820 public:
821 StarTask(narrowOop* p) {
822 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
823 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
824 }
825 StarTask(oop* p) {
826 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
827 _holder = (void*)p;
828 }
829 StarTask() { _holder = NULL; }
830 operator oop*() { return (oop*)_holder; }
831 operator narrowOop*() {
832 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
833 }
835 StarTask& operator=(const StarTask& t) {
836 _holder = t._holder;
837 return *this;
838 }
839 volatile StarTask& operator=(const volatile StarTask& t) volatile {
840 _holder = t._holder;
841 return *this;
842 }
844 bool is_narrow() const {
845 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
846 }
847 };
849 class ObjArrayTask
850 {
851 public:
852 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
853 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
854 assert(idx <= size_t(max_jint), "too big");
855 }
856 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
858 ObjArrayTask& operator =(const ObjArrayTask& t) {
859 _obj = t._obj;
860 _index = t._index;
861 return *this;
862 }
863 volatile ObjArrayTask&
864 operator =(const volatile ObjArrayTask& t) volatile {
865 (void)const_cast<oop&>(_obj = t._obj);
866 _index = t._index;
867 return *this;
868 }
870 inline oop obj() const { return _obj; }
871 inline int index() const { return _index; }
873 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
875 private:
876 oop _obj;
877 int _index;
878 };
880 #ifdef _MSC_VER
881 #pragma warning(pop)
882 #endif
884 typedef OverflowTaskQueue<StarTask, mtClass> OopStarTaskQueue;
885 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet;
887 typedef OverflowTaskQueue<size_t, mtInternal> RegionTaskQueue;
888 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass> RegionTaskQueueSet;
891 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP