1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/stringTable.hpp"
  27 #include "gc/cms/cmsHeap.inline.hpp"
  28 #include "gc/cms/compactibleFreeListSpace.hpp"
  29 #include "gc/cms/concurrentMarkSweepGeneration.hpp"
  30 #include "gc/cms/parNewGeneration.inline.hpp"
  31 #include "gc/cms/parOopClosures.inline.hpp"
  32 #include "gc/serial/defNewGeneration.inline.hpp"
  33 #include "gc/shared/adaptiveSizePolicy.hpp"
  34 #include "gc/shared/ageTable.inline.hpp"
  35 #include "gc/shared/copyFailedInfo.hpp"
  36 #include "gc/shared/gcHeapSummary.hpp"
  37 #include "gc/shared/gcTimer.hpp"
  38 #include "gc/shared/gcTrace.hpp"
  39 #include "gc/shared/gcTraceTime.inline.hpp"
  40 #include "gc/shared/genOopClosures.inline.hpp"
  41 #include "gc/shared/generation.hpp"
  42 #include "gc/shared/plab.inline.hpp"
  43 #include "gc/shared/preservedMarks.inline.hpp"
  44 #include "gc/shared/referencePolicy.hpp"
  45 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
  46 #include "gc/shared/space.hpp"
  47 #include "gc/shared/spaceDecorator.hpp"
  48 #include "gc/shared/strongRootsScope.hpp"
  49 #include "gc/shared/taskqueue.inline.hpp"
  50 #include "gc/shared/weakProcessor.hpp"
  51 #include "gc/shared/workgroup.hpp"
  52 #include "logging/log.hpp"
  53 #include "logging/logStream.hpp"
  54 #include "memory/iterator.inline.hpp"
  55 #include "memory/resourceArea.hpp"
  56 #include "oops/access.inline.hpp"
  57 #include "oops/compressedOops.inline.hpp"
  58 #include "oops/objArrayOop.hpp"
  59 #include "oops/oop.inline.hpp"
  60 #include "runtime/atomic.hpp"
  61 #include "runtime/handles.hpp"
  62 #include "runtime/handles.inline.hpp"
  63 #include "runtime/java.hpp"
  64 #include "runtime/thread.inline.hpp"
  65 #include "utilities/copy.hpp"
  66 #include "utilities/globalDefinitions.hpp"
  67 #include "utilities/stack.inline.hpp"
  68 
  69 ParScanThreadState::ParScanThreadState(Space* to_space_,
  70                                        ParNewGeneration* young_gen_,
  71                                        Generation* old_gen_,
  72                                        int thread_num_,
  73                                        ObjToScanQueueSet* work_queue_set_,
  74                                        Stack<oop, mtGC>* overflow_stacks_,
  75                                        PreservedMarks* preserved_marks_,
  76                                        size_t desired_plab_sz_,
  77                                        ParallelTaskTerminator& term_) :
  78   _to_space(to_space_),
  79   _old_gen(old_gen_),
  80   _young_gen(young_gen_),
  81   _thread_num(thread_num_),
  82   _work_queue(work_queue_set_->queue(thread_num_)),
  83   _to_space_full(false),
  84   _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
  85   _preserved_marks(preserved_marks_),
  86   _ageTable(false), // false ==> not the global age table, no perf data.
  87   _to_space_alloc_buffer(desired_plab_sz_),
  88   _to_space_closure(young_gen_, this),
  89   _old_gen_closure(young_gen_, this),
  90   _to_space_root_closure(young_gen_, this),
  91   _old_gen_root_closure(young_gen_, this),
  92   _older_gen_closure(young_gen_, this),
  93   _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
  94                       &_to_space_root_closure, young_gen_, &_old_gen_root_closure,
  95                       work_queue_set_, &term_),
  96   _is_alive_closure(young_gen_),
  97   _scan_weak_ref_closure(young_gen_, this),
  98   _keep_alive_closure(&_scan_weak_ref_closure),
  99   _strong_roots_time(0.0),
 100   _term_time(0.0)
 101 {
 102   #if TASKQUEUE_STATS
 103   _term_attempts = 0;
 104   _overflow_refills = 0;
 105   _overflow_refill_objs = 0;
 106   #endif // TASKQUEUE_STATS
 107 
 108   _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
 109   _hash_seed = 17;  // Might want to take time-based random value.
 110   _start = os::elapsedTime();
 111   _old_gen_closure.set_generation(old_gen_);
 112   _old_gen_root_closure.set_generation(old_gen_);
 113 }
 114 
 115 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
 116                                               size_t plab_word_size) {
 117   ChunkArray* sca = survivor_chunk_array();
 118   if (sca != NULL) {
 119     // A non-null SCA implies that we want the PLAB data recorded.
 120     sca->record_sample(plab_start, plab_word_size);
 121   }
 122 }
 123 
 124 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
 125   return new_obj->is_objArray() &&
 126          arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
 127          new_obj != old_obj;
 128 }
 129 
 130 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
 131   assert(old->is_objArray(), "must be obj array");
 132   assert(old->is_forwarded(), "must be forwarded");
 133   assert(CMSHeap::heap()->is_in_reserved(old), "must be in heap.");
 134   assert(!old_gen()->is_in(old), "must be in young generation.");
 135 
 136   objArrayOop obj = objArrayOop(old->forwardee());
 137   // Process ParGCArrayScanChunk elements now
 138   // and push the remainder back onto queue
 139   int start     = arrayOop(old)->length();
 140   int end       = obj->length();
 141   int remainder = end - start;
 142   assert(start <= end, "just checking");
 143   if (remainder > 2 * ParGCArrayScanChunk) {
 144     // Test above combines last partial chunk with a full chunk
 145     end = start + ParGCArrayScanChunk;
 146     arrayOop(old)->set_length(end);
 147     // Push remainder.
 148     bool ok = work_queue()->push(old);
 149     assert(ok, "just popped, push must be okay");
 150   } else {
 151     // Restore length so that it can be used if there
 152     // is a promotion failure and forwarding pointers
 153     // must be removed.
 154     arrayOop(old)->set_length(end);
 155   }
 156 
 157   // process our set of indices (include header in first chunk)
 158   // should make sure end is even (aligned to HeapWord in case of compressed oops)
 159   if ((HeapWord *)obj < young_old_boundary()) {
 160     // object is in to_space
 161     obj->oop_iterate_range(&_to_space_closure, start, end);
 162   } else {
 163     // object is in old generation
 164     obj->oop_iterate_range(&_old_gen_closure, start, end);
 165   }
 166 }
 167 
 168 void ParScanThreadState::trim_queues(int max_size) {
 169   ObjToScanQueue* queue = work_queue();
 170   do {
 171     while (queue->size() > (juint)max_size) {
 172       oop obj_to_scan;
 173       if (queue->pop_local(obj_to_scan)) {
 174         if ((HeapWord *)obj_to_scan < young_old_boundary()) {
 175           if (obj_to_scan->is_objArray() &&
 176               obj_to_scan->is_forwarded() &&
 177               obj_to_scan->forwardee() != obj_to_scan) {
 178             scan_partial_array_and_push_remainder(obj_to_scan);
 179           } else {
 180             // object is in to_space
 181             obj_to_scan->oop_iterate(&_to_space_closure);
 182           }
 183         } else {
 184           // object is in old generation
 185           obj_to_scan->oop_iterate(&_old_gen_closure);
 186         }
 187       }
 188     }
 189     // For the  case of compressed oops, we have a private, non-shared
 190     // overflow stack, so we eagerly drain it so as to more evenly
 191     // distribute load early. Note: this may be good to do in
 192     // general rather than delay for the final stealing phase.
