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