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/space.hpp"
  46 #include "gc/shared/spaceDecorator.hpp"
  47 #include "gc/shared/strongRootsScope.hpp"
  48 #include "gc/shared/taskqueue.inline.hpp"
  49 #include "gc/shared/weakProcessor.hpp"
  50 #include "gc/shared/workgroup.hpp"
  51 #include "logging/log.hpp"
  52 #include "logging/logStream.hpp"
  53 #include "memory/resourceArea.hpp"
  54 #include "oops/access.inline.hpp"
  55 #include "oops/compressedOops.inline.hpp"
  56 #include "oops/objArrayOop.hpp"
  57 #include "oops/oop.inline.hpp"
  58 #include "runtime/atomic.hpp"
  59 #include "runtime/handles.hpp"
  60 #include "runtime/handles.inline.hpp"
  61 #include "runtime/java.hpp"
  62 #include "runtime/thread.inline.hpp"
  63 #include "utilities/copy.hpp"
  64 #include "utilities/globalDefinitions.hpp"
  65 #include "utilities/stack.inline.hpp"
  66 
  67 ParScanThreadState::ParScanThreadState(Space* to_space_,
  68                                        ParNewGeneration* young_gen_,
  69                                        Generation* old_gen_,
  70                                        int thread_num_,
  71                                        ObjToScanQueueSet* work_queue_set_,
  72                                        Stack<oop, mtGC>* overflow_stacks_,
  73                                        PreservedMarks* preserved_marks_,
  74                                        size_t desired_plab_sz_,
  75                                        ParallelTaskTerminator& term_) :
  76   _to_space(to_space_),
  77   _old_gen(old_gen_),
  78   _young_gen(young_gen_),
  79   _thread_num(thread_num_),
  80   _work_queue(work_queue_set_->queue(thread_num_)),
  81   _to_space_full(false),
  82   _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
  83   _preserved_marks(preserved_marks_),
  84   _ageTable(false), // false ==> not the global age table, no perf data.
  85   _to_space_alloc_buffer(desired_plab_sz_),
  86   _to_space_closure(young_gen_, this),
  87   _old_gen_closure(young_gen_, this),
  88   _to_space_root_closure(young_gen_, this),
  89   _old_gen_root_closure(young_gen_, this),
  90   _older_gen_closure(young_gen_, this),
  91   _evacuate_followers(this, &_to_space_closure, &_old_gen_closure,
  92                       &_to_space_root_closure, young_gen_, &_old_gen_root_closure,
  93                       work_queue_set_, &term_),
  94   _is_alive_closure(young_gen_),
  95   _scan_weak_ref_closure(young_gen_, this),
  96   _keep_alive_closure(&_scan_weak_ref_closure),
  97   _strong_roots_time(0.0),
  98   _term_time(0.0)
  99 {
 100   #if TASKQUEUE_STATS
 101   _term_attempts = 0;
 102   _overflow_refills = 0;
 103   _overflow_refill_objs = 0;
 104   #endif // TASKQUEUE_STATS
 105 
 106   _survivor_chunk_array = (ChunkArray*) old_gen()->get_data_recorder(thread_num());
 107   _hash_seed = 17;  // Might want to take time-based random value.
 108   _start = os::elapsedTime();
 109   _old_gen_closure.set_generation(old_gen_);
 110   _old_gen_root_closure.set_generation(old_gen_);
 111 }
 112 
 113 void ParScanThreadState::record_survivor_plab(HeapWord* plab_start,
 114                                               size_t plab_word_size) {
 115   ChunkArray* sca = survivor_chunk_array();
 116   if (sca != NULL) {
 117     // A non-null SCA implies that we want the PLAB data recorded.
 118     sca->record_sample(plab_start, plab_word_size);
 119   }
 120 }
 121 
 122 bool ParScanThreadState::should_be_partially_scanned(oop new_obj, oop old_obj) const {
 123   return new_obj->is_objArray() &&
 124          arrayOop(new_obj)->length() > ParGCArrayScanChunk &&
 125          new_obj != old_obj;
 126 }
 127 
 128 void ParScanThreadState::scan_partial_array_and_push_remainder(oop old) {
 129   assert(old->is_objArray(), "must be obj array");
 130   assert(old->is_forwarded(), "must be forwarded");
 131   assert(CMSHeap::heap()->is_in_reserved(old), "must be in heap.");
 132   assert(!old_gen()->is_in(old), "must be in young generation.");
 133 
 134   objArrayOop obj = objArrayOop(old->forwardee());
 135   // Process ParGCArrayScanChunk elements now
 136   // and push the remainder back onto queue
 137   int start     = arrayOop(old)->length();
 138   int end       = obj->length();
 139   int remainder = end - start;
 140   assert(start <= end, "just checking");
 141   if (remainder > 2 * ParGCArrayScanChunk) {
 142     // Test above combines last partial chunk with a full chunk
 143     end = start + ParGCArrayScanChunk;
 144     arrayOop(old)->set_length(end);
 145     // Push remainder.
 146     bool ok = work_queue()->push(old);
 147     assert(ok, "just popped, push must be okay");
 148   } else {
 149     // Restore length so that it can be used if there
 150     // is a promotion failure and forwarding pointers
 151     // must be removed.
 152     arrayOop(old)->set_length(end);
 153   }
 154 
 155   // process our set of indices (include header in first chunk)
 156   // should make sure end is even (aligned to HeapWord in case of compressed oops)
 157   if ((HeapWord *)obj < young_old_boundary()) {
 158     // object is in to_space
 159     obj->oop_iterate_range(&_to_space_closure, start, end);
 160   } else {
 161     // object is in old generation
 162     obj->oop_iterate_range(&_old_gen_closure, start, end);
 163   }
 164 }
 165 
 166 void ParScanThreadState::trim_queues(int max_size) {
 167   ObjToScanQueue* queue = work_queue();
 168   do {
 169     while (queue->size() > (juint)max_size) {
 170       oop obj_to_scan;
 171       if (queue->pop_local(obj_to_scan)) {
 172         if ((HeapWord *)obj_to_scan < young_old_boundary()) {
 173           if (obj_to_scan->is_objArray() &&
 174               obj_to_scan->is_forwarded() &&
 175               obj_to_scan->forwardee() != obj_to_scan) {
 176             scan_partial_array_and_push_remainder(obj_to_scan);
 177           } else {
 178             // object is in to_space
 179             obj_to_scan->oop_iterate(&_to_space_closure);
 180           }
 181         } else {
 182           // object is in old generation
 183           obj_to_scan->oop_iterate(&_old_gen_closure);
 184         }
 185       }
 186     }
 187     // For the  case of compressed oops, we have a private, non-shared
 188     // overflow stack, so we eagerly drain it so as to more evenly
 189     // distribute load early. Note: this may be good to do in
 190     // general rather than delay for the final stealing phase.
 191     // If applicable, we'll transfer a set of objects over to our
 192     // work queue, allowing them to be stolen and draining our
 193     // private overflow stack.
 194   } while (ParGCTrimOverflow && young_gen()->take_from_overflow_list(this));
 195 }
 196 
 197 bool ParScanThreadState::take_from_overflow_stack() {
 198   assert(ParGCUseLocalOverflow, "Else should not call");
 199   assert(young_gen()->overflow_list() == NULL, "Error");
 200   ObjToScanQueue* queue = work_queue();
 201   Stack<oop, mtGC>* const of_stack = overflow_stack();
 202   const size_t num_overflow_elems = of_stack->size();
 203   const size_t space_available = queue->max_elems() - queue->size();
 204   const size_t num_take_elems = MIN3(space_available / 4,
 205                                      ParGCDesiredObjsFromOverflowList,
 206                                      num_overflow_elems);
 207   // Transfer the most recent num_take_elems from the overflow
 208   // stack to our work queue.
 209   for (size_t i = 0; i != num_take_elems; i++) {
 210     oop cur = of_stack->pop();
 211     oop obj_to_push = cur->forwardee();
 212     assert(CMSHeap::heap()->is_in_reserved(cur), "Should be in heap");
 213     assert(!old_gen()->is_in_reserved(cur), "Should be in young gen");
 214     assert(CMSHeap::heap()->is_in_reserved(obj_to_push), "Should be in heap");
 215     if (should_be_partially_scanned(obj_to_push, cur)) {
 216       assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
 217       obj_to_push = cur;
 218     }
 219     bool ok = queue->push(obj_to_push);
 220     assert(ok, "Should have succeeded");
 221   }
 222   assert(young_gen()->overflow_list() == NULL, "Error");
 223   return num_take_elems > 0;  // was something transferred?
 224 }
 225 
 226 void ParScanThreadState::push_on_overflow_stack(oop p) {
 227   assert(ParGCUseLocalOverflow, "Else should not call");
 228   overflow_stack()->push(p);
 229   assert(young_gen()->overflow_list() == NULL, "Error");
 230 }
 231 
 232 HeapWord* ParScanThreadState::alloc_in_to_space_slow(size_t word_sz) {
 233   // If the object is small enough, try to reallocate the buffer.
 234   HeapWord* obj = NULL;
 235   if (!_to_space_full) {
 236     PLAB* const plab = to_space_alloc_buffer();
 237     Space* const sp  = to_space();
 238     if (word_sz * 100 < ParallelGCBufferWastePct * plab->word_sz()) {
 239       // Is small enough; abandon this buffer and start a new one.
 240       plab->retire();
 241       // The minimum size has to be twice SurvivorAlignmentInBytes to
 242       // allow for padding used in the alignment of 1 word.  A padding
 243       // of 1 is too small for a filler word so the padding size will
 244       // be increased by SurvivorAlignmentInBytes.
 245       size_t min_usable_size = 2 * static_cast<size_t>(SurvivorAlignmentInBytes >> LogHeapWordSize);
 246       size_t buf_size = MAX2(plab->word_sz(), min_usable_size);
 247       HeapWord* buf_space = sp->par_allocate(buf_size);
 248       if (buf_space == NULL) {
 249         const size_t min_bytes = MAX2(PLAB::min_size(), min_usable_size) << LogHeapWordSize;
 250         size_t free_bytes = sp->free();
 251         while(buf_space == NULL && free_bytes >= min_bytes) {
 252           buf_size = free_bytes >> LogHeapWordSize;
 253           assert(buf_size == (size_t)align_object_size(buf_size), "Invariant");
 254           buf_space  = sp->par_allocate(buf_size);
 255           free_bytes = sp->free();
 256         }
 257       }
 258       if (buf_space != NULL) {
 259         plab->set_buf(buf_space, buf_size);
 260         record_survivor_plab(buf_space, buf_size);
 261         obj = plab->allocate_aligned(word_sz, SurvivorAlignmentInBytes);
 262         // Note that we cannot compare buf_size < word_sz below
 263         // because of AlignmentReserve (see PLAB::allocate()).
 264         assert(obj != NULL || plab->words_remaining() < word_sz,
 265                "Else should have been able to allocate requested object size "
 266                SIZE_FORMAT ", PLAB size " SIZE_FORMAT ", SurvivorAlignmentInBytes "
 267                SIZE_FORMAT ", words_remaining " SIZE_FORMAT,
 268                word_sz, buf_size, SurvivorAlignmentInBytes, plab->words_remaining());
 269         // It's conceivable that we may be able to use the
 270         // buffer we just grabbed for subsequent small requests
 271         // even if not for this one.
 272       } else {
 273         // We're used up.
 274         _to_space_full = true;
 275       }
 276     } else {
 277       // Too large; allocate the object individually.
 278       obj = sp->par_allocate(word_sz);
 279     }
 280   }
 281   return obj;
 282 }
 283 
 284 void ParScanThreadState::undo_alloc_in_to_space(HeapWord* obj, size_t word_sz) {
 285   to_space_alloc_buffer()->undo_allocation(obj, word_sz);
 286 }
 287 
 288 void ParScanThreadState::print_promotion_failure_size() {
 289   if (_promotion_failed_info.has_failed()) {
 290     log_trace(gc, promotion)(" (%d: promotion failure size = " SIZE_FORMAT ") ",
 291                              _thread_num, _promotion_failed_info.first_size());
 292   }
 293 }
 294 
 295 class ParScanThreadStateSet: StackObj {
 296 public:
 297   // Initializes states for the specified number of threads;
 298   ParScanThreadStateSet(int                     num_threads,
 299                         Space&                  to_space,
 300                         ParNewGeneration&       young_gen,
 301                         Generation&             old_gen,
 302                         ObjToScanQueueSet&      queue_set,
 303                         Stack<oop, mtGC>*       overflow_stacks_,
 304                         PreservedMarksSet&      preserved_marks_set,
 305                         size_t                  desired_plab_sz,
 306                         ParallelTaskTerminator& term);
 307 
 308   ~ParScanThreadStateSet() { TASKQUEUE_STATS_ONLY(reset_stats()); }
 309 
 310   inline ParScanThreadState& thread_state(int i);
 311 
 312   void trace_promotion_failed(const YoungGCTracer* gc_tracer);
 313   void reset(uint active_workers, bool promotion_failed);
 314   void flush();
 315 
 316   #if TASKQUEUE_STATS
 317   static void
 318     print_termination_stats_hdr(outputStream* const st);
 319   void print_termination_stats();
 320   static void
 321     print_taskqueue_stats_hdr(outputStream* const st);
 322   void print_taskqueue_stats();
 323   void reset_stats();
 324   #endif // TASKQUEUE_STATS
 325 
 326 private:
 327   ParallelTaskTerminator& _term;
 328   ParNewGeneration&       _young_gen;
 329   Generation&             _old_gen;
 330   ParScanThreadState*     _per_thread_states;
 331   const int               _num_threads;
 332  public:
 333   bool is_valid(int id) const { return id < _num_threads; }
 334   ParallelTaskTerminator* terminator() { return &_term; }
 335 };
 336 
 337 ParScanThreadStateSet::ParScanThreadStateSet(int num_threads,
 338                                              Space& to_space,
 339                                              ParNewGeneration& young_gen,
 340                                              Generation& old_gen,
 341                                              ObjToScanQueueSet& queue_set,
 342                                              Stack<oop, mtGC>* overflow_stacks,
 343                                              PreservedMarksSet& preserved_marks_set,
 344                                              size_t desired_plab_sz,
 345                                              ParallelTaskTerminator& term)
 346   : _young_gen(young_gen),
 347     _old_gen(old_gen),
 348     _term(term),
 349     _per_thread_states(NEW_RESOURCE_ARRAY(ParScanThreadState, num_threads)),
 350     _num_threads(num_threads)
 351 {
 352   assert(num_threads > 0, "sanity check!");
 353   assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
 354          "overflow_stack allocation mismatch");
 355   // Initialize states.
 356   for (int i = 0; i < num_threads; ++i) {
 357     new(_per_thread_states + i)
 358       ParScanThreadState(&to_space, &young_gen, &old_gen, i, &queue_set,
 359                          overflow_stacks, preserved_marks_set.get(i),
 360                          desired_plab_sz, term);
 361   }
 362 }
 363 
 364 inline ParScanThreadState& ParScanThreadStateSet::thread_state(int i) {
 365   assert(i >= 0 && i < _num_threads, "sanity check!");
 366   return _per_thread_states[i];
 367 }
 368 
 369 void ParScanThreadStateSet::trace_promotion_failed(const YoungGCTracer* gc_tracer) {
 370   for (int i = 0; i < _num_threads; ++i) {
 371     if (thread_state(i).promotion_failed()) {
 372       gc_tracer->report_promotion_failed(thread_state(i).promotion_failed_info());
 373       thread_state(i).promotion_failed_info().reset();
 374     }
 375   }
 376 }
 377 
 378 void ParScanThreadStateSet::reset(uint active_threads, bool promotion_failed) {
 379   _term.reset_for_reuse(active_threads);
 380   if (promotion_failed) {
 381     for (int i = 0; i < _num_threads; ++i) {
 382       thread_state(i).print_promotion_failure_size();
 383     }
 384   }
 385 }
 386 
 387 #if TASKQUEUE_STATS
 388 void ParScanThreadState::reset_stats() {
 389   taskqueue_stats().reset();
 390   _term_attempts = 0;
 391   _overflow_refills = 0;
 392   _overflow_refill_objs = 0;
 393 }
 394 
 395 void ParScanThreadStateSet::reset_stats() {
 396   for (int i = 0; i < _num_threads; ++i) {
 397     thread_state(i).reset_stats();
 398   }
 399 }
 400 
 401 void ParScanThreadStateSet::print_termination_stats_hdr(outputStream* const st) {
 402   st->print_raw_cr("GC Termination Stats");
 403   st->print_raw_cr("     elapsed  --strong roots-- -------termination-------");
 404   st->print_raw_cr("thr     ms        ms       %       ms       %   attempts");
 405   st->print_raw_cr("--- --------- --------- ------ --------- ------ --------");
 406 }
 407 
 408 void ParScanThreadStateSet::print_termination_stats() {
 409   Log(gc, task, stats) log;
 410   if (!log.is_debug()) {
 411     return;
 412   }
 413 
 414   ResourceMark rm;
 415   LogStream ls(log.debug());
 416   outputStream* st = &ls;
 417 
 418   print_termination_stats_hdr(st);
 419 
 420   for (int i = 0; i < _num_threads; ++i) {
 421     const ParScanThreadState & pss = thread_state(i);
 422     const double elapsed_ms = pss.elapsed_time() * 1000.0;
 423     const double s_roots_ms = pss.strong_roots_time() * 1000.0;
 424     const double term_ms = pss.term_time() * 1000.0;
 425     st->print_cr("%3d %9.2f %9.2f %6.2f %9.2f %6.2f " SIZE_FORMAT_W(8),
 426                  i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms,
 427                  term_ms, term_ms * 100 / elapsed_ms, pss.term_attempts());
 428   }
 429 }
 430 
 431 // Print stats related to work queue activity.
 432 void ParScanThreadStateSet::print_taskqueue_stats_hdr(outputStream* const st) {
 433   st->print_raw_cr("GC Task Stats");
 434   st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr();
 435   st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr();
 436 }
 437 
 438 void ParScanThreadStateSet::print_taskqueue_stats() {
 439   if (!log_develop_is_enabled(Trace, gc, task, stats)) {
 440     return;
 441   }
 442   Log(gc, task, stats) log;
 443   ResourceMark rm;
 444   LogStream ls(log.trace());
 445   outputStream* st = &ls;
 446   print_taskqueue_stats_hdr(st);
 447 
 448   TaskQueueStats totals;
 449   for (int i = 0; i < _num_threads; ++i) {
 450     const ParScanThreadState & pss = thread_state(i);
 451     const TaskQueueStats & stats = pss.taskqueue_stats();
 452     st->print("%3d ", i); stats.print(st); st->cr();
 453     totals += stats;
 454 
 455     if (pss.overflow_refills() > 0) {
 456       st->print_cr("    " SIZE_FORMAT_W(10) " overflow refills    "
 457                    SIZE_FORMAT_W(10) " overflow objects",
 458                    pss.overflow_refills(), pss.overflow_refill_objs());
 459     }
 460   }
 461   st->print("tot "); totals.print(st); st->cr();
 462 
 463   DEBUG_ONLY(totals.verify());
 464 }
 465 #endif // TASKQUEUE_STATS
 466 
 467 void ParScanThreadStateSet::flush() {
 468   // Work in this loop should be kept as lightweight as
 469   // possible since this might otherwise become a bottleneck
 470   // to scaling. Should we add heavy-weight work into this
 471   // loop, consider parallelizing the loop into the worker threads.
 472   for (int i = 0; i < _num_threads; ++i) {
 473     ParScanThreadState& par_scan_state = thread_state(i);
 474 
 475     // Flush stats related to To-space PLAB activity and
 476     // retire the last buffer.
 477     par_scan_state.to_space_alloc_buffer()->flush_and_retire_stats(_young_gen.plab_stats());
 478 
 479     // Every thread has its own age table.  We need to merge
 480     // them all into one.
 481     AgeTable *local_table = par_scan_state.age_table();
 482     _young_gen.age_table()->merge(local_table);
 483 
 484     // Inform old gen that we're done.
 485     _old_gen.par_promote_alloc_done(i);
 486   }
 487 
 488   if (UseConcMarkSweepGC) {
 489     // We need to call this even when ResizeOldPLAB is disabled
 490     // so as to avoid breaking some asserts. While we may be able
 491     // to avoid this by reorganizing the code a bit, I am loathe
 492     // to do that unless we find cases where ergo leads to bad
 493     // performance.
 494     CompactibleFreeListSpaceLAB::compute_desired_plab_size();
 495   }
 496 }
 497 
 498 ParScanClosure::ParScanClosure(ParNewGeneration* g,
 499                                ParScanThreadState* par_scan_state) :
 500   OopsInClassLoaderDataOrGenClosure(g), _par_scan_state(par_scan_state), _g(g) {
 501   _boundary = _g->reserved().end();
 502 }
 503 
 504 void ParScanWithBarrierClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, true, false); }
 505 void ParScanWithBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, false); }
 506 
 507 void ParScanWithoutBarrierClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, false, false); }
 508 void ParScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, false); }
 509 
 510 void ParRootScanWithBarrierTwoGensClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, true, true); }
 511 void ParRootScanWithBarrierTwoGensClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, true, true); }
 512 
 513 void ParRootScanWithoutBarrierClosure::do_oop(oop* p)       { ParScanClosure::do_oop_work(p, false, true); }
 514 void ParRootScanWithoutBarrierClosure::do_oop(narrowOop* p) { ParScanClosure::do_oop_work(p, false, true); }
 515 
 516 ParScanWeakRefClosure::ParScanWeakRefClosure(ParNewGeneration* g,
 517                                              ParScanThreadState* par_scan_state)
 518   : ScanWeakRefClosure(g), _par_scan_state(par_scan_state)
 519 {}
 520 
 521 void ParScanWeakRefClosure::do_oop(oop* p)       { ParScanWeakRefClosure::do_oop_work(p); }
 522 void ParScanWeakRefClosure::do_oop(narrowOop* p) { ParScanWeakRefClosure::do_oop_work(p); }
 523 
 524 #ifdef WIN32
 525 #pragma warning(disable: 4786) /* identifier was truncated to '255' characters in the browser information */
 526 #endif
 527 
 528 ParEvacuateFollowersClosure::ParEvacuateFollowersClosure(
 529     ParScanThreadState* par_scan_state_,
 530     ParScanWithoutBarrierClosure* to_space_closure_,
 531     ParScanWithBarrierClosure* old_gen_closure_,
 532     ParRootScanWithoutBarrierClosure* to_space_root_closure_,
 533     ParNewGeneration* par_gen_,
 534     ParRootScanWithBarrierTwoGensClosure* old_gen_root_closure_,
 535     ObjToScanQueueSet* task_queues_,
 536     ParallelTaskTerminator* terminator_) :
 537 
 538     _par_scan_state(par_scan_state_),
 539     _to_space_closure(to_space_closure_),
 540     _old_gen_closure(old_gen_closure_),
 541     _to_space_root_closure(to_space_root_closure_),
 542     _old_gen_root_closure(old_gen_root_closure_),
 543     _par_gen(par_gen_),
 544     _task_queues(task_queues_),
 545     _terminator(terminator_)
 546 {}
 547 
 548 void ParEvacuateFollowersClosure::do_void() {
 549   ObjToScanQueue* work_q = par_scan_state()->work_queue();
 550 
 551   while (true) {
 552     // Scan to-space and old-gen objs until we run out of both.