 193     // If applicable, we'll transfer a set of objects over to our
 194     // work queue, allowing them to be stolen and draining our
 195     // private overflow stack.
 196   } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
 197 }
 198 
 199 bool ParScanThreadState::take_from_overflow_stack() {
 200   assert(ParGCUseLocalOverflow, "Else should not call");
 201   assert(young_gen()->overflow_list() == NULL, "Error");
 202   ObjToScanQueue* queue = work_queue();
 203   Stack<oop, mtGC>* const of_stack = overflow_stack();
 204   const size_t num_overflow_elems = of_stack->size();
 205   const size_t space_available = queue->max_elems() - queue->size();
 206   const size_t num_take_elems = MIN3(space_available / 4,
 207                                      ParGCDesiredObjsFromOverflowList,
 208                                      num_overflow_elems);
 209   // Transfer the most recent num_take_elems from the overflow
 210   // stack to our work queue.
 211   for (size_t i = 0; i != num_take_elems; i++) {
 212     oop cur = of_stack->pop();
 213     oop obj_to_push = cur->forwardee();
 214     assert(CMSHeap::heap()->is_in_reserved(cur), "Should be in heap");
 215     assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
 216     assert(CMSHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
 217     if (should_be_partially_scanned(obj_to_push, cur)) {
 218       assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
 219       obj_to_push = cur;
 220     }
 221     bool ok = queue->push(obj_to_push);
 222     assert(ok, "Should have succeeded");
 223   }
 224   assert(young_gen()->overflow_list() == NULL, "Error");
 225   return num_take_elems > 0;  // was something transferred?
 226 }
 227 
 228 void ParScanThreadState::push_on_overflow_stack(oop p) {
 229   assert(ParGCUseLocalOverflow, "Else should not call");
 230   overflow_stack()->push(p);
 231   assert(young_gen()->overflow_list() == NULL, "Error");
 232 }
 233 
 234 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
 235   // If the object is small enough, try to reallocate the buffer.
 236   HeapWord* obj = NULL;
 237   if (!_to_space_full) {
 238     PLAB* const plab = to_space_alloc_buffer();
 239     Space* const sp  = to_space();
 240     if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
 241       // Is small enough; abandon this buffer and start a new one.
 242       plab->retire();
 243       // The minimum size has to be twice SurvivorAlignmentInBytes to
 244       // allow for padding used in the alignment of 1 word.  A padding
 245       // of 1 is too small for a filler word so the padding size will
 246       // be increased by SurvivorAlignmentInBytes.
 247       size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize);
 248       size_t buf_size = MAX2(plab->word_sz(), min_usable_size);
 249       HeapWord* buf_space = sp->par_allocate(buf_size);
 250       if (buf_space == NULL) {
 251         const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize;
 252         size_t free_bytes = sp->free();
 253         while(buf_space == NULL && free_bytes >= min_bytes) {
 254           buf_size = free_bytes >> LogHeapWordSize;
 255           assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
 256           buf_space  = sp->par_allocate(buf_size);
 257           free_bytes = sp->free();
 258         }
 259       }
 260       if (buf_space != NULL) {
 261         plab->set_buf(buf_space, buf_size);
 262         record_survivor_plab(buf_space, buf_size);
 263         obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
 264         // Note that we cannot compare buf_size < word_sz below
 265         // because of AlignmentReserve (see PLAB::allocate()).
 266         assert(obj != NULL || plab->words_remaining() < word_sz,
 267                "Else should have been able to allocate requested object size "
 268                SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes "
 269                SIZE_FORMAT ", words_remaining " SIZE_FORMAT,
 270                word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining());
 271         // It's conceivable that we may be able to use the
 272         // buffer we just grabbed for subsequent small requests
 273         // even if not for this one.
 274       } else {
 275         // We're used up.
 276         _to_space_full = true;
 277       }
 278     } else {
 279       // Too large; allocate the object individually.
 280       obj = sp->par_allocate(word_sz);
 281     }
 282   }
 283   return obj;
 284 }
 285 
 286 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
 287   to_space_alloc_buffer()->undo_allocation(obj, word_sz);
 288 }
 289 
 290 void ParScanThreadState::print_promotion_failure_size() {
 291   if (_promotion_failed_info.has_failed()) {
 292     log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ",
 293                              _thread_num, _promotion_failed_info.first_size());
 294   }
 295 }
 296 
 297 class ParScanThreadStateSet: StackObj {
 298 public:
 299   // Initializes states for the specified number of threads;
 300   ParScanThreadStateSet(int                     num_threads,
 301                         Space&                  to_space,
 302                         ParNewGeneration&       young_gen,
 303                         Generation&             old_gen,
 304                         ObjToScanQueueSet&      queue_set,
 305                         Stack<oop, mtGC>*       overflow_stacks_,
 306                         PreservedMarksSet&      preserved_marks_set,
 307                         size_t                  desired_plab_sz,
 308                         ParallelTaskTerminator& term);
 309 
 310   ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
 311 
 312   inline ParScanThreadState& thread_state(int i);
 313 
 314   void trace_promotion_failed(const YoungGCTracer* gc_tracer);
 315   void reset(uint active_workers, bool promotion_failed);
 316   void flush();
 317 
 318   #if TASKQUEUE_STATS
 319   static void
 320     print_termination_stats_hdr(outputStream* const st);
 321   void print_termination_stats();
 322   static void
 323     print_taskqueue_stats_hdr(outputStream* const st);
 324   void print_taskqueue_stats();
 325   void reset_stats();
 326   #endif // TASKQUEUE_STATS
 327 
 328 private:
 329   ParallelTaskTerminator& _term;
 330   ParNewGeneration&       _young_gen;
 331   Generation&             _old_gen;
 332   ParScanThreadState*     _per_thread_states;
 333   const int               _num_threads;
 334  public:
 335   bool is_valid(int id) const { return id < _num_threads; }
 336   ParallelTaskTerminator* terminator() { return &_term; }
 337 };
 338 
 339 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
 340                                              Space& to_space,
 341                                              ParNewGeneration& young_gen,
 342                                              Generation& old_gen,
 343                                              ObjToScanQueueSet& queue_set,
 344                                              Stack<oop, mtGC>* overflow_stacks,
 345                                              PreservedMarksSet& preserved_marks_set,
 346                                              size_t desired_plab_sz,
 347                                              ParallelTaskTerminator& term)
 348   : _young_gen(young_gen),
 349     _old_gen(old_gen),
 350     _term(term),
 351     _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)),
 352     _num_threads(num_threads)
 353 {
 354   assert(num_threads > 0, "sanity check!");
 355   assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
 356          "overflow_stack allocation mismatch");
 357   // Initialize states.
