Wed, 12 Oct 2016 02:29:05 -0400
Sync during TaskSteal for 3A2000.
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 "utilities/stack.hpp"
32 #ifdef TARGET_OS_ARCH_linux_x86
33 # include "orderAccess_linux_x86.inline.hpp"
34 #endif
35 #ifdef TARGET_OS_ARCH_linux_sparc
36 # include "orderAccess_linux_sparc.inline.hpp"
37 #endif
38 #ifdef TARGET_OS_ARCH_linux_zero
39 # include "orderAccess_linux_zero.inline.hpp"
40 #endif
41 #ifdef TARGET_OS_ARCH_solaris_x86
42 # include "orderAccess_solaris_x86.inline.hpp"
43 #endif
44 #ifdef TARGET_OS_ARCH_solaris_sparc
45 # include "orderAccess_solaris_sparc.inline.hpp"
46 #endif
47 #ifdef TARGET_OS_ARCH_windows_x86
48 # include "orderAccess_windows_x86.inline.hpp"
49 #endif
50 #ifdef TARGET_OS_ARCH_linux_arm
51 # include "orderAccess_linux_arm.inline.hpp"
52 #endif
53 #ifdef TARGET_OS_ARCH_linux_ppc
54 # include "orderAccess_linux_ppc.inline.hpp"
55 #endif
56 #ifdef TARGET_OS_ARCH_aix_ppc
57 # include "orderAccess_aix_ppc.inline.hpp"
58 #endif
59 #ifdef TARGET_OS_ARCH_bsd_x86
60 # include "orderAccess_bsd_x86.inline.hpp"
61 #endif
62 #ifdef TARGET_OS_ARCH_bsd_zero
63 # include "orderAccess_bsd_zero.inline.hpp"
64 #endif
66 // Simple TaskQueue stats that are collected by default in debug builds.
68 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
69 #define TASKQUEUE_STATS 1
70 #elif !defined(TASKQUEUE_STATS)
71 #define TASKQUEUE_STATS 0
72 #endif
74 #if TASKQUEUE_STATS
75 #define TASKQUEUE_STATS_ONLY(code) code
76 #else
77 #define TASKQUEUE_STATS_ONLY(code)
78 #endif // TASKQUEUE_STATS
80 #if TASKQUEUE_STATS
81 class TaskQueueStats {
82 public:
83 enum StatId {
84 push, // number of taskqueue pushes
85 pop, // number of taskqueue pops
86 pop_slow, // subset of taskqueue pops that were done slow-path
87 steal_attempt, // number of taskqueue steal attempts
88 steal, // number of taskqueue steals
89 overflow, // number of overflow pushes
90 overflow_max_len, // max length of overflow stack
91 last_stat_id
92 };
94 public:
95 inline TaskQueueStats() { reset(); }
97 inline void record_push() { ++_stats[push]; }
98 inline void record_pop() { ++_stats[pop]; }
99 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
100 inline void record_steal(bool success);
101 inline void record_overflow(size_t new_length);
103 TaskQueueStats & operator +=(const TaskQueueStats & addend);
105 inline size_t get(StatId id) const { return _stats[id]; }
106 inline const size_t* get() const { return _stats; }
108 inline void reset();
110 // Print the specified line of the header (does not include a line separator).
111 static void print_header(unsigned int line, outputStream* const stream = tty,
112 unsigned int width = 10);
113 // Print the statistics (does not include a line separator).
114 void print(outputStream* const stream = tty, unsigned int width = 10) const;
116 DEBUG_ONLY(void verify() const;)
118 private:
119 size_t _stats[last_stat_id];
120 static const char * const _names[last_stat_id];
121 };
123 void TaskQueueStats::record_steal(bool success) {
124 ++_stats[steal_attempt];
125 if (success) ++_stats[steal];
126 }
128 void TaskQueueStats::record_overflow(size_t new_len) {
129 ++_stats[overflow];
130 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
131 }
133 void TaskQueueStats::reset() {
134 memset(_stats, 0, sizeof(_stats));
135 }
136 #endif // TASKQUEUE_STATS
138 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances.
