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 #include "precompiled.hpp"
  26 #include "gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.hpp"
  27 #include "gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp"
  28 #include "gc_implementation/parNew/parNewGeneration.hpp"
  29 #include "gc_implementation/parNew/parOopClosures.inline.hpp"
  30 #include "gc_implementation/shared/adaptiveSizePolicy.hpp"
  31 #include "gc_implementation/shared/ageTable.hpp"
  32 #include "gc_implementation/shared/copyFailedInfo.hpp"
  33 #include "gc_implementation/shared/gcHeapSummary.hpp"
  34 #include "gc_implementation/shared/gcTimer.hpp"
  35 #include "gc_implementation/shared/gcTrace.hpp"
  36 #include "gc_implementation/shared/gcTraceTime.hpp"
  37 #include "gc_implementation/shared/parGCAllocBuffer.inline.hpp"
  38 #include "gc_implementation/shared/spaceDecorator.hpp"
  39 #include "memory/defNewGeneration.inline.hpp"
  40 #include "memory/genCollectedHeap.hpp"
  41 #include "memory/genOopClosures.inline.hpp"
  42 #include "memory/generation.hpp"
  43 #include "memory/referencePolicy.hpp"
  44 #include "memory/resourceArea.hpp"
  45 #include "memory/sharedHeap.hpp"
  46 #include "memory/space.hpp"
  47 #include "oops/objArrayOop.hpp"
  48 #include "oops/oop.inline.hpp"
  49 #include "runtime/atomic.inline.hpp"
  50 #include "runtime/handles.hpp"
  51 #include "runtime/handles.inline.hpp"
  52 #include "runtime/java.hpp"
  53 #include "runtime/thread.inline.hpp"
  54 #include "utilities/copy.hpp"
  55 #include "utilities/globalDefinitions.hpp"
  56 #include "utilities/workgroup.hpp"
  57 
  58 #ifdef _MSC_VER
  59 #pragma warning( push )
  60 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
  61 #endif
  62 ParScanThreadState::ParScanThreadState(Space* to_space_,
  63                                        ParNewGeneration* gen_,
  64                                        Generation* old_gen_,
  65                                        int thread_num_,
  66                                        ObjToScanQueueSet* work_queue_set_,
  67                                        Stack<oop, mtGC>* overflow_stacks_,
  68                                        size_t desired_plab_sz_,
  69                                        ParallelTaskTerminator& term_) :
  70   _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
  71   _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
  72   _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
  73   _ageTable(false), // false ==> not the global age table, no perf data.
  74   _to_space_alloc_buffer(desired_plab_sz_),
  75   _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
  76   _to_space_root_closure(gen_, this), _old_gen_root_closure(gen_, this),
  77   _older_gen_closure(gen_, this),
  78   _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
  79                       &_to_space_root_closure, gen_, &_old_gen_root_closure,
  80                       work_queue_set_, &term_),
  81   _is_alive_closure(gen_), _scan_weak_ref_closure(gen_, this),
  82   _keep_alive_closure(&_scan_weak_ref_closure),
  83   _strong_roots_time(0.0), _term_time(0.0)
  84 {
  85   #if TASKQUEUE_STATS
  86   _term_attempts = 0;
  87   _overflow_refills = 0;
  88   _overflow_refill_objs = 0;
  89   #endif // TASKQUEUE_STATS
  90 
  91   _survivor_chunk_array =
  92     (ChunkArray*) old_gen()->get_data_recorder(thread_num());
  93   _hash_seed = 17;  // Might want to take time-based random value.
  94   _start = os::elapsedTime();
  95   _old_gen_closure.set_generation(old_gen_);
  96   _old_gen_root_closure.set_generation(old_gen_);
  97 }
  98 #ifdef _MSC_VER
  99 #pragma warning( pop )
 100 #endif
 101 
 102 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
 103                                               size_t plab_word_size) {
 104   ChunkArray* sca = survivor_chunk_array();
 105   if (sca != NULL) {
 106     // A non-null SCA implies that we want the PLAB data recorded.
 107     sca->record_sample(plab_start, plab_word_size);
 108   }
 109 }
 110 
 111 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
 112   return new_obj->is_objArray() &&
 113          arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
 114          new_obj != old_obj;
 115 }
 116 
 117 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
 118   assert(old->is_objArray(), "must be obj array");
 119   assert(old->is_forwarded(), "must be forwarded");
 120   assert(Universe::heap()->is_in_reserved(old), "must be in heap.");
 121   assert(!old_gen()->is_in(old), "must be in young generation.");
 122 
 123   objArrayOop obj = objArrayOop(old->forwardee());
 124   // Process ParGCArrayScanChunk elements now
 125   // and push the remainder back onto queue
 126   int start     = arrayOop(old)->length();
 127   int end       = obj->length();
 128   int remainder = end - start;
 129   assert(start <= end, "just checking");
 130   if (remainder > 2 * ParGCArrayScanChunk) {
 131     // Test above combines last partial chunk with a full chunk
 132     end = start + ParGCArrayScanChunk;
 133     arrayOop(old)->set_length(end);
 134     // Push remainder.
 135     bool ok = work_queue()->push(old);
 136     assert(ok, "just popped, push must be okay");
 137   } else {
 138     // Restore length so that it can be used if there
 139     // is a promotion failure and forwarding pointers
 140     // must be removed.
 141     arrayOop(old)->set_length(end);
 142   }
 143 
 144   // process our set of indices (include header in first chunk)
 145   // should make sure end is even (aligned to HeapWord in case of compressed oops)
 146   if ((HeapWord *)obj < young_old_boundary()) {
 147     // object is in to_space
 148     obj->oop_iterate_range(&_to_space_closure, start, end);
 149   } else {
 150     // object is in old generation
 151     obj->oop_iterate_range(&_old_gen_closure, start, end);
 152   }
 153 }
 154 
 155 
 156 void ParScanThreadState::trim_queues(int max_size) {
 157   ObjToScanQueue* queue = work_queue();
 158   do {
 159     while (queue->size() > (juint)max_size) {
 160       oop obj_to_scan;
 161       if (queue->pop_local(obj_to_scan)) {
 162         if ((HeapWord *)obj_to_scan < young_old_boundary()) {
 163           if (obj_to_scan->is_objArray() &&
 164               obj_to_scan->is_forwarded() &&
 165               obj_to_scan->forwardee() != obj_to_scan) {
 166             scan_partial_array_and_push_remainder(obj_to_scan);
 167           } else {
 168             // object is in to_space
 169             obj_to_scan->oop_iterate(&_to_space_closure);
 170           }
 171         } else {
 172           // object is in old generation
 173           obj_to_scan->oop_iterate(&_old_gen_closure);
 174         }
 175       }
 176     }
 177     // For the  case of compressed oops, we have a private, non-shared
 178     // overflow stack, so we eagerly drain it so as to more evenly
 179     // distribute load early. Note: this may be good to do in
 180     // general rather than delay for the final stealing phase.
 181     // If applicable, we'll transfer a set of objects over to our
 182     // work queue, allowing them to be stolen and draining our
 183     // private overflow stack.
 184   } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
 185 }
 186 
 187 bool ParScanThreadState::take_from_overflow_stack() {
 188   assert(ParGCUseLocalOverflow, "Else should not call");
 189   assert(young_gen()->overflow_list() == NULL, "Error");
 190   ObjToScanQueue* queue = work_queue();
 191   Stack<oop, mtGC>* const of_stack = overflow_stack();
 192   const size_t num_overflow_elems = of_stack->size();
 193   const size_t space_available = queue->max_elems() - queue->size();
 194   const size_t num_take_elems = MIN3(space_available / 4,
 195                                      ParGCDesiredObjsFromOverflowList,
 196                                      num_overflow_elems);
 197   // Transfer the most recent num_take_elems from the overflow
 198   // stack to our work queue.
 199   for (size_t i = 0; i != num_take_elems; i++) {
 200     oop cur = of_stack->pop();
 201     oop obj_to_push = cur->forwardee();
 202     assert(Universe::heap()->is_in_reserved(cur), "Should be in heap");
 203     assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
 204     assert(Universe::heap()->is_in_reserved(obj_to_push), "Should be in heap");
 205     if (should_be_partially_scanned(obj_to_push, cur)) {
 206       assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
 207       obj_to_push = cur;
 208     }
 209     bool ok = queue->push(obj_to_push);
 210     assert(ok, "Should have succeeded");
 211   }
 212   assert(young_gen()->overflow_list() == NULL, "Error");
 213   return num_take_elems > 0;  // was something transferred?
 214 }
 215 
 216 void ParScanThreadState::push_on_overflow_stack(oop p) {
 217   assert(ParGCUseLocalOverflow, "Else should not call");
 218   overflow_stack()->push(p);
 219   assert(young_gen()->overflow_list() == NULL, "Error");
 220 }
 221 
 222 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
 223 
 224   // Otherwise, if the object is small enough, try to reallocate the
 225   // buffer.
 226   HeapWord* obj = NULL;
 227   if (!_to_space_full) {
 228     ParGCAllocBuffer* const plab = to_space_alloc_buffer();
 229     Space*            const sp   = to_space();
 230     if (word_sz * 100 <
 231         ParallelGCBufferWastePct * plab->word_sz()) {
 232       // Is small enough; abandon this buffer and start a new one.
 233       plab->retire();
 234       size_t buf_size = plab->word_sz();
 235       HeapWord* buf_space = sp->par_allocate(buf_size);
 236       if (buf_space == NULL) {
 237         const size_t min_bytes =
 238           ParGCAllocBuffer::min_size() << LogHeapWordSize;
 239         size_t free_bytes = sp->free();
 240         while(buf_space == NULL && free_bytes >= min_bytes) {
 241           buf_size = free_bytes >> LogHeapWordSize;
 242           assert(buf_size == (size_t)align_object_size(buf_size),
 243                  "Invariant");
 244           buf_space  = sp->par_allocate(buf_size);
 245           free_bytes = sp->free();
 246         }
 247       }
 248       if (buf_space != NULL) {
 249         plab->set_word_size(buf_size);
 250         plab->set_buf(buf_space);
 251         record_survivor_plab(buf_space, buf_size);
 252         obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
 253         // Note that we cannot compare buf_size < word_sz below
 254         // because of AlignmentReserve (see ParGCAllocBuffer::allocate()).
 255         assert(obj != NULL || plab->words_remaining() < word_sz,
 256                "Else should have been able to allocate");
 257         // It's conceivable that we may be able to use the
 258         // buffer we just grabbed for subsequent small requests
 259         // even if not for this one.
 260       } else {
 261         // We're used up.
 262         _to_space_full = true;
 263       }
 264 
 265     } else {
 266       // Too large; allocate the object individually.
 267       obj = sp->par_allocate(word_sz);
 268     }
 269   }
 270   return obj;
 271 }
 272 
 273 
 274 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj,
 275                                                 size_t word_sz) {
 276   // Is the alloc in the current alloc buffer?
 277   if (to_space_alloc_buffer()->contains(obj)) {
 278     assert(to_space_alloc_buffer()->contains(obj + word_sz - 1),
 279            "Should contain whole object.");
 280     to_space_alloc_buffer()->undo_allocation(obj, word_sz);
 281   } else {
 282     CollectedHeap::fill_with_object(obj, word_sz);
 283   }
 284 }
 285 
 286 void ParScanThreadState::print_promotion_failure_size() {
 287   if (_promotion_failed_info.has_failed() && PrintPromotionFailure) {
 288     gclog_or_tty->print(" (%d: promotion failure size = " SIZE_FORMAT ") ",
 289                         _thread_num, _promotion_failed_info.first_size());
 290   }
 291 }
 292 
 293 class ParScanThreadStateSet: private ResourceArray {
 294 public:
 295   // Initializes states for the specified number of threads;
 296   ParScanThreadStateSet(int                     num_threads,
 297                         Space&                  to_space,
 298                         ParNewGeneration&       gen,
 299                         Generation&             old_gen,
 300                         ObjToScanQueueSet&      queue_set,
 301                         Stack<oop, mtGC>*       overflow_stacks_,
 302                         size_t                  desired_plab_sz,
 303                         ParallelTaskTerminator& term);
 304 
 305   ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
 306 
 307   inline ParScanThreadState& thread_state(int i);
 308 
 309   void trace_promotion_failed(const YoungGCTracer* gc_tracer);
 310   void reset(int active_workers, bool promotion_failed);
 311   void flush();
 312 
 313   #if TASKQUEUE_STATS
 314   static void
 315     print_termination_stats_hdr(outputStream* const st = gclog_or_tty);
 316   void print_termination_stats(outputStream* const st = gclog_or_tty);
 317   static void
 318     print_taskqueue_stats_hdr(outputStream* const st = gclog_or_tty);
 319   void print_taskqueue_stats(outputStream* const st = gclog_or_tty);
 320   void reset_stats();
 321   #endif // TASKQUEUE_STATS
 322 
 323 private:
 324   ParallelTaskTerminator& _term;
 325   ParNewGeneration&       _gen;
 326   Generation&             _old_gen;
 327  public:
 328   bool is_valid(int id) const { return id < length(); }
 329   ParallelTaskTerminator* terminator() { return &_term; }
 330 };
 331 
 332 
 333 ParScanThreadStateSet::ParScanThreadStateSet(
 334   int num_threads, Space& to_space, ParNewGeneration& gen,
 335   Generation& old_gen, ObjToScanQueueSet& queue_set,
 336   Stack<oop, mtGC>* overflow_stacks,
 337   size_t desired_plab_sz, ParallelTaskTerminator& term)
 338   : ResourceArray(sizeof(ParScanThreadState), num_threads),
 339     _gen(gen), _old_gen(old_gen), _term(term)
 340 {
 341   assert(num_threads > 0, "sanity check!");
 342   assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
 343          "overflow_stack allocation mismatch");
 344   // Initialize states.