 553     oop obj_to_scan;
 554     par_scan_state()->trim_queues(0);
 555 
 556     // We have no local work, attempt to steal from other threads.
 557 
 558     // Attempt to steal work from promoted.
 559     if (task_queues()->steal(par_scan_state()->thread_num(),
 560                              par_scan_state()->hash_seed(),
 561                              obj_to_scan)) {
 562       bool res = work_q->push(obj_to_scan);
 563       assert(res, "Empty queue should have room for a push.");
 564 
 565       // If successful, goto Start.
 566       continue;
 567 
 568       // Try global overflow list.
 569     } else if (par_gen()->take_from_overflow_list(par_scan_state())) {
 570       continue;
 571     }
 572 
 573     // Otherwise, offer termination.
 574     par_scan_state()->start_term_time();
 575     if (terminator()->offer_termination()) break;
 576     par_scan_state()->end_term_time();
 577   }
 578   assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
 579          "Broken overflow list?");
 580   // Finish the last termination pause.
 581   par_scan_state()->end_term_time();
 582 }
 583 
 584 ParNewGenTask::ParNewGenTask(ParNewGeneration* young_gen,
 585                              Generation* old_gen,
 586                              HeapWord* young_old_boundary,
 587                              ParScanThreadStateSet* state_set,
 588                              StrongRootsScope* strong_roots_scope) :
 589     AbstractGangTask("ParNewGeneration collection"),
 590     _young_gen(young_gen), _old_gen(old_gen),
 591     _young_old_boundary(young_old_boundary),
 592     _state_set(state_set),
 593     _strong_roots_scope(strong_roots_scope),
 594     _par_state_string(StringTable::weak_storage())
 595 {}
 596 
 597 void ParNewGenTask::work(uint worker_id) {
 598   CMSHeap* heap = CMSHeap::heap();
 599   // Since this is being done in a separate thread, need new resource
 600   // and handle marks.
 601   ResourceMark rm;
 602   HandleMark hm;
 603 
 604   ParScanThreadState& par_scan_state = _state_set->thread_state(worker_id);
 605   assert(_state_set->is_valid(worker_id), "Should not have been called");
 606 
 607   par_scan_state.set_young_old_boundary(_young_old_boundary);
 608 
 609   CLDScanClosure cld_scan_closure(&par_scan_state.to_space_root_closure(),
 610                                   heap->rem_set()->cld_rem_set()->accumulate_modified_oops());
 611 
 612   par_scan_state.start_strong_roots();
 613   heap->young_process_roots(_strong_roots_scope,
 614                            &par_scan_state.to_space_root_closure(),
 615                            &par_scan_state.older_gen_closure(),
 616                            &cld_scan_closure,
 617                            &_par_state_string);
 618 
 619   par_scan_state.end_strong_roots();
 620 
 621   // "evacuate followers".
 622   par_scan_state.evacuate_followers_closure().do_void();
 623 
 624   // This will collapse this worker's promoted object list that's
 625   // created during the main ParNew parallel phase of ParNew. This has
 626   // to be called after all workers have finished promoting objects
 627   // and scanning promoted objects. It should be safe calling it from
 628   // here, given that we can only reach here after all thread have
 629   // offered termination, i.e., after there is no more work to be
 630   // done. It will also disable promotion tracking for the rest of
 631   // this GC as it's not necessary to be on during reference processing.
 632   _old_gen->par_oop_since_save_marks_iterate_done((int) worker_id);
 633 }
 634 
 635 ParNewGeneration::ParNewGeneration(ReservedSpace rs, size_t initial_byte_size)
 636   : DefNewGeneration(rs, initial_byte_size, "PCopy"),
 637   _overflow_list(NULL),
 638   _is_alive_closure(this),
 639   _plab_stats("Young", YoungPLABSize, PLABWeight)
 640 {
 641   NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
 642   NOT_PRODUCT(_num_par_pushes = 0;)
 643   _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
 644   guarantee(_task_queues != NULL, "task_queues allocation failure.");
 645 
 646   for (uint i = 0; i < ParallelGCThreads; i++) {
 647     ObjToScanQueue *q = new ObjToScanQueue();
 648     guarantee(q != NULL, "work_queue Allocation failure.");
 649     _task_queues->register_queue(i, q);
 650   }
 651 
 652   for (uint i = 0; i < ParallelGCThreads; i++) {
 653     _task_queues->queue(i)->initialize();
 654   }
 655 
 656   _overflow_stacks = NULL;
 657   if (ParGCUseLocalOverflow) {
 658     // typedef to workaround NEW_C_HEAP_ARRAY macro, which can not deal with ','
 659     typedef Stack<oop, mtGC> GCOopStack;
 660 
 661     _overflow_stacks = NEW_C_HEAP_ARRAY(GCOopStack, ParallelGCThreads, mtGC);
 662     for (size_t i = 0; i < ParallelGCThreads; ++i) {
 663       new (_overflow_stacks + i) Stack<oop, mtGC>();
 664     }
 665   }
 666 
 667   if (UsePerfData) {
 668     EXCEPTION_MARK;
 669     ResourceMark rm;
 670 
 671     const char* cname =
 672          PerfDataManager::counter_name(_gen_counters->name_space(), "threads");
 673     PerfDataManager::create_constant(SUN_GC, cname, PerfData::U_None,
 674                                      ParallelGCThreads, CHECK);
 675   }
 676 }
 677 
 678 // ParNewGeneration::
 679 ParKeepAliveClosure::ParKeepAliveClosure(ParScanWeakRefClosure* cl) :
 680   DefNewGeneration::KeepAliveClosure(cl), _par_cl(cl) {}
 681 
 682 template <class T>
 683 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop_work(T* p) {
 684 #ifdef ASSERT
 685   {
 686     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
 687     // We never expect to see a null reference being processed
 688     // as a weak reference.
 689     assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
 690   }
 691 #endif // ASSERT
 692 
 693   _par_cl->do_oop_nv(p);
 694 
 695   if (CMSHeap::heap()->is_in_reserved(p)) {
 696     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);;
 697     _rs->write_ref_field_gc_par(p, obj);
 698   }
 699 }
 700 
 701 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(oop* p)       { ParKeepAliveClosure::do_oop_work(p); }
 702 void /*ParNewGeneration::*/ParKeepAliveClosure::do_oop(narrowOop* p) { ParKeepAliveClosure::do_oop_work(p); }
 703 
 704 // ParNewGeneration::
 705 KeepAliveClosure::KeepAliveClosure(ScanWeakRefClosure* cl) :
 706   DefNewGeneration::KeepAliveClosure(cl) {}
 707 
 708 template <class T>
 709 void /*ParNewGeneration::*/KeepAliveClosure::do_oop_work(T* p) {
 710 #ifdef ASSERT
 711   {
 712     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
 713     // We never expect to see a null reference being processed
 714     // as a weak reference.
 715     assert(oopDesc::is_oop(obj), "expected an oop while scanning weak refs");
 716   }
 717 #endif // ASSERT
 718 
 719   _cl->do_oop_nv(p);
 720 
 721   if (CMSHeap::heap()->is_in_reserved(p)) {
 722     oop obj = RawAccess<OOP_NOT_NULL>::oop_load(p);
 723     _rs->write_ref_field_gc_par(p, obj);
 724   }
 725 }
 726 
 727 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(oop* p)       { KeepAliveClosure::do_oop_work(p); }
 728 void /*ParNewGeneration::*/KeepAliveClosure::do_oop(narrowOop* p) { KeepAliveClosure::do_oop_work(p); }
 729 
 730 template <class T> void ScanClosureWithParBarrier::do_oop_work(T* p) {
 731   T heap_oop = RawAccess<>::oop_load(p);
 732   if (!CompressedOops::is_null(heap_oop)) {
 733     oop obj = CompressedOops::decode_not_null(heap_oop);
 734     if ((HeapWord*)obj < _boundary) {
 735       assert(!_g->to()->is_in_reserved(obj), "Scanning field twice?");
 736       oop new_obj = obj->is_forwarded()
 737                       ? obj->forwardee()
 738                       : _g->DefNewGeneration::copy_to_survivor_space(obj);
 739       RawAccess<OOP_NOT_NULL>::oop_store(p, new_obj);
 740     }
 741     if (_gc_barrier) {
 742       // If p points to a younger generation, mark the card.
