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 "oops/oopsHierarchy.hpp" 30 #include "utilities/ostream.hpp" 31 #include "utilities/stack.hpp" 32 33 // Simple TaskQueue stats that are collected by default in debug builds. 34 35 #if !defined(TASKQUEUE_STATS) && defined(ASSERT) 36 #define TASKQUEUE_STATS 1 37 #elif !defined(TASKQUEUE_STATS) 38 #define TASKQUEUE_STATS 0 39 #endif 40 41 #if TASKQUEUE_STATS 42 #define TASKQUEUE_STATS_ONLY(code) code 43 #else 44 #define TASKQUEUE_STATS_ONLY(code) 45 #endif // TASKQUEUE_STATS 46 47 #if TASKQUEUE_STATS 48 class TaskQueueStats { 49 public: 50 enum StatId { 51 push, // number of taskqueue pushes 52 pop, // number of taskqueue pops 53 pop_slow, // subset of taskqueue pops that were done slow-path 54 steal_attempt, // number of taskqueue steal attempts 55 steal, // number of taskqueue steals 56 overflow, // number of overflow pushes 57 overflow_max_len, // max length of overflow stack 58 last_stat_id 59 }; 60 61 public: 62 inline TaskQueueStats() { reset(); } 63 64 inline void record_push() { ++_stats[push]; } 65 inline void record_pop() { ++_stats[pop]; } 66 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; } 67 inline void record_steal(bool success); 68 inline void record_overflow(size_t new_length); 69 70 TaskQueueStats & operator +=(const TaskQueueStats & addend); 71 72 inline size_t get(StatId id) const { return _stats[id]; } 73 inline const size_t* get() const { return _stats; } 74 75 inline void reset(); 76 77 // Print the specified line of the header (does not include a line separator). 78 static void print_header(unsigned int line, outputStream* const stream = tty, 79 unsigned int width = 10); 80 // Print the statistics (does not include a line separator). 81 void print(outputStream* const stream = tty, unsigned int width = 10) const; 82 83 DEBUG_ONLY(void verify() const;) 84 85 private: 86 size_t _stats[last_stat_id]; 87 static const char * const _names[last_stat_id]; 88 }; 89 90 void TaskQueueStats::record_steal(bool success) { 91 ++_stats[steal_attempt]; 92 if (success) ++_stats[steal]; 93 } 94 95 void TaskQueueStats::record_overflow(size_t new_len) { 96 ++_stats[overflow]; 97 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len; 98 } 99 100 void TaskQueueStats::reset() { 101 memset(_stats, 0, sizeof(_stats)); 102 } 103 #endif // TASKQUEUE_STATS 104 105 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances. 106 107 template <unsigned int N, MEMFLAGS F> 108 class TaskQueueSuper: public CHeapObj<F> { 109 protected: 110 // Internal type for indexing the queue; also used for the tag. 111 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t; 112 113 // The first free element after the last one pushed (mod N). 114 volatile uint _bottom; 115 116 enum { MOD_N_MASK = N - 1 }; 117 118 class Age { 119 public: 120 Age(size_t data = 0) { _data = data; } 121 Age(const Age& age) { _data = age._data; } 122 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; } 123 124 Age get() const volatile { return _data; } 125 void set(Age age) volatile { _data = age._data; } 126 127 idx_t top() const volatile { return _fields._top; } 128 idx_t tag() const volatile { return _fields._tag; } 129 130 // Increment top; if it wraps, increment tag also. 131 void increment() { 132 _fields._top = increment_index(_fields._top); 133 if (_fields._top == 0) ++_fields._tag; 134 } 135 136 Age cmpxchg(const Age new_age, const Age old_age) volatile; 137 138 bool operator ==(const Age& other) const { return _data == other._data; } 139 140 private: 141 struct fields { 142 idx_t _top; 143 idx_t _tag; 144 }; 145 union { 146 size_t _data; 147 fields _fields; 148 }; 149 }; 150 151 volatile Age _age; 152 153 // These both operate mod N. 154 static uint increment_index(uint ind) { 155 return (ind + 1) & MOD_N_MASK; 156 } 157 static uint decrement_index(uint ind) { 158 return (ind - 1) & MOD_N_MASK; 159 } 160 161 // Returns a number in the range [0..N). If the result is "N-1", it should be 162 // interpreted as 0. 163 uint dirty_size(uint bot, uint top) const { 164 return (bot - top) & MOD_N_MASK; 165 } 166 167 // Returns the size corresponding to the given "bot" and "top". 168 uint size(uint bot, uint top) const { 169 uint sz = dirty_size(bot, top); 170 // Has the queue "wrapped", so that bottom is less than top? There's a 171 // complicated special case here. A pair of threads could perform pop_local 172 // and pop_global operations concurrently, starting from a state in which 173 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom, 174 // and the pop_global in incrementing _top (in which case the pop_global 175 // will be awarded the contested queue element.) The resulting state must 176 // be interpreted as an empty queue. (We only need to worry about one such 177 // event: only the queue owner performs pop_local's, and several concurrent 178 // threads attempting to perform the pop_global will all perform the same 179 // CAS, and only one can succeed.) Any stealing thread that reads after 180 // either the increment or decrement will see an empty queue, and will not 181 // join the competitors. The "sz == -1 || sz == N-1" state will not be 182 // modified by concurrent queues, so the owner thread can reset the state to 183 // _bottom == top so subsequent pushes will be performed normally. 184 return (sz == N - 1) ? 0 : sz; 185 } 186 187 public: 188 TaskQueueSuper() : _bottom(0), _age() {} 189 190 // Return true if the TaskQueue contains/does not contain any tasks. 191 bool peek() const { return _bottom != _age.top(); } 192 bool is_empty() const { return size() == 0; } 193 194 // Return an estimate of the number of elements in the queue. 195 // The "careful" version admits the possibility of pop_local/pop_global 196 // races. 197 uint size() const { 198 return size(_bottom, _age.top()); 199 } 200 201 uint dirty_size() const { 202 return dirty_size(_bottom, _age.top()); 203 } 204 205 void set_empty() { 206 _bottom = 0; 207 _age.set(0); 208 } 209 210 // Maximum number of elements allowed in the queue. This is two less 211 // than the actual queue size, for somewhat complicated reasons. 212 uint max_elems() const { return N - 2; } 213 214 // Total size of queue. 215 static const uint total_size() { return N; } 216 217 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;) 218 }; 219 220 // 221 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double- 222 // ended-queue (deque), intended for use in work stealing. Queue operations 223 // are non-blocking. 224 // 225 // A queue owner thread performs push() and pop_local() operations on one end 226 // of the queue, while other threads may steal work using the pop_global() 227 // method. 228 // 229 // The main difference to the original algorithm is that this 230 // implementation allows wrap-around at the end of its allocated 231 // storage, which is an array. 232 // 233 // The original paper is: 234 // 235 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G. 236 // Thread scheduling for multiprogrammed multiprocessors. 237 // Theory of Computing Systems 34, 2 (2001), 115-144. 238 // 239 // The following paper provides an correctness proof and an 240 // implementation for weakly ordered memory models including (pseudo-) 241 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is 242 // similar to ABP, with the main difference that it allows resizing of the 243 // underlying storage: 244 // 245 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z. 246 // Correct and efficient work-stealing for weak memory models 247 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and 248 // practice of parallel programming (PPoPP 2013), 69-80 249 // 250 251 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 252 class GenericTaskQueue: public TaskQueueSuper<N, F> { 253 protected: 254 typedef typename TaskQueueSuper<N, F>::Age Age; 255 typedef typename TaskQueueSuper<N, F>::idx_t idx_t; 256 257 using TaskQueueSuper<N, F>::_bottom; 258 using TaskQueueSuper<N, F>::_age; 259 using TaskQueueSuper<N, F>::increment_index; 260 using TaskQueueSuper<N, F>::decrement_index; 261 using TaskQueueSuper<N, F>::dirty_size; 262 263 public: 264 using TaskQueueSuper<N, F>::max_elems; 265 using TaskQueueSuper<N, F>::size; 266 267 #if TASKQUEUE_STATS 268 using TaskQueueSuper<N, F>::stats; 269 #endif 270 271 private: 272 // Slow paths for push, pop_local. (pop_global has no fast path.) 273 bool push_slow(E t, uint dirty_n_elems); 274 bool pop_local_slow(uint localBot, Age oldAge); 275 276 public: 277 typedef E element_type; 278 279 // Initializes the queue to empty. 280 GenericTaskQueue(); 281 282 void initialize(); 283 284 // Push the task "t" on the queue. Returns "false" iff the queue is full. 285 inline bool push(E t); 286 287 // Attempts to claim a task from the "local" end of the queue (the most 288 // recently pushed) as long as the number of entries exceeds the threshold. 289 // If successful, returns true and sets t to the task; otherwise, returns false 290 // (the queue is empty or the number of elements below the threshold). 291 inline bool pop_local(volatile E& t, uint threshold = 0); 292 293 // Like pop_local(), but uses the "global" end of the queue (the least 294 // recently pushed). 295 bool pop_global(volatile E& t); 296 297 // Delete any resource associated with the queue. 298 ~GenericTaskQueue(); 299 300 // Apply fn to each element in the task queue. The queue must not 301 // be modified while iterating. 302 template<typename Fn> void iterate(Fn fn); 303 304 private: 305 // Element array. 306 volatile E* _elems; 307 }; 308 309 template<class E, MEMFLAGS F, unsigned int N> 310 GenericTaskQueue<E, F, N>::GenericTaskQueue() { 311 assert(sizeof(Age) == sizeof(size_t), "Depends on this."); 312 } 313 314 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for 315 // elements that do not fit in the TaskQueue. 316 // 317 // This class hides two methods from super classes: 318 // 319 // push() - push onto the task queue or, if that fails, onto the overflow stack 320 // is_empty() - return true if both the TaskQueue and overflow stack are empty 321 // 322 // Note that size() is not hidden--it returns the number of elements in the 323 // TaskQueue, and does not include the size of the overflow stack. This 324 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues. 325 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 326 class OverflowTaskQueue: public GenericTaskQueue<E, F, N> 327 { 328 public: 329 typedef Stack<E, F> overflow_t; 330 typedef GenericTaskQueue<E, F, N> taskqueue_t; 331 332 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;) 333 334 // Push task t onto the queue or onto the overflow stack. Return true. 335 inline bool push(E t); 336 // Try to push task t onto the queue only. Returns true if successful, false otherwise. 337 inline bool try_push_to_taskqueue(E t); 338 339 // Attempt to pop from the overflow stack; return true if anything was popped. 340 inline bool pop_overflow(E& t); 341 342 inline overflow_t* overflow_stack() { return &_overflow_stack; } 343 344 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); } 345 inline bool overflow_empty() const { return _overflow_stack.is_empty(); } 346 inline bool is_empty() const { 347 return taskqueue_empty() && overflow_empty(); 348 } 349 350 private: 351 overflow_t _overflow_stack; 352 }; 353 354 class TaskQueueSetSuper { 355 protected: 356 static int randomParkAndMiller(int* seed0); 357 public: 358 // Returns "true" if some TaskQueue in the set contains a task. 359 virtual bool peek() = 0; 360 }; 361 362 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper { 363 }; 364 365 template<class T, MEMFLAGS F> 366 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> { 367 private: 368 uint _n; 369 T** _queues; 370 371 public: 372 typedef typename T::element_type E; 373 374 GenericTaskQueueSet(int n); 375 ~GenericTaskQueueSet(); 376 377 bool steal_best_of_2(uint queue_num, int* seed, E& t); 378 379 void register_queue(uint i, T* q); 380 381 T* queue(uint n); 382 383 // The thread with queue number "queue_num" (and whose random number seed is 384 // at "seed") is trying to steal a task from some other queue. (It may try 385 // several queues, according to some configuration parameter.) If some steal 386 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns 387 // false. 388 bool steal(uint queue_num, int* seed, E& t); 389 390 bool peek(); 391 392 uint size() const { return _n; } 393 }; 394 395 template<class T, MEMFLAGS F> void 396 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) { 397 assert(i < _n, "index out of range."); 398 _queues[i] = q; 399 } 400 401 template<class T, MEMFLAGS F> T* 402 GenericTaskQueueSet<T, F>::queue(uint i) { 403 return _queues[i]; 404 } 405 406 template<class T, MEMFLAGS F> 407 bool GenericTaskQueueSet<T, F>::peek() { 408 // Try all the queues. 409 for (uint j = 0; j < _n; j++) { 410 if (_queues[j]->peek()) 411 return true; 412 } 413 return false; 414 } 415 416 // When to terminate from the termination protocol. 417 class TerminatorTerminator: public CHeapObj<mtInternal> { 418 public: 419 virtual bool should_exit_termination() = 0; 420 }; 421 422 // A class to aid in the termination of a set of parallel tasks using 423 // TaskQueueSet's for work stealing. 424 425 #undef TRACESPINNING 426 427 class ParallelTaskTerminator: public StackObj { 428 private: 429 uint _n_threads; 430 TaskQueueSetSuper* _queue_set; 431 volatile uint _offered_termination; 432 433 #ifdef TRACESPINNING 434 static uint _total_yields; 435 static uint _total_spins; 436 static uint _total_peeks; 437 #endif 438 439 bool peek_in_queue_set(); 440 protected: 441 virtual void yield(); 442 void sleep(uint millis); 443 444 public: 445 446 // "n_threads" is the number of threads to be terminated. "queue_set" is a 447 // queue sets of work queues of other threads. 448 ParallelTaskTerminator(uint n_threads, TaskQueueSetSuper* queue_set); 449 450 // The current thread has no work, and is ready to terminate if everyone 451 // else is. If returns "true", all threads are terminated. If returns 452 // "false", available work has been observed in one of the task queues, 453 // so the global task is not complete. 454 bool offer_termination() { 455 return offer_termination(NULL); 456 } 457 458 // As above, but it also terminates if the should_exit_termination() 459 // method of the terminator parameter returns true. If terminator is 460 // NULL, then it is ignored. 461 bool offer_termination(TerminatorTerminator* terminator); 462 463 // Reset the terminator, so that it may be reused again. 464 // The caller is responsible for ensuring that this is done 465 // in an MT-safe manner, once the previous round of use of 466 // the terminator is finished. 467 void reset_for_reuse(); 468 // Same as above but the number of parallel threads is set to the 469 // given number. 470 void reset_for_reuse(uint n_threads); 471 472 #ifdef TRACESPINNING 473 static uint total_yields() { return _total_yields; } 474 static uint total_spins() { return _total_spins; } 475 static uint total_peeks() { return _total_peeks; } 476 static void print_termination_counts(); 477 #endif 478 }; 479 480 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue; 481 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet; 482 483 #ifdef _MSC_VER 484 #pragma warning(push) 485 // warning C4522: multiple assignment operators specified 486 #pragma warning(disable:4522) 487 #endif 488 489 // This is a container class for either an oop* or a narrowOop*. 490 // Both are pushed onto a task queue and the consumer will test is_narrow() 491 // to determine which should be processed. 492 class StarTask { 493 void* _holder; // either union oop* or narrowOop* 494 495 enum { COMPRESSED_OOP_MASK = 1 }; 496 497 public: 498 StarTask(narrowOop* p) { 499 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 500 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK); 501 } 502 StarTask(oop* p) { 503 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 504 _holder = (void*)p; 505 } 506 StarTask() { _holder = NULL; } 507 operator oop*() { return (oop*)_holder; } 508 operator narrowOop*() { 509 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK); 510 } 511 512 StarTask& operator=(const StarTask& t) { 513 _holder = t._holder; 514 return *this; 515 } 516 volatile StarTask& operator=(const volatile StarTask& t) volatile { 517 _holder = t._holder; 518 return *this; 519 } 520 521 bool is_narrow() const { 522 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0); 523 } 524 }; 525 526 class ObjArrayTask 527 { 528 public: 529 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { } 530 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) { 531 assert(idx <= size_t(max_jint), "too big"); 532 } 533 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { } 534 535 ObjArrayTask& operator =(const ObjArrayTask& t) { 536 _obj = t._obj; 537 _index = t._index; 538 return *this; 539 } 540 volatile ObjArrayTask& 541 operator =(const volatile ObjArrayTask& t) volatile { 542 (void)const_cast<oop&>(_obj = t._obj); 543 _index = t._index; 544 return *this; 545 } 546 547 inline oop obj() const { return _obj; } 548 inline int index() const { return _index; } 549 550 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid. 551 552 private: 553 oop _obj; 554 int _index; 555 }; 556 557 #ifdef _MSC_VER 558 #pragma warning(pop) 559 #endif 560 561 typedef OverflowTaskQueue<StarTask, mtGC> OopStarTaskQueue; 562 typedef GenericTaskQueueSet<OopStarTaskQueue, mtGC> OopStarTaskQueueSet; 563 564 typedef OverflowTaskQueue<size_t, mtGC> RegionTaskQueue; 565 typedef GenericTaskQueueSet<RegionTaskQueue, mtGC> RegionTaskQueueSet; 566 567 #endif // SHARE_VM_GC_SHARED_TASKQUEUE_HPP