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(bool success); 69 inline void record_overflow(size_t new_length); 70 71 TaskQueueStats & operator +=(const TaskQueueStats & addend); 72 73 inline size_t get(StatId id) const { return _stats[id]; } 74 inline const size_t* get() const { return _stats; } 75 76 inline void reset(); 77 78 // Print the specified line of the header (does not include a line separator). 79 static void print_header(unsigned int line, outputStream* const stream = tty, 80 unsigned int width = 10); 81 // Print the statistics (does not include a line separator). 82 void print(outputStream* const stream = tty, unsigned int width = 10) const; 83 84 DEBUG_ONLY(void verify() const;) 85 86 private: 87 size_t _stats[last_stat_id]; 88 static const char * const _names[last_stat_id]; 89 }; 90 91 void TaskQueueStats::record_steal(bool success) { 92 ++_stats[steal_attempt]; 93 if (success) ++_stats[steal]; 94 } 95 96 void TaskQueueStats::record_overflow(size_t new_len) { 97 ++_stats[overflow]; 98 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len; 99 } 100 101 void TaskQueueStats::reset() { 102 memset(_stats, 0, sizeof(_stats)); 103 } 104 #endif // TASKQUEUE_STATS 105 106 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances. 107 108 template <unsigned int N, MEMFLAGS F> 109 class TaskQueueSuper: public CHeapObj<F> { 110 protected: 111 // Internal type for indexing the queue; also used for the tag. 112 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t; 113 114 // The first free element after the last one pushed (mod N). 115 volatile uint _bottom; 116 117 enum { MOD_N_MASK = N - 1 }; 118 119 class Age { 120 public: 121 Age(size_t data = 0) { _data = data; } 122 Age(const Age& age) { _data = age._data; } 123 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; } 124 125 Age get() const volatile { return _data; } 126 void set(Age age) volatile { _data = age._data; } 127 128 idx_t top() const volatile { return _fields._top; } 129 idx_t tag() const volatile { return _fields._tag; } 130 131 // Increment top; if it wraps, increment tag also. 132 void increment() { 133 _fields._top = increment_index(_fields._top); 134 if (_fields._top == 0) ++_fields._tag; 135 } 136 137 Age cmpxchg(const Age new_age, const Age old_age) volatile; 138 139 bool operator ==(const Age& other) const { return _data == other._data; } 140 141 private: 142 struct fields { 143 idx_t _top; 144 idx_t _tag; 145 }; 146 union { 147 size_t _data; 148 fields _fields; 149 }; 150 }; 151 152 volatile Age _age; 153 154 // These both operate mod N. 155 static uint increment_index(uint ind) { 156 return (ind + 1) & MOD_N_MASK; 157 } 158 static uint decrement_index(uint ind) { 159 return (ind - 1) & MOD_N_MASK; 160 } 161 162 // Returns a number in the range [0..N). If the result is "N-1", it should be 163 // interpreted as 0. 164 uint dirty_size(uint bot, uint top) const { 165 return (bot - top) & MOD_N_MASK; 166 } 167 168 // Returns the size corresponding to the given "bot" and "top". 169 uint size(uint bot, uint top) const { 170 uint sz = dirty_size(bot, top); 171 // Has the queue "wrapped", so that bottom is less than top? There's a 172 // complicated special case here. A pair of threads could perform pop_local 173 // and pop_global operations concurrently, starting from a state in which 174 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom, 175 // and the pop_global in incrementing _top (in which case the pop_global 176 // will be awarded the contested queue element.) The resulting state must 177 // be interpreted as an empty queue. (We only need to worry about one such 178 // event: only the queue owner performs pop_local's, and several concurrent 179 // threads attempting to perform the pop_global will all perform the same 180 // CAS, and only one can succeed.) Any stealing thread that reads after 181 // either the increment or decrement will see an empty queue, and will not 182 // join the competitors. The "sz == -1 || sz == N-1" state will not be 183 // modified by concurrent queues, so the owner thread can reset the state to 184 // _bottom == top so subsequent pushes will be performed normally. 185 return (sz == N - 1) ? 