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 // Add paddings to reduce false-sharing cache contention between _bottom and _age
153 DEFINE_PAD_MINUS_SIZE(0, DEFAULT_CACHE_LINE_SIZE, sizeof(_bottom));
154 volatile Age _age;
155
156 // These both operate mod N.
157 static uint increment_index(uint ind) {
158 return (ind + 1) & MOD_N_MASK;
159 }
160 static uint decrement_index(uint ind) {
161 return (ind - 1) & MOD_N_MASK;
162 }
163
164 // Returns a number in the range [0..N). If the result is "N-1", it should be
165 // interpreted as 0.
166 uint dirty_size(uint bot, uint top) const {
167 return (bot - top) & MOD_N_MASK;
168 }
169
170 // Returns the size corresponding to the given "bot" and "top".
171 uint size(uint bot, uint top) const {
172 uint sz = dirty_size(bot, top);
173 // Has the queue "wrapped", so that bottom is less than top? There's a
174 // complicated special case here. A pair of threads could perform pop_local
175 // and pop_global operations concurrently, starting from a state in which
176 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
177 // and the pop_global in incrementing _top (in which case the pop_global
178 // will be awarded the contested queue element.) The resulting state must
179 // be interpreted as an empty queue. (We only need to worry about one such
180 // event: only the queue owner performs pop_local's, and several concurrent
181 // threads attempting to perform the pop_global will all perform the same
182 // CAS, and only one can succeed.) Any stealing thread that reads after
183 // either the increment or decrement will see an empty queue, and will not
184 // join the competitors. The "sz == -1 || sz == N-1" state will not be
185 // modified by concurrent queues, so the owner thread can reset the state to
186 // _bottom == top so subsequent pushes will be performed normally.
187 return (sz == N - 1) ? 0 : sz;
188 }
189
190 public:
191 TaskQueueSuper() : _bottom(0), _age() {}
192
193 // Return true if the TaskQueue contains/does not contain any tasks.
194 bool peek() const { return _bottom != _age.top(); }
195 bool is_empty() const { return size() == 0; }
196
197 // Return an estimate of the number of elements in the queue.
198 // The "careful" version admits the possibility of pop_local/pop_global
199 // races.
200 uint size() const {
201 return size(_bottom, _age.top());
202 }
203
204 uint dirty_size() const {
205 return dirty_size(_bottom, _age.top());
206 }
207
208 void set_empty() {
209 _bottom = 0;
210 _age.set(0);
211 }
212
213 // Maximum number of elements allowed in the queue. This is two less
214 // than the actual queue size, for somewhat complicated reasons.
215 uint max_elems() const { return N - 2; }
216
217 // Total size of queue.
218 static const uint total_size() { return N; }
219
220 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
221 };
222
223 //
224 // GenericTaskQueue implements an ABP, Aurora-Blumofe-Plaxton, double-
225 // ended-queue (deque), intended for use in work stealing. Queue operations
226 // are non-blocking.
227 //
228 // A queue owner thread performs push() and pop_local() operations on one end
229 // of the queue, while other threads may steal work using the pop_global()
230 // method.
231 //
232 // The main difference to the original algorithm is that this
233 // implementation allows wrap-around at the end of its allocated
234 // storage, which is an array.
235 //
236 // The original paper is:
237 //
238 // Arora, N. S., Blumofe, R. D., and Plaxton, C. G.
239 // Thread scheduling for multiprogrammed multiprocessors.
240 // Theory of Computing Systems 34, 2 (2001), 115-144.
241 //
242 // The following paper provides an correctness proof and an
243 // implementation for weakly ordered memory models including (pseudo-)
244 // code containing memory barriers for a Chase-Lev deque. Chase-Lev is
245 // similar to ABP, with the main difference that it allows resizing of the
246 // underlying storage:
247 //
248 // Le, N. M., Pop, A., Cohen A., and Nardell, F. Z.
249 // Correct and efficient work-stealing for weak memory models
250 // Proceedings of the 18th ACM SIGPLAN symposium on Principles and
251 // practice of parallel programming (PPoPP 2013), 69-80
252 //
253
254 template <class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
255 class GenericTaskQueue: public TaskQueueSuper<N, F> {
256 protected:
257 typedef typename TaskQueueSuper<N, F>::Age Age;
258 typedef typename TaskQueueSuper<N, F>::idx_t idx_t;
259
260 using TaskQueueSuper<N, F>::_bottom;
261 using TaskQueueSuper<N, F>::_age;
262 using TaskQueueSuper<N, F>::increment_index;
263 using TaskQueueSuper<N, F>::decrement_index;
264 using TaskQueueSuper<N, F>::dirty_size;
265
266 public:
267 using TaskQueueSuper<N, F>::max_elems;
268 using TaskQueueSuper<N, F>::size;
269
270 #if TASKQUEUE_STATS
271 using TaskQueueSuper<N, F>::stats;
272 #endif
273
274 private:
275 // Slow paths for push, pop_local. (pop_global has no fast path.)
