1 /* 2 * Copyright (c) 2001, 2018, 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_GC_SHARED_TASKQUEUE_HPP 26 #define SHARE_VM_GC_SHARED_TASKQUEUE_HPP 27 28 #include "memory/allocation.hpp" 29 #include "memory/padded.hpp" 30 #include "oops/oopsHierarchy.hpp" 31 #include "utilities/ostream.hpp" 32 #include "utilities/stack.hpp" 33 34 // Simple TaskQueue stats that are collected by default in debug builds. 35 36 #if !defined(TASKQUEUE_STATS) && defined(ASSERT) 37 #define TASKQUEUE_STATS 1 38 #elif !defined(TASKQUEUE_STATS) 39 #define TASKQUEUE_STATS 0 40 #endif 41 42 #if TASKQUEUE_STATS 43 #define TASKQUEUE_STATS_ONLY(code) code 44 #else 45 #define TASKQUEUE_STATS_ONLY(code) 46 #endif // TASKQUEUE_STATS 47 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 }; 61 62 public: 63 inline TaskQueueStats() { reset(); } 64 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_attempt() { ++_stats[steal_attempt]; } 69 inline void record_steal() { ++_stats[steal]; } 70 inline void record_overflow(size_t new_length); 71 72 TaskQueueStats & operator +=(const TaskQueueStats & addend); 73 74 inline size_t get(StatId id) const { return _stats[id]; } 75 inline const size_t* get() const { return _stats; } 76 77 inline void reset(); 78 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; 84 85 DEBUG_ONLY(void verify() const;) 86 87 private: 88 size_t _stats[last_stat_id]; 89 static const char * const _names[last_stat_id]; 90 }; 91 92 void TaskQueueStats::record_overflow(size_t new_len) { 93 ++_stats[overflow]; 94 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len; 95 } 96 97 void TaskQueueStats::reset() { 98 memset(_stats, 0, sizeof(_stats)); 99 } 100 #endif // TASKQUEUE_STATS 101 102 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances. 103 104 template <unsigned int N, MEMFLAGS F> 105 class TaskQueueSuper: public CHeapObj<F> { 106 protected: 107 // Internal type for indexing the queue; also used for the tag. 108 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t; 109 110 // The first free element after the last one pushed (mod N). 111 volatile uint _bottom; 112 113 enum { MOD_N_MASK = N - 1 }; 114 115 class Age { 116 public: 117 Age(size_t data = 0) { _data = data; } 118 Age(const Age& age) { _data = age._data; } 119 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; } 120 121 Age get() const volatile { return _data; } 122 void set(Age age) volatile { _data = age._data; } 123 124 idx_t top() const volatile { return _fields._top; } 125 idx_t tag() const volatile { return _fields._tag; } 126 127 // Increment top; if it wraps, increment tag also. 128 void increment() { 129 _fields._top = increment_index(_fields._top); 130 if (_fields._top == 0) ++_fields._tag; 131 } 132 133 Age cmpxchg(const Age new_age, const Age old_age) volatile; 134 135 bool operator ==(const Age& other) const { return _data == other._data; } 136 137 private: 138 struct fields { 139 idx_t _top; 140 idx_t _tag; 141 }; 142 union { 143 size_t _data; 144 fields _fields; 145 }; 146 }; 147 148 volatile Age _age; 149 150 // These both operate mod N. 151 static uint increment_index(uint ind) { 152 return (ind + 1) & MOD_N_MASK; 153 } 154 static uint decrement_index(uint ind) { 155 return (ind - 1) & MOD_N_MASK; 156 } 157 158 // Returns a number in the range [0..N). If the result is "N-1", it should be 159 // interpreted as 0. 160 uint dirty_size(uint bot, uint top) const { 161 return (bot - top) & MOD_N_MASK; 162 } 163 164 // Returns the size corresponding to the given "bot" and "top". 165 uint size(uint bot, uint top) const { 166 uint sz = dirty_size(bot, top); 167 // Has the queue "wrapped", so that bottom is less than top? There's a 168 // complicated special case here. A pair of threads could perform pop_local 169 // and pop_global operations concurrently, starting from a state in which 170 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom, 171 // and the pop_global in incrementing _top (in which case the pop_global 172 // will be awarded the contested queue element.) The resulting state must 173 // be interpreted as an empty queue. (We only need to worry about one such 174 // event: only the queue owner performs pop_local's, and several concurrent 175 // threads attempting to perform the pop_global will all perform the same 176 // CAS, and only one can succeed.) Any stealing thread that reads after 177 // either the increment or decrement will see an empty queue, and will not 178 // join the competitors. The "sz == -1 || sz == N-1" state will not be 179 // modified by concurrent queues, so the owner thread can reset the state to 180 // _bottom == top so subsequent pushes will be performed normally. 