src/share/vm/utilities/taskqueue.hpp

Tue, 05 Apr 2011 14:12:31 -0700

author
trims
date
Tue, 05 Apr 2011 14:12:31 -0700
changeset 2708
1d1603768966
parent 2508
b92c45f2bc75
child 3156
f08d439fab8c
permissions
-rw-r--r--

7010070: Update all 2010 Oracle-changed OpenJDK files to have the proper copyright dates - second pass
Summary: Update the copyright to be 2010 on all changed files in OpenJDK
Reviewed-by: ohair

duke@435 1 /*
trims@2708 2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
duke@435 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@435 4 *
duke@435 5 * This code is free software; you can redistribute it and/or modify it
duke@435 6 * under the terms of the GNU General Public License version 2 only, as
duke@435 7 * published by the Free Software Foundation.
duke@435 8 *
duke@435 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@435 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@435 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@435 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@435 13 * accompanied this code).
duke@435 14 *
duke@435 15 * You should have received a copy of the GNU General Public License version
duke@435 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@435 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@435 18 *
trims@1907 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1907 20 * or visit www.oracle.com if you need additional information or have any
trims@1907 21 * questions.
duke@435 22 *
duke@435 23 */
duke@435 24
stefank@2314 25 #ifndef SHARE_VM_UTILITIES_TASKQUEUE_HPP
stefank@2314 26 #define SHARE_VM_UTILITIES_TASKQUEUE_HPP
stefank@2314 27
stefank@2314 28 #include "memory/allocation.hpp"
stefank@2314 29 #include "memory/allocation.inline.hpp"
stefank@2314 30 #include "runtime/mutex.hpp"
stefank@2314 31 #include "utilities/stack.hpp"
stefank@2314 32 #ifdef TARGET_OS_ARCH_linux_x86
stefank@2314 33 # include "orderAccess_linux_x86.inline.hpp"
stefank@2314 34 #endif
stefank@2314 35 #ifdef TARGET_OS_ARCH_linux_sparc
stefank@2314 36 # include "orderAccess_linux_sparc.inline.hpp"
stefank@2314 37 #endif
stefank@2314 38 #ifdef TARGET_OS_ARCH_linux_zero
stefank@2314 39 # include "orderAccess_linux_zero.inline.hpp"
stefank@2314 40 #endif
stefank@2314 41 #ifdef TARGET_OS_ARCH_solaris_x86
stefank@2314 42 # include "orderAccess_solaris_x86.inline.hpp"
stefank@2314 43 #endif
stefank@2314 44 #ifdef TARGET_OS_ARCH_solaris_sparc
stefank@2314 45 # include "orderAccess_solaris_sparc.inline.hpp"
stefank@2314 46 #endif
stefank@2314 47 #ifdef TARGET_OS_ARCH_windows_x86
stefank@2314 48 # include "orderAccess_windows_x86.inline.hpp"
stefank@2314 49 #endif
bobv@2508 50 #ifdef TARGET_OS_ARCH_linux_arm
bobv@2508 51 # include "orderAccess_linux_arm.inline.hpp"
bobv@2508 52 #endif
bobv@2508 53 #ifdef TARGET_OS_ARCH_linux_ppc
bobv@2508 54 # include "orderAccess_linux_ppc.inline.hpp"
bobv@2508 55 #endif
stefank@2314 56
jcoomes@2020 57 // Simple TaskQueue stats that are collected by default in debug builds.
jcoomes@2020 58
jcoomes@2020 59 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
jcoomes@2020 60 #define TASKQUEUE_STATS 1
jcoomes@2020 61 #elif !defined(TASKQUEUE_STATS)
jcoomes@2020 62 #define TASKQUEUE_STATS 0
jcoomes@2020 63 #endif
jcoomes@2020 64
jcoomes@2020 65 #if TASKQUEUE_STATS
jcoomes@2020 66 #define TASKQUEUE_STATS_ONLY(code) code
jcoomes@2020 67 #else
jcoomes@2020 68 #define TASKQUEUE_STATS_ONLY(code)
jcoomes@2020 69 #endif // TASKQUEUE_STATS
jcoomes@2020 70
jcoomes@2020 71 #if TASKQUEUE_STATS
jcoomes@2020 72 class TaskQueueStats {
jcoomes@2020 73 public:
jcoomes@2020 74 enum StatId {
jcoomes@2020 75 push, // number of taskqueue pushes
jcoomes@2020 76 pop, // number of taskqueue pops
jcoomes@2020 77 pop_slow, // subset of taskqueue pops that were done slow-path
jcoomes@2020 78 steal_attempt, // number of taskqueue steal attempts
jcoomes@2020 79 steal, // number of taskqueue steals
jcoomes@2020 80 overflow, // number of overflow pushes
jcoomes@2020 81 overflow_max_len, // max length of overflow stack
jcoomes@2020 82 last_stat_id
jcoomes@2020 83 };
jcoomes@2020 84
jcoomes@2020 85 public:
jcoomes@2020 86 inline TaskQueueStats() { reset(); }
jcoomes@2020 87
jcoomes@2020 88 inline void record_push() { ++_stats[push]; }
jcoomes@2020 89 inline void record_pop() { ++_stats[pop]; }
jcoomes@2020 90 inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
jcoomes@2020 91 inline void record_steal(bool success);
jcoomes@2020 92 inline void record_overflow(size_t new_length);
jcoomes@2020 93
jcoomes@2064 94 TaskQueueStats & operator +=(const TaskQueueStats & addend);
jcoomes@2064 95
jcoomes@2020 96 inline size_t get(StatId id) const { return _stats[id]; }
jcoomes@2020 97 inline const size_t* get() const { return _stats; }
jcoomes@2020 98
jcoomes@2020 99 inline void reset();
jcoomes@2020 100
jcoomes@2064 101 // Print the specified line of the header (does not include a line separator).
