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