 743       if ((HeapWord*)obj < _gen_boundary) {
 744         _rs->write_ref_field_gc_par(p, obj);
 745       }
 746     }
 747   }
 748 }
 749 
 750 void ScanClosureWithParBarrier::do_oop(oop* p)       { ScanClosureWithParBarrier::do_oop_work(p); }
 751 void ScanClosureWithParBarrier::do_oop(narrowOop* p) { ScanClosureWithParBarrier::do_oop_work(p); }
 752 
 753 class ParNewRefProcTaskProxy: public AbstractGangTask {
 754   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
 755 public:
 756   ParNewRefProcTaskProxy(ProcessTask& task,
 757                          ParNewGeneration& young_gen,
 758                          Generation& old_gen,
 759                          HeapWord* young_old_boundary,
 760                          ParScanThreadStateSet& state_set);
 761 
 762 private:
 763   virtual void work(uint worker_id);
 764 private:
 765   ParNewGeneration&      _young_gen;
 766   ProcessTask&           _task;
 767   Generation&            _old_gen;
 768   HeapWord*              _young_old_boundary;
 769   ParScanThreadStateSet& _state_set;
 770 };
 771 
 772 ParNewRefProcTaskProxy::ParNewRefProcTaskProxy(ProcessTask& task,
 773                                                ParNewGeneration& young_gen,
 774                                                Generation& old_gen,
 775                                                HeapWord* young_old_boundary,
 776                                                ParScanThreadStateSet& state_set)
 777   : AbstractGangTask("ParNewGeneration parallel reference processing"),
 778     _young_gen(young_gen),
 779     _task(task),
 780     _old_gen(old_gen),
 781     _young_old_boundary(young_old_boundary),
 782     _state_set(state_set)
 783 { }
 784 
 785 void ParNewRefProcTaskProxy::work(uint worker_id) {
 786   ResourceMark rm;
 787   HandleMark hm;
 788   ParScanThreadState& par_scan_state = _state_set.thread_state(worker_id);
 789   par_scan_state.set_young_old_boundary(_young_old_boundary);
 790   _task.work(worker_id, par_scan_state.is_alive_closure(),
 791              par_scan_state.keep_alive_closure(),
 792              par_scan_state.evacuate_followers_closure());
 793 }
 794 
 795 void ParNewRefProcTaskExecutor::execute(ProcessTask& task) {
 796   CMSHeap* gch = CMSHeap::heap();
 797   WorkGang* workers = gch->workers();
 798   assert(workers != NULL, "Need parallel worker threads.");
 799   _state_set.reset(workers->active_workers(), _young_gen.promotion_failed());
 800   ParNewRefProcTaskProxy rp_task(task, _young_gen, _old_gen,
 801                                  _young_gen.reserved().end(), _state_set);
 802   workers->run_task(&rp_task);
 803   _state_set.reset(0 /* bad value in debug if not reset */,
 804                    _young_gen.promotion_failed());
 805 }
 806 
 807 void ParNewRefProcTaskExecutor::set_single_threaded_mode() {
 808   _state_set.flush();
 809   CMSHeap* heap = CMSHeap::heap();
 810   heap->save_marks();
 811 }
 812 
 813 ScanClosureWithParBarrier::
 814 ScanClosureWithParBarrier(ParNewGeneration* g, bool gc_barrier) :
 815   ScanClosure(g, gc_barrier)
 816 { }
 817 
 818 template <typename OopClosureType1, typename OopClosureType2>
 819 EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::
 820 EvacuateFollowersClosureGeneral(CMSHeap* heap,
 821                                 OopClosureType1* cur,
 822                                 OopClosureType2* older) :
 823   _heap(heap),
 824   _scan_cur_or_nonheap(cur), _scan_older(older)
 825 { }
 826 
 827 template <typename OopClosureType1, typename OopClosureType2>
 828 void EvacuateFollowersClosureGeneral<OopClosureType1, OopClosureType2>::do_void() {
 829   do {
 830     _heap->oop_since_save_marks_iterate(_scan_cur_or_nonheap,
 831                                         _scan_older);
 832   } while (!_heap->no_allocs_since_save_marks());
 833 }
 834 
 835 // A Generation that does parallel young-gen collection.
 836 
 837 void ParNewGeneration::handle_promotion_failed(CMSHeap* gch, ParScanThreadStateSet& thread_state_set) {
 838   assert(_promo_failure_scan_stack.is_empty(), "post condition");
 839   _promo_failure_scan_stack.clear(true); // Clear cached segments.
 840 
 841   remove_forwarding_pointers();
 842   log_info(gc, promotion)("Promotion failed");
 843   // All the spaces are in play for mark-sweep.
 844   swap_spaces();  // Make life simpler for CMS || rescan; see 6483690.
 845   from()->set_next_compaction_space(to());
 846   gch->set_incremental_collection_failed();
 847   // Inform the next generation that a promotion failure occurred.
 848   _old_gen->promotion_failure_occurred();
 849 
 850   // Trace promotion failure in the parallel GC threads
 851   thread_state_set.trace_promotion_failed(gc_tracer());
 852   // Single threaded code may have reported promotion failure to the global state
 853   if (_promotion_failed_info.has_failed()) {
 854     _gc_tracer.report_promotion_failed(_promotion_failed_info);
 855   }
 856   // Reset the PromotionFailureALot counters.
 857   NOT_PRODUCT(gch->reset_promotion_should_fail();)
 858 }
 859 
 860 void ParNewGeneration::collect(bool   full,
 861                                bool   clear_all_soft_refs,
 862                                size_t size,
 863                                bool   is_tlab) {
 864   assert(full || size > 0, "otherwise we don't want to collect");
 865 
 866   CMSHeap* gch = CMSHeap::heap();
 867 
 868   _gc_timer->register_gc_start();
 869 
 870   AdaptiveSizePolicy* size_policy = gch->size_policy();
 871   WorkGang* workers = gch->workers();
 872   assert(workers != NULL, "Need workgang for parallel work");
 873   uint active_workers =
 874        AdaptiveSizePolicy::calc_active_workers(workers->total_workers(),
 875                                                workers->active_workers(),
 876                                                Threads::number_of_non_daemon_threads());
 877   active_workers = workers->update_active_workers(active_workers);
 878   log_info(gc,task)("Using %u workers of %u for evacuation", active_workers, workers->total_workers());
 879 
 880   _old_gen = gch->old_gen();
 881 
 882   // If the next generation is too full to accommodate worst-case promotion
 883   // from this generation, pass on collection; let the next generation
 884   // do it.
 885   if (!collection_attempt_is_safe()) {
 886     gch->set_incremental_collection_failed();  // slight lie, in that we did not even attempt one
 887     return;
 888   }
 889   assert(to()->is_empty(), "Else not collection_attempt_is_safe");
 890 
 891   _gc_tracer.report_gc_start(gch->gc_cause(), _gc_timer->gc_start());
 892   gch->trace_heap_before_gc(gc_tracer());
 893 
 894   init_assuming_no_promotion_failure();
 895 
 896   if (UseAdaptiveSizePolicy) {
 897     set_survivor_overflow(false);
 898     size_policy->minor_collection_begin();
 899   }
 900 
 901   GCTraceTime(Trace, gc, phases) t1("ParNew", NULL, gch->gc_cause());
 902 
 903   age_table()->clear();
 904   to()->clear(SpaceDecorator::Mangle);
 905 
 906   gch->save_marks();
 907 
 908   // Set the correct parallelism (number of queues) in the reference processor
 909   ref_processor()->set_active_mt_degree(active_workers);
 910 
 911   // Need to initialize the preserved marks before the ThreadStateSet c'tor.
