1 /*
   2  * Copyright (c) 2001, 2011, 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_UTILITIES_TASKQUEUE_HPP
  26 #define SHARE_VM_UTILITIES_TASKQUEUE_HPP
  27 
  28 #include "memory/allocation.hpp"
  29 #include "memory/allocation.inline.hpp"
  30 #include "runtime/mutex.hpp"
  31 #include "utilities/stack.hpp"
  32 #ifdef TARGET_OS_ARCH_linux_x86
  33 # include "orderAccess_linux_x86.inline.hpp"
  34 #endif
  35 #ifdef TARGET_OS_ARCH_linux_sparc
  36 # include "orderAccess_linux_sparc.inline.hpp"
  37 #endif
  38 #ifdef TARGET_OS_ARCH_linux_zero
  39 # include "orderAccess_linux_zero.inline.hpp"
  40 #endif
  41 #ifdef TARGET_OS_ARCH_solaris_x86
  42 # include "orderAccess_solaris_x86.inline.hpp"
  43 #endif
  44 #ifdef TARGET_OS_ARCH_solaris_sparc
  45 # include "orderAccess_solaris_sparc.inline.hpp"
  46 #endif
  47 #ifdef TARGET_OS_ARCH_windows_x86
  48 # include "orderAccess_windows_x86.inline.hpp"
  49 #endif
  50 #ifdef TARGET_OS_ARCH_linux_arm
  51 # include "orderAccess_linux_arm.inline.hpp"
  52 #endif
  53 #ifdef TARGET_OS_ARCH_linux_ppc
  54 # include "orderAccess_linux_ppc.inline.hpp"
  55 #endif
  56 #ifdef TARGET_OS_ARCH_bsd_x86
  57 # include "orderAccess_bsd_x86.inline.hpp"
  58 #endif
  59 #ifdef TARGET_OS_ARCH_bsd_zero
  60 # include "orderAccess_bsd_zero.inline.hpp"
  61 #endif
  62 
  63 // Simple TaskQueue stats that are collected by default in debug builds.
  64 
  65 #if !defined(TASKQUEUE_STATS) && defined(ASSERT)
  66 #define TASKQUEUE_STATS 1
  67 #elif !defined(TASKQUEUE_STATS)
  68 #define TASKQUEUE_STATS 0
  69 #endif
  70 
  71 #if TASKQUEUE_STATS
  72 #define TASKQUEUE_STATS_ONLY(code) code
  73 #else
  74 #define TASKQUEUE_STATS_ONLY(code)
  75 #endif // TASKQUEUE_STATS
  76 
  77 #if TASKQUEUE_STATS
  78 class TaskQueueStats {
  79 public:
  80   enum StatId {
  81     push,             // number of taskqueue pushes
  82     pop,              // number of taskqueue pops
  83     pop_slow,         // subset of taskqueue pops that were done slow-path
  84     steal_attempt,    // number of taskqueue steal attempts
  85     steal,            // number of taskqueue steals
  86     overflow,         // number of overflow pushes
  87     overflow_max_len, // max length of overflow stack
  88     last_stat_id
  89   };
  90 
  91 public:
  92   inline TaskQueueStats()       { reset(); }
  93 
  94   inline void record_push()     { ++_stats[push]; }
  95   inline void record_pop()      { ++_stats[pop]; }
  96   inline void record_pop_slow() { record_pop(); ++_stats[pop_slow]; }
  97   inline void record_steal(bool success);
  98   inline void record_overflow(size_t new_length);
  99 
 100   TaskQueueStats & operator +=(const TaskQueueStats & addend);
 101 
 102   inline size_t get(StatId id) const { return _stats[id]; }
 103   inline const size_t* get() const   { return _stats; }
 104 
 105   inline void reset();
 106 
 107   // Print the specified line of the header (does not include a line separator).
 108   static void print_header(unsigned int line, outputStream* const stream = tty,
 109                            unsigned int width = 10);
 110   // Print the statistics (does not include a line separator).