0 : sz; 186 } 187 188 public: 189 TaskQueueSuper() : _bottom(0), _age() {} 190 191 // Return true if the TaskQueue contains/does not contain any tasks. 192 bool peek() const { return _bottom != _age.top(); } 193 bool is_empty() const { return size() == 0; } 194 195 // Return an estimate of the number of elements in the queue. 196 // The "careful" version admits the possibility of pop_local/pop_global 197 // races. 198 uint size() const { 199 return size(_bottom, _age.top()); 200 } 201 202 uint dirty_size() const { 203 return dirty_size(_bottom, _age.top()); 204 } 205 206 void set_empty() { 207 _bottom = 0; 208 _age.set(0); 209 } 210 211 // Maximum number of elements allowed in the queue. This is two less 212 // than the actual queue size, for somewhat complicated reasons. 213 uint max_elems() const { return N - 2; } 214 215 // Total size of queue. 216 static const uint total_size() { return N; } 217 218 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;) 219 }; 220 221 // 222 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double- 223 // ended-queue (deque), intended for use in work stealing. Queue operations 224 // are non-blocking. 225 // 226 // A queue owner thread performs push() and pop_local() operations on one end 227 // of the queue, while other threads may steal work using the pop_global() 228 // method. 229 // 230 // The main difference to the original algorithm is that this 231 // implementation allows wrap-around at the end of its allocated 232 // storage, which is an array. 233 // 234 // The original paper is: 235 // 236 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G. 237 // Thread scheduling for multiprogrammed multiprocessors. 238 // Theory of Computing Systems 34, 2 (2001), 115-144. 239 // 240 // The following paper provides an correctness proof and an 241 // implementation for weakly ordered memory models including (pseudo-) 242 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is 243 // similar to ABP, with the main difference that it allows resizing of the 244 // underlying storage: 245 // 246 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z. 247 // Correct and efficient work-stealing for weak memory models 248 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and 249 // practice of parallel programming (PPoPP 2013), 69-80 250 // 251 252 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 253 class GenericTaskQueue: public TaskQueueSuper<N, F> { 254 protected: 255 typedef typename TaskQueueSuper<N, F>::Age Age; 256 typedef typename TaskQueueSuper<N, F>::idx_t idx_t; 257 258 using TaskQueueSuper<N, F>::_bottom; 259 using TaskQueueSuper<N, F>::_age; 260 using TaskQueueSuper<N, F>::increment_index; 261 using TaskQueueSuper<N, F>::decrement_index; 262 using TaskQueueSuper<N, F>::dirty_size; 263 264 public: 265 using TaskQueueSuper<N, F>::max_elems; 266 using TaskQueueSuper<N, F>::size; 267 268 #if TASKQUEUE_STATS 269 using TaskQueueSuper<N, F>::stats; 270 #endif 271 272 private: 273 // Slow paths for push, pop_local. (pop_global has no fast path.) 274 bool push_slow(E t, uint dirty_n_elems); 275 bool pop_local_slow(uint localBot, Age oldAge); 276 277 public: 278 typedef E element_type; 279 280 // Initializes the queue to empty. 281 GenericTaskQueue(); 282 283 void initialize(); 284 285 // Push the task "t" on the queue. Returns "false" iff the queue is full. 286 inline bool push(E t); 287 288 // Attempts to claim a task from the "local" end of the queue (the most 289 // recently pushed) as long as the number of entries exceeds the threshold. 290 // If successful, returns true and sets t to the task; otherwise, returns false 291 // (the queue is empty or the number of elements below the threshold). 292 inline bool pop_local(volatile E& t, uint threshold = 0); 293 294 // Like pop_local(), but uses the "global" end of the queue (the least 295 // recently pushed). 296 bool pop_global(volatile E& t); 297 298 // Delete any resource associated with the queue. 299 ~GenericTaskQueue(); 300 301 // Apply fn to each element in the task queue. The queue must not 302 // be modified while iterating. 303 template<typename Fn> void iterate(Fn fn); 304 305 private: 306 // Element array. 