 912   _preserved_marks_set.init(active_workers);
 913 
 914   // Always set the terminator for the active number of workers
 915   // because only those workers go through the termination protocol.
 916   ParallelTaskTerminator _term(active_workers, task_queues());
 917   ParScanThreadStateSet thread_state_set(active_workers,
 918                                          *to(), *this, *_old_gen, *task_queues(),
 919                                          _overflow_stacks, _preserved_marks_set,
 920                                          desired_plab_sz(), _term);
 921 
 922   thread_state_set.reset(active_workers, promotion_failed());
 923 
 924   {
 925     StrongRootsScope srs(active_workers);
 926 
 927     ParNewGenTask tsk(this, _old_gen, reserved().end(), &thread_state_set, &srs);
 928     gch->rem_set()->prepare_for_younger_refs_iterate(true);
 929     // It turns out that even when we're using 1 thread, doing the work in a
 930     // separate thread causes wide variance in run times.  We can't help this
 931     // in the multi-threaded case, but we special-case n=1 here to get
 932     // repeatable measurements of the 1-thread overhead of the parallel code.
 933     // Might multiple workers ever be used?  If yes, initialization
 934     // has been done such that the single threaded path should not be used.
 935     if (workers->total_workers() > 1) {
 936       workers->run_task(&tsk);
 937     } else {
 938       tsk.work(0);
 939     }
 940   }
 941 
 942   thread_state_set.reset(0 /* Bad value in debug if not reset */,
 943                          promotion_failed());
 944 
 945   // Trace and reset failed promotion info.
 946   if (promotion_failed()) {
 947     thread_state_set.trace_promotion_failed(gc_tracer());
 948   }
 949 
 950   // Process (weak) reference objects found during scavenge.
 951   ReferenceProcessor* rp = ref_processor();
 952   IsAliveClosure is_alive(this);
 953   ScanWeakRefClosure scan_weak_ref(this);
 954   KeepAliveClosure keep_alive(&scan_weak_ref);
 955   ScanClosure               scan_without_gc_barrier(this, false);
 956   ScanClosureWithParBarrier scan_with_gc_barrier(this, true);
 957   set_promo_failure_scan_stack_closure(&scan_without_gc_barrier);
 958   EvacuateFollowersClosureGeneral<ScanClosure, ScanClosureWithParBarrier> evacuate_followers(
 959       gch, &scan_without_gc_barrier, &scan_with_gc_barrier);
 960   rp->setup_policy(clear_all_soft_refs);
 961   // Can  the mt_degree be set later (at run_task() time would be best)?
 962   rp->set_active_mt_degree(active_workers);
 963   ReferenceProcessorStats stats;
 964   ReferenceProcessorPhaseTimes pt(_gc_timer, rp->num_queues());
 965   if (rp->processing_is_mt()) {
 966     ParNewRefProcTaskExecutor task_executor(*this, *_old_gen, thread_state_set);
 967     stats = rp->process_discovered_references(&is_alive, &keep_alive,
 968                                               &evacuate_followers, &task_executor,
 969                                               &pt);
 970   } else {
 971     thread_state_set.flush();
 972     gch->save_marks();
 973     stats = rp->process_discovered_references(&is_alive, &keep_alive,
 974                                               &evacuate_followers, NULL,
 975                                               &pt);
 976   }
 977   _gc_tracer.report_gc_reference_stats(stats);
 978   _gc_tracer.report_tenuring_threshold(tenuring_threshold());
 979   pt.print_all_references();
 980 
 981   assert(gch->no_allocs_since_save_marks(), "evacuation should be done at this point");
 982 
 983   WeakProcessor::weak_oops_do(&is_alive, &keep_alive);
 984 
 985   // Verify that the usage of keep_alive only forwarded
 986   // the oops and did not find anything new to copy.
 987   assert(gch->no_allocs_since_save_marks(), "unexpectedly copied objects");
 988 
 989   if (!promotion_failed()) {
 990     // Swap the survivor spaces.
 991     eden()->clear(SpaceDecorator::Mangle);
 992     from()->clear(SpaceDecorator::Mangle);
 993     if (ZapUnusedHeapArea) {
 994       // This is now done here because of the piece-meal mangling which
 995       // can check for valid mangling at intermediate points in the
 996       // collection(s).  When a young collection fails to collect
 997       // sufficient space resizing of the young generation can occur
 998       // and redistribute the spaces in the young generation.  Mangle
 999       // here so that unzapped regions don't get distributed to
1000       // other spaces.
1001       to()->mangle_unused_area();
1002     }
1003     swap_spaces();
1004 
1005     // A successful scavenge should restart the GC time limit count which is
1006     // for full GC's.
1007     size_policy->reset_gc_overhead_limit_count();
1008 
1009     assert(to()->is_empty(), "to space should be empty now");
1010 
1011     adjust_desired_tenuring_threshold();
1012   } else {
1013     handle_promotion_failed(gch, thread_state_set);
1014   }
1015   _preserved_marks_set.reclaim();
1016   // set new iteration safe limit for the survivor spaces
1017   from()->set_concurrent_iteration_safe_limit(from()->top());
1018   to()->set_concurrent_iteration_safe_limit(to()->top());
1019 
1020   plab_stats()->adjust_desired_plab_sz();
1021 
1022   TASKQUEUE_STATS_ONLY(thread_state_set.print_termination_stats());
1023   TASKQUEUE_STATS_ONLY(thread_state_set.print_taskqueue_stats());
1024 
1025   if (UseAdaptiveSizePolicy) {
1026     size_policy->minor_collection_end(gch->gc_cause());
1027     size_policy->avg_survived()->sample(from()->used());
1028   }
1029 
1030   // We need to use a monotonically non-decreasing time in ms
1031   // or we will see time-warp warnings and os::javaTimeMillis()
1032   // does not guarantee monotonicity.
1033   jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC;
1034   update_time_of_last_gc(now);
1035 
1036   rp->set_enqueuing_is_done(true);
1037   rp->verify_no_references_recorded();
1038 
1039   gch->trace_heap_after_gc(gc_tracer());
1040 
1041   _gc_timer->register_gc_end();
1042 
1043   _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
1044 }
1045 
1046 size_t ParNewGeneration::desired_plab_sz() {
1047   return _plab_stats.desired_plab_sz(CMSHeap::heap()->workers()->active_workers());
1048 }
1049 
1050 static int sum;
1051 void ParNewGeneration::waste_some_time() {
1052   for (int i = 0; i < 100; i++) {
1053     sum += i;
1054   }
1055 }
1056 
1057 static const oop ClaimedForwardPtr = cast_to_oop<intptr_t>(0x4);
1058 
1059 // Because of concurrency, there are times where an object for which
1060 // "is_forwarded()" is true contains an "interim" forwarding pointer
1061 // value.  Such a value will soon be overwritten with a real value.
1062 // This method requires "obj" to have a forwarding pointer, and waits, if
1063 // necessary for a real one to be inserted, and returns it.
1064 
1065 oop ParNewGeneration::real_forwardee(oop obj) {
1066   oop forward_ptr = obj->forwardee();
1067   if (forward_ptr != ClaimedForwardPtr) {
1068     return forward_ptr;
1069   } else {
1070     return real_forwardee_slow(obj);
1071   }
1072 }
1073 
1074 oop ParNewGeneration::real_forwardee_slow(oop obj) {
1075   // Spin-read if it is claimed but not yet written by another thread.
1076   oop forward_ptr = obj->forwardee();
1077   while (forward_ptr == ClaimedForwardPtr) {
1078     waste_some_time();
1079     assert(obj->is_forwarded(), "precondition");
1080     forward_ptr = obj->forwardee();
1081   }
1082   return forward_ptr;
1083 }
1084 
1085 // Multiple GC threads may try to promote an object.  If the object
1086 // is successfully promoted, a forwarding pointer will be installed in
1087 // the object in the young generation.  This method claims the right
1088 // to install the forwarding pointer before it copies the object,
1089 // thus avoiding the need to undo the copy as in
1090 // copy_to_survivor_space_avoiding_with_undo.