181 return (sz == N - 1) ? 0 : sz; 182 } 183 184 public: 185 TaskQueueSuper() : _bottom(0), _age() {} 186 187 // Return true if the TaskQueue contains/does not contain any tasks. 188 bool peek() const { return _bottom != _age.top(); } 189 bool is_empty() const { return size() == 0; } 190 191 // Return an estimate of the number of elements in the queue. 192 // The "careful" version admits the possibility of pop_local/pop_global 193 // races. 194 uint size() const { 195 return size(_bottom, _age.top()); 196 } 197 198 uint dirty_size() const { 199 return dirty_size(_bottom, _age.top()); 200 } 201 202 void set_empty() { 203 _bottom = 0; 204 _age.set(0); 205 } 206 207 // Maximum number of elements allowed in the queue. This is two less 208 // than the actual queue size, for somewhat complicated reasons. 209 uint max_elems() const { return N - 2; } 210 211 // Total size of queue. 212 static const uint total_size() { return N; } 213 214 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;) 215 }; 216 217 // 218 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double- 219 // ended-queue (deque), intended for use in work stealing. Queue operations 220 // are non-blocking. 221 // 222 // A queue owner thread performs push() and pop_local() operations on one end 223 // of the queue, while other threads may steal work using the pop_global() 224 // method. 225 // 226 // The main difference to the original algorithm is that this 227 // implementation allows wrap-around at the end of its allocated 228 // storage, which is an array. 229 // 230 // The original paper is: 231 // 232 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G. 233 // Thread scheduling for multiprogrammed multiprocessors. 234 // Theory of Computing Systems 34, 2 (2001), 115-144. 235 // 236 // The following paper provides an correctness proof and an 237 // implementation for weakly ordered memory models including (pseudo-) 238 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is 239 // similar to ABP, with the main difference that it allows resizing of the 240 // underlying storage: 241 // 242 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z. 243 // Correct and efficient work-stealing for weak memory models 244 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and 245 // practice of parallel programming (PPoPP 2013), 69-80 246 // 247 248 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 249 class GenericTaskQueue: public TaskQueueSuper<N, F> { 250 protected: 251 typedef typename TaskQueueSuper<N, F>::Age Age; 252 typedef typename TaskQueueSuper<N, F>::idx_t idx_t; 253 254 using TaskQueueSuper<N, F>::_bottom; 255 using TaskQueueSuper<N, F>::_age; 256 using TaskQueueSuper<N, F>::increment_index; 257 using TaskQueueSuper<N, F>::decrement_index; 258 using TaskQueueSuper<N, F>::dirty_size; 259 260 public: 261 using TaskQueueSuper<N, F>::max_elems; 262 using TaskQueueSuper<N, F>::size; 263 264 #if TASKQUEUE_STATS 265 using TaskQueueSuper<N, F>::stats; 266 #endif 267 268 private: 269 // Slow paths for push, pop_local. (pop_global has no fast path.) 270 bool push_slow(E t, uint dirty_n_elems); 271 bool pop_local_slow(uint localBot, Age oldAge); 272 273 public: 274 typedef E element_type; 275 276 // Initializes the queue to empty. 277 GenericTaskQueue(); 278 279 void initialize(); 280 281 // Push the task "t" on the queue. Returns "false" iff the queue is full. 282 inline bool push(E t); 283 284 // Attempts to claim a task from the "local" end of the queue (the most 285 // recently pushed) as long as the number of entries exceeds the threshold. 286 // If successful, returns true and sets t to the task; otherwise, returns false 287 // (the queue is empty or the number of elements below the threshold). 288 inline bool pop_local(volatile E& t, uint threshold = 0); 289 290 // Like pop_local(), but uses the "global" end of the queue (the least 291 // recently pushed). 292 bool pop_global(volatile E& t); 293 294 // Delete any resource associated with the queue. 295 ~GenericTaskQueue(); 296 297 // Apply fn to each element in the task queue. The queue must not 298 // be modified while iterating. 299 template<typename Fn> void iterate(Fn fn); 300 301 private: 302 DEFINE_PAD_MINUS_SIZE(0, DEFAULT_CACHE_LINE_SIZE, 0); 303 // Element array. 304 volatile E* _elems; 305 306 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(E*)); 307 // Queue owner local variables. Not to be accessed by other threads. 