307 volatile E* _elems; 308 }; 309 310 template<class E, MEMFLAGS F, unsigned int N> 311 GenericTaskQueue<E, F, N>::GenericTaskQueue() { 312 assert(sizeof(Age) == sizeof(size_t), "Depends on this."); 313 } 314 315 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for 316 // elements that do not fit in the TaskQueue. 317 // 318 // This class hides two methods from super classes: 319 // 320 // push() - push onto the task queue or, if that fails, onto the overflow stack 321 // is_empty() - return true if both the TaskQueue and overflow stack are empty 322 // 323 // Note that size() is not hidden--it returns the number of elements in the 324 // TaskQueue, and does not include the size of the overflow stack. This 325 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues. 326 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE> 327 class OverflowTaskQueue: public GenericTaskQueue<E, F, N> 328 { 329 public: 330 typedef Stack<E, F> overflow_t; 331 typedef GenericTaskQueue<E, F, N> taskqueue_t; 332 333 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;) 334 335 // Push task t onto the queue or onto the overflow stack. Return true. 336 inline bool push(E t); 337 // Try to push task t onto the queue only. Returns true if successful, false otherwise. 338 inline bool try_push_to_taskqueue(E t); 339 340 // Attempt to pop from the overflow stack; return true if anything was popped. 341 inline bool pop_overflow(E& t); 342 343 inline overflow_t* overflow_stack() { return &_overflow_stack; } 344 345 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); } 346 inline bool overflow_empty() const { return _overflow_stack.is_empty(); } 347 inline bool is_empty() const { 348 return taskqueue_empty() && overflow_empty(); 349 } 350 351 private: 352 overflow_t _overflow_stack; 353 }; 354 355 class TaskQueueSetSuper { 356 protected: 357 static int randomParkAndMiller(int* seed0); 358 public: 359 // Returns "true" if some TaskQueue in the set contains a task. 360 virtual bool peek() = 0; 361 }; 362 363 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper { 364 }; 365 366 template<class T, MEMFLAGS F> 367 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> { 368 private: 369 uint _n; 370 T** _queues; 371 372 public: 373 typedef typename T::element_type E; 374 375 GenericTaskQueueSet(int n); 376 ~GenericTaskQueueSet(); 377 378 bool steal_best_of_2(uint queue_num, int* seed, E& t); 379 380 void register_queue(uint i, T* q); 381 382 T* queue(uint n); 383 384 // The thread with queue number "queue_num" (and whose random number seed is 385 // at "seed") is trying to steal a task from some other queue. (It may try 386 // several queues, according to some configuration parameter.) If some steal 387 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns 388 // false. 389 bool steal(uint queue_num, int* seed, E& t); 390 391 bool peek(); 392 393 uint size() const { return _n; } 394 }; 395 396 template<class T, MEMFLAGS F> void 397 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) { 398 assert(i < _n, "index out of range."); 399 _queues[i] = q; 400 } 401 402 template<class T, MEMFLAGS F> T* 403 GenericTaskQueueSet<T, F>::queue(uint i) { 404 return _queues[i]; 405 } 406 407 template<class T, MEMFLAGS F> 408 bool GenericTaskQueueSet<T, F>::peek() { 409 // Try all the queues. 410 for (uint j = 0; j < _n; j++) { 411 if (_queues[j]->peek()) 412 return true; 413 } 414 return false; 415 } 416 417 // When to terminate from the termination protocol. 418 class TerminatorTerminator: public CHeapObj<mtInternal> { 419 public: 420 virtual bool should_exit_termination() = 0; 421 }; 422 423 // A class to aid in the termination of a set of parallel tasks using 424 // TaskQueueSet's for work stealing. 425 426 #undef TRACESPINNING 427 428 class ParallelTaskTerminator: public StackObj { 429 private: 430 uint _n_threads; 431 TaskQueueSetSuper* _queue_set; 432 433 DEFINE_PAD_MINUS_SIZE(0, DEFAULT_CACHE_LINE_SIZE, 0); 434 volatile uint _offered_termination; 435 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile uint)); 436 437 #ifdef TRACESPINNING 438 static uint _total_yields; 439 static uint _total_spins; 440 static uint _total_peeks; 441 #endif 442 443 bool peek_in_queue_set(); 444 protected: 445 virtual void yield(); 446 void sleep(uint millis); 447 448 public: 449 450 // "n_threads" is the number of threads to be terminated. "queue_set" is a 451 // queue sets of work queues of other threads. 