jcoomes@2020 102 static void print_header(unsigned int line, outputStream* const stream = tty,
jcoomes@2020 103 unsigned int width = 10);
jcoomes@2064 104 // Print the statistics (does not include a line separator).
jcoomes@2020 105 void print(outputStream* const stream = tty, unsigned int width = 10) const;
jcoomes@2020 106
jcoomes@2064 107 DEBUG_ONLY(void verify() const;)
jcoomes@2064 108
jcoomes@2020 109 private:
jcoomes@2020 110 size_t _stats[last_stat_id];
jcoomes@2020 111 static const char * const _names[last_stat_id];
jcoomes@2020 112 };
jcoomes@2020 113
jcoomes@2020 114 void TaskQueueStats::record_steal(bool success) {
jcoomes@2020 115 ++_stats[steal_attempt];
jcoomes@2020 116 if (success) ++_stats[steal];
jcoomes@2020 117 }
jcoomes@2020 118
jcoomes@2020 119 void TaskQueueStats::record_overflow(size_t new_len) {
jcoomes@2020 120 ++_stats[overflow];
jcoomes@2020 121 if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
jcoomes@2020 122 }
jcoomes@2020 123
jcoomes@2020 124 void TaskQueueStats::reset() {
jcoomes@2020 125 memset(_stats, 0, sizeof(_stats));
jcoomes@2020 126 }
jcoomes@2020 127 #endif // TASKQUEUE_STATS
jcoomes@2020 128
jcoomes@1746 129 template <unsigned int N>
duke@435 130 class TaskQueueSuper: public CHeapObj {
duke@435 131 protected:
jcoomes@1342 132 // Internal type for indexing the queue; also used for the tag.
jcoomes@1342 133 typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
jcoomes@1342 134
jcoomes@1342 135 // The first free element after the last one pushed (mod N).
ysr@976 136 volatile uint _bottom;
duke@435 137
jcoomes@1746 138 enum { MOD_N_MASK = N - 1 };
duke@435 139
jcoomes@1342 140 class Age {
jcoomes@1342 141 public:
jcoomes@1342 142 Age(size_t data = 0) { _data = data; }
jcoomes@1342 143 Age(const Age& age) { _data = age._data; }
jcoomes@1342 144 Age(idx_t top, idx_t tag) { _fields._top = top; _fields._tag = tag; }
duke@435 145
jcoomes@1342 146 Age get() const volatile { return _data; }
jcoomes@1342 147 void set(Age age) volatile { _data = age._data; }
duke@435 148
jcoomes@1342 149 idx_t top() const volatile { return _fields._top; }
jcoomes@1342 150 idx_t tag() const volatile { return _fields._tag; }
duke@435 151
jcoomes@1342 152 // Increment top; if it wraps, increment tag also.
jcoomes@1342 153 void increment() {
jcoomes@1342 154 _fields._top = increment_index(_fields._top);
jcoomes@1342 155 if (_fields._top == 0) ++_fields._tag;
jcoomes@1342 156 }
duke@435 157
jcoomes@1342 158 Age cmpxchg(const Age new_age, const Age old_age) volatile {
jcoomes@1342 159 return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
jcoomes@1342 160 (volatile intptr_t *)&_data,
jcoomes@1342 161 (intptr_t)old_age._data);
duke@435 162 }
jcoomes@1342 163
jcoomes@1342 164 bool operator ==(const Age& other) const { return _data == other._data; }
jcoomes@1342 165
jcoomes@1342 166 private:
jcoomes@1342 167 struct fields {
jcoomes@1342 168 idx_t _top;
jcoomes@1342 169 idx_t _tag;
jcoomes@1342 170 };
jcoomes@1342 171 union {
jcoomes@1342 172 size_t _data;
jcoomes@1342 173 fields _fields;
jcoomes@1342 174 };
duke@435 175 };
jcoomes@1342 176
jcoomes@1342 177 volatile Age _age;
jcoomes@1342 178
jcoomes@1342 179 // These both operate mod N.
jcoomes@1342 180 static uint increment_index(uint ind) {
jcoomes@1342 181 return (ind + 1) & MOD_N_MASK;
duke@435 182 }
jcoomes@1342 183 static uint decrement_index(uint ind) {
jcoomes@1342 184 return (ind - 1) & MOD_N_MASK;
duke@435 185 }
duke@435 186
jcoomes@1342 187 // Returns a number in the range [0..N). If the result is "N-1", it should be
jcoomes@1342 188 // interpreted as 0.
jcoomes@1746 189 uint dirty_size(uint bot, uint top) const {
jcoomes@1342 190 return (bot - top) & MOD_N_MASK;
duke@435 191 }
duke@435 192
duke@435 193 // Returns the size corresponding to the given "bot" and "top".