 111   void print(outputStream* const stream = tty, unsigned int width = 10) const;
 112 
 113   DEBUG_ONLY(void verify() const;)
 114 
 115 private:
 116   size_t                    _stats[last_stat_id];
 117   static const char * const _names[last_stat_id];
 118 };
 119 
 120 void TaskQueueStats::record_steal(bool success) {
 121   ++_stats[steal_attempt];
 122   if (success) ++_stats[steal];
 123 }
 124 
 125 void TaskQueueStats::record_overflow(size_t new_len) {
 126   ++_stats[overflow];
 127   if (new_len > _stats[overflow_max_len]) _stats[overflow_max_len] = new_len;
 128 }
 129 
 130 void TaskQueueStats::reset() {
 131   memset(_stats, 0, sizeof(_stats));
 132 }
 133 #endif // TASKQUEUE_STATS
 134 
 135 template <unsigned int N>
 136 class TaskQueueSuper: public CHeapObj {
 137 protected:
 138   // Internal type for indexing the queue; also used for the tag.
 139   typedef NOT_LP64(uint16_t) LP64_ONLY(uint32_t) idx_t;
 140 
 141   // The first free element after the last one pushed (mod N).
 142   volatile uint _bottom;
 143 
 144   enum { MOD_N_MASK = N - 1 };
 145 
 146   class Age {
 147   public:
 148     Age(size_t data = 0)         { _data = data; }
 149     Age(const Age& age)          { _data = age._data; }
 150     Age(idx_t top, idx_t tag)    { _fields._top = top; _fields._tag = tag; }
 151 
 152     Age   get()        const volatile { return _data; }
 153     void  set(Age age) volatile       { _data = age._data; }
 154 
 155     idx_t top()        const volatile { return _fields._top; }
 156     idx_t tag()        const volatile { return _fields._tag; }
 157 
 158     // Increment top; if it wraps, increment tag also.
 159     void increment() {
 160       _fields._top = increment_index(_fields._top);
 161       if (_fields._top == 0) ++_fields._tag;
 162     }
 163 
 164     Age cmpxchg(const Age new_age, const Age old_age) volatile {
 165       return (size_t) Atomic::cmpxchg_ptr((intptr_t)new_age._data,
 166                                           (volatile intptr_t *)&_data,
 167                                           (intptr_t)old_age._data);
 168     }
 169 
 170     bool operator ==(const Age& other) const { return _data == other._data; }
 171 
 172   private:
 173     struct fields {
 174       idx_t _top;
 175       idx_t _tag;
 176     };
 177     union {
 178       size_t _data;
 179       fields _fields;
 180     };
 181   };
 182 
 183   volatile Age _age;
 184 
 185   // These both operate mod N.
 186   static uint increment_index(uint ind) {
 187     return (ind + 1) & MOD_N_MASK;
 188   }
 189   static uint decrement_index(uint ind) {
 190     return (ind - 1) & MOD_N_MASK;
 191   }
 192 
 193   // Returns a number in the range [0..N).  If the result is "N-1", it should be
 194   // interpreted as 0.
 195   uint dirty_size(uint bot, uint top) const {
 196     return (bot - top) & MOD_N_MASK;
 197   }
 198 
 199   // Returns the size corresponding to the given "bot" and "top".
 200   uint size(uint bot, uint top) const {
 201     uint sz = dirty_size(bot, top);
 202     // Has the queue "wrapped", so that bottom is less than top?  There's a
 203     // complicated special case here.  A pair of threads could perform pop_local
 204     // and pop_global operations concurrently, starting from a state in which
 205     // _bottom == _top+1.  The pop_local could succeed in decrementing _bottom,
 206     // and the pop_global in incrementing _top (in which case the pop_global
 207     // will be awarded the contested queue element.)  The resulting state must
 208     // be interpreted as an empty queue.  (We only need to worry about one such
 209     // event: only the queue owner performs pop_local's, and several concurrent
 210     // threads attempting to perform the pop_global will all perform the same
 211     // CAS, and only one can succeed.)  Any stealing thread that reads after
 212     // either the increment or decrement will see an empty queue, and will not
 213     // join the competitors.  The "sz == -1 || sz == N-1" state will not be
 214     // modified by concurrent queues, so the owner thread can reset the state to
 215     // _bottom == top so subsequent pushes will be performed normally.