140 template <unsigned int N, MEMFLAGS F>
141 class TaskQueueSuper: public CHeapObj<F> {
142 protected:
143 // Internal type for indexing the queue; also used for the tag.
144 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
146 // The first free element after the last one pushed (mod N).
147 volatile uint _bottom;
149 enum { MOD_N_MASK = N - 1 };
151 class Age {
152 public:
153 Age(size_t data = 0) { _data = data; }
154 Age(const Age& age) { _data = age._data; }
155 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; }
157 Age get() const volatile { return _data; }
158 void set(Age age) volatile { _data = age._data; }
160 idx_t top() const volatile { return _fields._top; }
161 idx_t tag() const volatile { return _fields._tag; }
163 // Increment top; if it wraps, increment tag also.
164 void increment() {
165 _fields._top = increment_index(_fields._top);
166 if (_fields._top == 0) ++_fields._tag;
167 }
169 Age cmpxchg(const Age new_age, const Age old_age) volatile {
170 return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
171 (volatile intptr_t *)&_data,
172 (intptr_t)old_age._data);
173 }
175 bool operator ==(const Age& other) const { return _data == other._data; }
177 private:
178 struct fields {
179 idx_t _top;
180 idx_t _tag;
181 };
182 union {
183 size_t _data;
184 fields _fields;
185 };
186 };
188 volatile Age _age;
190 // These both operate mod N.
191 static uint increment_index(uint ind) {
192 return (ind + 1) & MOD_N_MASK;
193 }
194 static uint decrement_index(uint ind) {
195 return (ind - 1) & MOD_N_MASK;
196 }
198 // Returns a number in the range [0..N). If the result is "N-1", it should be
199 // interpreted as 0.
200 uint dirty_size(uint bot, uint top) const {
201 return (bot - top) & MOD_N_MASK;
202 }
204 // Returns the size corresponding to the given "bot" and "top".
205 uint size(uint bot, uint top) const {
206 uint sz = dirty_size(bot, top);
207 // Has the queue "wrapped", so that bottom is less than top? There's a
208 // complicated special case here. A pair of threads could perform pop_local
209 // and pop_global operations concurrently, starting from a state in which
210 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
211 // and the pop_global in incrementing _top (in which case the pop_global
212 // will be awarded the contested queue element.) The resulting state must
213 // be interpreted as an empty queue. (We only need to worry about one such
214 // event: only the queue owner performs pop_local's, and several concurrent
215 // threads attempting to perform the pop_global will all perform the same
216 // CAS, and only one can succeed.) Any stealing thread that reads after
217 // either the increment or decrement will see an empty queue, and will not
218 // join the competitors. The "sz == -1 || sz == N-1" state will not be
219 // modified by concurrent queues, so the owner thread can reset the state to
220 // _bottom == top so subsequent pushes will be performed normally.
221 return (sz == N - 1) ? 0 : sz;
222 }
224 public:
225 TaskQueueSuper() : _bottom(0), _age() {}
227 // Return true if the TaskQueue contains/does not contain any tasks.
228 bool peek() const { return _bottom != _age.top(); }
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 return size(_bottom, _age.top());
236 }
238 uint dirty_size() const {
239 return dirty_size(_bottom, _age.top());
240 }
242 void set_empty() {
243 _bottom = 0;
244 _age.set(0);
245 }
247 // Maximum number of elements allowed in the queue. This is two less
248 // than the actual queue size, for somewhat complicated reasons.
249 uint max_elems() const { return N - 2; }
251 // Total size of queue.
252 static const uint total_size() { return N; }
254 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
255 };
257 //
258 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
259 // ended-queue (deque), intended for use in work stealing. Queue operations
260 // are non-blocking.
