1 /*
2 * Copyright (c) 2001, 2015, 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 "utilities/stack.hpp"
30
31 // Simple TaskQueue stats that are collected by default in debug builds.
32
33 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
34 #define TASKQUEUE_STATS 1
35 #elif !defined(TASKQUEUE_STATS)
36 #define TASKQUEUE_STATS 0
37 #endif
38
39 #if TASKQUEUE_STATS
40 #define TASKQUEUE_STATS_ONLY(code) code
41 #else
42 #define TASKQUEUE_STATS_ONLY(code)
43 #endif // TASKQUEUE_STATS
44
45 #if TASKQUEUE_STATS
46 class TaskQueueStats {
47 public:
48 enum StatId {
49 push, // number of taskqueue pushes
50 pop, // number of taskqueue pops
51 pop_slow, // subset of taskqueue pops that were done slow-path
52 steal_attempt, // number of taskqueue steal attempts
53 steal, // number of taskqueue steals
54 overflow, // number of overflow pushes
55 overflow_max_len, // max length of overflow stack
56 last_stat_id
57 };
58
59 public:
60 inline TaskQueueStats() { reset(); }
61
62 inline void record_push() { ++_stats[push]; }
63 inline void record_pop() { ++_stats[pop]; }
64 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
65 inline void record_steal(bool success);
66 inline void record_overflow(size_t new_length);
67
68 TaskQueueStats & operator +=(const TaskQueueStats & addend);
69
70 inline size_t get(StatId id) const { return _stats[id]; }
71 inline const size_t* get() const { return _stats; }
72
73 inline void reset();
74
75 // Print the specified line of the header (does not include a line separator).
76 static void print_header(unsigned int line, outputStream* const stream = tty,
77 unsigned int width = 10);
78 // Print the statistics (does not include a line separator).
79 void print(outputStream* const stream = tty, unsigned int width = 10) const;
80
81 DEBUG_ONLY(void verify() const;)
82 DEBUG_ONLY(void verify_only_pushes() const;)
83
84 private:
85 size_t _stats[last_stat_id];
86 static const char * const _names[last_stat_id];
87 };
88
89 void TaskQueueStats::record_steal(bool success) {
90 ++_stats[steal_attempt];
91 if (success) ++_stats[steal];
92 }
93
94 void TaskQueueStats::record_overflow(size_t new_len) {
95 ++_stats[overflow];
96 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
97 }
98
99 void TaskQueueStats::reset() {
100 memset(_stats, 0, sizeof(_stats));
101 }
102 #endif // TASKQUEUE_STATS
103
104 // TaskQueueSuper collects functionality common to all GenericTaskQueue instances.
105
106 template <unsigned int N, MEMFLAGS F>
107 class TaskQueueSuper: public CHeapObj<F> {
108 protected:
109 // Internal type for indexing the queue; also used for the tag.
110 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
111
112 // The first free element after the last one pushed (mod N).
113 volatile uint _bottom;
114
115 enum { MOD_N_MASK = N - 1 };
116
117 class Age {
118 public:
119 Age(size_t data = 0) { _data = data; }
120 Age(const Age& age) { _data = age._data; }
121 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; }
122
123 Age get() const volatile { return _data; }
124 void set(Age age) volatile { _data = age._data; }
125
126 idx_t top() const volatile { return _fields._top; }
127 idx_t tag() const volatile { return _fields._tag; }
128
129 // Increment top; if it wraps, increment tag also.
130 void increment() {
131 _fields._top = increment_index(_fields._top);
132 if (_fields._top == 0) ++_fields._tag;
133 }
134
135 Age cmpxchg(const Age new_age, const Age old_age) volatile;
136
137 bool operator ==(const Age& other) const { return _data == other._data; }
138
139 private:
140 struct fields {
141 idx_t _top;
142 idx_t _tag;
143 };
144 union {
145 size_t _data;
146 fields _fields;
147 };
148 };
149
150 volatile Age _age;
151
152 // These both operate mod N.