 358   for (int i = 0; i < num_threads; ++i) {
 359     new(_per_thread_states + i)
 360       ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
 361                          overflow_stacks, preserved_marks_set.get(i),
 362                          desired_plab_sz, term);
 363   }
 364 }
 365 
 366 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
 367   assert(i >= 0 && i < _num_threads, "sanity check!");
 368   return _per_thread_states[i];
 369 }
 370 
 371 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
 372   for (int i = 0; i < _num_threads; ++i) {
 373     if (thread_state(i).promotion_failed()) {
 374       gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
 375       thread_state(i).promotion_failed_info().reset();
 376     }
 377   }
 378 }
 379 
 380 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
 381   _term.reset_for_reuse(active_threads);
 382   if (promotion_failed) {
 383     for (int i = 0; i < _num_threads; ++i) {
 384       thread_state(i).print_promotion_failure_size();
 385     }
 386   }
 387 }
 388 
 389 #if TASKQUEUE_STATS
 390 void ParScanThreadState::reset_stats() {
 391   taskqueue_stats().reset();
 392   _term_attempts = 0;
 393   _overflow_refills = 0;
 394   _overflow_refill_objs = 0;
 395 }
 396 
 397 void ParScanThreadStateSet::reset_stats() {
 398   for (int i = 0; i < _num_threads; ++i) {
 399     thread_state(i).reset_stats();
 400   }
 401 }
 402 
 403 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
 404   st->print_raw_cr("GC Termination Stats");
 405   st->print_raw_cr("     elapsed  --strong roots-- -------termination-------");
 406   st->print_raw_cr("thr     ms        ms       %       ms       %   attempts");
 407   st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
 408 }
 409 
 410 void ParScanThreadStateSet::print_termination_stats() {
 411   Log(gc, task, stats) log;
 412   if (!log.is_debug()) {
 413     return;
 414   }
 415 
 416   ResourceMark rm;
 417   LogStream ls(log.debug());
 418   outputStream* st = &ls;
 419 
 420   print_termination_stats_hdr(st);
 421 
 422   for (int i = 0; i < _num_threads; ++i) {
 423     const ParScanThreadState & pss = thread_state(i);
 424     const double elapsed_ms = pss.elapsed_time() * 1000.0;
 425     const double s_roots_ms = pss.strong_roots_time() * 1000.0;
 426     const double term_ms = pss.term_time() * 1000.0;
 427     st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
 428                  i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
 429                  term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
 430   }
 431 }
 432 
 433 // Print stats related to work queue activity.
 434 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
 435   st->print_raw_cr("GC Task Stats");
 436   st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
 437   st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
 438 }
 439 
 440 void ParScanThreadStateSet::print_taskqueue_stats() {
 441   if (!log_develop_is_enabled(Trace, gc, task, stats)) {
 442     return;
 443   }
 444   Log(gc, task, stats) log;
 445   ResourceMark rm;
 446   LogStream ls(log.trace());
 447   outputStream* st = &ls;
 448   print_taskqueue_stats_hdr(st);
 449 
 450   TaskQueueStats totals;
 451   for (int i = 0; i < _num_threads; ++i) {
 452     const ParScanThreadState & pss = thread_state(i);
 453     const TaskQueueStats & stats = pss.taskqueue_stats();
 454     st->print("%3d ", i); stats.print(st); st->cr();
 455     totals += stats;
 456 
 457     if (pss.overflow_refills() > 0) {
 458       st->print_cr("    " SIZE_FORMAT_W(10) " overflow refills    "
 459                    SIZE_FORMAT_W(10) " overflow objects",
 460                    pss.overflow_refills(), pss.overflow_refill_objs());
 461     }
 462   }
 463   st->print("tot "); totals.print(st); st->cr();
 464 
 465   DEBUG_ONLY(totals.verify());
 466 }
 467 #endif // TASKQUEUE_STATS
 468 
 469 void ParScanThreadStateSet::flush() {
 470   // Work in this loop should be kept as lightweight as
 471   // possible since this might otherwise become a bottleneck
 472   // to scaling. Should we add heavy-weight work into this
 473   // loop, consider parallelizing the loop into the worker threads.
 474   for (int i = 0; i < _num_threads; ++i) {
 475     ParScanThreadState& par_scan_state = thread_state(i);
 476 
 477     // Flush stats related to To-space PLAB activity and
 478     // retire the last buffer.
 479     par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
 480 
 481     // Every thread has its own age table.  We need to merge
 482     // them all into one.
 483     AgeTable *local_table = par_scan_state.age_table();
 484     _young_gen.age_table()->merge(local_table);
 485 
 486     // Inform old gen that we're done.
 487     _old_gen.par_promote_alloc_done(i);
 488   }
 489 
 490   if (UseConcMarkSweepGC) {
 491     // We need to call this even when ResizeOldPLAB is disabled
 492     // so as to avoid breaking some asserts. While we may be able
 493     // to avoid this by reorganizing the code a bit, I am loathe
 494     // to do that unless we find cases where ergo leads to bad
 495     // performance.
 496     CompactibleFreeListSpaceLAB::compute_desired_plab_size();
 497   }
 498 }
 499 
 500 ParScanClosure::ParScanClosure(ParNewGeneration* g,
 501                                ParScanThreadState* par_scan_state) :
 502   OopsInClassLoaderDataOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
 503   _boundary = _g->reserved().end();
 504 }
 505 
 506 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, true, true); }
 507 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
 508 
 509 void ParRootScanWithoutBarrierClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, false, true); }
 510 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
 511 
 512 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
 513                                              ParScanThreadState* par_scan_state)
 514   : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
 515 {}
 516 
 517 #ifdef WIN32
 518 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
 519 #endif
 520 
 521 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
 522     ParScanThreadState* par_scan_state_,
 523     ParScanWithoutBarrierClosure* to_space_closure_,
 524     ParScanWithBarrierClosure* old_gen_closure_,
 525     ParRootScanWithoutBarrierClosure* to_space_root_closure_,
 526     ParNewGeneration* par_gen_,
 527     ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
 528     ObjToScanQueueSet* task_queues_,
 529     ParallelTaskTerminator* terminator_) :
 530 
 531     _par_scan_state(par_scan_state_),
 532     _to_space_closure(to_space_closure_),
 533     _old_gen_closure(old_gen_closure_),
 534     _to_space_root_closure(to_space_root_closure_),
 535     _old_gen_root_closure(old_gen_root_closure_),
 536     _par_gen(par_gen_),
 537     _task_queues(task_queues_),
 538     _terminator(terminator_)
 539 {}
 540 
 541 void ParEvacuateFollowersClosure::do_void() {
 542   ObjToScanQueue* work_q = par_scan_state()->work_queue();
 543 
 544   while (true) {
 545     // Scan to-space and old-gen objs until we run out of both.