261 //
262 // A queue owner thread performs push() and pop_local() operations on one end
263 // of the queue, while other threads may steal work using the pop_global()
264 // method.
265 //
266 // The main difference to the original algorithm is that this
267 // implementation allows wrap-around at the end of its allocated
268 // storage, which is an array.
269 //
270 // The original paper is:
271 //
272 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
273 // Thread scheduling for multiprogrammed multiprocessors.
274 // Theory of Computing Systems 34, 2 (2001), 115-144.
275 //
276 // The following paper provides an correctness proof and an
277 // implementation for weakly ordered memory models including (pseudo-)
278 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
279 // similar to ABP, with the main difference that it allows resizing of the
280 // underlying storage:
281 //
282 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
283 // Correct and efficient work-stealing for weak memory models
284 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
285 // practice of parallel programming (PPoPP 2013), 69-80
286 //
288 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
289 class GenericTaskQueue: public TaskQueueSuper<N, F> {
290 ArrayAllocator<E, F> _array_allocator;
291 protected:
292 typedef typename TaskQueueSuper<N, F>::Age Age;
293 typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
295 using TaskQueueSuper<N, F>::_bottom;
296 using TaskQueueSuper<N, F>::_age;
297 using TaskQueueSuper<N, F>::increment_index;
298 using TaskQueueSuper<N, F>::decrement_index;
299 using TaskQueueSuper<N, F>::dirty_size;
301 public:
302 using TaskQueueSuper<N, F>::max_elems;
303 using TaskQueueSuper<N, F>::size;
305 #if TASKQUEUE_STATS
306 using TaskQueueSuper<N, F>::stats;
307 #endif
309 private:
310 // Slow paths for push, pop_local. (pop_global has no fast path.)
311 bool push_slow(E t, uint dirty_n_elems);
312 bool pop_local_slow(uint localBot, Age oldAge);
314 public:
315 typedef E element_type;
317 // Initializes the queue to empty.
318 GenericTaskQueue();
320 void initialize();
322 // Push the task "t" on the queue. Returns "false" iff the queue is full.
323 inline bool push(E t);
325 // Attempts to claim a task from the "local" end of the queue (the most
326 // recently pushed). If successful, returns true and sets t to the task;
327 // otherwise, returns false (the queue is empty).
328 inline bool pop_local(volatile E& t);
330 // Like pop_local(), but uses the "global" end of the queue (the least
331 // recently pushed).
332 bool pop_global(volatile E& t);
334 // Delete any resource associated with the queue.
335 ~GenericTaskQueue();
337 // apply the closure to all elements in the task queue
338 void oops_do(OopClosure* f);
340 private:
341 // Element array.
342 volatile E* _elems;
343 };
345 template<class E, MEMFLAGS F, unsigned int N>
346 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
347 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
348 }
350 template<class E, MEMFLAGS F, unsigned int N>
351 void GenericTaskQueue<E, F, N>::initialize() {
352 _elems = _array_allocator.allocate(N);
353 }
355 template<class E, MEMFLAGS F, unsigned int N>
356 void GenericTaskQueue<E, F, N>::oops_do(OopClosure* f) {
357 // tty->print_cr("START OopTaskQueue::oops_do");
358 uint iters = size();
359 uint index = _bottom;
360 for (uint i = 0; i < iters; ++i) {
361 index = decrement_index(index);
362 // tty->print_cr(" doing entry %d," INTPTR_T " -> " INTPTR_T,
363 // index, &_elems[index], _elems[index]);
364 E* t = (E*)&_elems[index]; // cast away volatility
365 oop* p = (oop*)t;
366 assert((*t)->is_oop_or_null(), "Not an oop or null");
367 f->do_oop(p);
368 }
369 // tty->print_cr("END OopTaskQueue::oops_do");
370 }
372 template<class E, MEMFLAGS F, unsigned int N>
373 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) {
374 if (dirty_n_elems == N - 1) {
375 // Actually means 0, so do the push.