 216     return (sz == N - 1) ? 0 : sz;
 217   }
 218 
 219 public:
 220   TaskQueueSuper() : _bottom(0), _age() {}
 221 
 222   // Return true if the TaskQueue contains/does not contain any tasks.
 223   bool peek()     const { return _bottom != _age.top(); }
 224   bool is_empty() const { return size() == 0; }
 225 
 226   // Return an estimate of the number of elements in the queue.
 227   // The "careful" version admits the possibility of pop_local/pop_global
 228   // races.
 229   uint size() const {
 230     return size(_bottom, _age.top());
 231   }
 232 
 233   uint dirty_size() const {
 234     return dirty_size(_bottom, _age.top());
 235   }
 236 
 237   void set_empty() {
 238     _bottom = 0;
 239     _age.set(0);
 240   }
 241 
 242   // Maximum number of elements allowed in the queue.  This is two less
 243   // than the actual queue size, for somewhat complicated reasons.
 244   uint max_elems() const { return N - 2; }
 245 
 246   // Total size of queue.
 247   static const uint total_size() { return N; }
 248 
 249   TASKQUEUE_STATS_ONLY(TaskQueueStats stats;)
 250 };
 251 
 252 template<class E, unsigned int N = TASKQUEUE_SIZE>
 253 class GenericTaskQueue: public TaskQueueSuper<N> {
 254 protected:
 255   typedef typename TaskQueueSuper<N>::Age Age;
 256   typedef typename TaskQueueSuper<N>::idx_t idx_t;
 257 
 258   using TaskQueueSuper<N>::_bottom;
 259   using TaskQueueSuper<N>::_age;
 260   using TaskQueueSuper<N>::increment_index;
 261   using TaskQueueSuper<N>::decrement_index;
 262   using TaskQueueSuper<N>::dirty_size;
 263 
 264 public:
 265   using TaskQueueSuper<N>::max_elems;
 266   using TaskQueueSuper<N>::size;
 267   TASKQUEUE_STATS_ONLY(using TaskQueueSuper<N>::stats;)
 268 
 269 private:
 270   // Slow paths for push, pop_local.  (pop_global has no fast path.)
 271   bool push_slow(E t, uint dirty_n_elems);
 272   bool pop_local_slow(uint localBot, Age oldAge);
 273 
 274 public:
 275   typedef E element_type;
 276 
 277   // Initializes the queue to empty.
 278   GenericTaskQueue();
 279 
 280   void initialize();
 281 
 282   // Push the task "t" on the queue.  Returns "false" iff the queue is full.
 283   inline bool push(E t);
 284 
 285   // Attempts to claim a task from the "local" end of the queue (the most
 286   // recently pushed).  If successful, returns true and sets t to the task;
 287   // otherwise, returns false (the queue is empty).
 288   inline bool pop_local(E& t);
 289 
 290   // Like pop_local(), but uses the "global" end of the queue (the least
 291   // recently pushed).
 292   bool pop_global(E& t);
 293 
 294   // Delete any resource associated with the queue.
 295   ~GenericTaskQueue();
 296 
 297   // apply the closure to all elements in the task queue
 298   void oops_do(OopClosure* f);
 299 
 300 private:
 301   // Element array.
 302   volatile E* _elems;
 303 };
 304 
 305 template<class E, unsigned int N>
 306 GenericTaskQueue<E, N>::GenericTaskQueue() {
 307   assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
 308 }
 309 
 310 template<class E, unsigned int N>
 311 void GenericTaskQueue<E, N>::initialize() {
 312   _elems = NEW_C_HEAP_ARRAY(E, N);
 313 }
 314 
 315 template<class E, unsigned int N>
 316 void GenericTaskQueue<E, N>::oops_do(OopClosure* f) {
 317   // tty->print_cr("START OopTaskQueue::oops_do");
 318   uint iters = size();
 319   uint index = _bottom;
 320   for (uint i = 0; i < iters; ++i) {
 321     index = decrement_index(index);
 322     // tty->print_cr("  doing entry %d," INTPTR_T " -> " INTPTR_T,
 323     //            index, &_elems[index], _elems[index]);
 324     E* t = (E*)&_elems[index];      // cast away volatility
 325     oop* p = (oop*)t;
 326     assert((*t)->is_oop_or_null(), "Not an oop or null");
 327     f->do_oop(p);
 328   }
 329   // tty->print_cr("END OopTaskQueue::oops_do");
 330 }
 331 
 332 template<class E, unsigned int N>
 333 bool GenericTaskQueue<E, N>::push_slow(E t, uint dirty_n_elems) {
 334   if (dirty_n_elems == N - 1) {
 335     // Actually means 0, so do the push.
 336     uint localBot = _bottom;
 337     // g++ complains if the volatile result of the assignment is unused.
 338     const_cast<E&>(_elems[localBot] = t);
 339     OrderAccess::release_store(&_bottom, increment_index(localBot));
 340     TASKQUEUE_STATS_ONLY(stats.record_push());
 341     return true;
 342   }
 343   return false;
 344 }
 345 
 346 // pop_local_slow() is done by the owning thread and is trying to
 347 // get the last task in the queue.  It will compete with pop_global()
 348 // that will be used by other threads.  The tag age is incremented
 349 // whenever the queue goes empty which it will do here if this thread
 350 // gets the last task or in pop_global() if the queue wraps (top == 0
 351 // and pop_global() succeeds, see pop_global()).
 352 template<class E, unsigned int N>
 353 bool GenericTaskQueue<E, N>::pop_local_slow(uint localBot, Age oldAge) {
 354   // This queue was observed to contain exactly one element; either this
 355   // thread will claim it, or a competing "pop_global".  In either case,
 356   // the queue will be logically empty afterwards.  Create a new Age value
 357   // that represents the empty queue for the given value of "_bottom".  (We
 358   // must also increment "tag" because of the case where "bottom == 1",
 359   // "top == 0".  A pop_global could read the queue element in that case,
 360   // then have the owner thread do a pop followed by another push.  Without
 361   // the incrementing of "tag", the pop_global's CAS could succeed,
 362   // allowing it to believe it has claimed the stale element.)
 363   Age newAge((idx_t)localBot, oldAge.tag() + 1);
 364   // Perhaps a competing pop_global has already incremented "top", in which
 365   // case it wins the element.
 366   if (localBot == oldAge.top()) {
 367     // No competing pop_global has yet incremented "top"; we'll try to
 368     // install new_age, thus claiming the element.
 369     Age tempAge = _age.cmpxchg(newAge, oldAge);
 370     if (tempAge == oldAge) {
 371       // We win.
 372       assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
 373       TASKQUEUE_STATS_ONLY(stats.record_pop_slow());
 374       return true;
 375     }
 376   }
 377   // We lose; a completing pop_global gets the element.  But the queue is empty
 378   // and top is greater than bottom.  Fix this representation of the empty queue
 379   // to become the canonical one.
 380   _age.set(newAge);
 381   assert(dirty_size(localBot, _age.top()) != N - 1, "sanity");
 382   return false;
 383 }
 384 
 385 template<class E, unsigned int N>
 386 bool GenericTaskQueue<E, N>::pop_global(E& t) {
 387   Age oldAge = _age.get();
 388   uint localBot = _bottom;
 389   uint n_elems = size(localBot, oldAge.top());
 390   if (n_elems == 0) {
 391     return false;
 392   }
 393 
 394   const_cast<E&>(t = _elems[oldAge.top()]);
 395   Age newAge(oldAge);
 396   newAge.increment();
 397   Age resAge = _age.cmpxchg(newAge, oldAge);
 398 
 399   // Note that using "_bottom" here might fail, since a pop_local might
 400   // have decremented it.