 546     oop obj_to_scan;
 547     par_scan_state()->trim_queues(0);
 548 
 549     // We have no local work, attempt to steal from other threads.
 550 
 551     // Attempt to steal work from promoted.
 552     if (task_queues()->steal(par_scan_state()->thread_num(),
 553                              par_scan_state()->hash_seed(),
 554                              obj_to_scan)) {
 555       bool res = work_q->push(obj_to_scan);
 556       assert(res, "Empty queue should have room for a push.");
 557 
 558       // If successful, goto Start.
 559       continue;
 560 
 561       // Try global overflow list.
 562     } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
 563       continue;
 564     }
 565 
 566     // Otherwise, offer termination.
 567     par_scan_state()->start_term_time();
 568     if (terminator()->offer_termination()) break;
 569     par_scan_state()->end_term_time();
 570   }
 571   assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
 572          "Broken overflow list?");
 573   // Finish the last termination pause.
 574   par_scan_state()->end_term_time();
 575 }
 576 
 577 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
 578                              Generation* old_gen,
 579                              HeapWord* young_old_boundary,
 580                              ParScanThreadStateSet* state_set,
 581                              StrongRootsScope* strong_roots_scope) :
 582     AbstractGangTask("ParNewGeneration collection"),
 583     _young_gen(young_gen), _old_gen(old_gen),
 584     _young_old_boundary(young_old_boundary),
 585     _state_set(state_set),
 586     _strong_roots_scope(strong_roots_scope),
 587     _par_state_string(StringTable::weak_storage())
 588 {}
 589 
 590 void ParNewGenTask::work(uint worker_id) {
 591   CMSHeap* heap = CMSHeap::heap();
 592   // Since this is being done in a separate thread, need new resource
 593   // and handle marks.
 594   ResourceMark rm;
 595   HandleMark hm;
 596 
 597   ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
 598   assert(_state_set->is_valid(worker_id), "Should not have been called");
 599 
 600   par_scan_state.set_young_old_boundary(_young_old_boundary);
 601 
 602   CLDScanClosure cld_scan_closure(&par_scan_state.to_space_root_closure(),
 603                                   heap->rem_set()->cld_rem_set()->accumulate_modified_oops());
 604 
 605   par_scan_state.start_strong_roots();
 606   heap->young_process_roots(_strong_roots_scope,
 607                            &par_scan_state.to_space_root_closure(),
 608                            &par_scan_state.older_gen_closure(),
 609                            &cld_scan_closure,
 610                            &_par_state_string);
 611 
 612   par_scan_state.end_strong_roots();
 613 
 614   // "evacuate followers".
 615   par_scan_state.evacuate_followers_closure().do_void();
 616 
 617   // This will collapse this worker's promoted object list that's
 618   // created during the main ParNew parallel phase of ParNew. This has
 619   // to be called after all workers have finished promoting objects
 620   // and scanning promoted objects. It should be safe calling it from
 621   // here, given that we can only reach here after all thread have
 622   // offered termination, i.e., after there is no more work to be
 623   // done. It will also disable promotion tracking for the rest of
 624   // this GC as it's not necessary to be on during reference processing.
 625   _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id);
 626 }
 627 
 628 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
 629   : DefNewGeneration(rs, initial_byte_size, "PCopy"),
 630   _overflow_list(NULL),
 631   _is_alive_closure(this),
 632   _plab_stats("Young", YoungPLABSize, PLABWeight)
 633 {
 634   NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
 635   NOT_PRODUCT(_num_par_pushes = 0;)
 636   _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
 637   guarantee(_task_queues != NULL, "task_queues allocation failure.");
 638 
 639   for (uint i = 0; i < ParallelGCThreads; i++) {
 640     ObjToScanQueue *q = new ObjToScanQueue();
 641     guarantee(q != NULL, "work_queue Allocation failure.");
 642     _task_queues->register_queue(i, q);
 643   }
 644 
 645   for (uint i = 0; i < ParallelGCThreads; i++) {
 646     _task_queues->queue(i)->initialize();
 647   }
 648 
 649   _overflow_stacks = NULL;
 650   if (ParGCUseLocalOverflow) {
 651     // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
 652     typedef Stack<oop, mtGC> GCOopStack;
 653 
 654     _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
 655     for (size_t i = 0; i < ParallelGCThreads; ++i) {
 656       new (_overflow_stacks + i) Stack<oop, mtGC>();
 657     }
 658   }
 659 
 660   if (UsePerfData) {
 661     EXCEPTION_MARK;
 662     ResourceMark rm;
 663 
 664     const char* cname =
 665          PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
 666     PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
 667                                      ParallelGCThreads, CHECK);
 668   }
 669 }
 670 
 671 // ParNewGeneration::
 672 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
 673   DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
 674 
 675 template <class T>
 676 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
 677 #ifdef ASSERT
 678   {
 679     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
 680     // We never expect to see a null reference being processed
 681     // as a weak reference.
 682     assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
 683   }
 684 #endif // ASSERT
 685 
 686   Devirtualizer::do_oop_no_verify(_par_cl, p);
 687 
 688   if (CMSHeap::heap()->is_in_reserved(p)) {
 689     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);;
 690     _rs->write_ref_field_gc_par(p, obj);
 691   }
 692 }
 693 
 694 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p)       { ParKeepAliveClosure::do_oop_work(p); }
 695 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
 696 
 697 // ParNewGeneration::
 698 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
 699   DefNewGeneration::KeepAliveClosure(cl) {}
 700 
 701 template <class T>
 702 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
 703 #ifdef ASSERT
 704   {
 705     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
 706     // We never expect to see a null reference being processed
 707     // as a weak reference.
 708     assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
 709   }
 710 #endif // ASSERT
 711 
 712   Devirtualizer::do_oop_no_verify(_cl, p);
 713 
 714   if (CMSHeap::heap()->is_in_reserved(p)) {
 715     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
 716     _rs->write_ref_field_gc_par(p, obj);
 717   }
 718 }
 719 
 720 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p)       { KeepAliveClosure::do_oop_work(p); }
 721 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
 722 
 723 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
 724   T heap_oop = RawAccess<>::oop_load(p);
 725   if (!CompressedOops::is_null(heap_oop)) {
 726     oop obj = CompressedOops::decode_not_null(heap_oop);
 727     if ((HeapWord*)obj < _boundary) {
 728       assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
 729       oop new_obj = obj->is_forwarded()
 730                       ? obj->forwardee()
 731                       : _g->DefNewGeneration::copy_to_survivor_space(obj);
 732       RawAccess<OOP_NOT_NULL>::oop_store(p, new_obj);
 733     }
 734     if (_gc_barrier) {
 735       // If p points to a younger generation, mark the card.
 736       if ((HeapWord*)obj < _gen_boundary) {
 737         _rs->write_ref_field_gc_par(p, obj);
 738       }
 739     }
 740   }
 741 }
 742 
 743 void ScanClosureWithParBarrier::do_oop(oop* p)       { ScanClosureWithParBarrier::do_oop_work(p); }
 744 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
 745 
 746 class ParNewRefProcTaskProxy: public AbstractGangTask {
 747   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
 748 public:
 749   ParNewRefProcTaskProxy(ProcessTask& task,
 750                          ParNewGeneration& young_gen,
 751                          Generation& old_gen,
 752                          HeapWord* young_old_boundary,
 753                          ParScanThreadStateSet& state_set);
 754 
 755 private:
 756   virtual void work(uint worker_id);
 757 private:
 758   ParNewGeneration&      _young_gen;
 759   ProcessTask&           _task;
 760   Generation&            _old_gen;
 761   HeapWord*              _young_old_boundary;
 762   ParScanThreadStateSet& _state_set;
 763 };
 764 
 765 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
 766                                                ParNewGeneration& young_gen,
 767                                                Generation& old_gen,
 768                                                HeapWord* young_old_boundary,
 769                                                ParScanThreadStateSet& state_set)
 770   : AbstractGangTask("ParNewGeneration parallel reference processing"),
 771     _young_gen(young_gen),
 772     _task(task),
 773     _old_gen(old_gen),
 774     _young_old_boundary(young_old_boundary),
 775     _state_set(state_set)
 776 { }
 777 
 778 void ParNewRefProcTaskProxy::work(uint worker_id) {
 779   ResourceMark rm;
 780   HandleMark hm;
 781   ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
 782   par_scan_state.set_young_old_boundary(_young_old_boundary);
 783   _task.work(worker_id, par_scan_state.is_alive_closure(),
 784              par_scan_state.keep_alive_closure(),
 785              par_scan_state.evacuate_followers_closure());
 786 }
 787 
 788 void ParNewRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) {
 789   CMSHeap* gch = CMSHeap::heap();
 790   WorkGang* workers = gch->workers();
 791   assert(workers != NULL, "Need parallel worker threads.");
 792   assert(workers->active_workers() == ergo_workers,
 793          "Ergonomically chosen workers (%u) must be equal to active workers (%u)",
 794          ergo_workers, workers->active_workers());
 795   _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
 796   ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
 797                                  _young_gen.reserved().end(), _state_set);
 798   workers->run_task(&rp_task, workers->active_workers());
 799   _state_set.reset(0 /* bad value in debug if not reset */,
 800                    _young_gen.promotion_failed());
 801 }
 802 
 803 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
 804   _state_set.flush();
 805   CMSHeap* heap = CMSHeap::heap();
 806   heap->save_marks();
 807 }
 808 
 809 ScanClosureWithParBarrier::
 810 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
 811   OopsInClassLoaderDataOrGenClosure(g), _g(g), _boundary(g->reserved().end()), _gc_barrier(gc_barrier)
 812 { }
 813 
 814 template <typename OopClosureType1, typename OopClosureType2>
 815 EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::
 816 EvacuateFollowersClosureGeneral(CMSHeap* heap,
 817                                 OopClosureType1* cur,
 818                                 OopClosureType2* older) :
 819   _heap(heap),
 820   _scan_cur_or_nonheap(cur), _scan_older(older)
 821 { }
 822 
 823 template <typename OopClosureType1, typename OopClosureType2>
 824 void EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::do_void() {
 825   do {
 826     _heap->oop_since_save_marks_iterate(_scan_cur_or_nonheap,
 827                                         _scan_older);
 828   } while (!_heap->no_allocs_since_save_marks());
 829 }
 830 
 831 // A Generation that does parallel young-gen collection.
 832 
 833 void ParNewGeneration::handle_promotion_failed(CMSHeap* gch, ParScanThreadStateSet& thread_state_set) {
 834   assert(_promo_failure_scan_stack.is_empty(), "post condition");
 835   _promo_failure_scan_stack.clear(true); // Clear cached segments.
 836 
 837   remove_forwarding_pointers();
 838   log_info(gc, promotion)("Promotion failed");
 839   // All the spaces are in play for mark-sweep.
 840   swap_spaces();  // Make life simpler for CMS || rescan; see 6483690.
 841   from()->set_next_compaction_space(to());
 842   gch->set_incremental_collection_failed();
 843   // Inform the next generation that a promotion failure occurred.
 844   _old_gen->promotion_failure_occurred();
 845 
 846   // Trace promotion failure in the parallel GC threads
 847   thread_state_set.trace_promotion_failed(gc_tracer());
 848   // Single threaded code may have reported promotion failure to the global state
 849   if (_promotion_failed_info.has_failed()) {
 850     _gc_tracer.report_promotion_failed(_promotion_failed_info);
 851   }
 852   // Reset the PromotionFailureALot counters.
 853   NOT_PRODUCT(gch->reset_promotion_should_fail();)
 854 }
 855 
 856 void ParNewGeneration::collect(bool   full,
 857                                bool   clear_all_soft_refs,
 858                                size_t size,
 859                                bool   is_tlab) {
 860   assert(full || size > 0, "otherwise we don't want to collect");
 861 
 862   CMSHeap* gch = CMSHeap::heap();
 863 
 864   _gc_timer->register_gc_start();
 865 
 866   AdaptiveSizePolicy* size_policy = gch->size_policy();
 867   WorkGang* workers = gch->workers();
 868   assert(workers != NULL, "Need workgang for parallel work");
 869   uint active_workers =
 870        AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
 871                                                workers->active_workers(),
 872                                                Threads::number_of_non_daemon_threads());
 873   active_workers = workers->update_active_workers(active_workers);
 874   log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers->total_workers());
 875 
 876   _old_gen = gch->old_gen();
 877 
 878   // If the next generation is too full to accommodate worst-case promotion
 879   // from this generation, pass on collection; let the next generation
 880   // do it.
 881   if (!collection_attempt_is_safe()) {
 882     gch->set_incremental_collection_failed();  // slight lie, in that we did not even attempt one
 883     return;
 884   }
 885   assert(to()->is_empty(), "Else not collection_attempt_is_safe");
 886 
 887   _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
 888   gch->trace_heap_before_gc(gc_tracer());
 889 
 890   init_assuming_no_promotion_failure();
 891 
 892   if (UseAdaptiveSizePolicy) {
 893     set_survivor_overflow(false);
 894     size_policy->minor_collection_begin();
 895   }
 896 
 897   GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause());
 898 
 899   age_table()->clear();
 900   to()->clear(SpaceDecorator::Mangle);
 901 
 902   gch->save_marks();
 903 
 904   // Set the correct parallelism (number of queues) in the reference processor
 905   ref_processor()->set_active_mt_degree(active_workers);
 906 
 907   // Need to initialize the preserved marks before the ThreadStateSet c'tor.
 908   _preserved_marks_set.init(active_workers);
 909 
 910   // Always set the terminator for the active number of workers
 911   // because only those workers go through the termination protocol.