376 uint localBot = _bottom;
377 // g++ complains if the volatile result of the assignment is
378 // unused, so we cast the volatile away. We cannot cast directly
379 // to void, because gcc treats that as not using the result of the
380 // assignment. However, casting to E& means that we trigger an
381 // unused-value warning. So, we cast the E& to void.
382 (void)const_cast<E&>(_elems[localBot] = t);
383 OrderAccess::release_store(&_bottom, increment_index(localBot));
384 TASKQUEUE_STATS_ONLY(stats.record_push());
385 #ifdef MIPS64
386 if (Use3A2000) OrderAccess::fence();
387 #endif
388 return true;
389 }
390 return false;
391 }
393 // pop_local_slow() is done by the owning thread and is trying to
394 // get the last task in the queue. It will compete with pop_global()
395 // that will be used by other threads. The tag age is incremented
396 // whenever the queue goes empty which it will do here if this thread
397 // gets the last task or in pop_global() if the queue wraps (top == 0
398 // and pop_global() succeeds, see pop_global()).
399 template<class E, MEMFLAGS F, unsigned int N>
400 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) {
401 // This queue was observed to contain exactly one element; either this
402 // thread will claim it, or a competing "pop_global". In either case,
403 // the queue will be logically empty afterwards. Create a new Age value
404 // that represents the empty queue for the given value of "_bottom". (We
405 // must also increment "tag" because of the case where "bottom == 1",
406 // "top == 0". A pop_global could read the queue element in that case,
407 // then have the owner thread do a pop followed by another push. Without
408 // the incrementing of "tag", the pop_global's CAS could succeed,
409 // allowing it to believe it has claimed the stale element.)
410 Age newAge((idx_t)localBot, oldAge.tag() + 1);
411 // Perhaps a competing pop_global has already incremented "top", in which
412 // case it wins the element.
413 if (localBot == oldAge.top()) {
414 // No competing pop_global has yet incremented "top"; we'll try to
415 // install new_age, thus claiming the element.
416 Age tempAge = _age.cmpxchg(newAge, oldAge);
417 if (tempAge == oldAge) {
418 // We win.
419 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
420 TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
421 return true;
422 }
423 }
424 // We lose; a completing pop_global gets the element. But the queue is empty
425 // and top is greater than bottom. Fix this representation of the empty queue
426 // to become the canonical one.
427 _age.set(newAge);
428 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
429 return false;
430 }
432 template<class E, MEMFLAGS F, unsigned int N>
433 bool GenericTaskQueue<E, F, N>::pop_global(volatile E& t) {
434 Age oldAge = _age.get();
435 // Architectures with weak memory model require a barrier here
436 // to guarantee that bottom is not older than age,
437 // which is crucial for the correctness of the algorithm.
438 #if !(defined SPARC || defined IA32 || defined AMD64)
439 OrderAccess::fence();
440 #endif
441 uint localBot = OrderAccess::load_acquire((volatile juint*)&_bottom);
442 uint n_elems = size(localBot, oldAge.top());
443 if (n_elems == 0) {
444 return false;
445 }
447 // g++ complains if the volatile result of the assignment is
448 // unused, so we cast the volatile away. We cannot cast directly
449 // to void, because gcc treats that as not using the result of the
450 // assignment. However, casting to E& means that we trigger an
451 // unused-value warning. So, we cast the E& to void.
452 (void) const_cast<E&>(t = _elems[oldAge.top()]);
453 Age newAge(oldAge);
454 newAge.increment();
455 Age resAge = _age.cmpxchg(newAge, oldAge);
457 // Note that using "_bottom" here might fail, since a pop_local might
458 // have decremented it.