452 ParallelTaskTerminator(uint n_threads, TaskQueueSetSuper* queue_set); 453 454 // The current thread has no work, and is ready to terminate if everyone 455 // else is. If returns "true", all threads are terminated. If returns 456 // "false", available work has been observed in one of the task queues, 457 // so the global task is not complete. 458 bool offer_termination() { 459 return offer_termination(NULL); 460 } 461 462 // As above, but it also terminates if the should_exit_termination() 463 // method of the terminator parameter returns true. If terminator is 464 // NULL, then it is ignored. 465 bool offer_termination(TerminatorTerminator* terminator); 466 467 // Reset the terminator, so that it may be reused again. 468 // The caller is responsible for ensuring that this is done 469 // in an MT-safe manner, once the previous round of use of 470 // the terminator is finished. 471 void reset_for_reuse(); 472 // Same as above but the number of parallel threads is set to the 473 // given number. 474 void reset_for_reuse(uint n_threads); 475 476 #ifdef TRACESPINNING 477 static uint total_yields() { return _total_yields; } 478 static uint total_spins() { return _total_spins; } 479 static uint total_peeks() { return _total_peeks; } 480 static void print_termination_counts(); 481 #endif 482 }; 483 484 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue; 485 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet; 486 487 #ifdef _MSC_VER 488 #pragma warning(push) 489 // warning C4522: multiple assignment operators specified 490 #pragma warning(disable:4522) 491 #endif 492 493 // This is a container class for either an oop* or a narrowOop*. 494 // Both are pushed onto a task queue and the consumer will test is_narrow() 495 // to determine which should be processed. 496 class StarTask { 497 void* _holder; // either union oop* or narrowOop* 498 499 enum { COMPRESSED_OOP_MASK = 1 }; 500 501 public: 502 StarTask(narrowOop* p) { 503 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 504 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK); 505 } 506 StarTask(oop* p) { 507 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!"); 508 _holder = (void*)p; 509 } 510 StarTask() { _holder = NULL; } 511 operator oop*() { return (oop*)_holder; } 512 operator narrowOop*() { 513 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK); 514 } 515 516 StarTask& operator=(const StarTask& t) { 517 _holder = t._holder; 518 return *this; 519 } 520 volatile StarTask& operator=(const volatile StarTask& t) volatile { 521 _holder = t._holder; 522 return *this; 523 } 524 525 bool is_narrow() const { 526 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0); 527 } 528 }; 529 530 class ObjArrayTask 531 { 532 public: 533 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { } 534 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) { 535 assert(idx <= size_t(max_jint), "too big"); 536 } 537 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { } 538 539 ObjArrayTask& operator =(const ObjArrayTask& t) { 540 _obj = t._obj; 541 _index = t._index; 542 return *this; 543 } 544 volatile ObjArrayTask& 545 operator =(const volatile ObjArrayTask& t) volatile { 546 (void)const_cast<oop&>(_obj = t._obj); 547 _index = t._index; 548 return *this; 549 } 550 551 inline oop obj() const { return _obj; } 552 inline int index() const { return _index; } 553 554 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid. 555 556 private: 557 oop _obj; 558 int _index; 559 }; 560 561 #ifdef _MSC_VER 562 #pragma warning(pop) 563 #endif 564 565 typedef OverflowTaskQueue<StarTask, mtGC> OopStarTaskQueue; 566 typedef GenericTaskQueueSet<OopStarTaskQueue, mtGC> OopStarTaskQueueSet; 567 568 typedef OverflowTaskQueue<size_t, mtGC> RegionTaskQueue; 569 typedef GenericTaskQueueSet<RegionTaskQueue, mtGC> RegionTaskQueueSet; 570 571 #endif // SHARE_VM_GC_SHARED_TASKQUEUE_HPP