276 bool push_slow(E t, uint dirty_n_elems);
277 bool pop_local_slow(uint localBot, Age oldAge);
278
279 public:
280 typedef E element_type;
281
282 // Initializes the queue to empty.
283 GenericTaskQueue();
284
285 void initialize();
286
287 // Push the task "t" on the queue. Returns "false" iff the queue is full.
288 inline bool push(E t);
289
290 // Attempts to claim a task from the "local" end of the queue (the most
291 // recently pushed) as long as the number of entries exceeds the threshold.
292 // If successful, returns true and sets t to the task; otherwise, returns false
293 // (the queue is empty or the number of elements below the threshold).
294 inline bool pop_local(volatile E& t, uint threshold = 0);
295
296 // Like pop_local(), but uses the "global" end of the queue (the least
297 // recently pushed).
298 bool pop_global(volatile E& t);
299
300 // Delete any resource associated with the queue.
301 ~GenericTaskQueue();
302
303 // Apply fn to each element in the task queue. The queue must not
304 // be modified while iterating.
305 template<typename Fn> void iterate(Fn fn);
306
307 private:
308 // Element array.
309 volatile E* _elems;
310 };
311
312 template<class E, MEMFLAGS F, unsigned int N>
313 GenericTaskQueue<E, F, N>::GenericTaskQueue() {
314 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
315 }
316
317 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
318 // elements that do not fit in the TaskQueue.
319 //
320 // This class hides two methods from super classes:
321 //
322 // push() - push onto the task queue or, if that fails, onto the overflow stack
323 // is_empty() - return true if both the TaskQueue and overflow stack are empty
324 //
325 // Note that size() is not hidden--it returns the number of elements in the
326 // TaskQueue, and does not include the size of the overflow stack. This
327 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
328 template<class E, MEMFLAGS F, unsigned int N = TASKQUEUE_SIZE>
329 class OverflowTaskQueue: public GenericTaskQueue<E, F, N>
330 {
331 public:
332 typedef Stack<E, F> overflow_t;
333 typedef GenericTaskQueue<E, F, N> taskqueue_t;
334
335 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
336
337 // Push task t onto the queue or onto the overflow stack. Return true.
338 inline bool push(E t);
339 // Try to push task t onto the queue only. Returns true if successful, false otherwise.
340 inline bool try_push_to_taskqueue(E t);
341
342 // Attempt to pop from the overflow stack; return true if anything was popped.
343 inline bool pop_overflow(E& t);
344
345 inline overflow_t* overflow_stack() { return &_overflow_stack; }
346
347 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
348 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
349 inline bool is_empty() const {
350 return taskqueue_empty() && overflow_empty();
351 }
352
353 private:
354 overflow_t _overflow_stack;
355 };
356
357 class TaskQueueSetSuper {
358 protected:
359 static int randomParkAndMiller(int* seed0);
360 public:
361 // Returns "true" if some TaskQueue in the set contains a task.
362 virtual bool peek() = 0;
363 };
364
365 template <MEMFLAGS F> class TaskQueueSetSuperImpl: public CHeapObj<F>, public TaskQueueSetSuper {
366 };
367
368 template<class T, MEMFLAGS F>
369 class GenericTaskQueueSet: public TaskQueueSetSuperImpl<F> {
370 private:
371 uint _n;
372 T** _queues;
373
374 public:
375 typedef typename T::element_type E;
376
377 GenericTaskQueueSet(int n);
378 ~GenericTaskQueueSet();
379
380 bool steal_best_of_2(uint queue_num, int* seed, E& t);
381
382 void register_queue(uint i, T* q);
383
384 T* queue(uint n);
385
386 // The thread with queue number "queue_num" (and whose random number seed is
387 // at "seed") is trying to steal a task from some other queue. (It may try
388 // several queues, according to some configuration parameter.) If some steal
389 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
390 // false.
391 bool steal(uint queue_num, int* seed, E& t);
392
393 bool peek();
394
395 uint size() const { return _n; }
396 };
397
398 template<class T, MEMFLAGS F> void
399 GenericTaskQueueSet<T, F>::register_queue(uint i, T* q) {
400 assert(i < _n, "index out of range.");
401 _queues[i] = q;
402 }
403
404 template<class T, MEMFLAGS F> T*
405 GenericTaskQueueSet<T, F>::queue(uint i) {
406 return _queues[i];
407 }
408
409 template<class T, MEMFLAGS F>
410 bool GenericTaskQueueSet<T, F>::peek() {
411 // Try all the queues.
412 for (uint j = 0; j < _n; j++) {
413 if (_queues[j]->peek())
414 return true;
415 }
416 return false;
417 }
418
419 // When to terminate from the termination protocol.