 345   for (int i = 0; i < num_threads; ++i) {
 346     new ((ParScanThreadState*)_data + i)
 347         ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
 348                            overflow_stacks, desired_plab_sz, term);
 349   }
 350 }
 351 
 352 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i)
 353 {
 354   assert(i >= 0 && i < length(), "sanity check!");
 355   return ((ParScanThreadState*)_data)[i];
 356 }
 357 
 358 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
 359   for (int i = 0; i < length(); ++i) {
 360     if (thread_state(i).promotion_failed()) {
 361       gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
 362       thread_state(i).promotion_failed_info().reset();
 363     }
 364   }
 365 }
 366 
 367 void ParScanThreadStateSet::reset(int active_threads, bool promotion_failed)
 368 {
 369   _term.reset_for_reuse(active_threads);
 370   if (promotion_failed) {
 371     for (int i = 0; i < length(); ++i) {
 372       thread_state(i).print_promotion_failure_size();
 373     }
 374   }
 375 }
 376 
 377 #if TASKQUEUE_STATS
 378 void
 379 ParScanThreadState::reset_stats()
 380 {
 381   taskqueue_stats().reset();
 382   _term_attempts = 0;
 383   _overflow_refills = 0;
 384   _overflow_refill_objs = 0;
 385 }
 386 
 387 void ParScanThreadStateSet::reset_stats()
 388 {
 389   for (int i = 0; i < length(); ++i) {
 390     thread_state(i).reset_stats();
 391   }
 392 }
 393 
 394 void
 395 ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st)
 396 {
 397   st->print_raw_cr("GC Termination Stats");
 398   st->print_raw_cr("     elapsed  --strong roots-- "
 399                    "-------termination-------");
 400   st->print_raw_cr("thr     ms        ms       %   "
 401                    "    ms       %   attempts");
 402   st->print_raw_cr("--- --------- --------- ------ "
 403                    "--------- ------ --------");
 404 }
 405 
 406 void ParScanThreadStateSet::print_termination_stats(outputStream* const st)
 407 {
 408   print_termination_stats_hdr(st);
 409 
 410   for (int i = 0; i < length(); ++i) {
 411     const ParScanThreadState & pss = thread_state(i);
 412     const double elapsed_ms = pss.elapsed_time() * 1000.0;
 413     const double s_roots_ms = pss.strong_roots_time() * 1000.0;
 414     const double term_ms = pss.term_time() * 1000.0;
 415     st->print_cr("%3d %9.2f %9.2f %6.2f "
 416                  "%9.2f %6.2f " SIZE_FORMAT_W(8),
 417                  i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
 418                  term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
 419   }
 420 }
 421 
 422 // Print stats related to work queue activity.
 423 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st)
 424 {
 425   st->print_raw_cr("GC Task Stats");
 426   st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
 427   st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
 428 }
 429 
 430 void ParScanThreadStateSet::print_taskqueue_stats(outputStream* const st)
 431 {
 432   print_taskqueue_stats_hdr(st);
 433 
 434   TaskQueueStats totals;
 435   for (int i = 0; i < length(); ++i) {
 436     const ParScanThreadState & pss = thread_state(i);
 437     const TaskQueueStats & stats = pss.taskqueue_stats();
 438     st->print("%3d ", i); stats.print(st); st->cr();
 439     totals += stats;
 440 
 441     if (pss.overflow_refills() > 0) {
 442       st->print_cr("    " SIZE_FORMAT_W(10) " overflow refills    "
 443                    SIZE_FORMAT_W(10) " overflow objects",
 444                    pss.overflow_refills(), pss.overflow_refill_objs());
 445     }
 446   }
 447   st->print("tot "); totals.print(st); st->cr();
 448 
 449   DEBUG_ONLY(totals.verify());
 450 }
 451 #endif // TASKQUEUE_STATS
 452 
 453 void ParScanThreadStateSet::flush()
 454 {
 455   // Work in this loop should be kept as lightweight as
 456   // possible since this might otherwise become a bottleneck
 457   // to scaling. Should we add heavy-weight work into this
 458   // loop, consider parallelizing the loop into the worker threads.
 459   for (int i = 0; i < length(); ++i) {
 460     ParScanThreadState& par_scan_state = thread_state(i);
 461 
 462     // Flush stats related to To-space PLAB activity and
 463     // retire the last buffer.
 464     par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_gen.plab_stats());
 465 
 466     // Every thread has its own age table.  We need to merge
 467     // them all into one.
 468     ageTable *local_table = par_scan_state.age_table();
 469     _gen.age_table()->merge(local_table);
 470 
 471     // Inform old gen that we're done.
 472     _old_gen.par_promote_alloc_done(i);
 473     _old_gen.par_oop_since_save_marks_iterate_done(i);
 474   }
 475 
 476   if (UseConcMarkSweepGC) {
 477     // We need to call this even when ResizeOldPLAB is disabled
 478     // so as to avoid breaking some asserts. While we may be able
 479     // to avoid this by reorganizing the code a bit, I am loathe
 480     // to do that unless we find cases where ergo leads to bad
 481     // performance.
 482     CFLS_LAB::compute_desired_plab_size();
 483   }
 484 }
 485 
 486 ParScanClosure::ParScanClosure(ParNewGeneration* g,
 487                                ParScanThreadState* par_scan_state) :
 488   OopsInKlassOrGenClosure(g), _par_scan_state(par_scan_state), _g(g)
 489 {
 490   assert(_g->level() == 0, "Optimized for youngest generation");
 491   _boundary = _g->reserved().end();
 492 }
 493 
 494 void ParScanWithBarrierClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, true, false); }
 495 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
 496 
 497 void ParScanWithoutBarrierClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, false, false); }
 498 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
 499 
 500 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, true, true); }
 501 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
 502 
 503 void ParRootScanWithoutBarrierClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, false, true); }
 504 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
 505 
 506 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
 507                                              ParScanThreadState* par_scan_state)
 508   : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
 509 {}
 510 
 511 void ParScanWeakRefClosure::do_oop(oop* p)       { ParScanWeakRefClosure::do_oop_work(p); }
 512 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
 513 
 514 #ifdef WIN32
 515 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
 516 #endif
 517 
 518 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
 519     ParScanThreadState* par_scan_state_,
 520     ParScanWithoutBarrierClosure* to_space_closure_,
 521     ParScanWithBarrierClosure* old_gen_closure_,
 522     ParRootScanWithoutBarrierClosure* to_space_root_closure_,
 523     ParNewGeneration* par_gen_,
 524     ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
 525     ObjToScanQueueSet* task_queues_,
 526     ParallelTaskTerminator* terminator_) :
 527 
 528     _par_scan_state(par_scan_state_),
 529     _to_space_closure(to_space_closure_),
 530     _old_gen_closure(old_gen_closure_),
 531     _to_space_root_closure(to_space_root_closure_),
 532     _old_gen_root_closure(old_gen_root_closure_),
 533     _par_gen(par_gen_),
 534     _task_queues(task_queues_),
 535     _terminator(terminator_)
 536 {}
 537 
 538 void ParEvacuateFollowersClosure::do_void() {
 539   ObjToScanQueue* work_q = par_scan_state()->work_queue();
 540 
 541   while (true) {
 542 
 543     // Scan to-space and old-gen objs until we run out of both.
 544     oop obj_to_scan;
 545     par_scan_state()->trim_queues(0);
 546 
 547     // We have no local work, attempt to steal from other threads.
 548 
 549     // attempt to steal work from promoted.
 550     if (task_queues()->steal(par_scan_state()->thread_num(),
 551                              par_scan_state()->hash_seed(),
 552                              obj_to_scan)) {
 553       bool res = work_q->push(obj_to_scan);
 554       assert(res, "Empty queue should have room for a push.");
 555 
 556       //   if successful, goto Start.
 557       continue;
 558 
 559       // try global overflow list.
 560     } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
 561       continue;
 562     }
 563 
 564     // Otherwise, offer termination.
 565     par_scan_state()->start_term_time();
 566     if (terminator()->offer_termination()) break;
 567     par_scan_state()->end_term_time();
 568   }
 569   assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
 570          "Broken overflow list?");
 571   // Finish the last termination pause.
 572   par_scan_state()->end_term_time();
 573 }
 574 
 575 ParNewGenTask::ParNewGenTask(ParNewGeneration* gen, Generation* old_gen,
 576                              HeapWord* young_old_boundary, ParScanThreadStateSet* state_set) :
 577     AbstractGangTask("ParNewGeneration collection"),
 578     _gen(gen), _old_gen(old_gen),
 579     _young_old_boundary(young_old_boundary),
 580     _state_set(state_set)
 581   {}
 582 
 583 // Reset the terminator for the given number of
 584 // active threads.
 585 void ParNewGenTask::set_for_termination(int active_workers) {
 586   _state_set->reset(active_workers, _gen->promotion_failed());
 587   // Should the heap be passed in?  There's only 1 for now so
 588   // grab it instead.
 589   GenCollectedHeap* gch = GenCollectedHeap::heap();
 590   gch->set_n_termination(active_workers);
 591 }
 592 
 593 void ParNewGenTask::work(uint worker_id) {
 594   GenCollectedHeap* gch = GenCollectedHeap::heap();
 595   // Since this is being done in a separate thread, need new resource
 596   // and handle marks.
 597   ResourceMark rm;
 598   HandleMark hm;
 599 
 600   ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
 601   assert(_state_set->is_valid(worker_id), "Should not have been called");
 602 
 603   par_scan_state.set_young_old_boundary(_young_old_boundary);
 604 
 605   KlassScanClosure klass_scan_closure(&par_scan_state.to_space_root_closure(),
 606                                       gch->rem_set()->klass_rem_set());
 607   CLDToKlassAndOopClosure cld_scan_closure(&klass_scan_closure,
 608                                            &par_scan_state.to_space_root_closure(),
 609                                            false);
 610 
 611   par_scan_state.start_strong_roots();
 612   gch->gen_process_roots(_gen->level(),
 613                          true,  // Process younger gens, if any,
 614                                 // as strong roots.
 615                          false, // no scope; this is parallel code
 616                          GenCollectedHeap::SO_ScavengeCodeCache,
 617                          GenCollectedHeap::StrongAndWeakRoots,
 618                          &par_scan_state.to_space_root_closure(),
 619                          &par_scan_state.older_gen_closure(),
 620                          &cld_scan_closure);
 621 
 622   par_scan_state.end_strong_roots();
 623 
 624   // "evacuate followers".
 625   par_scan_state.evacuate_followers_closure().do_void();
 626 }
 627 
 628 #ifdef _MSC_VER
 629 #pragma warning( push )
 630 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
 631 #endif
 632 ParNewGeneration::
 633 ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
 634   : DefNewGeneration(rs, initial_byte_size, level, "PCopy"),
 635   _overflow_list(NULL),
 636   _is_alive_closure(this),
 637   _plab_stats(YoungPLABSize, PLABWeight)
 638 {
 639   NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
 640   NOT_PRODUCT(_num_par_pushes = 0;)
 641   _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
 642   guarantee(_task_queues != NULL, "task_queues allocation failure.");
 643 
 644   for (uint i1 = 0; i1 < ParallelGCThreads; i1++) {
 645     ObjToScanQueue *q = new ObjToScanQueue();
 646     guarantee(q != NULL, "work_queue Allocation failure.");
 647     _task_queues->register_queue(i1, q);
 648   }
 649 
 650   for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
 651     _task_queues->queue(i2)->initialize();
 652 
 653   _overflow_stacks = NULL;
 654   if (ParGCUseLocalOverflow) {
 655 
 656     // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal
 657     // with ','
 658     typedef Stack<oop, mtGC> GCOopStack;
 659 
 660     _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
 661     for (size_t i = 0; i < ParallelGCThreads; ++i) {
 662       new (_overflow_stacks + i) Stack<oop, mtGC>();
 663     }
 664   }
 665 
 666   if (UsePerfData) {
 667     EXCEPTION_MARK;
 668     ResourceMark rm;
 669 
 670     const char* cname =
 671          PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
 672     PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
 673                                      ParallelGCThreads, CHECK);
 674   }
 675 }
 676 #ifdef _MSC_VER
 677 #pragma warning( pop )
 678 #endif
 679 
 680 // ParNewGeneration::
 681 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
 682   DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
 683 
 684 template <class T>
 685 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
 686 #ifdef ASSERT
 687   {
 688     assert(!oopDesc::is_null(*p), "expected non-null ref");
 689     oop obj = oopDesc::load_decode_heap_oop_not_null(p);
 690     // We never expect to see a null reference being processed
 691     // as a weak reference.
 692     assert(obj->is_oop(), "expected an oop while scanning weak refs");
 693   }
 694 #endif // ASSERT
 695 
 696   _par_cl->do_oop_nv(p);
 697 
 698   if (Universe::heap()->is_in_reserved(p)) {
 699     oop obj = oopDesc::load_decode_heap_oop_not_null(p);
 700     _rs->write_ref_field_gc_par(p, obj);
 701   }
 702 }
 703 
 704 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p)       { ParKeepAliveClosure::do_oop_work(p); }
 705 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
 706 
 707 // ParNewGeneration::
 708 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
 709   DefNewGeneration::KeepAliveClosure(cl) {}
 710 
 711 template <class T>
 712 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
 713 #ifdef ASSERT
 714   {
 715     assert(!oopDesc::is_null(*p), "expected non-null ref");
 716     oop obj = oopDesc::load_decode_heap_oop_not_null(p);
 717     // We never expect to see a null reference being processed
 718     // as a weak reference.
 719     assert(obj->is_oop(), "expected an oop while scanning weak refs");
 720   }
 721 #endif // ASSERT
 722 
 723   _cl->do_oop_nv(p);
 724 
 725   if (Universe::heap()->is_in_reserved(p)) {
 726     oop obj = oopDesc::load_decode_heap_oop_not_null(p);
 727     _rs->write_ref_field_gc_par(p, obj);
 728   }
 729 }
 730 
 731 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p)       { KeepAliveClosure::do_oop_work(p); }
 732 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
 733 
 734 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
 735   T heap_oop = oopDesc::load_heap_oop(p);
 736   if (!oopDesc::is_null(heap_oop)) {
 737     oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
 738     if ((HeapWord*)obj < _boundary) {
 739       assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
 740       oop new_obj = obj->is_forwarded()
 741                       ? obj->forwardee()
 742                       : _g->DefNewGeneration::copy_to_survivor_space(obj);
 743       oopDesc::encode_store_heap_oop_not_null(p, new_obj);
 744     }
 745     if (_gc_barrier) {
 746       // If p points to a younger generation, mark the card.
 747       if ((HeapWord*)obj < _gen_boundary) {
 748         _rs->write_ref_field_gc_par(p, obj);
 749       }
 750     }
 751   }
 752 }
 753 
 754 void ScanClosureWithParBarrier::do_oop(oop* p)       { ScanClosureWithParBarrier::do_oop_work(p); }
 755 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
 756 
 757 class ParNewRefProcTaskProxy: public AbstractGangTask {
 758   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
 759 public:
 760   ParNewRefProcTaskProxy(ProcessTask& task,
 761                          ParNewGeneration& gen,
 762                          Generation& old_gen,
 763                          HeapWord* young_old_boundary,
 764                          ParScanThreadStateSet& state_set);
 765 
 766 private:
 767   virtual void work(uint worker_id);
 768   virtual void set_for_termination(int active_workers) {
 769     _state_set.terminator()->reset_for_reuse(active_workers);
 770   }
 771 private:
 772   ParNewGeneration&      _gen;
 773   ProcessTask&           _task;
 774   Generation&            _old_gen;
 775   HeapWord*              _young_old_boundary;
 776   ParScanThreadStateSet& _state_set;
 777 };
 778 
 779 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
 780                                                ParNewGeneration& gen,
 781                                                Generation& old_gen,
 782                                                HeapWord* young_old_boundary,
 783                                                ParScanThreadStateSet& state_set)
 784   : AbstractGangTask("ParNewGeneration parallel reference processing"),
 785     _gen(gen),
 786     _task(task),
 787     _old_gen(old_gen),
 788     _young_old_boundary(young_old_boundary),
 789     _state_set(state_set)
 790 {
 791 }
 792 
 793 void ParNewRefProcTaskProxy::work(uint worker_id)
 794 {
 795   ResourceMark rm;
 796   HandleMark hm;
 797   ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
 798   par_scan_state.set_young_old_boundary(_young_old_boundary);
 799   _task.work(worker_id, par_scan_state.is_alive_closure(),
 800              par_scan_state.keep_alive_closure(),
 801              par_scan_state.evacuate_followers_closure());
 802 }
 803 
 804 class ParNewRefEnqueueTaskProxy: public AbstractGangTask {
 805   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
 806   EnqueueTask& _task;
 807 
 808 public:
 809   ParNewRefEnqueueTaskProxy(EnqueueTask& task)
 810     : AbstractGangTask("ParNewGeneration parallel reference enqueue"),
 811       _task(task)
 812   { }
 813 
 814   virtual void work(uint worker_id)
 815   {
 816     _task.work(worker_id);
 817   }
 818 };
 819 
 820 
 821 void ParNewRefProcTaskExecutor::execute(ProcessTask& task)
 822 {
 823   GenCollectedHeap* gch = GenCollectedHeap::heap();
 824   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
 825          "not a generational heap");
 826   FlexibleWorkGang* workers = gch->workers();
 827   assert(workers != NULL, "Need parallel worker threads.");
 828   _state_set.reset(workers->active_workers(), _generation.promotion_failed());
 829   ParNewRefProcTaskProxy rp_task(task, _generation, *_generation.next_gen(),
 830                                  _generation.reserved().end(), _state_set);
 831   workers->run_task(&rp_task);
 832   _state_set.reset(0 /* bad value in debug if not reset */,
 833                    _generation.promotion_failed());
 834 }
 835 
 836 void ParNewRefProcTaskExecutor::execute(EnqueueTask& task)
 837 {
 838   GenCollectedHeap* gch = GenCollectedHeap::heap();
 839   FlexibleWorkGang* workers = gch->workers();
 840   assert(workers != NULL, "Need parallel worker threads.");
 841   ParNewRefEnqueueTaskProxy enq_task(task);
 842   workers->run_task(&enq_task);
 843 }
 844 
 845 void ParNewRefProcTaskExecutor::set_single_threaded_mode()
 846 {
 847   _state_set.flush();
 848   GenCollectedHeap* gch = GenCollectedHeap::heap();
 849   gch->set_par_threads(0);  // 0 ==> non-parallel.
 850   gch->save_marks();
 851 }
 852 
 853 ScanClosureWithParBarrier::
 854 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
 855   ScanClosure(g, gc_barrier) {}
 856 
 857 EvacuateFollowersClosureGeneral::
 858 EvacuateFollowersClosureGeneral(GenCollectedHeap* gch, int level,
 859                                 OopsInGenClosure* cur,
 860                                 OopsInGenClosure* older) :
 861   _gch(gch), _level(level),
 862   _scan_cur_or_nonheap(cur), _scan_older(older)
 863 {}
 864 
 865 void EvacuateFollowersClosureGeneral::do_void() {
 866   do {
 867     // Beware: this call will lead to closure applications via virtual
 868     // calls.
 869     _gch->oop_since_save_marks_iterate(_level,
 870                                        _scan_cur_or_nonheap,
 871                                        _scan_older);
 872   } while (!_gch->no_allocs_since_save_marks(_level));
 873 }
 874 
 875 
 876 // A Generation that does parallel young-gen collection.
 877 
 878 void ParNewGeneration::handle_promotion_failed(GenCollectedHeap* gch, ParScanThreadStateSet& thread_state_set) {
 879   assert(_promo_failure_scan_stack.is_empty(), "post condition");
 880   _promo_failure_scan_stack.clear(true); // Clear cached segments.
 881 
 882   remove_forwarding_pointers();
 883   if (PrintGCDetails) {
 884     gclog_or_tty->print(" (promotion failed)");
 885   }
 886   // All the spaces are in play for mark-sweep.
 887   swap_spaces();  // Make life simpler for CMS || rescan; see 6483690.
 888   from()->set_next_compaction_space(to());
 889   gch->set_incremental_collection_failed();
 890   // Inform the next generation that a promotion failure occurred.
 891   _old_gen->promotion_failure_occurred();
 892 
 893   // Trace promotion failure in the parallel GC threads
 894   thread_state_set.trace_promotion_failed(gc_tracer());
 895   // Single threaded code may have reported promotion failure to the global state
 896   if (_promotion_failed_info.has_failed()) {
 897     _gc_tracer.report_promotion_failed(_promotion_failed_info);
 898   }
 899   // Reset the PromotionFailureALot counters.
 900   NOT_PRODUCT(Universe::heap()->reset_promotion_should_fail();)
 901 }
 902 
 903 void ParNewGeneration::collect(bool   full,
 904                                bool   clear_all_soft_refs,
 905                                size_t size,
 906                                bool   is_tlab) {
 907   assert(full || size > 0, "otherwise we don't want to collect");
 908 
 909   GenCollectedHeap* gch = GenCollectedHeap::heap();
 910 
 911   _gc_timer->register_gc_start();
 912 
 913   assert(gch->kind() == CollectedHeap::GenCollectedHeap,
 914     "not a CMS generational heap");
 915   AdaptiveSizePolicy* size_policy = gch->gen_policy()->size_policy();
 916   FlexibleWorkGang* workers = gch->workers();
 917   assert(workers != NULL, "Need workgang for parallel work");
 918   int active_workers =
 919       AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
 920                                    workers->active_workers(),
 921                                    Threads::number_of_non_daemon_threads());
 922   workers->set_active_workers(active_workers);
 923   _old_gen = gch->old_gen();
 924 
 925   // If the next generation is too full to accommodate worst-case promotion
 926   // from this generation, pass on collection; let the next generation
 927   // do it.
 928   if (!collection_attempt_is_safe()) {
 929     gch->set_incremental_collection_failed();  // slight lie, in that we did not even attempt one
 930     return;
 931   }
 932   assert(to()->is_empty(), "Else not collection_attempt_is_safe");
 933 
 934   _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
 935   gch->trace_heap_before_gc(gc_tracer());
 936 
 937   init_assuming_no_promotion_failure();
 938 
 939   if (UseAdaptiveSizePolicy) {
 940     set_survivor_overflow(false);
 941     size_policy->minor_collection_begin();
 942   }
 943 
 944   GCTraceTime t1(GCCauseString("GC", gch->gc_cause()), PrintGC && !PrintGCDetails, true, NULL, _gc_tracer.gc_id());
 945   // Capture heap used before collection (for printing).
 946   size_t gch_prev_used = gch->used();
 947 
 948   age_table()->clear();
 949   to()->clear(SpaceDecorator::Mangle);
 950 
 951   gch->save_marks();
 952   assert(workers != NULL, "Need parallel worker threads.");
 953   int n_workers = active_workers;
 954 
 955   // Set the correct parallelism (number of queues) in the reference processor
 956   ref_processor()->set_active_mt_degree(n_workers);
 957 
 958   // Always set the terminator for the active number of workers
 959   // because only those workers go through the termination protocol.
 960   ParallelTaskTerminator _term(n_workers, task_queues());
 961   ParScanThreadStateSet thread_state_set(workers->active_workers(),
 962                                          *to(), *this, *_old_gen, *task_queues(),
 963                                          _overflow_stacks, desired_plab_sz(), _term);
 964 
 965   ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set);
 966   gch->set_par_threads(n_workers);
 967   gch->rem_set()->prepare_for_younger_refs_iterate(true);
 968   // It turns out that even when we're using 1 thread, doing the work in a
 969   // separate thread causes wide variance in run times.  We can't help this
 970   // in the multi-threaded case, but we special-case n=1 here to get
 971   // repeatable measurements of the 1-thread overhead of the parallel code.
 972   if (n_workers > 1) {
 973     GenCollectedHeap::StrongRootsScope srs(gch);
 974     workers->run_task(&tsk);
 975   } else {
 976     GenCollectedHeap::StrongRootsScope srs(gch);
 977     tsk.work(0);
 978   }
 979   thread_state_set.reset(0 /* Bad value in debug if not reset */,
 980                          promotion_failed());
 981 
 982   // Trace and reset failed promotion info.
 983   if (promotion_failed()) {
 984     thread_state_set.trace_promotion_failed(gc_tracer());
 985   }
 986 
 987   // Process (weak) reference objects found during scavenge.
 988   ReferenceProcessor* rp = ref_processor();
 989   IsAliveClosure is_alive(this);
 990   ScanWeakRefClosure scan_weak_ref(this);
 991   KeepAliveClosure keep_alive(&scan_weak_ref);
 992   ScanClosure               scan_without_gc_barrier(this, false);
 993   ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
 994   set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
 995   EvacuateFollowersClosureGeneral evacuate_followers(gch, _level,
 996     &scan_without_gc_barrier, &scan_with_gc_barrier);
 997   rp->setup_policy(clear_all_soft_refs);
 998   // Can  the mt_degree be set later (at run_task() time would be best)?
 999   rp->set_active_mt_degree(active_workers);
1000   ReferenceProcessorStats stats;
1001   if (rp->processing_is_mt()) {
1002     ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1003     stats = rp->process_discovered_references(&is_alive, &keep_alive,
1004                                               &evacuate_followers, &task_executor,
1005                                               _gc_timer, _gc_tracer.gc_id());
1006   } else {
1007     thread_state_set.flush();
1008     gch->set_par_threads(0);  // 0 ==> non-parallel.
1009     gch->save_marks();
1010     stats = rp->process_discovered_references(&is_alive, &keep_alive,
1011                                               &evacuate_followers, NULL,
1012                                               _gc_timer, _gc_tracer.gc_id());
1013   }
1014   _gc_tracer.report_gc_reference_stats(stats);
1015   if (!promotion_failed()) {
1016     // Swap the survivor spaces.
1017     eden()->clear(SpaceDecorator::Mangle);
1018     from()->clear(SpaceDecorator::Mangle);
1019     if (ZapUnusedHeapArea) {
1020       // This is now done here because of the piece-meal mangling which
1021       // can check for valid mangling at intermediate points in the
1022       // collection(s).  When a minor collection fails to collect
1023       // sufficient space resizing of the young generation can occur
1024       // an redistribute the spaces in the young generation.  Mangle
1025       // here so that unzapped regions don't get distributed to
1026       // other spaces.
1027       to()->mangle_unused_area();
1028     }
1029     swap_spaces();
1030 
1031     // A successful scavenge should restart the GC time limit count which is
1032     // for full GC's.
1033     size_policy->reset_gc_overhead_limit_count();
1034 
1035     assert(to()->is_empty(), "to space should be empty now");
1036 
1037     adjust_desired_tenuring_threshold();
1038   } else {
1039     handle_promotion_failed(gch, thread_state_set);
1040   }
1041   // set new iteration safe limit for the survivor spaces
1042   from()->set_concurrent_iteration_safe_limit(from()->top());
1043   to()->set_concurrent_iteration_safe_limit(to()->top());
1044 
1045   if (ResizePLAB) {
1046     plab_stats()->adjust_desired_plab_sz(n_workers);
1047   }
1048 
1049   if (PrintGC && !PrintGCDetails) {
1050     gch->print_heap_change(gch_prev_used);
1051   }
1052 
1053   TASKQUEUE_STATS_ONLY(if (PrintTerminationStats) thread_state_set.print_termination_stats());
1054   TASKQUEUE_STATS_ONLY(if (PrintTaskqueue) thread_state_set.print_taskqueue_stats());
1055 
1056   if (UseAdaptiveSizePolicy) {
1057     size_policy->minor_collection_end(gch->gc_cause());
1058     size_policy->avg_survived()->sample(from()->used());
1059   }
1060 
1061   // We need to use a monotonically non-decreasing time in ms
1062   // or we will see time-warp warnings and os::javaTimeMillis()
1063   // does not guarantee monotonicity.
1064   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1065   update_time_of_last_gc(now);
1066 
1067   rp->set_enqueuing_is_done(true);
1068   if (rp->processing_is_mt()) {
1069     ParNewRefProcTaskExecutor task_executor(*this, thread_state_set);
1070     rp->enqueue_discovered_references(&task_executor);
1071   } else {
1072     rp->enqueue_discovered_references(NULL);
1073   }
1074   rp->verify_no_references_recorded();
1075 
1076   gch->trace_heap_after_gc(gc_tracer());
1077   _gc_tracer.report_tenuring_threshold(tenuring_threshold());
1078 
1079   _gc_timer->register_gc_end();
1080 
1081   _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1082 }
1083 
1084 static int sum;
1085 void ParNewGeneration::waste_some_time() {
1086   for (int i = 0; i < 100; i++) {
1087     sum += i;
1088   }
1089 }
1090 
1091 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1092 
1093 // Because of concurrency, there are times where an object for which
1094 // "is_forwarded()" is true contains an "interim" forwarding pointer
1095 // value.  Such a value will soon be overwritten with a real value.
1096 // This method requires "obj" to have a forwarding pointer, and waits, if
1097 // necessary for a real one to be inserted, and returns it.
1098 
1099 oop ParNewGeneration::real_forwardee(oop obj) {
1100   oop forward_ptr = obj->forwardee();
1101   if (forward_ptr != ClaimedForwardPtr) {
1102     return forward_ptr;
1103   } else {
1104     return real_forwardee_slow(obj);
1105   }
1106 }
1107 
1108 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1109   // Spin-read if it is claimed but not yet written by another thread.
1110   oop forward_ptr = obj->forwardee();
1111   while (forward_ptr == ClaimedForwardPtr) {
1112     waste_some_time();
1113     assert(obj->is_forwarded(), "precondition");
1114     forward_ptr = obj->forwardee();
1115   }
1116   return forward_ptr;
1117 }
1118 
1119 void ParNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
1120   if (m->must_be_preserved_for_promotion_failure(obj)) {
1121     // We should really have separate per-worker stacks, rather
1122     // than use locking of a common pair of stacks.
1123     MutexLocker ml(ParGCRareEvent_lock);
1124     preserve_mark(obj, m);
1125   }
1126 }
1127 
1128 // Multiple GC threads may try to promote an object.  If the object
1129 // is successfully promoted, a forwarding pointer will be installed in
1130 // the object in the young generation.  This method claims the right
1131 // to install the forwarding pointer before it copies the object,
1132 // thus avoiding the need to undo the copy as in
1133 // copy_to_survivor_space_avoiding_with_undo.
1134 
1135 oop ParNewGeneration::copy_to_survivor_space(
1136         ParScanThreadState* par_scan_state, oop old, size_t sz, markOop m) {
1137   // In the sequential version, this assert also says that the object is
1138   // not forwarded.  That might not be the case here.  It is the case that
1139   // the caller observed it to be not forwarded at some time in the past.
1140   assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1141 
1142   // The sequential code read "old->age()" below.  That doesn't work here,
1143   // since the age is in the mark word, and that might be overwritten with
1144   // a forwarding pointer by a parallel thread.  So we must save the mark
1145   // word in a local and then analyze it.
1146   oopDesc dummyOld;
1147   dummyOld.set_mark(m);
1148   assert(!dummyOld.is_forwarded(),
1149          "should not be called with forwarding pointer mark word.");
1150 
1151   oop new_obj = NULL;
1152   oop forward_ptr;
1153 
1154   // Try allocating obj in to-space (unless too old)
1155   if (dummyOld.age() < tenuring_threshold()) {
1156     new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1157     if (new_obj == NULL) {
1158       set_survivor_overflow(true);
1159     }
1160   }
1161 
1162   if (new_obj == NULL) {
1163     // Either to-space is full or we decided to promote
1164     // try allocating obj tenured
1165 
1166     // Attempt to install a null forwarding pointer (atomically),
1167     // to claim the right to install the real forwarding pointer.
1168     forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1169     if (forward_ptr != NULL) {
1170       // someone else beat us to it.
1171         return real_forwardee(old);
1172     }
1173 
1174     if (!_promotion_failed) {
1175       new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1176                                       old, m, sz);
1177     }
1178 
1179     if (new_obj == NULL) {
1180       // promotion failed, forward to self
1181       _promotion_failed = true;
1182       new_obj = old;
1183 
1184       preserve_mark_if_necessary(old, m);
1185       par_scan_state->register_promotion_failure(sz);
1186     }
1187 
1188     old->forward_to(new_obj);
1189     forward_ptr = NULL;
1190   } else {
1191     // Is in to-space; do copying ourselves.
1192     Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1193     assert(Universe::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1194     forward_ptr = old->forward_to_atomic(new_obj);
1195     // Restore the mark word copied above.
1196     new_obj->set_mark(m);
1197     // Increment age if obj still in new generation
1198     new_obj->incr_age();
1199     par_scan_state->age_table()->add(new_obj, sz);
1200   }
1201   assert(new_obj != NULL, "just checking");
1202 
1203 #ifndef PRODUCT
1204   // This code must come after the CAS test, or it will print incorrect
1205   // information.
1206   if (TraceScavenge) {
1207     gclog_or_tty->print_cr("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1208        is_in_reserved(new_obj) ? "copying" : "tenuring",
1209        new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1210   }
1211 #endif
1212 
1213   if (forward_ptr == NULL) {
1214     oop obj_to_push = new_obj;
1215     if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1216       // Length field used as index of next element to be scanned.
1217       // Real length can be obtained from real_forwardee()
1218       arrayOop(old)->set_length(0);
1219       obj_to_push = old;
1220       assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1221              "push forwarded object");
1222     }
1223     // Push it on one of the queues of to-be-scanned objects.
1224     bool simulate_overflow = false;
1225     NOT_PRODUCT(
1226       if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1227         // simulate a stack overflow
1228         simulate_overflow = true;
1229       }
1230     )
1231     if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1232       // Add stats for overflow pushes.
1233       if (Verbose && PrintGCDetails) {
1234         gclog_or_tty->print("queue overflow!\n");
1235       }
1236       push_on_overflow_list(old, par_scan_state);
1237       TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1238     }
1239 
1240     return new_obj;
1241   }
1242 
1243   // Oops.  Someone beat us to it.  Undo the allocation.  Where did we
1244   // allocate it?
1245   if (is_in_reserved(new_obj)) {
1246     // Must be in to_space.
1247     assert(to()->is_in_reserved(new_obj), "Checking");
1248     if (forward_ptr == ClaimedForwardPtr) {
1249       // Wait to get the real forwarding pointer value.
1250       forward_ptr = real_forwardee(old);
1251     }
1252     par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1253   }
1254 
1255   return forward_ptr;
1256 }
1257 
1258 #ifndef PRODUCT
1259 // It's OK to call this multi-threaded;  the worst thing
1260 // that can happen is that we'll get a bunch of closely
1261 // spaced simulated overflows, but that's OK, in fact
1262 // probably good as it would exercise the overflow code
1263 // under contention.
1264 bool ParNewGeneration::should_simulate_overflow() {
1265   if (_overflow_counter-- <= 0) { // just being defensive
1266     _overflow_counter = ParGCWorkQueueOverflowInterval;
1267     return true;
1268   } else {
1269     return false;
1270   }
1271 }
1272 #endif
1273 
1274 // In case we are using compressed oops, we need to be careful.
1275 // If the object being pushed is an object array, then its length
1276 // field keeps track of the "grey boundary" at which the next
1277 // incremental scan will be done (see ParGCArrayScanChunk).
1278 // When using compressed oops, this length field is kept in the
1279 // lower 32 bits of the erstwhile klass word and cannot be used
1280 // for the overflow chaining pointer (OCP below). As such the OCP
1281 // would itself need to be compressed into the top 32-bits in this
1282 // case. Unfortunately, see below, in the event that we have a
1283 // promotion failure, the node to be pushed on the list can be
1284 // outside of the Java heap, so the heap-based pointer compression
1285 // would not work (we would have potential aliasing between C-heap
1286 // and Java-heap pointers). For this reason, when using compressed
1287 // oops, we simply use a worker-thread-local, non-shared overflow
1288 // list in the form of a growable array, with a slightly different
1289 // overflow stack draining strategy. If/when we start using fat
1290 // stacks here, we can go back to using (fat) pointer chains
1291 // (although some performance comparisons would be useful since
1292 // single global lists have their own performance disadvantages
1293 // as we were made painfully aware not long ago, see 6786503).
1294 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1295 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1296   assert(is_in_reserved(from_space_obj), "Should be from this generation");
1297   if (ParGCUseLocalOverflow) {
1298     // In the case of compressed oops, we use a private, not-shared
1299     // overflow stack.
1300     par_scan_state->push_on_overflow_stack(from_space_obj);
1301   } else {
1302     assert(!UseCompressedOops, "Error");
1303     // if the object has been forwarded to itself, then we cannot
1304     // use the klass pointer for the linked list.  Instead we have
1305     // to allocate an oopDesc in the C-Heap and use that for the linked list.