308 309 static const uint InvalidQueueId = uint(-1); 310 uint _last_stolen_queue_id; // The id of the queue we last stole from 311 312 int _seed; // Current random seed used for selecting a random queue during stealing. 313 314 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(uint) + sizeof(int)); 315 public: 316 int next_random_queue_id(); 317 318 void set_last_stolen_queue_id(uint id) { _last_stolen_queue_id = id; } 319 uint last_stolen_queue_id() const { return _last_stolen_queue_id; } 320 bool is_last_stolen_queue_id_valid() const { return _last_stolen_queue_id != InvalidQueueId; } 321 void invalidate_last_stolen_queue_id() { _last_stolen_queue_id = InvalidQueueId; } 322 }; 323 324 template<class E, MEMFLAGS F, unsigned int N> 325 GenericTaskQueue<E, F, N>::GenericTaskQueue() : _last_stolen_queue_id(InvalidQueueId), _seed(17 /* random number */) { 326 assert(sizeof(Age) == sizeof(size_t), "Depends on this."); 327 } 328 329 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for 330 // elements that do not fit in the TaskQueue. 331 // 332 // This class hides two methods from super classes: 333 // 334 // push() - push onto the task queue or, if that fails, onto the overflow stack 335 // is_empty() - return true if both the TaskQueue and overflow stack are empty 336 // 337 // Note that size() is not hidden--it returns the number of elements in the 338 // TaskQueue, and does not include the size of the overflow stack. This 339 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues. 340 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 341 class OverflowTaskQueue: public GenericTaskQueue<E, F, N> 342 { 343 public: 344 typedef Stack<E, F> overflow_t; 345 typedef GenericTaskQueue<E, F, N> taskqueue_t; 346 347 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;) 348 349 // Push task t onto the queue or onto the overflow stack. Return true. 350 inline bool push(E t); 351 // Try to push task t onto the queue only. Returns true if successful, false otherwise. 352 inline bool try_push_to_taskqueue(E t); 353 354 // Attempt to pop from the overflow stack; return true if anything was popped. 355 inline bool pop_overflow(E& t); 356 357 inline overflow_t* overflow_stack() { return &_overflow_stack; } 358 359 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); } 360 inline bool overflow_empty() const { return _overflow_stack.is_empty(); } 361 inline bool is_empty() const { 362 return taskqueue_empty() && overflow_empty(); 363 } 364 365 private: 366 overflow_t _overflow_stack; 367 }; 368 369 class TaskQueueSetSuper { 370 public: 371 // Returns "true" if some TaskQueue in the set contains a task. 372 virtual bool peek() = 0; 373 // Tasks in queue 374 virtual uint tasks() const = 0; 375 }; 376 377 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper { 378 }; 379 380 template<class T, MEMFLAGS F> 381 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> { 382 public: 383 typedef typename T::element_type E; 384 385 private: 386 uint _n; 387 T** _queues; 388 389 bool steal_best_of_2(uint queue_num, E& t); 390 391 public: 392 GenericTaskQueueSet(uint n); 393 ~GenericTaskQueueSet(); 394 395 void register_queue(uint i, T* q); 396 397 T* queue(uint n); 398 399 // Try to steal a task from some other queue than queue_num. It may perform several attempts at doing so. 400 // Returns if stealing succeeds, and sets "t" to the stolen task. 401 bool steal(uint queue_num, E& t); 402 403 bool peek(); 404 uint tasks() const; 405 406 uint size() const { return _n; } 407 }; 408 409 template<class T, MEMFLAGS F> void 410 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) { 411 assert(i < _n, "index out of range."); 412 _queues[i] = q; 413 } 414 415 template<class T, MEMFLAGS F> T* 416 GenericTaskQueueSet<T, F>::queue(uint i) { 417 return _queues[i]; 418 } 419 420 template<class T, MEMFLAGS F> 421 bool GenericTaskQueueSet<T, F>::peek() { 422 // Try all the queues. 423 for (uint j = 0; j < _n; j++) { 424 if (_queues[j]->peek()) 425 return true; 426 } 427 return false; 428 } 429 430 template<class T, MEMFLAGS F> 431 uint GenericTaskQueueSet<T, F>::tasks() const { 432 uint n = 0; 433 for (uint j = 0; j < _n; j++) { 434 n += _queues[j]->size(); 435 } 436 return n; 437 } 438 439 // When to terminate from the termination protocol. 440 class TerminatorTerminator: public CHeapObj<mtInternal> { 441 public: 442 virtual bool should_exit_termination() = 0; 443 }; 444 445 // A class to aid in the termination of a set of parallel tasks using 446 // TaskQueueSet's for work stealing. 