459 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
460 return resAge == oldAge;
461 }
463 template<class E, MEMFLAGS F, unsigned int N>
464 GenericTaskQueue<E, F, N>::~GenericTaskQueue() {
465 FREE_C_HEAP_ARRAY(E, _elems, F);
466 }
468 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
469 // elements that do not fit in the TaskQueue.
470 //
471 // This class hides two methods from super classes:
472 //
473 // push() - push onto the task queue or, if that fails, onto the overflow stack
474 // is_empty() - return true if both the TaskQueue and overflow stack are empty
475 //
476 // Note that size() is not hidden--it returns the number of elements in the
477 // TaskQueue, and does not include the size of the overflow stack. This
478 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
479 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
480 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
481 {
482 public:
483 typedef Stack<E, F> overflow_t;
484 typedef GenericTaskQueue<E, F, N> taskqueue_t;
486 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
488 // Push task t onto the queue or onto the overflow stack. Return true.
489 inline bool push(E t);
491 // Attempt to pop from the overflow stack; return true if anything was popped.
492 inline bool pop_overflow(E& t);
494 inline overflow_t* overflow_stack() { return &_overflow_stack; }
496 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
497 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
498 inline bool is_empty() const {
499 return taskqueue_empty() && overflow_empty();
500 }
502 private:
503 overflow_t _overflow_stack;
504 };
506 template <class E, MEMFLAGS F, unsigned int N>
507 bool OverflowTaskQueue<E, F, N>::push(E t)
508 {
509 if (!taskqueue_t::push(t)) {
510 overflow_stack()->push(t);
511 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
512 #ifdef MIPS64
513 if (Use3A2000) OrderAccess::fence();
514 #endif
515 }
516 return true;
517 }
519 template <class E, MEMFLAGS F, unsigned int N>
520 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t)
521 {
522 if (overflow_empty()) return false;
523 t = overflow_stack()->pop();
524 return true;
525 }
527 class TaskQueueSetSuper {
528 protected:
529 static int randomParkAndMiller(int* seed0);
530 public:
531 // Returns "true" if some TaskQueue in the set contains a task.
532 virtual bool peek() = 0;
533 };
535 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
536 };
538 template<class T, MEMFLAGS F>
539 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
540 private:
541 uint _n;
542 T** _queues;
544 public:
545 typedef typename T::element_type E;
547 GenericTaskQueueSet(int n) : _n(n) {
548 typedef T* GenericTaskQueuePtr;
549 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F);
550 for (int i = 0; i < n; i++) {
551 _queues[i] = NULL;
552 }
553 }
555 bool steal_best_of_2(uint queue_num, int* seed, E& t);
557 void register_queue(uint i, T* q);
559 T* queue(uint n);
561 // The thread with queue number "queue_num" (and whose random number seed is
562 // at "seed") is trying to steal a task from some other queue. (It may try
563 // several queues, according to some configuration parameter.) If some steal
564 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
565 // false.
566 bool steal(uint queue_num, int* seed, E& t);
568 bool peek();
569 };
571 template<class T, MEMFLAGS F> void
572 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
573 assert(i < _n, "index out of range.");
574 _queues[i] = q;
575 }
577 template<class T, MEMFLAGS F> T*
578 GenericTaskQueueSet<T, F>::queue(uint i) {
579 return _queues[i];
580 }
582 template<class T, MEMFLAGS F> bool
583 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
584 for (uint i = 0; i < 2 * _n; i++) {
585 if (steal_best_of_2(queue_num, seed, t)) {
586 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
587 #ifdef MIPS64
588 if (Use3A2000) OrderAccess::fence();
589 #endif
590 return true;
591 }
592 }
593 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
594 #ifdef MIPS64
595 if (Use3A2000) OrderAccess::fence();
596 #endif
597 return false;
598 }
600 template<class T, MEMFLAGS F> bool
601 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) {
602 if (_n > 2) {
603 uint k1 = queue_num;
604 while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
605 uint k2 = queue_num;
606 while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
607 // Sample both and try the larger.