1306     // XXX This is horribly inefficient when a promotion failure occurs
1307     // and should be fixed. XXX FIX ME !!!
1308 #ifndef PRODUCT
1309     Atomic::inc_ptr(&_num_par_pushes);
1310     assert(_num_par_pushes > 0, "Tautology");
1311 #endif
1312     if (from_space_obj->forwardee() == from_space_obj) {
1313       oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1314       listhead->forward_to(from_space_obj);
1315       from_space_obj = listhead;
1316     }
1317     oop observed_overflow_list = _overflow_list;
1318     oop cur_overflow_list;
1319     do {
1320       cur_overflow_list = observed_overflow_list;
1321       if (cur_overflow_list != BUSY) {
1322         from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1323       } else {
1324         from_space_obj->set_klass_to_list_ptr(NULL);
1325       }
1326       observed_overflow_list =
1327         (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
1328     } while (cur_overflow_list != observed_overflow_list);
1329   }
1330 }
1331 
1332 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1333   bool res;
1334 
1335   if (ParGCUseLocalOverflow) {
1336     res = par_scan_state->take_from_overflow_stack();
1337   } else {
1338     assert(!UseCompressedOops, "Error");
1339     res = take_from_overflow_list_work(par_scan_state);
1340   }
1341   return res;
1342 }
1343 
1344 
1345 // *NOTE*: The overflow list manipulation code here and
1346 // in CMSCollector:: are very similar in shape,
1347 // except that in the CMS case we thread the objects
1348 // directly into the list via their mark word, and do
1349 // not need to deal with special cases below related
1350 // to chunking of object arrays and promotion failure
1351 // handling.
1352 // CR 6797058 has been filed to attempt consolidation of
1353 // the common code.
1354 // Because of the common code, if you make any changes in
1355 // the code below, please check the CMS version to see if
1356 // similar changes might be needed.
1357 // See CMSCollector::par_take_from_overflow_list() for
1358 // more extensive documentation comments.
1359 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1360   ObjToScanQueue* work_q = par_scan_state->work_queue();
1361   // How many to take?
1362   size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1363                                  (size_t)ParGCDesiredObjsFromOverflowList);
1364 
1365   assert(!UseCompressedOops, "Error");
1366   assert(par_scan_state->overflow_stack() == NULL, "Error");
1367   if (_overflow_list == NULL) return false;
1368 
1369   // Otherwise, there was something there; try claiming the list.
1370   oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1371   // Trim off a prefix of at most objsFromOverflow items
1372   Thread* tid = Thread::current();
1373   size_t spin_count = (size_t)ParallelGCThreads;
1374   size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1375   for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1376     // someone grabbed it before we did ...
1377     // ... we spin for a short while...
1378     os::sleep(tid, sleep_time_millis, false);
1379     if (_overflow_list == NULL) {
1380       // nothing left to take
1381       return false;
1382     } else if (_overflow_list != BUSY) {
1383      // try and grab the prefix
1384      prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list));
1385     }
1386   }
1387   if (prefix == NULL || prefix == BUSY) {
1388      // Nothing to take or waited long enough
1389      if (prefix == NULL) {
1390        // Write back the NULL in case we overwrote it with BUSY above
1391        // and it is still the same value.
1392        (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1393      }
1394      return false;
1395   }
1396   assert(prefix != NULL && prefix != BUSY, "Error");
1397   size_t i = 1;
1398   oop cur = prefix;
1399   while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
1400     i++; cur = cur->list_ptr_from_klass();
1401   }
1402 
1403   // Reattach remaining (suffix) to overflow list
1404   if (cur->klass_or_null() == NULL) {
1405     // Write back the NULL in lieu of the BUSY we wrote
1406     // above and it is still the same value.
1407     if (_overflow_list == BUSY) {
1408       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
1409     }
1410   } else {
1411     assert(cur->klass_or_null() != (Klass*)(address)BUSY, "Error");
1412     oop suffix = cur->list_ptr_from_klass();       // suffix will be put back on global list
1413     cur->set_klass_to_list_ptr(NULL);     // break off suffix
1414     // It's possible that the list is still in the empty(busy) state
1415     // we left it in a short while ago; in that case we may be
1416     // able to place back the suffix.
1417     oop observed_overflow_list = _overflow_list;
1418     oop cur_overflow_list = observed_overflow_list;
1419     bool attached = false;
1420     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1421       observed_overflow_list =
1422         (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1423       if (cur_overflow_list == observed_overflow_list) {
1424         attached = true;
1425         break;
1426       } else cur_overflow_list = observed_overflow_list;
1427     }
1428     if (!attached) {
1429       // Too bad, someone else got in in between; we'll need to do a splice.
1430       // Find the last item of suffix list
1431       oop last = suffix;
1432       while (last->klass_or_null() != NULL) {
1433         last = last->list_ptr_from_klass();
1434       }
1435       // Atomically prepend suffix to current overflow list
1436       observed_overflow_list = _overflow_list;
1437       do {
1438         cur_overflow_list = observed_overflow_list;
1439         if (cur_overflow_list != BUSY) {
1440           // Do the splice ...
1441           last->set_klass_to_list_ptr(cur_overflow_list);
1442         } else { // cur_overflow_list == BUSY
1443           last->set_klass_to_list_ptr(NULL);
1444         }
1445         observed_overflow_list =
1446           (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
1447       } while (cur_overflow_list != observed_overflow_list);
1448     }
1449   }
1450 
1451   // Push objects on prefix list onto this thread's work queue
1452   assert(prefix != NULL && prefix != BUSY, "program logic");
1453   cur = prefix;
1454   ssize_t n = 0;
1455   while (cur != NULL) {
1456     oop obj_to_push = cur->forwardee();
1457     oop next        = cur->list_ptr_from_klass();
1458     cur->set_klass(obj_to_push->klass());
1459     // This may be an array object that is self-forwarded. In that case, the list pointer
1460     // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1461     if (!is_in_reserved(cur)) {
1462       // This can become a scaling bottleneck when there is work queue overflow coincident
1463       // with promotion failure.
1464       oopDesc* f = cur;
1465       FREE_C_HEAP_ARRAY(oopDesc, f);
1466     } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1467       assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1468       obj_to_push = cur;
1469     }
1470     bool ok = work_q->push(obj_to_push);
1471     assert(ok, "Should have succeeded");
1472     cur = next;
1473     n++;
1474   }
1475   TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1476 #ifndef PRODUCT
1477   assert(_num_par_pushes >= n, "Too many pops?");
1478   Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
1479 #endif
1480   return true;
1481 }
1482 #undef BUSY
1483 
1484 void ParNewGeneration::ref_processor_init() {
1485   if (_ref_processor == NULL) {
1486     // Allocate and initialize a reference processor
1487     _ref_processor =
1488       new ReferenceProcessor(_reserved,                  // span
1489                              ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1490                              (int) ParallelGCThreads,    // mt processing degree
1491                              refs_discovery_is_mt(),     // mt discovery
1492                              (int) ParallelGCThreads,    // mt discovery degree
1493                              refs_discovery_is_atomic(), // atomic_discovery
1494                              NULL);                      // is_alive_non_header
1495   }
1496 }
1497 
1498 const char* ParNewGeneration::name() const {
1499   return "par new generation";
1500 }