447 448 #undef TRACESPINNING 449 450 class ParallelTaskTerminator: public StackObj { 451 private: 452 uint _n_threads; 453 TaskQueueSetSuper* _queue_set; 454 volatile uint _offered_termination; 455 456 #ifdef TRACESPINNING 457 static uint _total_yields; 458 static uint _total_spins; 459 static uint _total_peeks; 460 #endif 461 462 bool peek_in_queue_set(); 463 protected: 464 virtual void yield(); 465 void sleep(uint millis); 466 467 public: 468 469 // "n_threads" is the number of threads to be terminated. "queue_set" is a 470 // queue sets of work queues of other threads. 471 ParallelTaskTerminator(uint n_threads, TaskQueueSetSuper* queue_set); 472 473 // The current thread has no work, and is ready to terminate if everyone 474 // else is. If returns "true", all threads are terminated. If returns 475 // "false", available work has been observed in one of the task queues, 476 // so the global task is not complete. 477 bool offer_termination() { 478 return offer_termination(NULL); 479 } 480 481 // As above, but it also terminates if the should_exit_termination() 482 // method of the terminator parameter returns true. If terminator is 483 // NULL, then it is ignored. 484 bool offer_termination(TerminatorTerminator* terminator); 485 486 // Reset the terminator, so that it may be reused again. 487 // The caller is responsible for ensuring that this is done 488 // in an MT-safe manner, once the previous round of use of 489 // the terminator is finished. 490 void reset_for_reuse(); 491 // Same as above but the number of parallel threads is set to the 492 // given number. 493 void reset_for_reuse(uint n_threads); 494 495 #ifdef TRACESPINNING 496 static uint total_yields() { return _total_yields; } 497 static uint total_spins() { return _total_spins; } 498 static uint total_peeks() { return _total_peeks; } 499 static void print_termination_counts(); 500 #endif 501 }; 502 503 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue; 504 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet; 505 506 #ifdef _MSC_VER 507 #pragma warning(push) 508 // warning C4522: multiple assignment operators specified 509 #pragma warning(disable:4522) 510 #endif 511 512 // This is a container class for either an oop* or a narrowOop*. 513 // Both are pushed onto a task queue and the consumer will test is_narrow() 514 // to determine which should be processed. 515 class StarTask { 516 void* _holder; // either union oop* or narrowOop* 517 518 enum { COMPRESSED_OOP_MASK = 1 }; 519 520 public: 521 StarTask(narrowOop* p) { 522 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 523 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK); 524 } 525 StarTask(oop* p) { 526 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 527 _holder = (void*)p; 528 } 529 StarTask() { _holder = NULL; } 530 operator oop*() { return (oop*)_holder; } 531 operator narrowOop*() { 532 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK); 533 } 534 535 StarTask& operator=(const StarTask& t) { 536 _holder = t._holder; 537 return *this; 538 } 539 volatile StarTask& operator=(const volatile StarTask& t) volatile { 540 _holder = t._holder; 541 return *this; 542 } 543 544 bool is_narrow() const { 545 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0); 546 } 547 }; 548 549 class ObjArrayTask 550 { 551 public: 552 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { } 553 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) { 554 assert(idx <= size_t(max_jint), "too big"); 555 } 556 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { } 557 558 ObjArrayTask& operator =(const ObjArrayTask& t) { 559 _obj = t._obj; 560 _index = t._index; 561 return *this; 562 } 563 volatile ObjArrayTask& 564 operator =(const volatile ObjArrayTask& t) volatile { 565 (void)const_cast<oop&>(_obj = t._obj); 566 _index = t._index; 567 return *this; 568 } 569 570 inline oop obj() const { return _obj; } 571 inline int index() const { return _index; } 572 573 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid. 574 575 private: 576 oop _obj; 577 int _index; 578 }; 579 580 #ifdef _MSC_VER 581 #pragma warning(pop) 582 #endif 583 584 typedef OverflowTaskQueue<StarTask, mtGC> OopStarTaskQueue; 585 typedef GenericTaskQueueSet<OopStarTaskQueue, mtGC> OopStarTaskQueueSet; 586 587 typedef OverflowTaskQueue<size_t, mtGC> RegionTaskQueue; 588 typedef GenericTaskQueueSet<RegionTaskQueue, mtGC> RegionTaskQueueSet; 589 590 #endif // SHARE_VM_GC_SHARED_TASKQUEUE_HPP