608 uint sz1 = _queues[k1]->size();
609 uint sz2 = _queues[k2]->size();
610 if (sz2 > sz1) return _queues[k2]->pop_global(t);
611 else return _queues[k1]->pop_global(t);
612 } else if (_n == 2) {
613 // Just try the other one.
614 uint k = (queue_num + 1) % 2;
615 return _queues[k]->pop_global(t);
616 } else {
617 assert(_n == 1, "can't be zero.");
618 return false;
619 }
620 }
622 template<class T, MEMFLAGS F>
623 bool GenericTaskQueueSet<T, F>::peek() {
624 // Try all the queues.
625 for (uint j = 0; j < _n; j++) {
626 if (_queues[j]->peek())
627 return true;
628 }
629 return false;
630 }
632 // When to terminate from the termination protocol.
633 class TerminatorTerminator: public CHeapObj<mtInternal> {
634 public:
635 virtual bool should_exit_termination() = 0;
636 };
638 // A class to aid in the termination of a set of parallel tasks using
639 // TaskQueueSet's for work stealing.
641 #undef TRACESPINNING
643 class ParallelTaskTerminator: public StackObj {
644 private:
645 int _n_threads;
646 TaskQueueSetSuper* _queue_set;
647 int _offered_termination;
649 #ifdef TRACESPINNING
650 static uint _total_yields;
651 static uint _total_spins;
652 static uint _total_peeks;
653 #endif
655 bool peek_in_queue_set();
656 protected:
657 virtual void yield();
658 void sleep(uint millis);
660 public:
662 // "n_threads" is the number of threads to be terminated. "queue_set" is a
663 // queue sets of work queues of other threads.
664 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
666 // The current thread has no work, and is ready to terminate if everyone
667 // else is. If returns "true", all threads are terminated. If returns
668 // "false", available work has been observed in one of the task queues,
669 // so the global task is not complete.
670 bool offer_termination() {
671 return offer_termination(NULL);
672 }
674 // As above, but it also terminates if the should_exit_termination()
675 // method of the terminator parameter returns true. If terminator is
676 // NULL, then it is ignored.
677 bool offer_termination(TerminatorTerminator* terminator);
679 // Reset the terminator, so that it may be reused again.
680 // The caller is responsible for ensuring that this is done
681 // in an MT-safe manner, once the previous round of use of
682 // the terminator is finished.
683 void reset_for_reuse();
684 // Same as above but the number of parallel threads is set to the
685 // given number.
686 void reset_for_reuse(int n_threads);
688 #ifdef TRACESPINNING
689 static uint total_yields() { return _total_yields; }
690 static uint total_spins() { return _total_spins; }
691 static uint total_peeks() { return _total_peeks; }
692 static void print_termination_counts();
693 #endif
694 };
696 template<class E, MEMFLAGS F, unsigned int N> inline bool
697 GenericTaskQueue<E, F, N>::push(E t) {
698 uint localBot = _bottom;
699 assert(localBot < N, "_bottom out of range.");
700 idx_t top = _age.top();
701 uint dirty_n_elems = dirty_size(localBot, top);
702 assert(dirty_n_elems < N, "n_elems out of range.");
703 if (dirty_n_elems < max_elems()) {
704 // g++ complains if the volatile result of the assignment is
705 // unused, so we cast the volatile away. We cannot cast directly
706 // to void, because gcc treats that as not using the result of the
707 // assignment. However, casting to E& means that we trigger an
708 // unused-value warning. So, we cast the E& to void.
709 (void) const_cast<E&>(_elems[localBot] = t);
710 OrderAccess::release_store(&_bottom, increment_index(localBot));
711 TASKQUEUE_STATS_ONLY(stats.record_push());
712 #ifdef MIPS64
713 if (Use3A2000) OrderAccess::fence();
714 #endif
715 return true;
716 } else {
717 return push_slow(t, dirty_n_elems);
718 }
719 }
721 template<class E, MEMFLAGS F, unsigned int N> inline bool
722 GenericTaskQueue<E, F, N>::pop_local(volatile E& t) {
723 uint localBot = _bottom;
724 // This value cannot be N-1. That can only occur as a result of
725 // the assignment to bottom in this method. If it does, this method
726 // resets the size to 0 before the next call (which is sequential,
727 // since this is pop_local.)