1091 
1092 oop ParNewGeneration::copy_to_survivor_space(ParScanThreadState* par_scan_state,
1093                                              oop old,
1094                                              size_t sz,
1095                                              markOop m) {
1096   // In the sequential version, this assert also says that the object is
1097   // not forwarded.  That might not be the case here.  It is the case that
1098   // the caller observed it to be not forwarded at some time in the past.
1099   assert(is_in_reserved(old), "shouldn't be scavenging this oop");
1100 
1101   // The sequential code read "old->age()" below.  That doesn't work here,
1102   // since the age is in the mark word, and that might be overwritten with
1103   // a forwarding pointer by a parallel thread.  So we must save the mark
1104   // word in a local and then analyze it.
1105   oopDesc dummyOld;
1106   dummyOld.set_mark_raw(m);
1107   assert(!dummyOld.is_forwarded(),
1108          "should not be called with forwarding pointer mark word.");
1109 
1110   oop new_obj = NULL;
1111   oop forward_ptr;
1112 
1113   // Try allocating obj in to-space (unless too old)
1114   if (dummyOld.age() < tenuring_threshold()) {
1115     new_obj = (oop)par_scan_state->alloc_in_to_space(sz);
1116     if (new_obj == NULL) {
1117       set_survivor_overflow(true);
1118     }
1119   }
1120 
1121   if (new_obj == NULL) {
1122     // Either to-space is full or we decided to promote try allocating obj tenured
1123 
1124     // Attempt to install a null forwarding pointer (atomically),
1125     // to claim the right to install the real forwarding pointer.
1126     forward_ptr = old->forward_to_atomic(ClaimedForwardPtr);
1127     if (forward_ptr != NULL) {
1128       // someone else beat us to it.
1129         return real_forwardee(old);
1130     }
1131 
1132     if (!_promotion_failed) {
1133       new_obj = _old_gen->par_promote(par_scan_state->thread_num(),
1134                                       old, m, sz);
1135     }
1136 
1137     if (new_obj == NULL) {
1138       // promotion failed, forward to self
1139       _promotion_failed = true;
1140       new_obj = old;
1141 
1142       par_scan_state->preserved_marks()->push_if_necessary(old, m);
1143       par_scan_state->register_promotion_failure(sz);
1144     }
1145 
1146     old->forward_to(new_obj);
1147     forward_ptr = NULL;
1148   } else {
1149     // Is in to-space; do copying ourselves.
1150     Copy::aligned_disjoint_words((HeapWord*)old, (HeapWord*)new_obj, sz);
1151     assert(CMSHeap::heap()->is_in_reserved(new_obj), "illegal forwarding pointer value.");
1152     forward_ptr = old->forward_to_atomic(new_obj);
1153     // Restore the mark word copied above.
1154     new_obj->set_mark_raw(m);
1155     // Increment age if obj still in new generation
1156     new_obj->incr_age();
1157     par_scan_state->age_table()->add(new_obj, sz);
1158   }
1159   assert(new_obj != NULL, "just checking");
1160 
1161   // This code must come after the CAS test, or it will print incorrect
1162   // information.
1163   log_develop_trace(gc, scavenge)("{%s %s " PTR_FORMAT " -> " PTR_FORMAT " (%d)}",
1164                                   is_in_reserved(new_obj) ? "copying" : "tenuring",
1165                                   new_obj->klass()->internal_name(), p2i(old), p2i(new_obj), new_obj->size());
1166 
1167   if (forward_ptr == NULL) {
1168     oop obj_to_push = new_obj;
1169     if (par_scan_state->should_be_partially_scanned(obj_to_push, old)) {
1170       // Length field used as index of next element to be scanned.
1171       // Real length can be obtained from real_forwardee()
1172       arrayOop(old)->set_length(0);
1173       obj_to_push = old;
1174       assert(obj_to_push->is_forwarded() && obj_to_push->forwardee() != obj_to_push,
1175              "push forwarded object");
1176     }
1177     // Push it on one of the queues of to-be-scanned objects.
1178     bool simulate_overflow = false;
1179     NOT_PRODUCT(
1180       if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
1181         // simulate a stack overflow
1182         simulate_overflow = true;
1183       }
1184     )
1185     if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
1186       // Add stats for overflow pushes.
1187       log_develop_trace(gc)("Queue Overflow");
1188       push_on_overflow_list(old, par_scan_state);
1189       TASKQUEUE_STATS_ONLY(par_scan_state->taskqueue_stats().record_overflow(0));
1190     }
1191 
1192     return new_obj;
1193   }
1194 
1195   // Oops.  Someone beat us to it.  Undo the allocation.  Where did we
1196   // allocate it?
1197   if (is_in_reserved(new_obj)) {
1198     // Must be in to_space.
1199     assert(to()->is_in_reserved(new_obj), "Checking");
1200     if (forward_ptr == ClaimedForwardPtr) {
1201       // Wait to get the real forwarding pointer value.
1202       forward_ptr = real_forwardee(old);
1203     }
1204     par_scan_state->undo_alloc_in_to_space((HeapWord*)new_obj, sz);
1205   }
1206 
1207   return forward_ptr;
1208 }
1209 
1210 #ifndef PRODUCT
1211 // It's OK to call this multi-threaded;  the worst thing
1212 // that can happen is that we'll get a bunch of closely
1213 // spaced simulated overflows, but that's OK, in fact
1214 // probably good as it would exercise the overflow code
1215 // under contention.
1216 bool ParNewGeneration::should_simulate_overflow() {
1217   if (_overflow_counter-- <= 0) { // just being defensive
1218     _overflow_counter = ParGCWorkQueueOverflowInterval;
1219     return true;
1220   } else {
1221     return false;
1222   }
1223 }
1224 #endif
1225 
1226 // In case we are using compressed oops, we need to be careful.
1227 // If the object being pushed is an object array, then its length
1228 // field keeps track of the "grey boundary" at which the next
1229 // incremental scan will be done (see ParGCArrayScanChunk).
1230 // When using compressed oops, this length field is kept in the
1231 // lower 32 bits of the erstwhile klass word and cannot be used
1232 // for the overflow chaining pointer (OCP below). As such the OCP
1233 // would itself need to be compressed into the top 32-bits in this
1234 // case. Unfortunately, see below, in the event that we have a
1235 // promotion failure, the node to be pushed on the list can be
1236 // outside of the Java heap, so the heap-based pointer compression
1237 // would not work (we would have potential aliasing between C-heap
1238 // and Java-heap pointers). For this reason, when using compressed
1239 // oops, we simply use a worker-thread-local, non-shared overflow
1240 // list in the form of a growable array, with a slightly different
1241 // overflow stack draining strategy. If/when we start using fat
1242 // stacks here, we can go back to using (fat) pointer chains
1243 // (although some performance comparisons would be useful since
1244 // single global lists have their own performance disadvantages
1245 // as we were made painfully aware not long ago, see 6786503).
1246 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff))
1247 void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
1248   assert(is_in_reserved(from_space_obj), "Should be from this generation");
1249   if (ParGCUseLocalOverflow) {
1250     // In the case of compressed oops, we use a private, not-shared
1251     // overflow stack.
1252     par_scan_state->push_on_overflow_stack(from_space_obj);
1253   } else {
1254     assert(!UseCompressedOops, "Error");
1255     // if the object has been forwarded to itself, then we cannot
1256     // use the klass pointer for the linked list.  Instead we have
1257     // to allocate an oopDesc in the C-Heap and use that for the linked list.
1258     // XXX This is horribly inefficient when a promotion failure occurs
1259     // and should be fixed. XXX FIX ME !!!