jcoomes@1746 194 uint size(uint bot, uint top) const {
ysr@976 195 uint sz = dirty_size(bot, top);
jcoomes@1342 196 // Has the queue "wrapped", so that bottom is less than top? There's a
jcoomes@1342 197 // complicated special case here. A pair of threads could perform pop_local
jcoomes@1342 198 // and pop_global operations concurrently, starting from a state in which
jcoomes@1342 199 // _bottom == _top+1. The pop_local could succeed in decrementing _bottom,
jcoomes@1342 200 // and the pop_global in incrementing _top (in which case the pop_global
jcoomes@1342 201 // will be awarded the contested queue element.) The resulting state must
jcoomes@1342 202 // be interpreted as an empty queue. (We only need to worry about one such
jcoomes@1342 203 // event: only the queue owner performs pop_local's, and several concurrent
jcoomes@1342 204 // threads attempting to perform the pop_global will all perform the same
jcoomes@1342 205 // CAS, and only one can succeed.) Any stealing thread that reads after
jcoomes@1342 206 // either the increment or decrement will see an empty queue, and will not
jcoomes@1342 207 // join the competitors. The "sz == -1 || sz == N-1" state will not be
jcoomes@1342 208 // modified by concurrent queues, so the owner thread can reset the state to
jcoomes@1342 209 // _bottom == top so subsequent pushes will be performed normally.
jcoomes@1342 210 return (sz == N - 1) ? 0 : sz;
duke@435 211 }
duke@435 212
duke@435 213 public:
duke@435 214 TaskQueueSuper() : _bottom(0), _age() {}
duke@435 215
jcoomes@1993 216 // Return true if the TaskQueue contains/does not contain any tasks.
jcoomes@1993 217 bool peek() const { return _bottom != _age.top(); }
jcoomes@1993 218 bool is_empty() const { return size() == 0; }
duke@435 219
duke@435 220 // Return an estimate of the number of elements in the queue.
duke@435 221 // The "careful" version admits the possibility of pop_local/pop_global
duke@435 222 // races.
jcoomes@1746 223 uint size() const {
jcoomes@1342 224 return size(_bottom, _age.top());
duke@435 225 }
duke@435 226
jcoomes@1746 227 uint dirty_size() const {
jcoomes@1342 228 return dirty_size(_bottom, _age.top());
duke@435 229 }
duke@435 230
ysr@777 231 void set_empty() {
ysr@777 232 _bottom = 0;
jcoomes@1342 233 _age.set(0);
ysr@777 234 }
ysr@777 235
duke@435 236 // Maximum number of elements allowed in the queue. This is two less
duke@435 237 // than the actual queue size, for somewhat complicated reasons.
jcoomes@1746 238 uint max_elems() const { return N - 2; }
jmasa@1719 239
jmasa@1719 240 // Total size of queue.
jmasa@1719 241 static const uint total_size() { return N; }
jcoomes@2020 242
jcoomes@2020 243 TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
duke@435 244 };
duke@435 245
jcoomes@1746 246 template<class E, unsigned int N = TASKQUEUE_SIZE>
jcoomes@1746 247 class GenericTaskQueue: public TaskQueueSuper<N> {
jcoomes@1746 248 protected:
jcoomes@1746 249 typedef typename TaskQueueSuper<N>::Age Age;
jcoomes@1746 250 typedef typename TaskQueueSuper<N>::idx_t idx_t;
jcoomes@1746 251
jcoomes@1746 252 using TaskQueueSuper<N>::_bottom;
jcoomes@1746 253 using TaskQueueSuper<N>::_age;
jcoomes@1746 254 using TaskQueueSuper<N>::increment_index;
jcoomes@1746 255 using TaskQueueSuper<N>::decrement_index;
jcoomes@1746 256 using TaskQueueSuper<N>::dirty_size;
jcoomes@1746 257
jcoomes@1746 258 public:
jcoomes@1746 259 using TaskQueueSuper<N>::max_elems;
jcoomes@1746 260 using TaskQueueSuper<N>::size;
jcoomes@2020 261 TASKQUEUE_STATS_ONLY(using TaskQueueSuper<N>::stats;)
jcoomes@1746 262
duke@435 263 private:
duke@435 264 // Slow paths for push, pop_local. (pop_global has no fast path.)
ysr@976 265 bool push_slow(E t, uint dirty_n_elems);
ysr@976 266 bool pop_local_slow(uint localBot, Age oldAge);
duke@435 267
duke@435 268 public:
jcoomes@1746 269 typedef E element_type;
jcoomes@1746 270
duke@435 271 // Initializes the queue to empty.
duke@435 272 GenericTaskQueue();
duke@435 273
duke@435 274 void initialize();
duke@435 275
jcoomes@1993 276 // Push the task "t" on the queue. Returns "false" iff the queue is full.
duke@435 277 inline bool push(E t);
duke@435 278
jcoomes@1993 279 // Attempts to claim a task from the "local" end of the queue (the most
jcoomes@1993 280 // recently pushed). If successful, returns true and sets t to the task;
jcoomes@1993 281 // otherwise, returns false (the queue is empty).
duke@435 282 inline bool pop_local(E& t);
duke@435 283
jcoomes@1993 284 // Like pop_local(), but uses the "global" end of the queue (the least
jcoomes@1993 285 // recently pushed).
duke@435 286 bool pop_global(E& t);
duke@435 287
duke@435 288 // Delete any resource associated with the queue.
duke@435 289 ~GenericTaskQueue();
duke@435 290
ysr@777 291 // apply the closure to all elements in the task queue
ysr@777 292 void oops_do(OopClosure* f);
ysr@777 293
duke@435 294 private:
duke@435 295 // Element array.
duke@435 296 volatile E* _elems;
duke@435 297 };
duke@435 298
jcoomes@1746 299 template<class E, unsigned int N>
jcoomes@1746 300 GenericTaskQueue<E, N>::GenericTaskQueue() {
jcoomes@1342 301 assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
duke@435 302 }
duke@435 303
jcoomes@1746 304 template<class E, unsigned int N>
jcoomes@1746 305 void GenericTaskQueue<E, N>::initialize() {
jcoomes@1342 306 _elems = NEW_C_HEAP_ARRAY(E, N);
duke@435 307 }
duke@435 308
jcoomes@1746 309 template<class E, unsigned int N>
jcoomes@1746 310 void GenericTaskQueue<E, N>::oops_do(OopClosure* f) {
ysr@777 311 // tty->print_cr("START OopTaskQueue::oops_do");
ysr@976 312 uint iters = size();
ysr@976 313 uint index = _bottom;
ysr@976 314 for (uint i = 0; i < iters; ++i) {
ysr@777 315 index = decrement_index(index);
ysr@777 316 // tty->print_cr(" doing entry %d," INTPTR_T " -> " INTPTR_T,
ysr@777 317 // index, &_elems[index], _elems[index]);
ysr@777 318 E* t = (E*)&_elems[index]; // cast away volatility
ysr@777 319 oop* p = (oop*)t;
ysr@777 320 assert((*t)->is_oop_or_null(), "Not an oop or null");
ysr@777 321 f->do_oop(p);
ysr@777 322 }
ysr@777 323 // tty->print_cr("END OopTaskQueue::oops_do");
ysr@777 324 }
ysr@777 325
jcoomes@1746 326 template<class E, unsigned int N>
jcoomes@1746 327 bool GenericTaskQueue<E, N>::push_slow(E t, uint dirty_n_elems) {
jcoomes@1342 328 if (dirty_n_elems == N - 1) {
duke@435 329 // Actually means 0, so do the push.
ysr@976 330 uint localBot = _bottom;
jcoomes@1746 331 // g++ complains if the volatile result of the assignment is unused.
jcoomes@1746 332 const_cast<E&>(_elems[localBot] = t);
bobv@1459 333 OrderAccess::release_store(&_bottom, increment_index(localBot));
jcoomes@2020 334 TASKQUEUE_STATS_ONLY(stats.record_push());
duke@435 335 return true;
jcoomes@1342 336 }
jcoomes@1342 337 return false;
duke@435 338 }
duke@435 339
jmasa@2188 340 // pop_local_slow() is done by the owning thread and is trying to
jmasa@2188 341 // get the last task in the queue. It will compete with pop_global()
jmasa@2188 342 // that will be used by other threads. The tag age is incremented
jmasa@2188 343 // whenever the queue goes empty which it will do here if this thread
jmasa@2188 344 // gets the last task or in pop_global() if the queue wraps (top == 0
jmasa@2188 345 // and pop_global() succeeds, see pop_global()).
jcoomes@1746 346 template<class E, unsigned int N>
jcoomes@2020 347 bool GenericTaskQueue<E, N>::pop_local_slow(uint localBot, Age oldAge) {
duke@435 348 // This queue was observed to contain exactly one element; either this
duke@435 349 // thread will claim it, or a competing "pop_global". In either case,
duke@435 350 // the queue will be logically empty afterwards. Create a new Age value
duke@435 351 // that represents the empty queue for the given value of "_bottom". (We
duke@435 352 // must also increment "tag" because of the case where "bottom == 1",
duke@435 353 // "top == 0". A pop_global could read the queue element in that case,
duke@435 354 // then have the owner thread do a pop followed by another push. Without
duke@435 355 // the incrementing of "tag", the pop_global's CAS could succeed,
duke@435 356 // allowing it to believe it has claimed the stale element.)
jcoomes@1342 357 Age newAge((idx_t)localBot, oldAge.tag() + 1);
duke@435 358 // Perhaps a competing pop_global has already incremented "top", in which
duke@435 359 // case it wins the element.
duke@435 360 if (localBot == oldAge.top()) {
duke@435 361 // No competing pop_global has yet incremented "top"; we'll try to
duke@435 362 // install new_age, thus claiming the element.
jcoomes@1342 363 Age tempAge = _age.cmpxchg(newAge, oldAge);
duke@435 364 if (tempAge == oldAge) {
duke@435 365 // We win.
jcoomes@1342 366 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
jcoomes@2020 367 TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
duke@435 368 return true;
duke@435 369 }
duke@435 370 }
jcoomes@1342 371 // We lose; a completing pop_global gets the element. But the queue is empty
jcoomes@1342 372 // and top is greater than bottom. Fix this representation of the empty queue
jcoomes@1342 373 // to become the canonical one.
jcoomes@1342 374 _age.set(newAge);
jcoomes@1342 375 assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
duke@435 376 return false;
duke@435 377 }
duke@435 378
jcoomes@1746 379 template<class E, unsigned int N>
jcoomes@1746 380 bool GenericTaskQueue<E, N>::pop_global(E& t) {
jcoomes@1342 381 Age oldAge = _age.get();
ysr@976 382 uint localBot = _bottom;
ysr@976 383 uint n_elems = size(localBot, oldAge.top());
duke@435 384 if (n_elems == 0) {
duke@435 385 return false;
duke@435 386 }
jcoomes@1342 387
jcoomes@1746 388 const_cast<E&>(t = _elems[oldAge.top()]);
jcoomes@1342 389 Age newAge(oldAge);
jcoomes@1342 390 newAge.increment();
jcoomes@1342 391 Age resAge = _age.cmpxchg(newAge, oldAge);
jcoomes@1342 392
duke@435 393 // Note that using "_bottom" here might fail, since a pop_local might
duke@435 394 // have decremented it.
jcoomes@1342 395 assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
jcoomes@1342 396 return resAge == oldAge;
duke@435 397 }
duke@435 398
jcoomes@1746 399 template<class E, unsigned int N>
jcoomes@1746 400 GenericTaskQueue<E, N>::~GenericTaskQueue() {
duke@435 401 FREE_C_HEAP_ARRAY(E, _elems);
duke@435 402 }
duke@435 403
jcoomes@1993 404 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
jcoomes@1993 405 // elements that do not fit in the TaskQueue.
jcoomes@1993 406 //
jcoomes@2191 407 // This class hides two methods from super classes:
jcoomes@1993 408 //
jcoomes@1993 409 // push() - push onto the task queue or, if that fails, onto the overflow stack
jcoomes@1993 410 // is_empty() - return true if both the TaskQueue and overflow stack are empty
jcoomes@1993 411 //
jcoomes@2191 412 // Note that size() is not hidden--it returns the number of elements in the
jcoomes@1993 413 // TaskQueue, and does not include the size of the overflow stack. This
jcoomes@1993 414 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
jcoomes@1993 415 template<class E, unsigned int N = TASKQUEUE_SIZE>
jcoomes@1993 416 class OverflowTaskQueue: public GenericTaskQueue<E, N>
jcoomes@1993 417 {
jcoomes@1993 418 public:
jcoomes@2191 419 typedef Stack<E> overflow_t;
jcoomes@1993 420 typedef GenericTaskQueue<E, N> taskqueue_t;
jcoomes@1993 421
jcoomes@2020 422 TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
jcoomes@2020 423
jcoomes@1993 424 // Push task t onto the queue or onto the overflow stack. Return true.
jcoomes@1993 425 inline bool push(E t);
jcoomes@1993 426
jcoomes@1993 427 // Attempt to pop from the overflow stack; return true if anything was popped.
jcoomes@1993 428 inline bool pop_overflow(E& t);
jcoomes@1993 429
jcoomes@2191 430 inline overflow_t* overflow_stack() { return &_overflow_stack; }
jcoomes@2191 431
jcoomes@1993 432 inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
jcoomes@2191 433 inline bool overflow_empty() const { return _overflow_stack.is_empty(); }
jcoomes@1993 434 inline bool is_empty() const {
jcoomes@1993 435 return taskqueue_empty() && overflow_empty();
jcoomes@1993 436 }
jcoomes@1993 437
jcoomes@1993 438 private:
jcoomes@2191 439 overflow_t _overflow_stack;
jcoomes@1993 440 };
jcoomes@1993 441
jcoomes@1993 442 template <class E, unsigned int N>
jcoomes@1993 443 bool OverflowTaskQueue<E, N>::push(E t)
jcoomes@1993 444 {
jcoomes@1993 445 if (!taskqueue_t::push(t)) {
jcoomes@1993 446 overflow_stack()->push(t);
jcoomes@2191 447 TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
jcoomes@1993 448 }
jcoomes@1993 449 return true;
jcoomes@1993 450 }
jcoomes@1993 451
jcoomes@1993 452 template <class E, unsigned int N>
jcoomes@1993 453 bool OverflowTaskQueue<E, N>::pop_overflow(E& t)
jcoomes@1993 454 {
jcoomes@1993 455 if (overflow_empty()) return false;
jcoomes@1993 456 t = overflow_stack()->pop();
jcoomes@1993 457 return true;
jcoomes@1993 458 }
jcoomes@1993 459
duke@435 460 class TaskQueueSetSuper: public CHeapObj {
duke@435 461 protected:
duke@435 462 static int randomParkAndMiller(int* seed0);
duke@435 463 public:
duke@435 464 // Returns "true" if some TaskQueue in the set contains a task.
duke@435 465 virtual bool peek() = 0;
duke@435 466 };
duke@435 467
jcoomes@1746 468 template<class T>
jcoomes@1746 469 class GenericTaskQueueSet: public TaskQueueSetSuper {
duke@435 470 private:
ysr@976 471 uint _n;
jcoomes@1746 472 T** _queues;
duke@435 473
duke@435 474 public:
jcoomes@1746 475 typedef typename T::element_type E;
jcoomes@1746 476
duke@435 477 GenericTaskQueueSet(int n) : _n(n) {
jcoomes@1746 478 typedef T* GenericTaskQueuePtr;
duke@435 479 _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n);
duke@435 480 for (int i = 0; i < n; i++) {
duke@435 481 _queues[i] = NULL;
duke@435 482 }
duke@435 483 }
duke@435 484
ysr@976 485 bool steal_1_random(uint queue_num, int* seed, E& t);
ysr@976 486 bool steal_best_of_2(uint queue_num, int* seed, E& t);
ysr@976 487 bool steal_best_of_all(uint queue_num, int* seed, E& t);
duke@435 488
jcoomes@1746 489 void register_queue(uint i, T* q);
duke@435 490
jcoomes@1746 491 T* queue(uint n);
duke@435 492
jcoomes@1993 493 // The thread with queue number "queue_num" (and whose random number seed is
jcoomes@1993 494 // at "seed") is trying to steal a task from some other queue. (It may try
jcoomes@1993 495 // several queues, according to some configuration parameter.) If some steal
jcoomes@1993 496 // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
jcoomes@1993 497 // false.
ysr@976 498 bool steal(uint queue_num, int* seed, E& t);
duke@435 499
duke@435 500 bool peek();
duke@435 501 };
duke@435 502
jcoomes@1746 503 template<class T> void
jcoomes@1746 504 GenericTaskQueueSet<T>::register_queue(uint i, T* q) {
ysr@976 505 assert(i < _n, "index out of range.");
duke@435 506 _queues[i] = q;
duke@435 507 }
duke@435 508
jcoomes@1746 509 template<class T> T*
jcoomes@1746 510 GenericTaskQueueSet<T>::queue(uint i) {
duke@435 511 return _queues[i];
duke@435 512 }
duke@435 513
jcoomes@1746 514 template<class T> bool
jcoomes@1746 515 GenericTaskQueueSet<T>::steal(uint queue_num, int* seed, E& t) {
jcoomes@2020 516 for (uint i = 0; i < 2 * _n; i++) {
jcoomes@2020 517 if (steal_best_of_2(queue_num, seed, t)) {
jcoomes@2020 518 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
duke@435 519 return true;
jcoomes@2020 520 }
jcoomes@2020 521 }
jcoomes@2020 522 TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
duke@435 523 return false;
duke@435 524 }
duke@435 525
jcoomes@1746 526 template<class T> bool
jcoomes@1746 527 GenericTaskQueueSet<T>::steal_best_of_all(uint queue_num, int* seed, E& t) {
duke@435 528 if (_n > 2) {
duke@435 529 int best_k;
ysr@976 530 uint best_sz = 0;
ysr@976 531 for (uint k = 0; k < _n; k++) {
duke@435 532 if (k == queue_num) continue;
ysr@976 533 uint sz = _queues[k]->size();
duke@435 534 if (sz > best_sz) {
duke@435 535 best_sz = sz;
duke@435 536 best_k = k;
duke@435 537 }
duke@435 538 }
duke@435 539 return best_sz > 0 && _queues[best_k]->pop_global(t);
duke@435 540 } else if (_n == 2) {
duke@435 541 // Just try the other one.
duke@435 542 int k = (queue_num + 1) % 2;
duke@435 543 return _queues[k]->pop_global(t);
duke@435 544 } else {
duke@435 545 assert(_n == 1, "can't be zero.");
duke@435 546 return false;
duke@435 547 }
duke@435 548 }
duke@435 549
jcoomes@1746 550 template<class T> bool
jcoomes@1746 551 GenericTaskQueueSet<T>::steal_1_random(uint queue_num, int* seed, E& t) {
duke@435 552 if (_n > 2) {
ysr@976 553 uint k = queue_num;
duke@435 554 while (k == queue_num) k = randomParkAndMiller(seed) % _n;
duke@435 555 return _queues[2]->pop_global(t);
duke@435 556 } else if (_n == 2) {
duke@435 557 // Just try the other one.
duke@435 558 int k = (queue_num + 1) % 2;
duke@435 559 return _queues[k]->pop_global(t);
duke@435 560 } else {
duke@435 561 assert(_n == 1, "can't be zero.");
duke@435 562 return false;
duke@435 563 }
duke@435 564 }
duke@435 565
jcoomes@1746 566 template<class T> bool
jcoomes@1746 567 GenericTaskQueueSet<T>::steal_best_of_2(uint queue_num, int* seed, E& t) {
duke@435 568 if (_n > 2) {
ysr@976 569 uint k1 = queue_num;
duke@435 570 while (k1 == queue_num) k1 = randomParkAndMiller(seed) % _n;
ysr@976 571 uint k2 = queue_num;
duke@435 572 while (k2 == queue_num || k2 == k1) k2 = randomParkAndMiller(seed) % _n;
duke@435 573 // Sample both and try the larger.
ysr@976 574 uint sz1 = _queues[k1]->size();
ysr@976 575 uint sz2 = _queues[k2]->size();
duke@435 576 if (sz2 > sz1) return _queues[k2]->pop_global(t);
duke@435 577 else return _queues[k1]->pop_global(t);
duke@435 578 } else if (_n == 2) {
duke@435 579 // Just try the other one.
ysr@976 580 uint k = (queue_num + 1) % 2;
duke@435 581 return _queues[k]->pop_global(t);
duke@435 582 } else {
duke@435 583 assert(_n == 1, "can't be zero.");
duke@435 584 return false;
duke@435 585 }
duke@435 586 }
duke@435 587
jcoomes@1746 588 template<class T>
jcoomes@1746 589 bool GenericTaskQueueSet<T>::peek() {
duke@435 590 // Try all the queues.
ysr@976 591 for (uint j = 0; j < _n; j++) {
duke@435 592 if (_queues[j]->peek())
duke@435 593 return true;
duke@435 594 }
duke@435 595 return false;
duke@435 596 }
duke@435 597
ysr@777 598 // When to terminate from the termination protocol.
ysr@777 599 class TerminatorTerminator: public CHeapObj {
ysr@777 600 public:
ysr@777 601 virtual bool should_exit_termination() = 0;
ysr@777 602 };
ysr@777 603
duke@435 604 // A class to aid in the termination of a set of parallel tasks using
duke@435 605 // TaskQueueSet's for work stealing.
duke@435 606
jmasa@981 607 #undef TRACESPINNING
jmasa@981 608
duke@435 609 class ParallelTaskTerminator: public StackObj {
duke@435 610 private:
duke@435 611 int _n_threads;
duke@435 612 TaskQueueSetSuper* _queue_set;
ysr@976 613 int _offered_termination;
duke@435 614
jmasa@981 615 #ifdef TRACESPINNING
jmasa@981 616 static uint _total_yields;
jmasa@981 617 static uint _total_spins;
jmasa@981 618 static uint _total_peeks;
jmasa@981 619 #endif
jmasa@981 620
duke@435 621 bool peek_in_queue_set();
duke@435 622 protected:
duke@435 623 virtual void yield();
duke@435 624 void sleep(uint millis);
duke@435 625
duke@435 626 public:
duke@435 627
duke@435 628 // "n_threads" is the number of threads to be terminated. "queue_set" is a
duke@435 629 // queue sets of work queues of other threads.
duke@435 630 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
duke@435 631
duke@435 632 // The current thread has no work, and is ready to terminate if everyone
duke@435 633 // else is. If returns "true", all threads are terminated. If returns
duke@435 634 // "false", available work has been observed in one of the task queues,
duke@435 635 // so the global task is not complete.
ysr@777 636 bool offer_termination() {
ysr@777 637 return offer_termination(NULL);
ysr@777 638 }
ysr@777 639
jcoomes@1342 640 // As above, but it also terminates if the should_exit_termination()
ysr@777 641 // method of the terminator parameter returns true. If terminator is
ysr@777 642 // NULL, then it is ignored.
ysr@777 643 bool offer_termination(TerminatorTerminator* terminator);
duke@435 644
duke@435 645 // Reset the terminator, so that it may be reused again.
duke@435 646 // The caller is responsible for ensuring that this is done
duke@435 647 // in an MT-safe manner, once the previous round of use of
duke@435 648 // the terminator is finished.
duke@435 649 void reset_for_reuse();
jmasa@2188 650 // Same as above but the number of parallel threads is set to the
jmasa@2188 651 // given number.
jmasa@2188 652 void reset_for_reuse(int n_threads);
duke@435 653
jmasa@981 654 #ifdef TRACESPINNING
jmasa@981 655 static uint total_yields() { return _total_yields; }
jmasa@981 656 static uint total_spins() { return _total_spins; }
jmasa@981 657 static uint total_peeks() { return _total_peeks; }
jmasa@981 658 static void print_termination_counts();
jmasa@981 659 #endif
duke@435 660 };
duke@435 661
jcoomes@1746 662 template<class E, unsigned int N> inline bool
jcoomes@1746 663 GenericTaskQueue<E, N>::push(E t) {
ysr@976 664 uint localBot = _bottom;
jcoomes@1342 665 assert((localBot >= 0) && (localBot < N), "_bottom out of range.");
jcoomes@1342 666 idx_t top = _age.top();
ysr@976 667 uint dirty_n_elems = dirty_size(localBot, top);
jcoomes@1746 668 assert(dirty_n_elems < N, "n_elems out of range.");
duke@435 669 if (dirty_n_elems < max_elems()) {
jcoomes@1746 670 // g++ complains if the volatile result of the assignment is unused.
jcoomes@1746 671 const_cast<E&>(_elems[localBot] = t);
bobv@1459 672 OrderAccess::release_store(&_bottom, increment_index(localBot));
jcoomes@2020 673 TASKQUEUE_STATS_ONLY(stats.record_push());
duke@435 674 return true;
duke@435 675 } else {
duke@435 676 return push_slow(t, dirty_n_elems);
duke@435 677 }
duke@435 678 }
duke@435 679
jcoomes@1746 680 template<class E, unsigned int N> inline bool
jcoomes@1746 681 GenericTaskQueue<E, N>::pop_local(E& t) {
ysr@976 682 uint localBot = _bottom;
jcoomes@1342 683 // This value cannot be N-1. That can only occur as a result of
duke@435 684 // the assignment to bottom in this method. If it does, this method
jcoomes@1993 685 // resets the size to 0 before the next call (which is sequential,
duke@435 686 // since this is pop_local.)
jcoomes@1342 687 uint dirty_n_elems = dirty_size(localBot, _age.top());
jcoomes@1342 688 assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
duke@435 689 if (dirty_n_elems == 0) return false;
duke@435 690 localBot = decrement_index(localBot);
duke@435 691 _bottom = localBot;
duke@435 692 // This is necessary to prevent any read below from being reordered
duke@435 693 // before the store just above.
duke@435 694 OrderAccess::fence();
jcoomes@1746 695 const_cast<E&>(t = _elems[localBot]);
duke@435 696 // This is a second read of "age"; the "size()" above is the first.
duke@435 697 // If there's still at least one element in the queue, based on the
duke@435 698 // "_bottom" and "age" we've read, then there can be no interference with
duke@435 699 // a "pop_global" operation, and we're done.
jcoomes@1342 700 idx_t tp = _age.top(); // XXX
duke@435 701 if (size(localBot, tp) > 0) {
jcoomes@1342 702 assert(dirty_size(localBot, tp) != N - 1, "sanity");
jcoomes@2020 703 TASKQUEUE_STATS_ONLY(stats.record_pop());
duke@435 704 return true;
duke@435 705 } else {
duke@435 706 // Otherwise, the queue contained exactly one element; we take the slow
duke@435 707 // path.
jcoomes@1342 708 return pop_local_slow(localBot, _age.get());
duke@435 709 }
duke@435 710 }
duke@435 711
jcoomes@1993 712 typedef GenericTaskQueue<oop> OopTaskQueue;
jcoomes@1746 713 typedef GenericTaskQueueSet<OopTaskQueue> OopTaskQueueSet;
duke@435 714
jcoomes@1746 715 #ifdef _MSC_VER
jcoomes@1746 716 #pragma warning(push)
jcoomes@1746 717 // warning C4522: multiple assignment operators specified
jcoomes@1746 718 #pragma warning(disable:4522)
jcoomes@1746 719 #endif
coleenp@548 720
coleenp@548 721 // This is a container class for either an oop* or a narrowOop*.
coleenp@548 722 // Both are pushed onto a task queue and the consumer will test is_narrow()
coleenp@548 723 // to determine which should be processed.
coleenp@548 724 class StarTask {
coleenp@548 725 void* _holder; // either union oop* or narrowOop*
jcoomes@1746 726
jcoomes@1746 727 enum { COMPRESSED_OOP_MASK = 1 };
jcoomes@1746 728
coleenp@548 729 public:
ysr@1280 730 StarTask(narrowOop* p) {
ysr@1280 731 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
ysr@1280 732 _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
ysr@1280 733 }
ysr@1280 734 StarTask(oop* p) {
ysr@1280 735 assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
ysr@1280 736 _holder = (void*)p;
ysr@1280 737 }
coleenp@548 738 StarTask() { _holder = NULL; }
coleenp@548 739 operator oop*() { return (oop*)_holder; }
coleenp@548 740 operator narrowOop*() {
coleenp@548 741 return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
coleenp@548 742 }
coleenp@548 743
jcoomes@1746 744 StarTask& operator=(const StarTask& t) {
jcoomes@1746 745 _holder = t._holder;
jcoomes@1746 746 return *this;
jcoomes@1746 747 }
jcoomes@1746 748 volatile StarTask& operator=(const volatile StarTask& t) volatile {
jcoomes@1746 749 _holder = t._holder;
jcoomes@1746 750 return *this;
jcoomes@1746 751 }
coleenp@548 752
coleenp@548 753 bool is_narrow() const {
coleenp@548 754 return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
coleenp@548 755 }
coleenp@548 756 };
coleenp@548 757
jcoomes@1746 758 class ObjArrayTask
jcoomes@1746 759 {
jcoomes@1746 760 public:
jcoomes@1746 761 ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
jcoomes@1746 762 ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
jcoomes@1746 763 assert(idx <= size_t(max_jint), "too big");
jcoomes@1746 764 }
jcoomes@1746 765 ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
jcoomes@1746 766
jcoomes@1746 767 ObjArrayTask& operator =(const ObjArrayTask& t) {
jcoomes@1746 768 _obj = t._obj;
jcoomes@1746 769 _index = t._index;
jcoomes@1746 770 return *this;
jcoomes@1746 771 }
jcoomes@1746 772 volatile ObjArrayTask&
jcoomes@1746 773 operator =(const volatile ObjArrayTask& t) volatile {
jcoomes@1746 774 _obj = t._obj;
jcoomes@1746 775 _index = t._index;
jcoomes@1746 776 return *this;
jcoomes@1746 777 }
jcoomes@1746 778
jcoomes@1746 779 inline oop obj() const { return _obj; }
jcoomes@1746 780 inline int index() const { return _index; }
jcoomes@1746 781
jcoomes@1746 782 DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
jcoomes@1746 783
jcoomes@1746 784 private:
jcoomes@1746 785 oop _obj;
jcoomes@1746 786 int _index;
jcoomes@1746 787 };
jcoomes@1746 788
jcoomes@1746 789 #ifdef _MSC_VER
jcoomes@1746 790 #pragma warning(pop)
jcoomes@1746 791 #endif
jcoomes@1746 792
jcoomes@1993 793 typedef OverflowTaskQueue<StarTask> OopStarTaskQueue;
jcoomes@1746 794 typedef GenericTaskQueueSet<OopStarTaskQueue> OopStarTaskQueueSet;
duke@435 795
jcoomes@1993 796 typedef OverflowTaskQueue<size_t> RegionTaskQueue;
jcoomes@1993 797 typedef GenericTaskQueueSet<RegionTaskQueue> RegionTaskQueueSet;
jmasa@2188 798
stefank@2314 799
stefank@2314 800 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP

mercurial