 401   assert(dirty_size(localBot, newAge.top()) != N - 1, "sanity");
 402   return resAge == oldAge;
 403 }
 404 
 405 template<class E, unsigned int N>
 406 GenericTaskQueue<E, N>::~GenericTaskQueue() {
 407   FREE_C_HEAP_ARRAY(E, _elems);
 408 }
 409 
 410 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
 411 // elements that do not fit in the TaskQueue.
 412 //
 413 // This class hides two methods from super classes:
 414 //
 415 // push() - push onto the task queue or, if that fails, onto the overflow stack
 416 // is_empty() - return true if both the TaskQueue and overflow stack are empty
 417 //
 418 // Note that size() is not hidden--it returns the number of elements in the
 419 // TaskQueue, and does not include the size of the overflow stack.  This
 420 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
 421 template<class E, unsigned int N = TASKQUEUE_SIZE>
 422 class OverflowTaskQueue: public GenericTaskQueue<E, N>
 423 {
 424 public:
 425   typedef Stack<E>               overflow_t;
 426   typedef GenericTaskQueue<E, N> taskqueue_t;
 427 
 428   TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
 429 
 430   // Push task t onto the queue or onto the overflow stack.  Return true.
 431   inline bool push(E t);
 432 
 433   // Attempt to pop from the overflow stack; return true if anything was popped.
 434   inline bool pop_overflow(E& t);
 435 
 436   inline overflow_t* overflow_stack() { return &_overflow_stack; }
 437 
 438   inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
 439   inline bool overflow_empty()  const { return _overflow_stack.is_empty(); }
 440   inline bool is_empty()        const {
 441     return taskqueue_empty() && overflow_empty();
 442   }
 443 
 444 private:
 445   overflow_t _overflow_stack;
 446 };
 447 
 448 template <class E, unsigned int N>
 449 bool OverflowTaskQueue<E, N>::push(E t)
 450 {
 451   if (!taskqueue_t::push(t)) {
 452     overflow_stack()->push(t);
 453     TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
 454   }
 455   return true;
 456 }
 457 
 458 template <class E, unsigned int N>
 459 bool OverflowTaskQueue<E, N>::pop_overflow(E& t)
 460 {
 461   if (overflow_empty()) return false;
 462   t = overflow_stack()->pop();
 463   return true;
 464 }
 465 
 466 class TaskQueueSetSuper: public CHeapObj {
 467 protected:
 468   static int randomParkAndMiller(int* seed0);
 469 public:
 470   // Returns "true" if some TaskQueue in the set contains a task.
 471   virtual bool peek() = 0;
 472 };
 473 
 474 template<class T>
 475 class GenericTaskQueueSet: public TaskQueueSetSuper {
 476 private:
 477   uint _n;
 478   T** _queues;
 479 
 480 public:
 481   typedef typename T::element_type E;
 482 
 483   GenericTaskQueueSet(int n) : _n(n) {
 484     typedef T* GenericTaskQueuePtr;
 485     _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n);
 486     for (int i = 0; i < n; i++) {
 487       _queues[i] = NULL;
 488     }
 489   }
 490 
 491   bool steal_1_random(uint queue_num, int* seed, E& t);
 492   bool steal_best_of_2(uint queue_num, int* seed, E& t);
 493   bool steal_best_of_all(uint queue_num, int* seed, E& t);
 494 
 495   void register_queue(uint i, T* q);
 496 
 497   T* queue(uint n);
 498 
 499   // The thread with queue number "queue_num" (and whose random number seed is
 500   // at "seed") is trying to steal a task from some other queue.  (It may try
 501   // several queues, according to some configuration parameter.)  If some steal
 502   // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
 503   // false.
 504   bool steal(uint queue_num, int* seed, E& t);
 505 
 506   bool peek();
 507 };
 508 
 509 template<class T> void
 510 GenericTaskQueueSet<T>::register_queue(uint i, T* q) {
 511   assert(i < _n, "index out of range.");
 512   _queues[i] = q;
 513 }
 514 
 515 template<class T> T*
 516 GenericTaskQueueSet<T>::queue(uint i) {
 517   return _queues[i];
 518 }
 519 
 520 template<class T> bool
 521 GenericTaskQueueSet<T>::steal(uint queue_num, int* seed, E& t) {
 522   for (uint i = 0; i < 2 * _n; i++) {
 523     if (steal_best_of_2(queue_num, seed, t)) {
 524       TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
 525       return true;
 526     }
 527   }
 528   TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(false));
 529   return false;
 530 }
 531 
 532 template<class T> bool
 533 GenericTaskQueueSet<T>::steal_best_of_all(uint queue_num, int* seed, E& t) {
 534   if (_n > 2) {
 535     int best_k;
 536     uint best_sz = 0;
 537     for (uint k = 0; k < _n; k++) {
 538       if (k == queue_num) continue;
 539       uint sz = _queues[k]->size();
 540       if (sz > best_sz) {
 541         best_sz = sz;
 542         best_k = k;
 543       }
 544     }
 545     return best_sz > 0 && _queues[best_k]->pop_global(t);
 546   } else if (_n == 2) {
 547     // Just try the other one.
 548     int k = (queue_num + 1) % 2;
 549     return _queues[k]->pop_global(t);
 550   } else {
 551     assert(_n == 1, "can't be zero.");
 552     return false;
 553   }
 554 }
 555 
 556 template<class T> bool
 557 GenericTaskQueueSet<T>::steal_1_random(uint queue_num, int* seed, E& t) {
 558   if (_n > 2) {
 559     uint k = queue_num;
 560     while (k == queue_num) k = randomParkAndMiller(seed) % _n;
 561     return _queues[2]->pop_global(t);
 562   } else if (_n == 2) {
 563     // Just try the other one.
 564     int k = (queue_num + 1) % 2;
 565     return _queues[k]->pop_global(t);
 566   } else {
 567     assert(_n == 1, "can't be zero.");
 568     return false;
 569   }
 570 }
 571 
 572 template<class T> bool
 573 GenericTaskQueueSet<T>::steal_best_of_2(uint queue_num, int* seed, E& t) {
 574   if (_n > 2) {
 575     uint k1 = queue_num;
 576     while (k1 == queue_num) k1 = randomParkAndMiller(seed) % _n;
 577     uint k2 = queue_num;
 578     while (k2 == queue_num || k2 == k1) k2 = randomParkAndMiller(seed) % _n;
 579     // Sample both and try the larger.
 580     uint sz1 = _queues[k1]->size();
 581     uint sz2 = _queues[k2]->size();
 582     if (sz2 > sz1) return _queues[k2]->pop_global(t);
 583     else return _queues[k1]->pop_global(t);
 584   } else if (_n == 2) {
 585     // Just try the other one.
 586     uint k = (queue_num + 1) % 2;
 587     return _queues[k]->pop_global(t);
 588   } else {
 589     assert(_n == 1, "can't be zero.");
 590     return false;
 591   }
 592 }
 593 
 594 template<class T>
 595 bool GenericTaskQueueSet<T>::peek() {
 596   // Try all the queues.
 597   for (uint j = 0; j < _n; j++) {
 598     if (_queues[j]->peek())
 599       return true;
 600   }
 601   return false;
 602 }
 603 
 604 // When to terminate from the termination protocol.
 605 class TerminatorTerminator: public CHeapObj {
 606 public:
 607   virtual bool should_exit_termination() = 0;
 608 };
 609 
 610 // A class to aid in the termination of a set of parallel tasks using
 611 // TaskQueueSet's for work stealing.
 612 
 613 #undef TRACESPINNING
 614 
 615 class ParallelTaskTerminator: public StackObj {
 616 private:
 617   int _n_threads;
 618   TaskQueueSetSuper* _queue_set;
 619   int _offered_termination;
 620 
 621 #ifdef TRACESPINNING
 622   static uint _total_yields;
 623   static uint _total_spins;
 624   static uint _total_peeks;
 625 #endif
 626 
 627   bool peek_in_queue_set();
 628 protected:
 629   virtual void yield();
 630   void sleep(uint millis);
 631 
 632 public:
 633 
 634   // "n_threads" is the number of threads to be terminated.  "queue_set" is a
 635   // queue sets of work queues of other threads.
 636   ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set);
 637 
 638   // The current thread has no work, and is ready to terminate if everyone
 639   // else is.  If returns "true", all threads are terminated.  If returns
 640   // "false", available work has been observed in one of the task queues,
 641   // so the global task is not complete.
 642   bool offer_termination() {
 643     return offer_termination(NULL);
 644   }
 645 
 646   // As above, but it also terminates if the should_exit_termination()
 647   // method of the terminator parameter returns true. If terminator is
 648   // NULL, then it is ignored.
 649   bool offer_termination(TerminatorTerminator* terminator);
 650 
 651   // Reset the terminator, so that it may be reused again.
 652   // The caller is responsible for ensuring that this is done
 653   // in an MT-safe manner, once the previous round of use of
 654   // the terminator is finished.
 655   void reset_for_reuse();
 656   // Same as above but the number of parallel threads is set to the
 657   // given number.
 658   void reset_for_reuse(int n_threads);
 659 
 660 #ifdef TRACESPINNING
 661   static uint total_yields() { return _total_yields; }
 662   static uint total_spins() { return _total_spins; }
 663   static uint total_peeks() { return _total_peeks; }
 664   static void print_termination_counts();
 665 #endif
 666 };
 667 
 668 template<class E, unsigned int N> inline bool
 669 GenericTaskQueue<E, N>::push(E t) {
 670   uint localBot = _bottom;
 671   assert((localBot >= 0) && (localBot < N), "_bottom out of range.");
 672   idx_t top = _age.top();
 673   uint dirty_n_elems = dirty_size(localBot, top);
 674   assert(dirty_n_elems < N, "n_elems out of range.");
 675   if (dirty_n_elems < max_elems()) {
 676     // g++ complains if the volatile result of the assignment is unused.
 677     const_cast<E&>(_elems[localBot] = t);
 678     OrderAccess::release_store(&_bottom, increment_index(localBot));
 679     TASKQUEUE_STATS_ONLY(stats.record_push());
 680     return true;
 681   } else {
 682     return push_slow(t, dirty_n_elems);
 683   }
 684 }
 685 
 686 template<class E, unsigned int N> inline bool
 687 GenericTaskQueue<E, N>::pop_local(E& t) {
 688   uint localBot = _bottom;
 689   // This value cannot be N-1.  That can only occur as a result of
 690   // the assignment to bottom in this method.  If it does, this method
 691   // resets the size to 0 before the next call (which is sequential,
 692   // since this is pop_local.)
 693   uint dirty_n_elems = dirty_size(localBot, _age.top());
 694   assert(dirty_n_elems != N - 1, "Shouldn't be possible...");
 695   if (dirty_n_elems == 0) return false;
 696   localBot = decrement_index(localBot);
 697   _bottom = localBot;
 698   // This is necessary to prevent any read below from being reordered
 699   // before the store just above.
 700   OrderAccess::fence();
 701   const_cast<E&>(t = _elems[localBot]);
 702   // This is a second read of "age"; the "size()" above is the first.
 703   // If there's still at least one element in the queue, based on the
 704   // "_bottom" and "age" we've read, then there can be no interference with
 705   // a "pop_global" operation, and we're done.
 706   idx_t tp = _age.top();    // XXX
 707   if (size(localBot, tp) > 0) {
 708     assert(dirty_size(localBot, tp) != N - 1, "sanity");
 709     TASKQUEUE_STATS_ONLY(stats.record_pop());
 710     return true;
 711   } else {
 712     // Otherwise, the queue contained exactly one element; we take the slow
 713     // path.
 714     return pop_local_slow(localBot, _age.get());
 715   }
 716 }
 717 
 718 typedef GenericTaskQueue<oop>             OopTaskQueue;
 719 typedef GenericTaskQueueSet<OopTaskQueue> OopTaskQueueSet;
 720 
 721 #ifdef _MSC_VER
 722 #pragma warning(push)
 723 // warning C4522: multiple assignment operators specified
 724 #pragma warning(disable:4522)
 725 #endif
 726 
 727 // This is a container class for either an oop* or a narrowOop*.
 728 // Both are pushed onto a task queue and the consumer will test is_narrow()
 729 // to determine which should be processed.
 730 class StarTask {
 731   void*  _holder;        // either union oop* or narrowOop*
 732 
 733   enum { COMPRESSED_OOP_MASK = 1 };
 734 
 735  public:
 736   StarTask(narrowOop* p) {
 737     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
 738     _holder = (void *)((uintptr_t)p | COMPRESSED_OOP_MASK);
 739   }
 740   StarTask(oop* p)       {
 741     assert(((uintptr_t)p & COMPRESSED_OOP_MASK) == 0, "Information loss!");
 742     _holder = (void*)p;
 743   }
 744   StarTask()             { _holder = NULL; }
 745   operator oop*()        { return (oop*)_holder; }
 746   operator narrowOop*()  {
 747     return (narrowOop*)((uintptr_t)_holder & ~COMPRESSED_OOP_MASK);
 748   }
 749 
 750   StarTask& operator=(const StarTask& t) {
 751     _holder = t._holder;
 752     return *this;
 753   }
 754   volatile StarTask& operator=(const volatile StarTask& t) volatile {
 755     _holder = t._holder;
 756     return *this;
 757   }
 758 
 759   bool is_narrow() const {
 760     return (((uintptr_t)_holder & COMPRESSED_OOP_MASK) != 0);
 761   }
 762 };
 763 
 764 class ObjArrayTask
 765 {
 766 public:
 767   ObjArrayTask(oop o = NULL, int idx = 0): _obj(o), _index(idx) { }
 768   ObjArrayTask(oop o, size_t idx): _obj(o), _index(int(idx)) {
 769     assert(idx <= size_t(max_jint), "too big");
 770   }
 771   ObjArrayTask(const ObjArrayTask& t): _obj(t._obj), _index(t._index) { }
 772 
 773   ObjArrayTask& operator =(const ObjArrayTask& t) {
 774     _obj = t._obj;
 775     _index = t._index;
 776     return *this;
 777   }
 778   volatile ObjArrayTask&
 779   operator =(const volatile ObjArrayTask& t) volatile {
 780     _obj = t._obj;
 781     _index = t._index;
 782     return *this;
 783   }
 784 
 785   inline oop obj()   const { return _obj; }
 786   inline int index() const { return _index; }
 787 
 788   DEBUG_ONLY(bool is_valid() const); // Tasks to be pushed/popped must be valid.
 789 
 790 private:
 791   oop _obj;
 792   int _index;
 793 };
 794 
 795 #ifdef _MSC_VER
 796 #pragma warning(pop)
 797 #endif
 798 
 799 typedef OverflowTaskQueue<StarTask>           OopStarTaskQueue;
 800 typedef GenericTaskQueueSet<OopStarTaskQueue> OopStarTaskQueueSet;
 801 
 802 typedef OverflowTaskQueue<size_t>             RegionTaskQueue;
 803 typedef GenericTaskQueueSet<RegionTaskQueue>  RegionTaskQueueSet;
 804 
 805 
 806 #endif // SHARE_VM_UTILITIES_TASKQUEUE_HPP