153 static uint increment_index(uint ind) {
154 return (ind + 1) & MOD_N_MASK;
155 }
156 static uint decrement_index(uint ind) {
157 return (ind - 1) & MOD_N_MASK;
158 }
159
160 // Returns a number in the range [0..N). If the result is "N-1", it should be
161 // interpreted as 0.
162 uint dirty_size(uint bot, uint top) const {
163 return (bot - top) & MOD_N_MASK;
164 }
165
166 // Returns the size corresponding to the given "bot" and "top".
167 uint size(uint bot, uint top) const {
168 uint sz = dirty_size(bot, top);
169 // Has the queue "wrapped", so that bottom is less than top? There's a
170 // complicated special case here. A pair of threads could perform pop_local
171 // and pop_global operations concurrently, starting from a state in which
172 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
173 // and the pop_global in incrementing _top (in which case the pop_global
174 // will be awarded the contested queue element.) The resulting state must
175 // be interpreted as an empty queue. (We only need to worry about one such
176 // event: only the queue owner performs pop_local's, and several concurrent
177 // threads attempting to perform the pop_global will all perform the same
178 // CAS, and only one can succeed.) Any stealing thread that reads after
179 // either the increment or decrement will see an empty queue, and will not
180 // join the competitors. The "sz == -1 || sz == N-1" state will not be
181 // modified by concurrent queues, so the owner thread can reset the state to
182 // _bottom == top so subsequent pushes will be performed normally.
183 return (sz == N - 1) ? 0 : sz;
184 }
185
186 public:
187 TaskQueueSuper() : _bottom(0), _age() {}
188
189 // Return true if the TaskQueue contains/does not contain any tasks.
190 bool peek() const { return _bottom != _age.top(); }
191 bool is_empty() const { return size() == 0; }
192
193 // Return an estimate of the number of elements in the queue.
194 // The "careful" version admits the possibility of pop_local/pop_global
195 // races.
196 uint size() const {
197 return size(_bottom, _age.top());
198 }
199
200 uint dirty_size() const {
201 return dirty_size(_bottom, _age.top());
202 }
203
204 void set_empty() {
205 _bottom = 0;
206 _age.set(0);
207 }
208
209 // Maximum number of elements allowed in the queue. This is two less
210 // than the actual queue size, for somewhat complicated reasons.
211 uint max_elems() const { return N - 2; }
212
213 // Total size of queue.
214 static const uint total_size() { return N; }
215
216 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
217 };
218
219 //
220 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
221 // ended-queue (deque), intended for use in work stealing. Queue operations
222 // are non-blocking.
223 //
224 // A queue owner thread performs push() and pop_local() operations on one end
225 // of the queue, while other threads may steal work using the pop_global()
226 // method.
227 //
228 // The main difference to the original algorithm is that this
229 // implementation allows wrap-around at the end of its allocated
230 // storage, which is an array.
231 //
232 // The original paper is:
233 //
234 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
235 // Thread scheduling for multiprogrammed multiprocessors.
236 // Theory of Computing Systems 34, 2 (2001), 115-144.
237 //
238 // The following paper provides an correctness proof and an
239 // implementation for weakly ordered memory models including (pseudo-)
240 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
241 // similar to ABP, with the main difference that it allows resizing of the
242 // underlying storage:
243 //
244 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
245 // Correct and efficient work-stealing for weak memory models
246 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
247 // practice of parallel programming (PPoPP 2013), 69-80
248 //
249
250 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
251 class GenericTaskQueue: public TaskQueueSuper<N, F> {
252 ArrayAllocator<E, F> _array_allocator;
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). If successful, returns true and sets t to the task;
289 // otherwise, returns false (the queue is empty).
290 inline bool pop_local(volatile E& t);
291
292 // Like pop_local(), but uses the "global" end of the queue (the least
293 // recently pushed).