728 uint dirty_n_elems = dirty_size(localBot, _age.top());
729 assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
730 if (dirty_n_elems == 0) return false;
731 localBot = decrement_index(localBot);
732 _bottom = localBot;
733 // This is necessary to prevent any read below from being reordered
734 // before the store just above.
735 OrderAccess::fence();
736 // g++ complains if the volatile result of the assignment is
737 // unused, so we cast the volatile away. We cannot cast directly
738 // to void, because gcc treats that as not using the result of the
739 // assignment. However, casting to E& means that we trigger an
740 // unused-value warning. So, we cast the E& to void.
741 (void) const_cast<E&>(t = _elems[localBot]);
742 // This is a second read of "age"; the "size()" above is the first.
743 // If there's still at least one element in the queue, based on the
744 // "_bottom" and "age" we've read, then there can be no interference with
745 // a "pop_global" operation, and we're done.
746 idx_t tp = _age.top(); // XXX
747 if (size(localBot, tp) > 0) {
748 assert(dirty_size(localBot, tp) != N - 1, "sanity");
749 TASKQUEUE_STATS_ONLY(stats.record_pop());
750 return true;
751 } else {
752 // Otherwise, the queue contained exactly one element; we take the slow
753 // path.
754 return pop_local_slow(localBot, _age.get());
755 }
756 }
758 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue;
759 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
761 #ifdef _MSC_VER
762 #pragma warning(push)
763 // warning C4522: multiple assignment operators specified
764 #pragma warning(disable:4522)
765 #endif
767 // This is a container class for either an oop* or a narrowOop*.
768 // Both are pushed onto a task queue and the consumer will test is_narrow()
769 // to determine which should be processed.
770 class StarTask {
771 void* _holder; // either union oop* or narrowOop*
773 enum { COMPRESSED_OOP_MASK = 1 };
775 public:
776 StarTask(narrowOop* p) {
777 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
778 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
779 }
780 StarTask(oop* p) {
781 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
782 _holder = (void*)p;
783 }
784 StarTask() { _holder = NULL; }
785 operator oop*() { return (oop*)_holder; }
786 operator narrowOop*() {
787 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
788 }
790 StarTask& operator=(const StarTask& t) {
791 _holder = t._holder;
792 return *this;
793 }
794 volatile StarTask& operator=(const volatile StarTask& t) volatile {
795 _holder = t._holder;
796 return *this;
797 }
799 bool is_narrow() const {
800 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
801 }
802 };
804 class ObjArrayTask
805 {
806 public:
807 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
808 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
809 assert(idx <= size_t(max_jint), "too big");
810 }
811 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
813 ObjArrayTask& operator =(const ObjArrayTask& t) {
814 _obj = t._obj;
815 _index = t._index;
816 return *this;
817 }
818 volatile ObjArrayTask&
819 operator =(const volatile ObjArrayTask& t) volatile {
820 (void)const_cast<oop&>(_obj = t._obj);
821 _index = t._index;
822 return *this;
823 }
825 inline oop obj() const { return _obj; }
826 inline int index() const { return _index; }
828 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
830 private:
831 oop _obj;
832 int _index;
833 };
835 #ifdef _MSC_VER
836 #pragma warning(pop)
837 #endif
839 typedef OverflowTaskQueue<StarTask, mtClass> OopStarTaskQueue;
840 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet;
842 typedef OverflowTaskQueue<size_t, mtInternal> RegionTaskQueue;
843 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass> RegionTaskQueueSet;
846 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP