1 /* 2 * Copyright (c) 2001, 2011, 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 "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_bsd_x86 57 # include "orderAccess_bsd_x86.inline.hpp" 58 #endif 59 #ifdef TARGET_OS_ARCH_bsd_zero 60 # include "orderAccess_bsd_zero.inline.hpp" 61 #endif 62 63 // Simple TaskQueue stats that are collected by default in debug builds. 64 65 #if !defined(TASKQUEUE_STATS) && defined(ASSERT) 66 #define TASKQUEUE_STATS 1 67 #elif !defined(TASKQUEUE_STATS) 68 #define TASKQUEUE_STATS 0 69 #endif 70 71 #if TASKQUEUE_STATS 72 #define TASKQUEUE_STATS_ONLY(code) code 73 #else 74 #define TASKQUEUE_STATS_ONLY(code) 75 #endif // TASKQUEUE_STATS 76 77 #if TASKQUEUE_STATS 78 class TaskQueueStats { 79 public: 80 enum StatId { 81 push, // number of taskqueue pushes 82 pop, // number of taskqueue pops 83 pop_slow, // subset of taskqueue pops that were done slow-path 84 steal_attempt, // number of taskqueue steal attempts 85 steal, // number of taskqueue steals 86 overflow, // number of overflow pushes 87 overflow_max_len, // max length of overflow stack 88 last_stat_id 89 }; 90 91 public: 92 inline TaskQueueStats() { reset(); } 93 94 inline void record_push() { ++_stats[push]; } 95 inline void record_pop() { ++_stats[pop]; } 96 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; } 97 inline void record_steal(bool success); 98 inline void record_overflow(size_t new_length); 99 100 TaskQueueStats & operator +=(const TaskQueueStats & addend); 101 102 inline size_t get(StatId id) const { return _stats[id]; } 103 inline const size_t* get() const { return _stats; } 104 105 inline void reset(); 106 107 // Print the specified line of the header (does not include a line separator). 108 static void print_header(unsigned int line, outputStream* const stream = tty, 109 unsigned int width = 10); 110 // Print the statistics (does not include a line separator). 111 void print(outputStream* const stream = tty, unsigned int width = 10) const; 112 113 DEBUG_ONLY(void verify() const;) 114 115 private: 116 size_t _stats[last_stat_id]; 117 static const char * const _names[last_stat_id]; 118 }; 119 120 void TaskQueueStats::record_steal(bool success) { 121 ++_stats[steal_attempt]; 122 if (success) ++_stats[steal]; 123 } 124 125 void TaskQueueStats::record_overflow(size_t new_len) { 126 ++_stats[overflow]; 127 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len; 128 } 129 130 void TaskQueueStats::reset() { 131 memset(_stats, 0, sizeof(_stats)); 132 } 133 #endif // TASKQUEUE_STATS 134 135 template <unsigned int N, MEMFLAGS F> 136 class TaskQueueSuper: public CHeapObj<F> { 137 protected: 138 // Internal type for indexing the queue; also used for the tag. 139 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t; 140 141 // The first free element after the last one pushed (mod N). 142 volatile uint _bottom; 143 144 enum { MOD_N_MASK = N - 1 }; 145 146 class Age { 147 public: 148 Age(size_t data = 0) { _data = data; } 149 Age(const Age& age) { _data = age._data; } 150 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; } 151 152 Age get() const volatile { return _data; } 153 void set(Age age) volatile { _data = age._data; } 154 155 idx_t top() const volatile { return _fields._top; } 156 idx_t tag() const volatile { return _fields._tag; } 157 158 // Increment top; if it wraps, increment tag also. 159 void increment() { 160 _fields._top = increment_index(_fields._top); 161 if (_fields._top == 0) ++_fields._tag; 162 } 163 164 Age cmpxchg(const Age new_age, const Age old_age) volatile { 165 return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data, 166 (volatile intptr_t *)&_data, 167 (intptr_t)old_age._data); 168 } 169 170 bool operator ==(const Age& other) const { return _data == other._data; } 171 172 private: 173 struct fields { 174 idx_t _top; 175 idx_t _tag; 176 }; 177 union { 178 size_t _data; 179 fields _fields; 180 }; 181 }; 182 183 volatile Age _age; 184 185 // These both operate mod N. 186 static uint increment_index(uint ind) { 187 return (ind + 1) & MOD_N_MASK; 188 } 189 static uint decrement_index(uint ind) { 190 return (ind - 1) & MOD_N_MASK; 191 } 192 193 // Returns a number in the range [0..N). If the result is "N-1", it should be 194 // interpreted as 0. 195 uint dirty_size(uint bot, uint top) const { 196 return (bot - top) & MOD_N_MASK; 197 } 198 199 // Returns the size corresponding to the given "bot" and "top". 200 uint size(uint bot, uint top) const { 201 uint sz = dirty_size(bot, top); 202 // Has the queue "wrapped", so that bottom is less than top? There's a 203 // complicated special case here. A pair of threads could perform pop_local 204 // and pop_global operations concurrently, starting from a state in which 205 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom, 206 // and the pop_global in incrementing _top (in which case the pop_global 207 // will be awarded the contested queue element.) The resulting state must 208 // be interpreted as an empty queue. (We only need to worry about one such 209 // event: only the queue owner performs pop_local's, and several concurrent 210 // threads attempting to perform the pop_global will all perform the same 211 // CAS, and only one can succeed.) Any stealing thread that reads after 212 // either the increment or decrement will see an empty queue, and will not 213 // join the competitors. The "sz == -1 || sz == N-1" state will not be 214 // modified by concurrent queues, so the owner thread can reset the state to 215 // _bottom == top so subsequent pushes will be performed normally. 216 return (sz == N - 1) ? 0 : sz; 217 } 218 219 public: 220 TaskQueueSuper() : _bottom(0), _age() {} 221 222 // Return true if the TaskQueue contains/does not contain any tasks. 223 bool peek() const { return _bottom != _age.top(); } 224 bool is_empty() const { return size() == 0; } 225 226 // Return an estimate of the number of elements in the queue. 227 // The "careful" version admits the possibility of pop_local/pop_global 228 // races. 229 uint size() const { 230 return size(_bottom, _age.top()); 231 } 232 233 uint dirty_size() const { 234 return dirty_size(_bottom, _age.top()); 235 } 236 237 void set_empty() { 238 _bottom = 0; 239 _age.set(0); 240 } 241 242 // Maximum number of elements allowed in the queue. This is two less 243 // than the actual queue size, for somewhat complicated reasons. 244 uint max_elems() const { return N - 2; } 245 246 // Total size of queue. 247 static const uint total_size() { return N; } 248 249 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;) 250 }; 251 252 253 254 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 255 class GenericTaskQueue: public TaskQueueSuper<N, F> { 256 ArrayAllocator<E, F> _array_allocator; 257 protected: 258 typedef typename TaskQueueSuper<N, F>::Age Age; 259 typedef typename TaskQueueSuper<N, F>::idx_t idx_t; 260 261 using TaskQueueSuper<N, F>::_bottom; 262 using TaskQueueSuper<N, F>::_age; 263 using TaskQueueSuper<N, F>::increment_index; 264 using TaskQueueSuper<N, F>::decrement_index; 265 using TaskQueueSuper<N, F>::dirty_size; 266 267 public: 268 using TaskQueueSuper<N, F>::max_elems; 269 using TaskQueueSuper<N, F>::size; 270 271 #if TASKQUEUE_STATS 272 using TaskQueueSuper<N, F>::stats; 273 #endif 274 275 private: 276 // Slow paths for push, pop_local. (pop_global has no fast path.) 277 bool push_slow(E t, uint dirty_n_elems); 278 bool pop_local_slow(uint localBot, Age oldAge); 279 280 public: 281 typedef E element_type; 282 283 // Initializes the queue to empty. 284 GenericTaskQueue(); 285 286 void initialize(); 287 288 // Push the task "t" on the queue. Returns "false" iff the queue is full. 289 inline bool push(E t); 290 291 // Attempts to claim a task from the "local" end of the queue (the most 292 // recently pushed). If successful, returns true and sets t to the task; 293 // otherwise, returns false (the queue is empty). 294 inline bool pop_local(E& t); 295 296 // Like pop_local(), but uses the "global" end of the queue (the least 297 // recently pushed). 298 bool pop_global(E& t); 299 300 // Delete any resource associated with the queue. 301 ~GenericTaskQueue(); 302 303 // apply the closure to all elements in the task queue 304 void oops_do(OopClosure* f); 305 306 private: 307 // Element array. 308 volatile E* _elems; 309 }; 310 311 template<class E, MEMFLAGS F, unsigned int N> 312 GenericTaskQueue<E, F, N>::GenericTaskQueue() { 313 assert(sizeof(Age) == sizeof(size_t), "Depends on this."); 314 } 315 316 template<class E, MEMFLAGS F, unsigned int N> 317 void GenericTaskQueue<E, F, N>::initialize() { 318 _elems = _array_allocator.allocate(N); 319 } 320 321 template<class E, MEMFLAGS F, unsigned int N> 322 void GenericTaskQueue<E, F, N>::oops_do(OopClosure* f) { 323 // tty->print_cr("START OopTaskQueue::oops_do"); 324 uint iters = size(); 325 uint index = _bottom; 326 for (uint i = 0; i < iters; ++i) { 327 index = decrement_index(index); 328 // tty->print_cr(" doing entry %d," INTPTR_T " -> " INTPTR_T, 329 // index, &_elems[index], _elems[index]); 330 E* t = (E*)&_elems[index]; // cast away volatility 331 oop* p = (oop*)t; 332 assert((*t)->is_oop_or_null(), "Not an oop or null"); 333 f->do_oop(p); 334 } 335 // tty->print_cr("END OopTaskQueue::oops_do"); 336 } 337 338 template<class E, MEMFLAGS F, unsigned int N> 339 bool GenericTaskQueue<E, F, N>::push_slow(E t, uint dirty_n_elems) { 340 if (dirty_n_elems == N - 1) { 341 // Actually means 0, so do the push. 342 uint localBot = _bottom; 343 // g++ complains if the volatile result of the assignment is unused. 344 const_cast<E&>(_elems[localBot] = t); 345 OrderAccess::release_store(&_bottom, increment_index(localBot)); 346 TASKQUEUE_STATS_ONLY(stats.record_push()); 347 return true; 348 } 349 return false; 350 } 351 352 // pop_local_slow() is done by the owning thread and is trying to 353 // get the last task in the queue. It will compete with pop_global() 354 // that will be used by other threads. The tag age is incremented 355 // whenever the queue goes empty which it will do here if this thread 356 // gets the last task or in pop_global() if the queue wraps (top == 0 357 // and pop_global() succeeds, see pop_global()). 358 template<class E, MEMFLAGS F, unsigned int N> 359 bool GenericTaskQueue<E, F, N>::pop_local_slow(uint localBot, Age oldAge) { 360 // This queue was observed to contain exactly one element; either this 361 // thread will claim it, or a competing "pop_global". In either case, 362 // the queue will be logically empty afterwards. Create a new Age value 363 // that represents the empty queue for the given value of "_bottom". (We 364 // must also increment "tag" because of the case where "bottom == 1", 365 // "top == 0". A pop_global could read the queue element in that case, 366 // then have the owner thread do a pop followed by another push. Without 367 // the incrementing of "tag", the pop_global's CAS could succeed, 368 // allowing it to believe it has claimed the stale element.) 369 Age newAge((idx_t)localBot, oldAge.tag() + 1); 370 // Perhaps a competing pop_global has already incremented "top", in which 371 // case it wins the element. 372 if (localBot == oldAge.top()) { 373 // No competing pop_global has yet incremented "top"; we'll try to 374 // install new_age, thus claiming the element. 375 Age tempAge = _age.cmpxchg(newAge, oldAge); 376 if (tempAge == oldAge) { 377 // We win. 378 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); 379 TASKQUEUE_STATS_ONLY(stats.record_pop_slow()); 380 return true; 381 } 382 } 383 // We lose; a completing pop_global gets the element. But the queue is empty 384 // and top is greater than bottom. Fix this representation of the empty queue 385 // to become the canonical one. 386 _age.set(newAge); 387 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity"); 388 return false; 389 } 390 391 template<class E, MEMFLAGS F, unsigned int N> 392 bool GenericTaskQueue<E, F, N>::pop_global(E& t) { 393 Age oldAge = _age.get(); 394 uint localBot = _bottom; 395 uint n_elems = size(localBot, oldAge.top()); 396 if (n_elems == 0) { 397 return false; 398 } 399 400 const_cast<E&>(t = _elems[oldAge.top()]); 401 Age newAge(oldAge); 402 newAge.increment(); 403 Age resAge = _age.cmpxchg(newAge, oldAge); 404 405 // Note that using "_bottom" here might fail, since a pop_local might 406 // have decremented it. 407 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity"); 408 return resAge == oldAge; 409 } 410 411 template<class E, MEMFLAGS F, unsigned int N> 412 GenericTaskQueue<E, F, N>::~GenericTaskQueue() { 413 FREE_C_HEAP_ARRAY(E, _elems, F); 414 } 415 416 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for 417 // elements that do not fit in the TaskQueue. 