294 bool pop_global(volatile E& t);
295
296 // Delete any resource associated with the queue.
297 ~GenericTaskQueue();
298
299 // Apply fn to each element in the task queue. The queue must not
300 // be modified while iterating.
301 template<typename Fn> void iterate(Fn fn);
302
303 private:
304 // Element array.
305 volatile E* _elems;
306 };
307
308 template<class E, MEMFLAGS F, unsigned int N>
309 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
310 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
311 }
312
313 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
314 // elements that do not fit in the TaskQueue.
315 //
316 // This class hides two methods from super classes:
317 //
318 // push() - push onto the task queue or, if that fails, onto the overflow stack
319 // is_empty() - return true if both the TaskQueue and overflow stack are empty
320 //
321 // Note that size() is not hidden--it returns the number of elements in the
322 // TaskQueue, and does not include the size of the overflow stack. This
323 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
324 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
325 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
326 {
327 public:
328 typedef Stack<E, F> overflow_t;
329 typedef GenericTaskQueue<E, F, N> taskqueue_t;
330
331 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
332
333 // Push task t onto the queue or onto the overflow stack. Return true.
334 inline bool push(E t);
335
336 // Attempt to pop from the overflow stack; return true if anything was popped.
337 inline bool pop_overflow(E& t);
338
339 inline overflow_t* overflow_stack() { return &_overflow_stack; }
340
341 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
342 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
343 inline bool is_empty() const {
344 return taskqueue_empty() && overflow_empty();
345 }
346
347 private:
348 overflow_t _overflow_stack;
349 };
350
351 class TaskQueueSetSuper {
352 protected:
353 static int randomParkAndMiller(int* seed0);
354 public:
355 // Returns "true" if some TaskQueue in the set contains a task.
356 virtual bool peek() = 0;
357 };
358
359 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
360 };
361
362 template<class T, MEMFLAGS F>
363 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
364 private:
365 uint _n;
366 T** _queues;
367
368 public:
369 typedef typename T::element_type E;
370
371 GenericTaskQueueSet(int n);
372
373 bool steal_best_of_2(uint queue_num, int* seed, E& t);
374
375 void register_queue(uint i, T* q);
376
377 T* queue(uint n);
378
379 // The thread with queue number "queue_num" (and whose random number seed is
380 // at "seed") is trying to steal a task from some other queue. (It may try
381 // several queues, according to some configuration parameter.) If some steal
382 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
383 // false.
384 bool steal(uint queue_num, int* seed, E& t);
385
386 bool peek();
387
388 uint size() const { return _n; }
389 };
390
391 template<class T, MEMFLAGS F> void
392 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
393 assert(i < _n, "index out of range.");
394 _queues[i] = q;
395 }
396
397 template<class T, MEMFLAGS F> T*
398 GenericTaskQueueSet<T, F>::queue(uint i) {
399 return _queues[i];
400 }
401
402 template<class T, MEMFLAGS F>
403 bool GenericTaskQueueSet<T, F>::peek() {
404 // Try all the queues.
405 for (uint j = 0; j < _n; j++) {
406 if (_queues[j]->peek())
407 return true;
408 }
409 return false;
410 }
411
412 // When to terminate from the termination protocol.
413 class TerminatorTerminator: public CHeapObj<mtInternal> {
414 public:
415 virtual bool should_exit_termination() = 0;
416 };
417
418 // A class to aid in the termination of a set of parallel tasks using
419 // TaskQueueSet's for work stealing.
420
421 #undef TRACESPINNING
422
423 class ParallelTaskTerminator: public StackObj {
424 private:
425 uint _n_threads;
426 TaskQueueSetSuper* _queue_set;
427 uint _offered_termination;
428
429 #ifdef TRACESPINNING
430 static uint _total_yields;
431 static uint _total_spins;
432 static uint _total_peeks;
433 #endif
434
435 bool peek_in_queue_set();
436 protected:
437 virtual void yield();
438 void sleep(uint millis);
439
440 public:
441
442 // "n_threads" is the number of threads to be terminated. "queue_set" is a
443 // queue sets of work queues of other threads.