 912   ParallelTaskTerminator _term(active_workers, task_queues());
 913   ParScanThreadStateSet thread_state_set(active_workers,
 914                                          *to(), *this, *_old_gen, *task_queues(),
 915                                          _overflow_stacks, _preserved_marks_set,
 916                                          desired_plab_sz(), _term);
 917 
 918   thread_state_set.reset(active_workers, promotion_failed());
 919 
 920   {
 921     StrongRootsScope srs(active_workers);
 922 
 923     ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
 924     gch->rem_set()->prepare_for_younger_refs_iterate(true);
 925     // It turns out that even when we're using 1 thread, doing the work in a
 926     // separate thread causes wide variance in run times.  We can't help this
 927     // in the multi-threaded case, but we special-case n=1 here to get
 928     // repeatable measurements of the 1-thread overhead of the parallel code.
 929     // Might multiple workers ever be used?  If yes, initialization
 930     // has been done such that the single threaded path should not be used.
 931     if (workers->total_workers() > 1) {
 932       workers->run_task(&tsk);
 933     } else {
 934       tsk.work(0);
 935     }
 936   }
 937 
 938   thread_state_set.reset(0 /* Bad value in debug if not reset */,
 939                          promotion_failed());
 940 
 941   // Trace and reset failed promotion info.
 942   if (promotion_failed()) {
 943     thread_state_set.trace_promotion_failed(gc_tracer());
 944   }
 945 
 946   // Process (weak) reference objects found during scavenge.
 947   ReferenceProcessor* rp = ref_processor();
 948   IsAliveClosure is_alive(this);
 949   ScanWeakRefClosure scan_weak_ref(this);
 950   KeepAliveClosure keep_alive(&scan_weak_ref);
 951   ScanClosure               scan_without_gc_barrier(this, false);
 952   ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
 953   set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
 954   EvacuateFollowersClosureGeneral<ScanClosure, ScanClosureWithParBarrier> evacuate_followers(
 955       gch, &scan_without_gc_barrier, &scan_with_gc_barrier);
 956   rp->setup_policy(clear_all_soft_refs);
 957   // Can  the mt_degree be set later (at run_task() time would be best)?
 958   rp->set_active_mt_degree(active_workers);
 959   ReferenceProcessorStats stats;
 960   ReferenceProcessorPhaseTimes pt(_gc_timer, rp->max_num_queues());
 961   if (rp->processing_is_mt()) {
 962     ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
 963     stats = rp->process_discovered_references(&is_alive, &keep_alive,
 964                                               &evacuate_followers, &task_executor,
 965                                               &pt);
 966   } else {
 967     thread_state_set.flush();
 968     gch->save_marks();
 969     stats = rp->process_discovered_references(&is_alive, &keep_alive,
 970                                               &evacuate_followers, NULL,
 971                                               &pt);
 972   }
 973   _gc_tracer.report_gc_reference_stats(stats);
 974   _gc_tracer.report_tenuring_threshold(tenuring_threshold());
 975   pt.print_all_references();
 976 
 977   assert(gch->no_allocs_since_save_marks(), "evacuation should be done at this point");
 978 
 979   WeakProcessor::weak_oops_do(&is_alive, &keep_alive);
 980 
 981   // Verify that the usage of keep_alive only forwarded
 982   // the oops and did not find anything new to copy.
 983   assert(gch->no_allocs_since_save_marks(), "unexpectedly copied objects");
 984 
 985   if (!promotion_failed()) {
 986     // Swap the survivor spaces.
 987     eden()->clear(SpaceDecorator::Mangle);
 988     from()->clear(SpaceDecorator::Mangle);
 989     if (ZapUnusedHeapArea) {
 990       // This is now done here because of the piece-meal mangling which
 991       // can check for valid mangling at intermediate points in the
 992       // collection(s).  When a young collection fails to collect
 993       // sufficient space resizing of the young generation can occur
 994       // and redistribute the spaces in the young generation.  Mangle
 995       // here so that unzapped regions don't get distributed to
 996       // other spaces.
 997       to()->mangle_unused_area();
 998     }
 999     swap_spaces();
1000 
1001     // A successful scavenge should restart the GC time limit count which is
1002     // for full GC's.
1003     size_policy->reset_gc_overhead_limit_count();
1004 
1005     assert(to()->is_empty(), "to space should be empty now");
1006 
1007     adjust_desired_tenuring_threshold();
1008   } else {
1009     handle_promotion_failed(gch, thread_state_set);
1010   }
1011   _preserved_marks_set.reclaim();
1012   // set new iteration safe limit for the survivor spaces
1013   from()->set_concurrent_iteration_safe_limit(from()->top());
1014   to()->set_concurrent_iteration_safe_limit(to()->top());
1015 
1016   plab_stats()->adjust_desired_plab_sz();
1017 
1018   TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1019   TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1020 
1021   if (UseAdaptiveSizePolicy) {
1022     size_policy->minor_collection_end(gch->gc_cause());
1023     size_policy->avg_survived()->sample(from()->used());
1024   }
1025 
1026   // We need to use a monotonically non-decreasing time in ms
1027   // or we will see time-warp warnings and os::javaTimeMillis()
1028   // does not guarantee monotonicity.
1029   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1030   update_time_of_last_gc(now);
1031 
1032   rp->set_enqueuing_is_done(true);
1033   rp->verify_no_references_recorded();
1034 
1035   gch->trace_heap_after_gc(gc_tracer());
1036 
1037   _gc_timer->register_gc_end();
1038 
1039   _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1040 }
1041 
1042 size_t ParNewGeneration::desired_plab_sz() {
1043   return _plab_stats.desired_plab_sz(CMSHeap::heap()->workers()->active_workers());
1044 }
1045 
1046 static int sum;
1047 void ParNewGeneration::waste_some_time() {
1048   for (int i = 0; i < 100; i++) {
1049     sum += i;
1050   }
1051 }
1052 
1053 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1054 
1055 // Because of concurrency, there are times where an object for which
1056 // "is_forwarded()" is true contains an "interim" forwarding pointer
1057 // value.  Such a value will soon be overwritten with a real value.
1058 // This method requires "obj" to have a forwarding pointer, and waits, if
1059 // necessary for a real one to be inserted, and returns it.
1060 
1061 oop ParNewGeneration::real_forwardee(oop obj) {
1062   oop forward_ptr = obj->forwardee();
1063   if (forward_ptr != ClaimedForwardPtr) {
1064     return forward_ptr;
1065   } else {
1066     return real_forwardee_slow(obj);
1067   }
1068 }
1069 
1070 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1071   // Spin-read if it is claimed but not yet written by another thread.
1072   oop forward_ptr = obj->forwardee();
1073   while (forward_ptr == ClaimedForwardPtr) {
1074     waste_some_time();
1075     assert(obj->is_forwarded(), "precondition");
1076     forward_ptr = obj->forwardee();
1077   }
1078   return forward_ptr;
1079 }
1080 
1081 // Multiple GC threads may try to promote an object.  If the object
1082 // is successfully promoted, a forwarding pointer will be installed in
1083 // the object in the young generation.  This method claims the right
1084 // to install the forwarding pointer before it copies the object,
1085 // thus avoiding the need to undo the copy as in
1086 // copy_to_survivor_space_avoiding_with_undo.
1087 
1088 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1089                                              oop old,
1090                                              size_t sz,
1091                                              markOop m) {
1092   // In the sequential version, this assert also says that the object is
1093   // not forwarded.  That might not be the case here.  It is the case that
1094   // the caller observed it to be not forwarded at some time in the past.
1095   assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1096 
1097   // The sequential code read "old->age()" below.  That doesn't work here,
1098   // since the age is in the mark word, and that might be overwritten with
1099   // a forwarding pointer by a parallel thread.  So we must save the mark
1100   // word in a local and then analyze it.
1101   oopDesc dummyOld;
1102   dummyOld.set_mark_raw(m);
1103   assert(!dummyOld.is_forwarded(),
1104          "should not be called with forwarding pointer mark word.");
1105 
1106   oop new_obj = NULL;
1107   oop forward_ptr;
1108 
1109   // Try allocating obj in to-space (unless too old)
1110   if (dummyOld.age() < tenuring_threshold()) {
1111     new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1112     if (new_obj == NULL) {
1113       set_survivor_overflow(true);
1114     }
1115   }
1116 
1117   if (new_obj == NULL) {
1118     // Either to-space is full or we decided to promote try allocating obj tenured
1119 
1120     // Attempt to install a null forwarding pointer (atomically),
1121     // to claim the right to install the real forwarding pointer.
1122     forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1123     if (forward_ptr != NULL) {
1124       // someone else beat us to it.
1125         return real_forwardee(old);
1126     }
1127 
1128     if (!_promotion_failed) {
1129       new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1130                                       old, m, sz);
1131     }
1132 
1133     if (new_obj == NULL) {
1134       // promotion failed, forward to self
1135       _promotion_failed = true;
1136       new_obj = old;
1137 
1138       par_scan_state->preserved_marks()->push_if_necessary(old, m);
1139       par_scan_state->register_promotion_failure(sz);
1140     }
1141 
1142     old->forward_to(new_obj);
1143     forward_ptr = NULL;
1144   } else {
1145     // Is in to-space; do copying ourselves.
1146     Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1147     assert(CMSHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1148     forward_ptr = old->forward_to_atomic(new_obj);
1149     // Restore the mark word copied above.
1150     new_obj->set_mark_raw(m);
1151     // Increment age if obj still in new generation
1152     new_obj->incr_age();
1153     par_scan_state->age_table()->add(new_obj, sz);
1154   }
1155   assert(new_obj != NULL, "just checking");
1156 
1157   // This code must come after the CAS test, or it will print incorrect
1158   // information.
1159   log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1160                                   is_in_reserved(new_obj) ? "copying" : "tenuring",
1161                                   new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1162 
1163   if (forward_ptr == NULL) {
1164     oop obj_to_push = new_obj;
1165     if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1166       // Length field used as index of next element to be scanned.
1167       // Real length can be obtained from real_forwardee()
1168       arrayOop(old)->set_length(0);
1169       obj_to_push = old;
1170       assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1171              "push forwarded object");
1172     }
1173     // Push it on one of the queues of to-be-scanned objects.
1174     bool simulate_overflow = false;
1175     NOT_PRODUCT(
1176       if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1177         // simulate a stack overflow
1178         simulate_overflow = true;
1179       }
1180     )
1181     if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1182       // Add stats for overflow pushes.
1183       log_develop_trace(gc)("Queue Overflow");
1184       push_on_overflow_list(old, par_scan_state);
1185       TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1186     }
1187 
1188     return new_obj;
1189   }
1190 
1191   // Oops.  Someone beat us to it.  Undo the allocation.  Where did we
1192   // allocate it?
1193   if (is_in_reserved(new_obj)) {
1194     // Must be in to_space.
1195     assert(to()->is_in_reserved(new_obj), "Checking");
1196     if (forward_ptr == ClaimedForwardPtr) {
1197       // Wait to get the real forwarding pointer value.
1198       forward_ptr = real_forwardee(old);
1199     }
1200     par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1201   }
1202 
1203   return forward_ptr;
1204 }
1205 
1206 #ifndef PRODUCT
1207 // It's OK to call this multi-threaded;  the worst thing
1208 // that can happen is that we'll get a bunch of closely
1209 // spaced simulated overflows, but that's OK, in fact
1210 // probably good as it would exercise the overflow code
1211 // under contention.
1212 bool ParNewGeneration::should_simulate_overflow() {
1213   if (_overflow_counter-- <= 0) { // just being defensive
1214     _overflow_counter = ParGCWorkQueueOverflowInterval;
1215     return true;
1216   } else {
1217     return false;
1218   }
1219 }
1220 #endif
1221 
1222 // In case we are using compressed oops, we need to be careful.
1223 // If the object being pushed is an object array, then its length
1224 // field keeps track of the "grey boundary" at which the next
1225 // incremental scan will be done (see ParGCArrayScanChunk).
1226 // When using compressed oops, this length field is kept in the
1227 // lower 32 bits of the erstwhile klass word and cannot be used
1228 // for the overflow chaining pointer (OCP below). As such the OCP
1229 // would itself need to be compressed into the top 32-bits in this
1230 // case. Unfortunately, see below, in the event that we have a
1231 // promotion failure, the node to be pushed on the list can be
1232 // outside of the Java heap, so the heap-based pointer compression
1233 // would not work (we would have potential aliasing between C-heap
1234 // and Java-heap pointers). For this reason, when using compressed
1235 // oops, we simply use a worker-thread-local, non-shared overflow
1236 // list in the form of a growable array, with a slightly different
1237 // overflow stack draining strategy. If/when we start using fat
1238 // stacks here, we can go back to using (fat) pointer chains
1239 // (although some performance comparisons would be useful since
1240 // single global lists have their own performance disadvantages
1241 // as we were made painfully aware not long ago, see 6786503).
1242 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1243 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1244   assert(is_in_reserved(from_space_obj), "Should be from this generation");
1245   if (ParGCUseLocalOverflow) {
1246     // In the case of compressed oops, we use a private, not-shared
1247     // overflow stack.
1248     par_scan_state->push_on_overflow_stack(from_space_obj);
1249   } else {
1250     assert(!UseCompressedOops, "Error");
1251     // if the object has been forwarded to itself, then we cannot
1252     // use the klass pointer for the linked list.  Instead we have
1253     // to allocate an oopDesc in the C-Heap and use that for the linked list.
1254     // XXX This is horribly inefficient when a promotion failure occurs
1255     // and should be fixed. XXX FIX ME !!!
1256 #ifndef PRODUCT
1257     Atomic::inc(&_num_par_pushes);
1258     assert(_num_par_pushes > 0, "Tautology");
1259 #endif
1260     if (from_space_obj->forwardee() == from_space_obj) {
1261       oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1262       listhead->forward_to(from_space_obj);
1263       from_space_obj = listhead;
1264     }
1265     oop observed_overflow_list = _overflow_list;
1266     oop cur_overflow_list;
1267     do {
1268       cur_overflow_list = observed_overflow_list;
1269       if (cur_overflow_list != BUSY) {
1270         from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1271       } else {
1272         from_space_obj->set_klass_to_list_ptr(NULL);
1273       }
1274       observed_overflow_list =
1275         Atomic::cmpxchg((oopDesc*)from_space_obj, &_overflow_list, (oopDesc*)cur_overflow_list);
1276     } while (cur_overflow_list != observed_overflow_list);
1277   }
1278 }
1279 
1280 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1281   bool res;
1282 
1283   if (ParGCUseLocalOverflow) {
1284     res = par_scan_state->take_from_overflow_stack();
1285   } else {
1286     assert(!UseCompressedOops, "Error");
1287     res = take_from_overflow_list_work(par_scan_state);
1288   }
1289   return res;
1290 }
1291 
1292 
1293 // *NOTE*: The overflow list manipulation code here and
1294 // in CMSCollector:: are very similar in shape,
1295 // except that in the CMS case we thread the objects
1296 // directly into the list via their mark word, and do
1297 // not need to deal with special cases below related
1298 // to chunking of object arrays and promotion failure
1299 // handling.
1300 // CR 6797058 has been filed to attempt consolidation of
1301 // the common code.
1302 // Because of the common code, if you make any changes in
1303 // the code below, please check the CMS version to see if
1304 // similar changes might be needed.
1305 // See CMSCollector::par_take_from_overflow_list() for
1306 // more extensive documentation comments.
1307 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1308   ObjToScanQueue* work_q = par_scan_state->work_queue();
1309   // How many to take?
1310   size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1311                                  (size_t)ParGCDesiredObjsFromOverflowList);
1312 
1313   assert(!UseCompressedOops, "Error");
1314   assert(par_scan_state->overflow_stack() == NULL, "Error");
1315   if (_overflow_list == NULL) return false;
1316 
1317   // Otherwise, there was something there; try claiming the list.
1318   oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1319   // Trim off a prefix of at most objsFromOverflow items
1320   Thread* tid = Thread::current();
1321   size_t spin_count = ParallelGCThreads;
1322   size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1323   for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1324     // someone grabbed it before we did ...
1325     // ... we spin for a short while...
1326     os::sleep(tid, sleep_time_millis, false);
1327     if (_overflow_list == NULL) {
1328       // nothing left to take
1329       return false;
1330     } else if (_overflow_list != BUSY) {
1331      // try and grab the prefix
1332      prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1333     }
1334   }
1335   if (prefix == NULL || prefix == BUSY) {
1336      // Nothing to take or waited long enough
1337      if (prefix == NULL) {
1338        // Write back the NULL in case we overwrote it with BUSY above
1339        // and it is still the same value.
1340        (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1341      }
1342      return false;
1343   }
1344   assert(prefix != NULL && prefix != BUSY, "Error");
1345   oop cur = prefix;
1346   for (size_t i = 1; i < objsFromOverflow; ++i) {
1347     oop next = cur->list_ptr_from_klass();
1348     if (next == NULL) break;
1349     cur = next;
1350   }
1351   assert(cur != NULL, "Loop postcondition");
1352 
1353   // Reattach remaining (suffix) to overflow list
1354   oop suffix = cur->list_ptr_from_klass();
1355   if (suffix == NULL) {
1356     // Write back the NULL in lieu of the BUSY we wrote
1357     // above and it is still the same value.
1358     if (_overflow_list == BUSY) {
1359       (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1360     }
1361   } else {
1362     assert(suffix != BUSY, "Error");
1363     // suffix will be put back on global list
1364     cur->set_klass_to_list_ptr(NULL);     // break off suffix
1365     // It's possible that the list is still in the empty(busy) state
1366     // we left it in a short while ago; in that case we may be
1367     // able to place back the suffix.
1368     oop observed_overflow_list = _overflow_list;
1369     oop cur_overflow_list = observed_overflow_list;
1370     bool attached = false;
1371     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1372       observed_overflow_list =
1373         Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1374       if (cur_overflow_list == observed_overflow_list) {
1375         attached = true;
1376         break;
1377       } else cur_overflow_list = observed_overflow_list;
1378     }
1379     if (!attached) {
1380       // Too bad, someone else got in in between; we'll need to do a splice.
1381       // Find the last item of suffix list
1382       oop last = suffix;
1383       while (true) {
1384         oop next = last->list_ptr_from_klass();
1385         if (next == NULL) break;
1386         last = next;
1387       }
1388       // Atomically prepend suffix to current overflow list
1389       observed_overflow_list = _overflow_list;
1390       do {
1391         cur_overflow_list = observed_overflow_list;
1392         if (cur_overflow_list != BUSY) {
1393           // Do the splice ...
1394           last->set_klass_to_list_ptr(cur_overflow_list);
1395         } else { // cur_overflow_list == BUSY
1396           last->set_klass_to_list_ptr(NULL);
1397         }
1398         observed_overflow_list =
1399           Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1400       } while (cur_overflow_list != observed_overflow_list);
1401     }
1402   }
1403 
1404   // Push objects on prefix list onto this thread's work queue
1405   assert(prefix != NULL && prefix != BUSY, "program logic");
1406   cur = prefix;
1407   ssize_t n = 0;
1408   while (cur != NULL) {
1409     oop obj_to_push = cur->forwardee();
1410     oop next        = cur->list_ptr_from_klass();
1411     cur->set_klass(obj_to_push->klass());
1412     // This may be an array object that is self-forwarded. In that case, the list pointer
1413     // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1414     if (!is_in_reserved(cur)) {
1415       // This can become a scaling bottleneck when there is work queue overflow coincident
1416       // with promotion failure.
1417       oopDesc* f = cur;
1418       FREE_C_HEAP_ARRAY(oopDesc, f);
1419     } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1420       assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1421       obj_to_push = cur;
1422     }
1423     bool ok = work_q->push(obj_to_push);
1424     assert(ok, "Should have succeeded");
1425     cur = next;
1426     n++;
1427   }
1428   TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1429 #ifndef PRODUCT
1430   assert(_num_par_pushes >= n, "Too many pops?");
1431   Atomic::sub(n, &_num_par_pushes);
1432 #endif
1433   return true;
1434 }
1435 #undef BUSY
1436 
1437 void ParNewGeneration::ref_processor_init() {
1438   if (_ref_processor == NULL) {
1439     // Allocate and initialize a reference processor
1440     _span_based_discoverer.set_span(_reserved);
1441     _ref_processor =
1442       new ReferenceProcessor(&_span_based_discoverer,    // span
1443                              ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1444                              ParallelGCThreads,          // mt processing degree
1445                              refs_discovery_is_mt(),     // mt discovery
1446                              ParallelGCThreads,          // mt discovery degree
1447                              refs_discovery_is_atomic(), // atomic_discovery
1448                              NULL,                       // is_alive_non_header
1449                              false);                     // disable adjusting number of processing threads
1450   }
1451 }
1452 
1453 const char* ParNewGeneration::name() const {
1454   return "par new generation";
1455 }
1456 
1457 void ParNewGeneration::restore_preserved_marks() {
1458   SharedRestorePreservedMarksTaskExecutor task_executor(CMSHeap::heap()->workers());
1459   _preserved_marks_set.restore(&task_executor);
1460 }