1260 #ifndef PRODUCT
1261     Atomic::inc(&_num_par_pushes);
1262     assert(_num_par_pushes > 0, "Tautology");
1263 #endif
1264     if (from_space_obj->forwardee() == from_space_obj) {
1265       oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1, mtGC);
1266       listhead->forward_to(from_space_obj);
1267       from_space_obj = listhead;
1268     }
1269     oop observed_overflow_list = _overflow_list;
1270     oop cur_overflow_list;
1271     do {
1272       cur_overflow_list = observed_overflow_list;
1273       if (cur_overflow_list != BUSY) {
1274         from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
1275       } else {
1276         from_space_obj->set_klass_to_list_ptr(NULL);
1277       }
1278       observed_overflow_list =
1279         Atomic::cmpxchg((oopDesc*)from_space_obj, &_overflow_list, (oopDesc*)cur_overflow_list);
1280     } while (cur_overflow_list != observed_overflow_list);
1281   }
1282 }
1283 
1284 bool ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
1285   bool res;
1286 
1287   if (ParGCUseLocalOverflow) {
1288     res = par_scan_state->take_from_overflow_stack();
1289   } else {
1290     assert(!UseCompressedOops, "Error");
1291     res = take_from_overflow_list_work(par_scan_state);
1292   }
1293   return res;
1294 }
1295 
1296 
1297 // *NOTE*: The overflow list manipulation code here and
1298 // in CMSCollector:: are very similar in shape,
1299 // except that in the CMS case we thread the objects
1300 // directly into the list via their mark word, and do
1301 // not need to deal with special cases below related
1302 // to chunking of object arrays and promotion failure
1303 // handling.
1304 // CR 6797058 has been filed to attempt consolidation of
1305 // the common code.
1306 // Because of the common code, if you make any changes in
1307 // the code below, please check the CMS version to see if
1308 // similar changes might be needed.
1309 // See CMSCollector::par_take_from_overflow_list() for
1310 // more extensive documentation comments.
1311 bool ParNewGeneration::take_from_overflow_list_work(ParScanThreadState* par_scan_state) {
1312   ObjToScanQueue* work_q = par_scan_state->work_queue();
1313   // How many to take?
1314   size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
1315                                  (size_t)ParGCDesiredObjsFromOverflowList);
1316 
1317   assert(!UseCompressedOops, "Error");
1318   assert(par_scan_state->overflow_stack() == NULL, "Error");
1319   if (_overflow_list == NULL) return false;
1320 
1321   // Otherwise, there was something there; try claiming the list.
1322   oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1323   // Trim off a prefix of at most objsFromOverflow items
1324   Thread* tid = Thread::current();
1325   size_t spin_count = ParallelGCThreads;
1326   size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
1327   for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
1328     // someone grabbed it before we did ...
1329     // ... we spin for a short while...
1330     os::sleep(tid, sleep_time_millis, false);
1331     if (_overflow_list == NULL) {
1332       // nothing left to take
1333       return false;
1334     } else if (_overflow_list != BUSY) {
1335      // try and grab the prefix
1336      prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list));
1337     }
1338   }
1339   if (prefix == NULL || prefix == BUSY) {
1340      // Nothing to take or waited long enough
1341      if (prefix == NULL) {
1342        // Write back the NULL in case we overwrote it with BUSY above
1343        // and it is still the same value.
1344        (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1345      }
1346      return false;
1347   }
1348   assert(prefix != NULL && prefix != BUSY, "Error");
1349   oop cur = prefix;
1350   for (size_t i = 1; i < objsFromOverflow; ++i) {
1351     oop next = cur->list_ptr_from_klass();
1352     if (next == NULL) break;
1353     cur = next;
1354   }
1355   assert(cur != NULL, "Loop postcondition");
1356 
1357   // Reattach remaining (suffix) to overflow list
1358   oop suffix = cur->list_ptr_from_klass();
1359   if (suffix == NULL) {
1360     // Write back the NULL in lieu of the BUSY we wrote
1361     // above and it is still the same value.
1362     if (_overflow_list == BUSY) {
1363       (void) Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY);
1364     }
1365   } else {
1366     assert(suffix != BUSY, "Error");
1367     // suffix will be put back on global list
1368     cur->set_klass_to_list_ptr(NULL);     // break off suffix
1369     // It's possible that the list is still in the empty(busy) state
1370     // we left it in a short while ago; in that case we may be
1371     // able to place back the suffix.
1372     oop observed_overflow_list = _overflow_list;
1373     oop cur_overflow_list = observed_overflow_list;
1374     bool attached = false;
1375     while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
1376       observed_overflow_list =
1377         Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1378       if (cur_overflow_list == observed_overflow_list) {
1379         attached = true;
1380         break;
1381       } else cur_overflow_list = observed_overflow_list;
1382     }
1383     if (!attached) {
1384       // Too bad, someone else got in in between; we'll need to do a splice.
1385       // Find the last item of suffix list
1386       oop last = suffix;
1387       while (true) {
1388         oop next = last->list_ptr_from_klass();
1389         if (next == NULL) break;
1390         last = next;
1391       }
1392       // Atomically prepend suffix to current overflow list
1393       observed_overflow_list = _overflow_list;
1394       do {
1395         cur_overflow_list = observed_overflow_list;
1396         if (cur_overflow_list != BUSY) {
1397           // Do the splice ...
1398           last->set_klass_to_list_ptr(cur_overflow_list);
1399         } else { // cur_overflow_list == BUSY
1400           last->set_klass_to_list_ptr(NULL);
1401         }
1402         observed_overflow_list =
1403           Atomic::cmpxchg((oopDesc*)suffix, &_overflow_list, (oopDesc*)cur_overflow_list);
1404       } while (cur_overflow_list != observed_overflow_list);
1405     }
1406   }
1407 
1408   // Push objects on prefix list onto this thread's work queue
1409   assert(prefix != NULL && prefix != BUSY, "program logic");
1410   cur = prefix;
1411   ssize_t n = 0;
1412   while (cur != NULL) {
1413     oop obj_to_push = cur->forwardee();
1414     oop next        = cur->list_ptr_from_klass();
1415     cur->set_klass(obj_to_push->klass());
1416     // This may be an array object that is self-forwarded. In that case, the list pointer
1417     // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
1418     if (!is_in_reserved(cur)) {
1419       // This can become a scaling bottleneck when there is work queue overflow coincident
1420       // with promotion failure.
1421       oopDesc* f = cur;
1422       FREE_C_HEAP_ARRAY(oopDesc, f);
1423     } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
1424       assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
1425       obj_to_push = cur;
1426     }
1427     bool ok = work_q->push(obj_to_push);
1428     assert(ok, "Should have succeeded");
1429     cur = next;
1430     n++;
1431   }
1432   TASKQUEUE_STATS_ONLY(par_scan_state->note_overflow_refill(n));
1433 #ifndef PRODUCT
1434   assert(_num_par_pushes >= n, "Too many pops?");
1435   Atomic::sub(n, &_num_par_pushes);
1436 #endif
1437   return true;
1438 }
1439 #undef BUSY
1440 
1441 void ParNewGeneration::ref_processor_init() {
1442   if (_ref_processor == NULL) {
1443     // Allocate and initialize a reference processor
1444     _span_based_discoverer.set_span(_reserved);
1445     _ref_processor =
1446       new ReferenceProcessor(&_span_based_discoverer,    // span
1447                              ParallelRefProcEnabled && (ParallelGCThreads > 1), // mt processing
1448                              ParallelGCThreads,          // mt processing degree
1449                              refs_discovery_is_mt(),     // mt discovery
1450                              ParallelGCThreads,          // mt discovery degree
1451                              refs_discovery_is_atomic(), // atomic_discovery
1452                              NULL);                      // is_alive_non_header
1453   }
1454 }
1455 
1456 const char* ParNewGeneration::name() const {
1457   return "par new generation";
1458 }
1459 
1460 void ParNewGeneration::restore_preserved_marks() {
1461   SharedRestorePreservedMarksTaskExecutor task_executor(CMSHeap::heap()->workers());
1462   _preserved_marks_set.restore(&task_executor);
1463 }