420 class TerminatorTerminator: public CHeapObj<mtInternal> {
421 public:
422 virtual bool should_exit_termination() = 0;
423 };
424
425 // A class to aid in the termination of a set of parallel tasks using
426 // TaskQueueSet's for work stealing.
427
428 #undef TRACESPINNING
429
430 class ParallelTaskTerminator: public StackObj {
431 private:
432 uint _n_threads;
433 TaskQueueSetSuper* _queue_set;
434
435 DEFINE_PAD_MINUS_SIZE(0, DEFAULT_CACHE_LINE_SIZE, 0);
436 volatile uint _offered_termination;
437 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile uint));
438
439 #ifdef TRACESPINNING
440 static uint _total_yields;
441 static uint _total_spins;
442 static uint _total_peeks;
443 #endif
444
445 bool peek_in_queue_set();
446 protected:
447 virtual void yield();
448 void sleep(uint millis);
449
450 public:
451
452 // "n_threads" is the number of threads to be terminated. "queue_set" is a
453 // queue sets of work queues of other threads.
454 ParallelTaskTerminator(uint n_threads, TaskQueueSetSuper* queue_set);
455
456 // The current thread has no work, and is ready to terminate if everyone
457 // else is. If returns "true", all threads are terminated. If returns
458 // "false", available work has been observed in one of the task queues,
459 // so the global task is not complete.
460 bool offer_termination() {
461 return offer_termination(NULL);
462 }
463
464 // As above, but it also terminates if the should_exit_termination()
465 // method of the terminator parameter returns true. If terminator is
466 // NULL, then it is ignored.
467 bool offer_termination(TerminatorTerminator* terminator);
468
469 // Reset the terminator, so that it may be reused again.
470 // The caller is responsible for ensuring that this is done
471 // in an MT-safe manner, once the previous round of use of
472 // the terminator is finished.
473 void reset_for_reuse();
474 // Same as above but the number of parallel threads is set to the
475 // given number.
476 void reset_for_reuse(uint n_threads);
477
478 #ifdef TRACESPINNING
479 static uint total_yields() { return _total_yields; }
480 static uint total_spins() { return _total_spins; }
481 static uint total_peeks() { return _total_peeks; }
482 static void print_termination_counts();
483 #endif
484 };
485
486 typedef GenericTaskQueue<oop, mtGC> OopTaskQueue;
487 typedef GenericTaskQueueSet<OopTaskQueue, mtGC> OopTaskQueueSet;
488
489 #ifdef _MSC_VER
490 #pragma warning(push)
491 // warning C4522: multiple assignment operators specified
492 #pragma warning(disable:4522)
493 #endif
494
495 // This is a container class for either an oop* or a narrowOop*.
496 // Both are pushed onto a task queue and the consumer will test is_narrow()
497 // to determine which should be processed.
498 class StarTask {
499 void* _holder; // either union oop* or narrowOop*
500
501 enum { COMPRESSED_OOP_MASK = 1 };
502
503 public:
504 StarTask(narrowOop* p) {
505 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
506 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
507 }
508 StarTask(oop* p) {
509 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
510 _holder = (void*)p;
511 }
512 StarTask() { _holder = NULL; }
513 operator oop*() { return (oop*)_holder; }
514 operator narrowOop*() {
515 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
516 }
517
518 StarTask& operator=(const StarTask& t) {
519 _holder = t._holder;
520 return *this;
521 }
522 volatile StarTask& operator=(const volatile StarTask& t) volatile {
523 _holder = t._holder;
524 return *this;
525 }
526
527 bool is_narrow() const {
528 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
529 }
530 };
531
532 class ObjArrayTask
533 {
534 public:
535 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
536 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
537 assert(idx <= size_t(max_jint), "too big");
538 }
539 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
540
541 ObjArrayTask& operator =(const ObjArrayTask& t) {
542 _obj = t._obj;
543 _index = t._index;
544 return *this;
545 }
546 volatile ObjArrayTask&
547 operator =(const volatile ObjArrayTask& t) volatile {
548 (void)const_cast<oop&>(_obj = t._obj);
549 _index = t._index;
550 return *this;
551 }
552
553 inline oop obj() const { return _obj; }
554 inline int index() const { return _index; }
555
556 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
557
558 private:
559 oop _obj;
560 int _index;
561 };
562
563 #ifdef _MSC_VER
564 #pragma warning(pop)
565 #endif
566
567 typedef OverflowTaskQueue<StarTask, mtGC> OopStarTaskQueue;
568 typedef GenericTaskQueueSet<OopStarTaskQueue, mtGC> OopStarTaskQueueSet;
569
570 typedef OverflowTaskQueue<size_t, mtGC> RegionTaskQueue;
571 typedef GenericTaskQueueSet<RegionTaskQueue, mtGC> RegionTaskQueueSet;
572
573 #endif // SHARE_VM_GC_SHARED_TASKQUEUE_HPP