444 ParallelTaskTerminator(uint n_threads, TaskQueueSetSuper* queue_set);
445
446 // The current thread has no work, and is ready to terminate if everyone
447 // else is. If returns "true", all threads are terminated. If returns
448 // "false", available work has been observed in one of the task queues,
449 // so the global task is not complete.
450 bool offer_termination() {
451 return offer_termination(NULL);
452 }
453
454 // As above, but it also terminates if the should_exit_termination()
455 // method of the terminator parameter returns true. If terminator is
456 // NULL, then it is ignored.
457 bool offer_termination(TerminatorTerminator* terminator);
458
459 // Reset the terminator, so that it may be reused again.
460 // The caller is responsible for ensuring that this is done
461 // in an MT-safe manner, once the previous round of use of
462 // the terminator is finished.
463 void reset_for_reuse();
464 // Same as above but the number of parallel threads is set to the
465 // given number.
466 void reset_for_reuse(uint n_threads);
467
468 #ifdef TRACESPINNING
469 static uint total_yields() { return _total_yields; }
470 static uint total_spins() { return _total_spins; }
471 static uint total_peeks() { return _total_peeks; }
472 static void print_termination_counts();
473 #endif
474 };
475
476 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue;
477 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
478
479 #ifdef _MSC_VER
480 #pragma warning(push)
481 // warning C4522: multiple assignment operators specified
482 #pragma warning(disable:4522)
483 #endif
484
485 // This is a container class for either an oop* or a narrowOop*.
486 // Both are pushed onto a task queue and the consumer will test is_narrow()
487 // to determine which should be processed.
488 class StarTask {
489 void* _holder; // either union oop* or narrowOop*
490
491 enum { COMPRESSED_OOP_MASK = 1 };
492
493 public:
494 StarTask(narrowOop* p) {
495 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
496 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
497 }
498 StarTask(oop* p) {
499 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
500 _holder = (void*)p;
501 }
502 StarTask() { _holder = NULL; }
503 operator oop*() { return (oop*)_holder; }
504 operator narrowOop*() {
505 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
506 }
507
508 StarTask& operator=(const StarTask& t) {
509 _holder = t._holder;
510 return *this;
511 }
512 volatile StarTask& operator=(const volatile StarTask& t) volatile {
513 _holder = t._holder;
514 return *this;
515 }
516
517 bool is_narrow() const {
518 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
519 }
520 };
521
522 class ObjArrayTask
523 {
524 public:
525 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
526 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
527 assert(idx <= size_t(max_jint), "too big");
528 }
529 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
530
531 ObjArrayTask& operator =(const ObjArrayTask& t) {
532 _obj = t._obj;
533 _index = t._index;
534 return *this;
535 }
536 volatile ObjArrayTask&
537 operator =(const volatile ObjArrayTask& t) volatile {
538 (void)const_cast<oop&>(_obj = t._obj);
539 _index = t._index;
540 return *this;
541 }
542
543 inline oop obj() const { return _obj; }
544 inline int index() const { return _index; }
545
546 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
547
548 private:
549 oop _obj;
550 int _index;
551 };
552
553 #ifdef _MSC_VER
554 #pragma warning(pop)
555 #endif
556
557 typedef OverflowTaskQueue<StarTask, mtClass> OopStarTaskQueue;
558 typedef GenericTaskQueueSet<OopStarTaskQueue, mtClass> OopStarTaskQueueSet;
559
560 typedef OverflowTaskQueue<size_t, mtInternal> RegionTaskQueue;
561 typedef GenericTaskQueueSet<RegionTaskQueue, mtClass> RegionTaskQueueSet;
562
563
564 #endif // SHARE_VM_GC_SHARED_TASKQUEUE_HPP