418 // 419 // This class hides two methods from super classes: 420 // 421 // push() - push onto the task queue or, if that fails, onto the overflow stack 422 // is_empty() - return true if both the TaskQueue and overflow stack are empty 423 // 424 // Note that size() is not hidden--it returns the number of elements in the 425 // TaskQueue, and does not include the size of the overflow stack. This 426 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues. 427 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 428 class OverflowTaskQueue: public GenericTaskQueue<E, F, N> 429 { 430 public: 431 typedef Stack<E, F> overflow_t; 432 typedef GenericTaskQueue<E, F, N> taskqueue_t; 433 434 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;) 435 436 // Push task t onto the queue or onto the overflow stack. Return true. 437 inline bool push(E t); 438 439 // Attempt to pop from the overflow stack; return true if anything was popped. 440 inline bool pop_overflow(E& t); 441 442 inline overflow_t* overflow_stack() { return &_overflow_stack; } 443 444 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); } 445 inline bool overflow_empty() const { return _overflow_stack.is_empty(); } 446 inline bool is_empty() const { 447 return taskqueue_empty() && overflow_empty(); 448 } 449 450 private: 451 overflow_t _overflow_stack; 452 }; 453 454 template <class E, MEMFLAGS F, unsigned int N> 455 bool OverflowTaskQueue<E, F, N>::push(E t) 456 { 457 if (!taskqueue_t::push(t)) { 458 overflow_stack()->push(t); 459 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size())); 460 } 461 return true; 462 } 463 464 template <class E, MEMFLAGS F, unsigned int N> 465 bool OverflowTaskQueue<E, F, N>::pop_overflow(E& t) 466 { 467 if (overflow_empty()) return false; 468 t = overflow_stack()->pop(); 469 return true; 470 } 471 472 class TaskQueueSetSuper { 473 protected: 474 static int randomParkAndMiller(int* seed0); 475 public: 476 // Returns "true" if some TaskQueue in the set contains a task. 477 virtual bool peek() = 0; 478 }; 479 480 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper { 481 }; 482 483 template<class T, MEMFLAGS F> 484 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> { 485 private: 486 uint _n; 487 T** _queues; 488 489 public: 490 typedef typename T::element_type E; 491 492 GenericTaskQueueSet(int n) : _n(n) { 493 typedef T* GenericTaskQueuePtr; 494 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F); 495 for (int i = 0; i < n; i++) { 496 _queues[i] = NULL; 497 } 498 } 499 500 bool steal_1_random(uint queue_num, int* seed, E& t); 501 bool steal_best_of_2(uint queue_num, int* seed, E& t); 502 bool steal_best_of_all(uint queue_num, int* seed, E& t); 503 504 void register_queue(uint i, T* q); 505 506 T* queue(uint n); 507 508 // The thread with queue number "queue_num" (and whose random number seed is 509 // at "seed") is trying to steal a task from some other queue. (It may try 510 // several queues, according to some configuration parameter.) If some steal 511 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns 512 // false. 513 bool steal(uint queue_num, int* seed, E& t); 514 515 bool peek(); 516 }; 517 518 template<class T, MEMFLAGS F> void 519 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) { 520 assert(i < _n, "index out of range."); 521 _queues[i] = q; 522 } 523 524 template<class T, MEMFLAGS F> T* 525 GenericTaskQueueSet<T, F>::queue(uint i) { 526 return _queues[i]; 527 } 528 529 template<class T, MEMFLAGS F> bool 530 GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) { 531 for (uint i = 0; i < 2 * _n; i++) { 532 if (steal_best_of_2(queue_num, seed, t)) { 533 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true)); 534 return true; 535 } 536 } 537 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false)); 538 return false; 539 } 540 541 template<class T, MEMFLAGS F> bool 542 GenericTaskQueueSet<T, F>::steal_best_of_all(uint queue_num, int* seed, E& t) { 543 if (_n > 2) { 544 int best_k; 545 uint best_sz = 0; 546 for (uint k = 0; k < _n; k++) { 547 if (k == queue_num) continue; 548 uint sz = _queues[k]->size(); 549 if (sz > best_sz) { 550 best_sz = sz; 551 best_k = k; 552 } 553 } 554 return best_sz > 0 && _queues[best_k]->pop_global(t); 555 } else if (_n == 2) { 556 // Just try the other one. 557 int k = (queue_num + 1) % 2; 558 return _queues[k]->pop_global(t); 559 } else { 560 assert(_n == 1, "can't be zero."); 561 return false; 562 } 563 } 564 565 template<class T, MEMFLAGS F> bool 566 GenericTaskQueueSet<T, F>::steal_1_random(uint queue_num, int* seed, E& t) { 567 if (_n > 2) { 568 uint k = queue_num; 569 while (k == queue_num) k = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; 570 return _queues[2]->pop_global(t); 571 } else if (_n == 2) { 572 // Just try the other one. 573 int k = (queue_num + 1) % 2; 574 return _queues[k]->pop_global(t); 575 } else { 576 assert(_n == 1, "can't be zero."); 577 return false; 578 } 579 } 580 581 template<class T, MEMFLAGS F> bool 582 GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) { 583 if (_n > 2) { 584 uint k1 = queue_num; 585 while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; 586 uint k2 = queue_num; 587 while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n; 588 // Sample both and try the larger. 589 uint sz1 = _queues[k1]->size(); 590 uint sz2 = _queues[k2]->size(); 591 if (sz2 > sz1) return _queues[k2]->pop_global(t); 592 else return _queues[k1]->pop_global(t); 593 } else if (_n == 2) { 594 // Just try the other one. 595 uint k = (queue_num + 1) % 2; 596 return _queues[k]->pop_global(t); 597 } else { 598 assert(_n == 1, "can't be zero."); 599 return false; 600 } 601 } 602 603 template<class T, MEMFLAGS F> 604 bool GenericTaskQueueSet<T, F>::peek() { 605 // Try all the queues. 606 for (uint j = 0; j < _n; j++) { 607 if (_queues[j]->peek()) 608 return true; 609 } 610 return false; 611 } 612 613 // When to terminate from the termination protocol. 614 class TerminatorTerminator: public CHeapObj<mtInternal> { 615 public: 616 virtual bool should_exit_termination() = 0; 617 }; 618 619 // A class to aid in the termination of a set of parallel tasks using 620 // TaskQueueSet's for work stealing. 621 622 #undef TRACESPINNING 623 624 class ParallelTaskTerminator: public StackObj { 625 private: 626 int _n_threads; 627 TaskQueueSetSuper* _queue_set; 628 int _offered_termination; 629 630 #ifdef TRACESPINNING 631 static uint _total_yields; 632 static uint _total_spins; 633 static uint _total_peeks; 634 #endif 635 636 bool peek_in_queue_set(); 637 protected: 638 virtual void yield(); 639 void sleep(uint millis); 640 641 public: 642 643 // "n_threads" is the number of threads to be terminated. "queue_set" is a 644 // queue sets of work queues of other threads. 645 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set); 646 647 // The current thread has no work, and is ready to terminate if everyone 648 // else is. If returns "true", all threads are terminated. If returns 649 // "false", available work has been observed in one of the task queues, 650 // so the global task is not complete. 651 bool offer_termination() { 652 return offer_termination(NULL); 653 } 654 655 // As above, but it also terminates if the should_exit_termination() 656 // method of the terminator parameter returns true. If terminator is 657 // NULL, then it is ignored. 658 bool offer_termination(TerminatorTerminator* terminator); 659 660 // Reset the terminator, so that it may be reused again. 661 // The caller is responsible for ensuring that this is done 662 // in an MT-safe manner, once the previous round of use of 663 // the terminator is finished. 664 void reset_for_reuse(); 665 // Same as above but the number of parallel threads is set to the 666 // given number. 667 void reset_for_reuse(int n_threads); 668 669 #ifdef TRACESPINNING 670 static uint total_yields() { return _total_yields; } 671 static uint total_spins() { return _total_spins; } 672 static uint total_peeks() { return _total_peeks; } 673 static void print_termination_counts(); 674 #endif 675 }; 676 677 template<class E, MEMFLAGS F, unsigned int N> inline bool 678 GenericTaskQueue<E, F, N>::push(E t) { 679 uint localBot = _bottom; 680 assert((localBot >= 0) && (localBot < N), "_bottom out of range."); 681 idx_t top = _age.top(); 682 uint dirty_n_elems = dirty_size(localBot, top); 683 assert(dirty_n_elems < N, "n_elems out of range."); 684 if (dirty_n_elems < max_elems()) { 685 // g++ complains if the volatile result of the assignment is unused. 686 const_cast<E&>(_elems[localBot] = t); 687 OrderAccess::release_store(&_bottom, increment_index(localBot)); 688 TASKQUEUE_STATS_ONLY(stats.record_push()); 689 return true; 690 } else { 691 return push_slow(t, dirty_n_elems); 692 } 693 } 694 695 template<class E, MEMFLAGS F, unsigned int N> inline bool 696 GenericTaskQueue<E, F, N>::pop_local(E& t) { 697 uint localBot = _bottom; 698 // This value cannot be N-1. That can only occur as a result of 699 // the assignment to bottom in this method. If it does, this method 700 // resets the size to 0 before the next call (which is sequential, 701 // since this is pop_local.) 702 uint dirty_n_elems = dirty_size(localBot, _age.top()); 703 assert(dirty_n_elems != N - 1, "Shouldn't be possible..."); 704 if (dirty_n_elems == 0) return false; 705 localBot = decrement_index(localBot); 706 _bottom = localBot; 707 // This is necessary to prevent any read below from being reordered 708 // before the store just above. 709 OrderAccess::fence(); 710 const_cast<E&>(t = _elems[localBot]); 711 // This is a second read of "age"; the "size()" above is the first. 712 // If there's still at least one element in the queue, based on the 713 // "_bottom" and "age" we've read, then there can be no interference with 714 // a "pop_global" operation, and we're done. 715 idx_t tp = _age.top(); // XXX 716 if (size(localBot, tp) > 0) { 717 assert(dirty_size(localBot, tp) != N - 1, "sanity"); 718 TASKQUEUE_STATS_ONLY(stats.record_pop()); 719 return true; 720 } else { 721 // Otherwise, the queue contained exactly one element; we take the slow 722 // path. 723 return pop_local_slow(localBot, _age.get()); 724 } 725 } 726 727 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue; 728 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet; 729 730 #ifdef _MSC_VER 731 #pragma warning(push) 732 // warning C4522: multiple assignment operators specified 733 #pragma warning(disable:4522) 734 #endif 735 736 // This is a container class for either an oop* or a narrowOop*. 737 // Both are pushed onto a task queue and the consumer will test is_narrow() 738 // to determine which should be processed. 739 class StarTask { 740 void* _holder; // either union oop* or narrowOop* 741 742 enum { COMPRESSED_OOP_MASK = 1 }; 743 744 public: 745 StarTask(narrowOop* p) { 746 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 747 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK); 748 } 749 StarTask(oop* p) { 750 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 751 _holder = (void*)p; 752 } 753 StarTask() { _holder = NULL; } 754 operator oop*() { return (oop*)_holder; } 755 operator narrowOop*() { 756 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK); 757 } 758 759 StarTask& operator=(const StarTask& t) { 760 _holder = t._holder; 761 return *this; 762 } 763 volatile StarTask& operator=(const volatile StarTask& t) volatile { 764 _holder = t._holder; 765 return *this; 766 } 767 768 bool is_narrow() const { 769 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0); 770 } 771 }; 772 773 class ObjArrayTask 774 { 775 public: 776 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { } 777 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) { 778 assert(idx <= size_t(max_jint), "too big"); 779 } 780 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { } 781 782 ObjArrayTask& operator =(const ObjArrayTask& t) { 783 _obj = t._obj; 784 _index = t._index; 785 return *this; 786 } 787 volatile ObjArrayTask& 788 operator =(const volatile ObjArrayTask& t) volatile { 789 _obj = t._obj; 790 _index = t._index; 791 return *this; 792 } 793 794 inline oop obj() const { return _obj; } 795 inline int index() const { return _index; } 796 797 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid. 798 799 private: 800 oop _obj; 801 int _index; 802 }; 803 804 #ifdef _MSC_VER 805 #pragma warning(pop) 806 #endif 807 808 typedef OverflowTaskQueue<StarTask, mtClass> OopStarTaskQueue; 809 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet; 810 811 typedef OverflowTaskQueue<size_t, mtInternal> RegionTaskQueue; 812 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass> RegionTaskQueueSet; 813 814 815 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP