1 /*
   2  * Copyright (c) 2001, 2012, 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/symbolTable.hpp"
  27 #include "gc_implementation/g1/concurrentMark.inline.hpp"
  28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
  29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
  30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
  31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
  32 #include "gc_implementation/g1/g1Log.hpp"
  33 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
  34 #include "gc_implementation/g1/g1RemSet.hpp"
  35 #include "gc_implementation/g1/heapRegion.inline.hpp"
  36 #include "gc_implementation/g1/heapRegionRemSet.hpp"
  37 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
  38 #include "gc_implementation/shared/vmGCOperations.hpp"
  39 #include "memory/genOopClosures.inline.hpp"
  40 #include "memory/referencePolicy.hpp"
  41 #include "memory/resourceArea.hpp"
  42 #include "oops/oop.inline.hpp"
  43 #include "runtime/handles.inline.hpp"
  44 #include "runtime/java.hpp"
  45 #include "services/memTracker.hpp"
  46 
  47 // Concurrent marking bit map wrapper
  48 
  49 CMBitMapRO::CMBitMapRO(ReservedSpace rs, int shifter) :
  50   _bm((uintptr_t*)NULL,0),
  51   _shifter(shifter) {
  52   _bmStartWord = (HeapWord*)(rs.base());
  53   _bmWordSize  = rs.size()/HeapWordSize;    // rs.size() is in bytes
  54   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
  55                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
  56 
  57   MemTracker::record_virtual_memory_type((address)brs.base(), mtGC);
  58 
  59   guarantee(brs.is_reserved(), "couldn't allocate concurrent marking bit map");
  60   // For now we'll just commit all of the bit map up fromt.
  61   // Later on we'll try to be more parsimonious with swap.
  62   guarantee(_virtual_space.initialize(brs, brs.size()),
  63             "couldn't reseve backing store for concurrent marking bit map");
  64   assert(_virtual_space.committed_size() == brs.size(),
  65          "didn't reserve backing store for all of concurrent marking bit map?");
  66   _bm.set_map((uintptr_t*)_virtual_space.low());
  67   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
  68          _bmWordSize, "inconsistency in bit map sizing");
  69   _bm.set_size(_bmWordSize >> _shifter);
  70 }
  71 
  72 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr,
  73                                                HeapWord* limit) const {
  74   // First we must round addr *up* to a possible object boundary.
  75   addr = (HeapWord*)align_size_up((intptr_t)addr,
  76                                   HeapWordSize << _shifter);
  77   size_t addrOffset = heapWordToOffset(addr);
  78   if (limit == NULL) {
  79     limit = _bmStartWord + _bmWordSize;
  80   }
  81   size_t limitOffset = heapWordToOffset(limit);
  82   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  83   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  84   assert(nextAddr >= addr, "get_next_one postcondition");
  85   assert(nextAddr == limit || isMarked(nextAddr),
  86          "get_next_one postcondition");
  87   return nextAddr;
  88 }
  89 
  90 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr,
  91                                                  HeapWord* limit) const {
  92   size_t addrOffset = heapWordToOffset(addr);
  93   if (limit == NULL) {
  94     limit = _bmStartWord + _bmWordSize;
  95   }
  96   size_t limitOffset = heapWordToOffset(limit);
  97   size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
  98   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  99   assert(nextAddr >= addr, "get_next_one postcondition");
 100   assert(nextAddr == limit || !isMarked(nextAddr),
 101          "get_next_one postcondition");
 102   return nextAddr;
 103 }
 104 
 105 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
 106   assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
 107   return (int) (diff >> _shifter);
 108 }
 109 
 110 #ifndef PRODUCT
 111 bool CMBitMapRO::covers(ReservedSpace rs) const {
 112   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
 113   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
 114          "size inconsistency");
 115   return _bmStartWord == (HeapWord*)(rs.base()) &&
 116          _bmWordSize  == rs.size()>>LogHeapWordSize;
 117 }
 118 #endif
 119 
 120 void CMBitMap::clearAll() {
 121   _bm.clear();
 122   return;
 123 }
 124 
 125 void CMBitMap::markRange(MemRegion mr) {
 126   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 127   assert(!mr.is_empty(), "unexpected empty region");
 128   assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
 129           ((HeapWord *) mr.end())),
 130          "markRange memory region end is not card aligned");
 131   // convert address range into offset range
 132   _bm.at_put_range(heapWordToOffset(mr.start()),
 133                    heapWordToOffset(mr.end()), true);
 134 }
 135 
 136 void CMBitMap::clearRange(MemRegion mr) {
 137   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 138   assert(!mr.is_empty(), "unexpected empty region");
 139   // convert address range into offset range
 140   _bm.at_put_range(heapWordToOffset(mr.start()),
 141                    heapWordToOffset(mr.end()), false);
 142 }
 143 
 144 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
 145                                             HeapWord* end_addr) {
 146   HeapWord* start = getNextMarkedWordAddress(addr);
 147   start = MIN2(start, end_addr);
 148   HeapWord* end   = getNextUnmarkedWordAddress(start);
 149   end = MIN2(end, end_addr);
 150   assert(start <= end, "Consistency check");
 151   MemRegion mr(start, end);
 152   if (!mr.is_empty()) {
 153     clearRange(mr);
 154   }
 155   return mr;
 156 }
 157 
 158 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
 159   _base(NULL), _cm(cm)
 160 #ifdef ASSERT
 161   , _drain_in_progress(false)
 162   , _drain_in_progress_yields(false)
 163 #endif
 164 {}
 165 
 166 void CMMarkStack::allocate(size_t size) {
 167   _base = NEW_C_HEAP_ARRAY(oop, size, mtGC);
 168   if (_base == NULL) {
 169     vm_exit_during_initialization("Failed to allocate CM region mark stack");
 170   }
 171   _index = 0;
 172   _capacity = (jint) size;
 173   _saved_index = -1;
 174   NOT_PRODUCT(_max_depth = 0);
 175 }
 176 
 177 CMMarkStack::~CMMarkStack() {
 178   if (_base != NULL) {
 179     FREE_C_HEAP_ARRAY(oop, _base, mtGC);
 180   }
 181 }
 182 
 183 void CMMarkStack::par_push(oop ptr) {
 184   while (true) {
 185     if (isFull()) {
 186       _overflow = true;
 187       return;
 188     }
 189     // Otherwise...
 190     jint index = _index;
 191     jint next_index = index+1;
 192     jint res = Atomic::cmpxchg(next_index, &_index, index);
 193     if (res == index) {
 194       _base[index] = ptr;
 195       // Note that we don't maintain this atomically.  We could, but it
 196       // doesn't seem necessary.
 197       NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 198       return;
 199     }
 200     // Otherwise, we need to try again.
 201   }
 202 }
 203 
 204 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
 205   while (true) {
 206     if (isFull()) {
 207       _overflow = true;
 208       return;
 209     }
 210     // Otherwise...
 211     jint index = _index;
 212     jint next_index = index + n;
 213     if (next_index > _capacity) {
 214       _overflow = true;
 215       return;
 216     }
 217     jint res = Atomic::cmpxchg(next_index, &_index, index);
 218     if (res == index) {
 219       for (int i = 0; i < n; i++) {
 220         int ind = index + i;
 221         assert(ind < _capacity, "By overflow test above.");
 222         _base[ind] = ptr_arr[i];
 223       }
 224       NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 225       return;
 226     }
 227     // Otherwise, we need to try again.
 228   }
 229 }
 230 
 231 
 232 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 233   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 234   jint start = _index;
 235   jint next_index = start + n;
 236   if (next_index > _capacity) {
 237     _overflow = true;
 238     return;
 239   }
 240   // Otherwise.
 241   _index = next_index;
 242   for (int i = 0; i < n; i++) {
 243     int ind = start + i;
 244     assert(ind < _capacity, "By overflow test above.");
 245     _base[ind] = ptr_arr[i];
 246   }
 247 }
 248 
 249 
 250 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 251   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 252   jint index = _index;
 253   if (index == 0) {
 254     *n = 0;
 255     return false;
 256   } else {
 257     int k = MIN2(max, index);
 258     jint new_ind = index - k;
 259     for (int j = 0; j < k; j++) {
 260       ptr_arr[j] = _base[new_ind + j];
 261     }
 262     _index = new_ind;
 263     *n = k;
 264     return true;
 265   }
 266 }
 267 
 268 template<class OopClosureClass>
 269 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
 270   assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
 271          || SafepointSynchronize::is_at_safepoint(),
 272          "Drain recursion must be yield-safe.");
 273   bool res = true;
 274   debug_only(_drain_in_progress = true);
 275   debug_only(_drain_in_progress_yields = yield_after);
 276   while (!isEmpty()) {
 277     oop newOop = pop();
 278     assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
 279     assert(newOop->is_oop(), "Expected an oop");
 280     assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
 281            "only grey objects on this stack");
 282     newOop->oop_iterate(cl);
 283     if (yield_after && _cm->do_yield_check()) {
 284       res = false;
 285       break;
 286     }
 287   }
 288   debug_only(_drain_in_progress = false);
 289   return res;
 290 }
 291 
 292 void CMMarkStack::note_start_of_gc() {
 293   assert(_saved_index == -1,
 294          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 295   _saved_index = _index;
 296 }
 297 
 298 void CMMarkStack::note_end_of_gc() {
 299   // This is intentionally a guarantee, instead of an assert. If we
 300   // accidentally add something to the mark stack during GC, it
 301   // will be a correctness issue so it's better if we crash. we'll
 302   // only check this once per GC anyway, so it won't be a performance
 303   // issue in any way.
 304   guarantee(_saved_index == _index,
 305             err_msg("saved index: %d index: %d", _saved_index, _index));
 306   _saved_index = -1;
 307 }
 308 
 309 void CMMarkStack::oops_do(OopClosure* f) {
 310   assert(_saved_index == _index,
 311          err_msg("saved index: %d index: %d", _saved_index, _index));
 312   for (int i = 0; i < _index; i += 1) {
 313     f->do_oop(&_base[i]);
 314   }
 315 }
 316 
 317 bool ConcurrentMark::not_yet_marked(oop obj) const {
 318   return (_g1h->is_obj_ill(obj)
 319           || (_g1h->is_in_permanent(obj)
 320               && !nextMarkBitMap()->isMarked((HeapWord*)obj)));
 321 }
 322 
 323 CMRootRegions::CMRootRegions() :
 324   _young_list(NULL), _cm(NULL), _scan_in_progress(false),
 325   _should_abort(false),  _next_survivor(NULL) { }
 326 
 327 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
 328   _young_list = g1h->young_list();
 329   _cm = cm;
 330 }
 331 
 332 void CMRootRegions::prepare_for_scan() {
 333   assert(!scan_in_progress(), "pre-condition");
 334 
 335   // Currently, only survivors can be root regions.
 336   assert(_next_survivor == NULL, "pre-condition");
 337   _next_survivor = _young_list->first_survivor_region();
 338   _scan_in_progress = (_next_survivor != NULL);
 339   _should_abort = false;
 340 }
 341 
 342 HeapRegion* CMRootRegions::claim_next() {
 343   if (_should_abort) {
 344     // If someone has set the should_abort flag, we return NULL to
 345     // force the caller to bail out of their loop.
 346     return NULL;
 347   }
 348 
 349   // Currently, only survivors can be root regions.
 350   HeapRegion* res = _next_survivor;
 351   if (res != NULL) {
 352     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 353     // Read it again in case it changed while we were waiting for the lock.
 354     res = _next_survivor;
 355     if (res != NULL) {
 356       if (res == _young_list->last_survivor_region()) {
 357         // We just claimed the last survivor so store NULL to indicate
 358         // that we're done.
 359         _next_survivor = NULL;
 360       } else {
 361         _next_survivor = res->get_next_young_region();
 362       }
 363     } else {
 364       // Someone else claimed the last survivor while we were trying
 365       // to take the lock so nothing else to do.
 366     }
 367   }
 368   assert(res == NULL || res->is_survivor(), "post-condition");
 369 
 370   return res;
 371 }
 372 
 373 void CMRootRegions::scan_finished() {
 374   assert(scan_in_progress(), "pre-condition");
 375 
 376   // Currently, only survivors can be root regions.
 377   if (!_should_abort) {
 378     assert(_next_survivor == NULL, "we should have claimed all survivors");
 379   }
 380   _next_survivor = NULL;
 381 
 382   {
 383     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 384     _scan_in_progress = false;
 385     RootRegionScan_lock->notify_all();
 386   }
 387 }
 388 
 389 bool CMRootRegions::wait_until_scan_finished() {
 390   if (!scan_in_progress()) return false;
 391 
 392   {
 393     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 394     while (scan_in_progress()) {
 395       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 396     }
 397   }
 398   return true;
 399 }
 400 
 401 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
 402 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
 403 #endif // _MSC_VER
 404 
 405 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 406   return MAX2((n_par_threads + 2) / 4, 1U);
 407 }
 408 
 409 ConcurrentMark::ConcurrentMark(ReservedSpace rs, uint max_regions) :
 410   _markBitMap1(rs, MinObjAlignment - 1),
 411   _markBitMap2(rs, MinObjAlignment - 1),
 412 
 413   _parallel_marking_threads(0),
 414   _max_parallel_marking_threads(0),
 415   _sleep_factor(0.0),
 416   _marking_task_overhead(1.0),
 417   _cleanup_sleep_factor(0.0),
 418   _cleanup_task_overhead(1.0),
 419   _cleanup_list("Cleanup List"),
 420   _region_bm((BitMap::idx_t) max_regions, false /* in_resource_area*/),
 421   _card_bm((rs.size() + CardTableModRefBS::card_size - 1) >>
 422            CardTableModRefBS::card_shift,
 423            false /* in_resource_area*/),
 424 
 425   _prevMarkBitMap(&_markBitMap1),
 426   _nextMarkBitMap(&_markBitMap2),
 427 
 428   _markStack(this),
 429   // _finger set in set_non_marking_state
 430 
 431   _max_task_num(MAX2((uint)ParallelGCThreads, 1U)),
 432   // _active_tasks set in set_non_marking_state
 433   // _tasks set inside the constructor
 434   _task_queues(new CMTaskQueueSet((int) _max_task_num)),
 435   _terminator(ParallelTaskTerminator((int) _max_task_num, _task_queues)),
 436 
 437   _has_overflown(false),
 438   _concurrent(false),
 439   _has_aborted(false),
 440   _restart_for_overflow(false),
 441   _concurrent_marking_in_progress(false),
 442 
 443   // _verbose_level set below
 444 
 445   _init_times(),
 446   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 447   _cleanup_times(),
 448   _total_counting_time(0.0),
 449   _total_rs_scrub_time(0.0),
 450 
 451   _parallel_workers(NULL),
 452 
 453   _count_card_bitmaps(NULL),
 454   _count_marked_bytes(NULL) {
 455   CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
 456   if (verbose_level < no_verbose) {
 457     verbose_level = no_verbose;
 458   }
 459   if (verbose_level > high_verbose) {
 460     verbose_level = high_verbose;
 461   }
 462   _verbose_level = verbose_level;
 463 
 464   if (verbose_low()) {
 465     gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
 466                            "heap end = "PTR_FORMAT, _heap_start, _heap_end);
 467   }
 468 
 469   _markStack.allocate(MarkStackSize);
 470 
 471   // Create & start a ConcurrentMark thread.
 472   _cmThread = new ConcurrentMarkThread(this);
 473   assert(cmThread() != NULL, "CM Thread should have been created");
 474   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 475 
 476   _g1h = G1CollectedHeap::heap();
 477   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 478   assert(_markBitMap1.covers(rs), "_markBitMap1 inconsistency");
 479   assert(_markBitMap2.covers(rs), "_markBitMap2 inconsistency");
 480 
 481   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 482   satb_qs.set_buffer_size(G1SATBBufferSize);
 483 
 484   _root_regions.init(_g1h, this);
 485 
 486   _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_task_num, mtGC);
 487   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_task_num, mtGC);
 488 
 489   _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap,  _max_task_num, mtGC);
 490   _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_task_num, mtGC);
 491 
 492   BitMap::idx_t card_bm_size = _card_bm.size();
 493 
 494   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 495   _active_tasks = _max_task_num;
 496   for (int i = 0; i < (int) _max_task_num; ++i) {
 497     CMTaskQueue* task_queue = new CMTaskQueue();
 498     task_queue->initialize();
 499     _task_queues->register_queue(i, task_queue);
 500 
 501     _count_card_bitmaps[i] = BitMap(card_bm_size, false);
 502     _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, (size_t) max_regions, mtGC);
 503 
 504     _tasks[i] = new CMTask(i, this,
 505                            _count_marked_bytes[i],
 506                            &_count_card_bitmaps[i],
 507                            task_queue, _task_queues);
 508 
 509     _accum_task_vtime[i] = 0.0;
 510   }
 511 
 512   // Calculate the card number for the bottom of the heap. Used
 513   // in biasing indexes into the accounting card bitmaps.
 514   _heap_bottom_card_num =
 515     intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
 516                                 CardTableModRefBS::card_shift);
 517 
 518   // Clear all the liveness counting data
 519   clear_all_count_data();
 520 
 521   if (ConcGCThreads > ParallelGCThreads) {
 522     vm_exit_during_initialization("Can't have more ConcGCThreads "
 523                                   "than ParallelGCThreads.");
 524   }
 525   if (ParallelGCThreads == 0) {
 526     // if we are not running with any parallel GC threads we will not
 527     // spawn any marking threads either
 528     _parallel_marking_threads =       0;
 529     _max_parallel_marking_threads =   0;
 530     _sleep_factor             =     0.0;
 531     _marking_task_overhead    =     1.0;
 532   } else {
 533     if (ConcGCThreads > 0) {
 534       // notice that ConcGCThreads overwrites G1MarkingOverheadPercent
 535       // if both are set
 536 
 537       _parallel_marking_threads = (uint) ConcGCThreads;
 538       _max_parallel_marking_threads = _parallel_marking_threads;
 539       _sleep_factor             = 0.0;
 540       _marking_task_overhead    = 1.0;
 541     } else if (G1MarkingOverheadPercent > 0) {
 542       // we will calculate the number of parallel marking threads
 543       // based on a target overhead with respect to the soft real-time
 544       // goal
 545 
 546       double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 547       double overall_cm_overhead =
 548         (double) MaxGCPauseMillis * marking_overhead /
 549         (double) GCPauseIntervalMillis;
 550       double cpu_ratio = 1.0 / (double) os::processor_count();
 551       double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 552       double marking_task_overhead =
 553         overall_cm_overhead / marking_thread_num *
 554                                                 (double) os::processor_count();
 555       double sleep_factor =
 556                          (1.0 - marking_task_overhead) / marking_task_overhead;
 557 
 558       _parallel_marking_threads = (uint) marking_thread_num;
 559       _max_parallel_marking_threads = _parallel_marking_threads;
 560       _sleep_factor             = sleep_factor;
 561       _marking_task_overhead    = marking_task_overhead;
 562     } else {
 563       _parallel_marking_threads = scale_parallel_threads((uint)ParallelGCThreads);
 564       _max_parallel_marking_threads = _parallel_marking_threads;
 565       _sleep_factor             = 0.0;
 566       _marking_task_overhead    = 1.0;
 567     }
 568 
 569     if (parallel_marking_threads() > 1) {
 570       _cleanup_task_overhead = 1.0;
 571     } else {
 572       _cleanup_task_overhead = marking_task_overhead();
 573     }
 574     _cleanup_sleep_factor =
 575                      (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
 576 
 577 #if 0
 578     gclog_or_tty->print_cr("Marking Threads          %d", parallel_marking_threads());
 579     gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
 580     gclog_or_tty->print_cr("CM Sleep Factor          %1.4lf", sleep_factor());
 581     gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
 582     gclog_or_tty->print_cr("CL Sleep Factor          %1.4lf", cleanup_sleep_factor());
 583 #endif
 584 
 585     guarantee(parallel_marking_threads() > 0, "peace of mind");
 586     _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
 587          _max_parallel_marking_threads, false, true);
 588     if (_parallel_workers == NULL) {
 589       vm_exit_during_initialization("Failed necessary allocation.");
 590     } else {
 591       _parallel_workers->initialize_workers();
 592     }
 593   }
 594 
 595   // so that the call below can read a sensible value
 596   _heap_start = (HeapWord*) rs.base();
 597   set_non_marking_state();
 598 }
 599 
 600 void ConcurrentMark::update_g1_committed(bool force) {
 601   // If concurrent marking is not in progress, then we do not need to
 602   // update _heap_end.
 603   if (!concurrent_marking_in_progress() && !force) return;
 604 
 605   MemRegion committed = _g1h->g1_committed();
 606   assert(committed.start() == _heap_start, "start shouldn't change");
 607   HeapWord* new_end = committed.end();
 608   if (new_end > _heap_end) {
 609     // The heap has been expanded.
 610 
 611     _heap_end = new_end;
 612   }
 613   // Notice that the heap can also shrink. However, this only happens
 614   // during a Full GC (at least currently) and the entire marking
 615   // phase will bail out and the task will not be restarted. So, let's
 616   // do nothing.
 617 }
 618 
 619 void ConcurrentMark::reset() {
 620   // Starting values for these two. This should be called in a STW
 621   // phase. CM will be notified of any future g1_committed expansions
 622   // will be at the end of evacuation pauses, when tasks are
 623   // inactive.
 624   MemRegion committed = _g1h->g1_committed();
 625   _heap_start = committed.start();
 626   _heap_end   = committed.end();
 627 
 628   // Separated the asserts so that we know which one fires.
 629   assert(_heap_start != NULL, "heap bounds should look ok");
 630   assert(_heap_end != NULL, "heap bounds should look ok");
 631   assert(_heap_start < _heap_end, "heap bounds should look ok");
 632 
 633   // reset all the marking data structures and any necessary flags
 634   clear_marking_state();
 635 
 636   if (verbose_low()) {
 637     gclog_or_tty->print_cr("[global] resetting");
 638   }
 639 
 640   // We do reset all of them, since different phases will use
 641   // different number of active threads. So, it's easiest to have all
 642   // of them ready.
 643   for (int i = 0; i < (int) _max_task_num; ++i) {
 644     _tasks[i]->reset(_nextMarkBitMap);
 645   }
 646 
 647   // we need this to make sure that the flag is on during the evac
 648   // pause with initial mark piggy-backed
 649   set_concurrent_marking_in_progress();
 650 }
 651 
 652 void ConcurrentMark::set_phase(uint active_tasks, bool concurrent) {
 653   assert(active_tasks <= _max_task_num, "we should not have more");
 654 
 655   _active_tasks = active_tasks;
 656   // Need to update the three data structures below according to the
 657   // number of active threads for this phase.
 658   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 659   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 660   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 661 
 662   _concurrent = concurrent;
 663   // We propagate this to all tasks, not just the active ones.
 664   for (int i = 0; i < (int) _max_task_num; ++i)
 665     _tasks[i]->set_concurrent(concurrent);
 666 
 667   if (concurrent) {
 668     set_concurrent_marking_in_progress();
 669   } else {
 670     // We currently assume that the concurrent flag has been set to
 671     // false before we start remark. At this point we should also be
 672     // in a STW phase.
 673     assert(!concurrent_marking_in_progress(), "invariant");
 674     assert(_finger == _heap_end, "only way to get here");
 675     update_g1_committed(true);
 676   }
 677 }
 678 
 679 void ConcurrentMark::set_non_marking_state() {
 680   // We set the global marking state to some default values when we're
 681   // not doing marking.
 682   clear_marking_state();
 683   _active_tasks = 0;
 684   clear_concurrent_marking_in_progress();
 685 }
 686 
 687 ConcurrentMark::~ConcurrentMark() {
 688   // The ConcurrentMark instance is never freed.
 689   ShouldNotReachHere();
 690 }
 691 
 692 void ConcurrentMark::clearNextBitmap() {
 693   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 694   G1CollectorPolicy* g1p = g1h->g1_policy();
 695 
 696   // Make sure that the concurrent mark thread looks to still be in
 697   // the current cycle.
 698   guarantee(cmThread()->during_cycle(), "invariant");
 699 
 700   // We are finishing up the current cycle by clearing the next
 701   // marking bitmap and getting it ready for the next cycle. During
 702   // this time no other cycle can start. So, let's make sure that this
 703   // is the case.
 704   guarantee(!g1h->mark_in_progress(), "invariant");
 705 
 706   // clear the mark bitmap (no grey objects to start with).
 707   // We need to do this in chunks and offer to yield in between
 708   // each chunk.
 709   HeapWord* start  = _nextMarkBitMap->startWord();
 710   HeapWord* end    = _nextMarkBitMap->endWord();
 711   HeapWord* cur    = start;
 712   size_t chunkSize = M;
 713   while (cur < end) {
 714     HeapWord* next = cur + chunkSize;
 715     if (next > end) {
 716       next = end;
 717     }
 718     MemRegion mr(cur,next);
 719     _nextMarkBitMap->clearRange(mr);
 720     cur = next;
 721     do_yield_check();
 722 
 723     // Repeat the asserts from above. We'll do them as asserts here to
 724     // minimize their overhead on the product. However, we'll have
 725     // them as guarantees at the beginning / end of the bitmap
 726     // clearing to get some checking in the product.
 727     assert(cmThread()->during_cycle(), "invariant");
 728     assert(!g1h->mark_in_progress(), "invariant");
 729   }
 730 
 731   // Clear the liveness counting data
 732   clear_all_count_data();
 733 
 734   // Repeat the asserts from above.
 735   guarantee(cmThread()->during_cycle(), "invariant");
 736   guarantee(!g1h->mark_in_progress(), "invariant");
 737 }
 738 
 739 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 740 public:
 741   bool doHeapRegion(HeapRegion* r) {
 742     if (!r->continuesHumongous()) {
 743       r->note_start_of_marking();
 744     }
 745     return false;
 746   }
 747 };
 748 
 749 void ConcurrentMark::checkpointRootsInitialPre() {
 750   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 751   G1CollectorPolicy* g1p = g1h->g1_policy();
 752 
 753   _has_aborted = false;
 754 
 755 #ifndef PRODUCT
 756   if (G1PrintReachableAtInitialMark) {
 757     print_reachable("at-cycle-start",
 758                     VerifyOption_G1UsePrevMarking, true /* all */);
 759   }
 760 #endif
 761 
 762   // Initialise marking structures. This has to be done in a STW phase.
 763   reset();
 764 
 765   // For each region note start of marking.
 766   NoteStartOfMarkHRClosure startcl;
 767   g1h->heap_region_iterate(&startcl);
 768 }
 769 
 770 
 771 void ConcurrentMark::checkpointRootsInitialPost() {
 772   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 773 
 774   // If we force an overflow during remark, the remark operation will
 775   // actually abort and we'll restart concurrent marking. If we always
 776   // force an oveflow during remark we'll never actually complete the
 777   // marking phase. So, we initilize this here, at the start of the
 778   // cycle, so that at the remaining overflow number will decrease at
 779   // every remark and we'll eventually not need to cause one.
 780   force_overflow_stw()->init();
 781 
 782   // Start Concurrent Marking weak-reference discovery.
 783   ReferenceProcessor* rp = g1h->ref_processor_cm();
 784   // enable ("weak") refs discovery
 785   rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
 786   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 787 
 788   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 789   // This is the start of  the marking cycle, we're expected all
 790   // threads to have SATB queues with active set to false.
 791   satb_mq_set.set_active_all_threads(true, /* new active value */
 792                                      false /* expected_active */);
 793 
 794   _root_regions.prepare_for_scan();
 795 
 796   // update_g1_committed() will be called at the end of an evac pause
 797   // when marking is on. So, it's also called at the end of the
 798   // initial-mark pause to update the heap end, if the heap expands
 799   // during it. No need to call it here.
 800 }
 801 
 802 /*
 803  * Notice that in the next two methods, we actually leave the STS
 804  * during the barrier sync and join it immediately afterwards. If we
 805  * do not do this, the following deadlock can occur: one thread could
 806  * be in the barrier sync code, waiting for the other thread to also
 807  * sync up, whereas another one could be trying to yield, while also
 808  * waiting for the other threads to sync up too.
 809  *
 810  * Note, however, that this code is also used during remark and in
 811  * this case we should not attempt to leave / enter the STS, otherwise
 812  * we'll either hit an asseert (debug / fastdebug) or deadlock
 813  * (product). So we should only leave / enter the STS if we are
 814  * operating concurrently.
 815  *
 816  * Because the thread that does the sync barrier has left the STS, it
 817  * is possible to be suspended for a Full GC or an evacuation pause
 818  * could occur. This is actually safe, since the entering the sync
 819  * barrier is one of the last things do_marking_step() does, and it
 820  * doesn't manipulate any data structures afterwards.
 821  */
 822 
 823 void ConcurrentMark::enter_first_sync_barrier(int task_num) {
 824   if (verbose_low()) {
 825     gclog_or_tty->print_cr("[%d] entering first barrier", task_num);
 826   }
 827 
 828   if (concurrent()) {
 829     ConcurrentGCThread::stsLeave();
 830   }
 831   _first_overflow_barrier_sync.enter();
 832   if (concurrent()) {
 833     ConcurrentGCThread::stsJoin();
 834   }
 835   // at this point everyone should have synced up and not be doing any
 836   // more work
 837 
 838   if (verbose_low()) {
 839     gclog_or_tty->print_cr("[%d] leaving first barrier", task_num);
 840   }
 841 
 842   // let task 0 do this
 843   if (task_num == 0) {
 844     // task 0 is responsible for clearing the global data structures
 845     // We should be here because of an overflow. During STW we should
 846     // not clear the overflow flag since we rely on it being true when
 847     // we exit this method to abort the pause and restart concurent
 848     // marking.
 849     clear_marking_state(concurrent() /* clear_overflow */);
 850     force_overflow()->update();
 851 
 852     if (G1Log::fine()) {
 853       gclog_or_tty->date_stamp(PrintGCDateStamps);
 854       gclog_or_tty->stamp(PrintGCTimeStamps);
 855       gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
 856     }
 857   }
 858 
 859   // after this, each task should reset its own data structures then
 860   // then go into the second barrier
 861 }
 862 
 863 void ConcurrentMark::enter_second_sync_barrier(int task_num) {
 864   if (verbose_low()) {
 865     gclog_or_tty->print_cr("[%d] entering second barrier", task_num);
 866   }
 867 
 868   if (concurrent()) {
 869     ConcurrentGCThread::stsLeave();
 870   }
 871   _second_overflow_barrier_sync.enter();
 872   if (concurrent()) {
 873     ConcurrentGCThread::stsJoin();
 874   }
 875   // at this point everything should be re-initialised and ready to go
 876 
 877   if (verbose_low()) {
 878     gclog_or_tty->print_cr("[%d] leaving second barrier", task_num);
 879   }
 880 }
 881 
 882 #ifndef PRODUCT
 883 void ForceOverflowSettings::init() {
 884   _num_remaining = G1ConcMarkForceOverflow;
 885   _force = false;
 886   update();
 887 }
 888 
 889 void ForceOverflowSettings::update() {
 890   if (_num_remaining > 0) {
 891     _num_remaining -= 1;
 892     _force = true;
 893   } else {
 894     _force = false;
 895   }
 896 }
 897 
 898 bool ForceOverflowSettings::should_force() {
 899   if (_force) {
 900     _force = false;
 901     return true;
 902   } else {
 903     return false;
 904   }
 905 }
 906 #endif // !PRODUCT
 907 
 908 class CMConcurrentMarkingTask: public AbstractGangTask {
 909 private:
 910   ConcurrentMark*       _cm;
 911   ConcurrentMarkThread* _cmt;
 912 
 913 public:
 914   void work(uint worker_id) {
 915     assert(Thread::current()->is_ConcurrentGC_thread(),
 916            "this should only be done by a conc GC thread");
 917     ResourceMark rm;
 918 
 919     double start_vtime = os::elapsedVTime();
 920 
 921     ConcurrentGCThread::stsJoin();
 922 
 923     assert(worker_id < _cm->active_tasks(), "invariant");
 924     CMTask* the_task = _cm->task(worker_id);
 925     the_task->record_start_time();
 926     if (!_cm->has_aborted()) {
 927       do {
 928         double start_vtime_sec = os::elapsedVTime();
 929         double start_time_sec = os::elapsedTime();
 930         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
 931 
 932         the_task->do_marking_step(mark_step_duration_ms,
 933                                   true /* do_stealing    */,
 934                                   true /* do_termination */);
 935 
 936         double end_time_sec = os::elapsedTime();
 937         double end_vtime_sec = os::elapsedVTime();
 938         double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
 939         double elapsed_time_sec = end_time_sec - start_time_sec;
 940         _cm->clear_has_overflown();
 941 
 942         bool ret = _cm->do_yield_check(worker_id);
 943 
 944         jlong sleep_time_ms;
 945         if (!_cm->has_aborted() && the_task->has_aborted()) {
 946           sleep_time_ms =
 947             (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
 948           ConcurrentGCThread::stsLeave();
 949           os::sleep(Thread::current(), sleep_time_ms, false);
 950           ConcurrentGCThread::stsJoin();
 951         }
 952         double end_time2_sec = os::elapsedTime();
 953         double elapsed_time2_sec = end_time2_sec - start_time_sec;
 954 
 955 #if 0
 956           gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
 957                                  "overhead %1.4lf",
 958                                  elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
 959                                  the_task->conc_overhead(os::elapsedTime()) * 8.0);
 960           gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
 961                                  elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
 962 #endif
 963       } while (!_cm->has_aborted() && the_task->has_aborted());
 964     }
 965     the_task->record_end_time();
 966     guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
 967 
 968     ConcurrentGCThread::stsLeave();
 969 
 970     double end_vtime = os::elapsedVTime();
 971     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 972   }
 973 
 974   CMConcurrentMarkingTask(ConcurrentMark* cm,
 975                           ConcurrentMarkThread* cmt) :
 976       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
 977 
 978   ~CMConcurrentMarkingTask() { }
 979 };
 980 
 981 // Calculates the number of active workers for a concurrent
 982 // phase.
 983 uint ConcurrentMark::calc_parallel_marking_threads() {
 984   if (G1CollectedHeap::use_parallel_gc_threads()) {
 985     uint n_conc_workers = 0;
 986     if (!UseDynamicNumberOfGCThreads ||
 987         (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 988          !ForceDynamicNumberOfGCThreads)) {
 989       n_conc_workers = max_parallel_marking_threads();
 990     } else {
 991       n_conc_workers =
 992         AdaptiveSizePolicy::calc_default_active_workers(
 993                                      max_parallel_marking_threads(),
 994                                      1, /* Minimum workers */
 995                                      parallel_marking_threads(),
 996                                      Threads::number_of_non_daemon_threads());
 997       // Don't scale down "n_conc_workers" by scale_parallel_threads() because
 998       // that scaling has already gone into "_max_parallel_marking_threads".
 999     }
1000     assert(n_conc_workers > 0, "Always need at least 1");
1001     return n_conc_workers;
1002   }
1003   // If we are not running with any parallel GC threads we will not
1004   // have spawned any marking threads either. Hence the number of
1005   // concurrent workers should be 0.
1006   return 0;
1007 }
1008 
1009 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1010   // Currently, only survivors can be root regions.
1011   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1012   G1RootRegionScanClosure cl(_g1h, this, worker_id);
1013 
1014   const uintx interval = PrefetchScanIntervalInBytes;
1015   HeapWord* curr = hr->bottom();
1016   const HeapWord* end = hr->top();
1017   while (curr < end) {
1018     Prefetch::read(curr, interval);
1019     oop obj = oop(curr);
1020     int size = obj->oop_iterate(&cl);
1021     assert(size == obj->size(), "sanity");
1022     curr += size;
1023   }
1024 }
1025 
1026 class CMRootRegionScanTask : public AbstractGangTask {
1027 private:
1028   ConcurrentMark* _cm;
1029 
1030 public:
1031   CMRootRegionScanTask(ConcurrentMark* cm) :
1032     AbstractGangTask("Root Region Scan"), _cm(cm) { }
1033 
1034   void work(uint worker_id) {
1035     assert(Thread::current()->is_ConcurrentGC_thread(),
1036            "this should only be done by a conc GC thread");
1037 
1038     CMRootRegions* root_regions = _cm->root_regions();
1039     HeapRegion* hr = root_regions->claim_next();
1040     while (hr != NULL) {
1041       _cm->scanRootRegion(hr, worker_id);
1042       hr = root_regions->claim_next();
1043     }
1044   }
1045 };
1046 
1047 void ConcurrentMark::scanRootRegions() {
1048   // scan_in_progress() will have been set to true only if there was
1049   // at least one root region to scan. So, if it's false, we
1050   // should not attempt to do any further work.
1051   if (root_regions()->scan_in_progress()) {
1052     _parallel_marking_threads = calc_parallel_marking_threads();
1053     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1054            "Maximum number of marking threads exceeded");
1055     uint active_workers = MAX2(1U, parallel_marking_threads());
1056 
1057     CMRootRegionScanTask task(this);
1058     if (parallel_marking_threads() > 0) {
1059       _parallel_workers->set_active_workers((int) active_workers);
1060       _parallel_workers->run_task(&task);
1061     } else {
1062       task.work(0);
1063     }
1064 
1065     // It's possible that has_aborted() is true here without actually
1066     // aborting the survivor scan earlier. This is OK as it's
1067     // mainly used for sanity checking.
1068     root_regions()->scan_finished();
1069   }
1070 }
1071 
1072 void ConcurrentMark::markFromRoots() {
1073   // we might be tempted to assert that:
1074   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1075   //        "inconsistent argument?");
1076   // However that wouldn't be right, because it's possible that
1077   // a safepoint is indeed in progress as a younger generation
1078   // stop-the-world GC happens even as we mark in this generation.
1079 
1080   _restart_for_overflow = false;
1081   force_overflow_conc()->init();
1082 
1083   // _g1h has _n_par_threads
1084   _parallel_marking_threads = calc_parallel_marking_threads();
1085   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1086     "Maximum number of marking threads exceeded");
1087 
1088   uint active_workers = MAX2(1U, parallel_marking_threads());
1089 
1090   // Parallel task terminator is set in "set_phase()"
1091   set_phase(active_workers, true /* concurrent */);
1092 
1093   CMConcurrentMarkingTask markingTask(this, cmThread());
1094   if (parallel_marking_threads() > 0) {
1095     _parallel_workers->set_active_workers((int)active_workers);
1096     // Don't set _n_par_threads because it affects MT in proceess_strong_roots()
1097     // and the decisions on that MT processing is made elsewhere.
1098     assert(_parallel_workers->active_workers() > 0, "Should have been set");
1099     _parallel_workers->run_task(&markingTask);
1100   } else {
1101     markingTask.work(0);
1102   }
1103   print_stats();
1104 }
1105 
1106 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1107   // world is stopped at this checkpoint
1108   assert(SafepointSynchronize::is_at_safepoint(),
1109          "world should be stopped");
1110 
1111   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1112 
1113   // If a full collection has happened, we shouldn't do this.
1114   if (has_aborted()) {
1115     g1h->set_marking_complete(); // So bitmap clearing isn't confused
1116     return;
1117   }
1118 
1119   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1120 
1121   if (VerifyDuringGC) {
1122     HandleMark hm;  // handle scope
1123     gclog_or_tty->print(" VerifyDuringGC:(before)");
1124     Universe::heap()->prepare_for_verify();
1125     Universe::verify(/* silent */ false,
1126                      /* option */ VerifyOption_G1UsePrevMarking);
1127   }
1128 
1129   G1CollectorPolicy* g1p = g1h->g1_policy();
1130   g1p->record_concurrent_mark_remark_start();
1131 
1132   double start = os::elapsedTime();
1133 
1134   checkpointRootsFinalWork();
1135 
1136   double mark_work_end = os::elapsedTime();
1137 
1138   weakRefsWork(clear_all_soft_refs);
1139 
1140   if (has_overflown()) {
1141     // Oops.  We overflowed.  Restart concurrent marking.
1142     _restart_for_overflow = true;
1143     // Clear the flag. We do not need it any more.
1144     clear_has_overflown();
1145     if (G1TraceMarkStackOverflow) {
1146       gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1147     }
1148   } else {
1149     // Aggregate the per-task counting data that we have accumulated
1150     // while marking.
1151     aggregate_count_data();
1152 
1153     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1154     // We're done with marking.
1155     // This is the end of  the marking cycle, we're expected all
1156     // threads to have SATB queues with active set to true.
1157     satb_mq_set.set_active_all_threads(false, /* new active value */
1158                                        true /* expected_active */);
1159 
1160     if (VerifyDuringGC) {
1161       HandleMark hm;  // handle scope
1162       gclog_or_tty->print(" VerifyDuringGC:(after)");
1163       Universe::heap()->prepare_for_verify();
1164       Universe::verify(/* silent */ false,
1165                        /* option */ VerifyOption_G1UseNextMarking);
1166     }
1167     assert(!restart_for_overflow(), "sanity");
1168   }
1169 
1170   // Reset the marking state if marking completed
1171   if (!restart_for_overflow()) {
1172     set_non_marking_state();
1173   }
1174 
1175 #if VERIFY_OBJS_PROCESSED
1176   _scan_obj_cl.objs_processed = 0;
1177   ThreadLocalObjQueue::objs_enqueued = 0;
1178 #endif
1179 
1180   // Statistics
1181   double now = os::elapsedTime();
1182   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1183   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1184   _remark_times.add((now - start) * 1000.0);
1185 
1186   g1p->record_concurrent_mark_remark_end();
1187 }
1188 
1189 // Base class of the closures that finalize and verify the
1190 // liveness counting data.
1191 class CMCountDataClosureBase: public HeapRegionClosure {
1192 protected:
1193   G1CollectedHeap* _g1h;
1194   ConcurrentMark* _cm;
1195   CardTableModRefBS* _ct_bs;
1196 
1197   BitMap* _region_bm;
1198   BitMap* _card_bm;
1199 
1200   // Takes a region that's not empty (i.e., it has at least one
1201   // live object in it and sets its corresponding bit on the region
1202   // bitmap to 1. If the region is "starts humongous" it will also set
1203   // to 1 the bits on the region bitmap that correspond to its
1204   // associated "continues humongous" regions.
1205   void set_bit_for_region(HeapRegion* hr) {
1206     assert(!hr->continuesHumongous(), "should have filtered those out");
1207 
1208     BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1209     if (!hr->startsHumongous()) {
1210       // Normal (non-humongous) case: just set the bit.
1211       _region_bm->par_at_put(index, true);
1212     } else {
1213       // Starts humongous case: calculate how many regions are part of
1214       // this humongous region and then set the bit range.
1215       BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1216       _region_bm->par_at_put_range(index, end_index, true);
1217     }
1218   }
1219 
1220 public:
1221   CMCountDataClosureBase(G1CollectedHeap* g1h,
1222                          BitMap* region_bm, BitMap* card_bm):
1223     _g1h(g1h), _cm(g1h->concurrent_mark()),
1224     _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1225     _region_bm(region_bm), _card_bm(card_bm) { }
1226 };
1227 
1228 // Closure that calculates the # live objects per region. Used
1229 // for verification purposes during the cleanup pause.
1230 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1231   CMBitMapRO* _bm;
1232   size_t _region_marked_bytes;
1233 
1234 public:
1235   CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1236                          BitMap* region_bm, BitMap* card_bm) :
1237     CMCountDataClosureBase(g1h, region_bm, card_bm),
1238     _bm(bm), _region_marked_bytes(0) { }
1239 
1240   bool doHeapRegion(HeapRegion* hr) {
1241 
1242     if (hr->continuesHumongous()) {
1243       // We will ignore these here and process them when their
1244       // associated "starts humongous" region is processed (see
1245       // set_bit_for_heap_region()). Note that we cannot rely on their
1246       // associated "starts humongous" region to have their bit set to
1247       // 1 since, due to the region chunking in the parallel region
1248       // iteration, a "continues humongous" region might be visited
1249       // before its associated "starts humongous".
1250       return false;
1251     }
1252 
1253     HeapWord* ntams = hr->next_top_at_mark_start();
1254     HeapWord* start = hr->bottom();
1255 
1256     assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1257            err_msg("Preconditions not met - "
1258                    "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1259                    start, ntams, hr->end()));
1260 
1261     // Find the first marked object at or after "start".
1262     start = _bm->getNextMarkedWordAddress(start, ntams);
1263 
1264     size_t marked_bytes = 0;
1265 
1266     while (start < ntams) {
1267       oop obj = oop(start);
1268       int obj_sz = obj->size();
1269       HeapWord* obj_end = start + obj_sz;
1270 
1271       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1272       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1273 
1274       // Note: if we're looking at the last region in heap - obj_end
1275       // could be actually just beyond the end of the heap; end_idx
1276       // will then correspond to a (non-existent) card that is also
1277       // just beyond the heap.
1278       if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1279         // end of object is not card aligned - increment to cover
1280         // all the cards spanned by the object
1281         end_idx += 1;
1282       }
1283 
1284       // Set the bits in the card BM for the cards spanned by this object.
1285       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1286 
1287       // Add the size of this object to the number of marked bytes.
1288       marked_bytes += (size_t)obj_sz * HeapWordSize;
1289 
1290       // Find the next marked object after this one.
1291       start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1292     }
1293 
1294     // Mark the allocated-since-marking portion...
1295     HeapWord* top = hr->top();
1296     if (ntams < top) {
1297       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1298       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1299 
1300       // Note: if we're looking at the last region in heap - top
1301       // could be actually just beyond the end of the heap; end_idx
1302       // will then correspond to a (non-existent) card that is also
1303       // just beyond the heap.
1304       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1305         // end of object is not card aligned - increment to cover
1306         // all the cards spanned by the object
1307         end_idx += 1;
1308       }
1309       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1310 
1311       // This definitely means the region has live objects.
1312       set_bit_for_region(hr);
1313     }
1314 
1315     // Update the live region bitmap.
1316     if (marked_bytes > 0) {
1317       set_bit_for_region(hr);
1318     }
1319 
1320     // Set the marked bytes for the current region so that
1321     // it can be queried by a calling verificiation routine
1322     _region_marked_bytes = marked_bytes;
1323 
1324     return false;
1325   }
1326 
1327   size_t region_marked_bytes() const { return _region_marked_bytes; }
1328 };
1329 
1330 // Heap region closure used for verifying the counting data
1331 // that was accumulated concurrently and aggregated during
1332 // the remark pause. This closure is applied to the heap
1333 // regions during the STW cleanup pause.
1334 
1335 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1336   G1CollectedHeap* _g1h;
1337   ConcurrentMark* _cm;
1338   CalcLiveObjectsClosure _calc_cl;
1339   BitMap* _region_bm;   // Region BM to be verified
1340   BitMap* _card_bm;     // Card BM to be verified
1341   bool _verbose;        // verbose output?
1342 
1343   BitMap* _exp_region_bm; // Expected Region BM values
1344   BitMap* _exp_card_bm;   // Expected card BM values
1345 
1346   int _failures;
1347 
1348 public:
1349   VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1350                                 BitMap* region_bm,
1351                                 BitMap* card_bm,
1352                                 BitMap* exp_region_bm,
1353                                 BitMap* exp_card_bm,
1354                                 bool verbose) :
1355     _g1h(g1h), _cm(g1h->concurrent_mark()),
1356     _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1357     _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1358     _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1359     _failures(0) { }
1360 
1361   int failures() const { return _failures; }
1362 
1363   bool doHeapRegion(HeapRegion* hr) {
1364     if (hr->continuesHumongous()) {
1365       // We will ignore these here and process them when their
1366       // associated "starts humongous" region is processed (see
1367       // set_bit_for_heap_region()). Note that we cannot rely on their
1368       // associated "starts humongous" region to have their bit set to
1369       // 1 since, due to the region chunking in the parallel region
1370       // iteration, a "continues humongous" region might be visited
1371       // before its associated "starts humongous".
1372       return false;
1373     }
1374 
1375     int failures = 0;
1376 
1377     // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1378     // this region and set the corresponding bits in the expected region
1379     // and card bitmaps.
1380     bool res = _calc_cl.doHeapRegion(hr);
1381     assert(res == false, "should be continuing");
1382 
1383     MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1384                     Mutex::_no_safepoint_check_flag);
1385 
1386     // Verify the marked bytes for this region.
1387     size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1388     size_t act_marked_bytes = hr->next_marked_bytes();
1389 
1390     // We're not OK if expected marked bytes > actual marked bytes. It means
1391     // we have missed accounting some objects during the actual marking.
1392     if (exp_marked_bytes > act_marked_bytes) {
1393       if (_verbose) {
1394         gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1395                                "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1396                                hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1397       }
1398       failures += 1;
1399     }
1400 
1401     // Verify the bit, for this region, in the actual and expected
1402     // (which was just calculated) region bit maps.
1403     // We're not OK if the bit in the calculated expected region
1404     // bitmap is set and the bit in the actual region bitmap is not.
1405     BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1406 
1407     bool expected = _exp_region_bm->at(index);
1408     bool actual = _region_bm->at(index);
1409     if (expected && !actual) {
1410       if (_verbose) {
1411         gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1412                                "expected: %s, actual: %s",
1413                                hr->hrs_index(),
1414                                BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1415       }
1416       failures += 1;
1417     }
1418 
1419     // Verify that the card bit maps for the cards spanned by the current
1420     // region match. We have an error if we have a set bit in the expected
1421     // bit map and the corresponding bit in the actual bitmap is not set.
1422 
1423     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1424     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1425 
1426     for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1427       expected = _exp_card_bm->at(i);
1428       actual = _card_bm->at(i);
1429 
1430       if (expected && !actual) {
1431         if (_verbose) {
1432           gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1433                                  "expected: %s, actual: %s",
1434                                  hr->hrs_index(), i,
1435                                  BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1436         }
1437         failures += 1;
1438       }
1439     }
1440 
1441     if (failures > 0 && _verbose)  {
1442       gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1443                              "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1444                              HR_FORMAT_PARAMS(hr), hr->next_top_at_mark_start(),
1445                              _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1446     }
1447 
1448     _failures += failures;
1449 
1450     // We could stop iteration over the heap when we
1451     // find the first violating region by returning true.
1452     return false;
1453   }
1454 };
1455 
1456 
1457 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1458 protected:
1459   G1CollectedHeap* _g1h;
1460   ConcurrentMark* _cm;
1461   BitMap* _actual_region_bm;
1462   BitMap* _actual_card_bm;
1463 
1464   uint    _n_workers;
1465 
1466   BitMap* _expected_region_bm;
1467   BitMap* _expected_card_bm;
1468 
1469   int  _failures;
1470   bool _verbose;
1471 
1472 public:
1473   G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1474                             BitMap* region_bm, BitMap* card_bm,
1475                             BitMap* expected_region_bm, BitMap* expected_card_bm)
1476     : AbstractGangTask("G1 verify final counting"),
1477       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1478       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1479       _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1480       _failures(0), _verbose(false),
1481       _n_workers(0) {
1482     assert(VerifyDuringGC, "don't call this otherwise");
1483 
1484     // Use the value already set as the number of active threads
1485     // in the call to run_task().
1486     if (G1CollectedHeap::use_parallel_gc_threads()) {
1487       assert( _g1h->workers()->active_workers() > 0,
1488         "Should have been previously set");
1489       _n_workers = _g1h->workers()->active_workers();
1490     } else {
1491       _n_workers = 1;
1492     }
1493 
1494     assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1495     assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1496 
1497     _verbose = _cm->verbose_medium();
1498   }
1499 
1500   void work(uint worker_id) {
1501     assert(worker_id < _n_workers, "invariant");
1502 
1503     VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1504                                             _actual_region_bm, _actual_card_bm,
1505                                             _expected_region_bm,
1506                                             _expected_card_bm,
1507                                             _verbose);
1508 
1509     if (G1CollectedHeap::use_parallel_gc_threads()) {
1510       _g1h->heap_region_par_iterate_chunked(&verify_cl,
1511                                             worker_id,
1512                                             _n_workers,
1513                                             HeapRegion::VerifyCountClaimValue);
1514     } else {
1515       _g1h->heap_region_iterate(&verify_cl);
1516     }
1517 
1518     Atomic::add(verify_cl.failures(), &_failures);
1519   }
1520 
1521   int failures() const { return _failures; }
1522 };
1523 
1524 // Closure that finalizes the liveness counting data.
1525 // Used during the cleanup pause.
1526 // Sets the bits corresponding to the interval [NTAMS, top]
1527 // (which contains the implicitly live objects) in the
1528 // card liveness bitmap. Also sets the bit for each region,
1529 // containing live data, in the region liveness bitmap.
1530 
1531 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1532  public:
1533   FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1534                               BitMap* region_bm,
1535                               BitMap* card_bm) :
1536     CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1537 
1538   bool doHeapRegion(HeapRegion* hr) {
1539 
1540     if (hr->continuesHumongous()) {
1541       // We will ignore these here and process them when their
1542       // associated "starts humongous" region is processed (see
1543       // set_bit_for_heap_region()). Note that we cannot rely on their
1544       // associated "starts humongous" region to have their bit set to
1545       // 1 since, due to the region chunking in the parallel region
1546       // iteration, a "continues humongous" region might be visited
1547       // before its associated "starts humongous".
1548       return false;
1549     }
1550 
1551     HeapWord* ntams = hr->next_top_at_mark_start();
1552     HeapWord* top   = hr->top();
1553 
1554     assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1555 
1556     // Mark the allocated-since-marking portion...
1557     if (ntams < top) {
1558       // This definitely means the region has live objects.
1559       set_bit_for_region(hr);
1560 
1561       // Now set the bits in the card bitmap for [ntams, top)
1562       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1563       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1564 
1565       // Note: if we're looking at the last region in heap - top
1566       // could be actually just beyond the end of the heap; end_idx
1567       // will then correspond to a (non-existent) card that is also
1568       // just beyond the heap.
1569       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1570         // end of object is not card aligned - increment to cover
1571         // all the cards spanned by the object
1572         end_idx += 1;
1573       }
1574 
1575       assert(end_idx <= _card_bm->size(),
1576              err_msg("oob: end_idx=  "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1577                      end_idx, _card_bm->size()));
1578       assert(start_idx < _card_bm->size(),
1579              err_msg("oob: start_idx=  "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1580                      start_idx, _card_bm->size()));
1581 
1582       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1583      }
1584 
1585     // Set the bit for the region if it contains live data
1586     if (hr->next_marked_bytes() > 0) {
1587       set_bit_for_region(hr);
1588     }
1589 
1590     return false;
1591   }
1592 };
1593 
1594 class G1ParFinalCountTask: public AbstractGangTask {
1595 protected:
1596   G1CollectedHeap* _g1h;
1597   ConcurrentMark* _cm;
1598   BitMap* _actual_region_bm;
1599   BitMap* _actual_card_bm;
1600 
1601   uint    _n_workers;
1602 
1603 public:
1604   G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1605     : AbstractGangTask("G1 final counting"),
1606       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1607       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1608       _n_workers(0) {
1609     // Use the value already set as the number of active threads
1610     // in the call to run_task().
1611     if (G1CollectedHeap::use_parallel_gc_threads()) {
1612       assert( _g1h->workers()->active_workers() > 0,
1613         "Should have been previously set");
1614       _n_workers = _g1h->workers()->active_workers();
1615     } else {
1616       _n_workers = 1;
1617     }
1618   }
1619 
1620   void work(uint worker_id) {
1621     assert(worker_id < _n_workers, "invariant");
1622 
1623     FinalCountDataUpdateClosure final_update_cl(_g1h,
1624                                                 _actual_region_bm,
1625                                                 _actual_card_bm);
1626 
1627     if (G1CollectedHeap::use_parallel_gc_threads()) {
1628       _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1629                                             worker_id,
1630                                             _n_workers,
1631                                             HeapRegion::FinalCountClaimValue);
1632     } else {
1633       _g1h->heap_region_iterate(&final_update_cl);
1634     }
1635   }
1636 };
1637 
1638 class G1ParNoteEndTask;
1639 
1640 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1641   G1CollectedHeap* _g1;
1642   int _worker_num;
1643   size_t _max_live_bytes;
1644   uint _regions_claimed;
1645   size_t _freed_bytes;
1646   FreeRegionList* _local_cleanup_list;
1647   OldRegionSet* _old_proxy_set;
1648   HumongousRegionSet* _humongous_proxy_set;
1649   HRRSCleanupTask* _hrrs_cleanup_task;
1650   double _claimed_region_time;
1651   double _max_region_time;
1652 
1653 public:
1654   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1655                              int worker_num,
1656                              FreeRegionList* local_cleanup_list,
1657                              OldRegionSet* old_proxy_set,
1658                              HumongousRegionSet* humongous_proxy_set,
1659                              HRRSCleanupTask* hrrs_cleanup_task) :
1660     _g1(g1), _worker_num(worker_num),
1661     _max_live_bytes(0), _regions_claimed(0),
1662     _freed_bytes(0),
1663     _claimed_region_time(0.0), _max_region_time(0.0),
1664     _local_cleanup_list(local_cleanup_list),
1665     _old_proxy_set(old_proxy_set),
1666     _humongous_proxy_set(humongous_proxy_set),
1667     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1668 
1669   size_t freed_bytes() { return _freed_bytes; }
1670 
1671   bool doHeapRegion(HeapRegion *hr) {
1672     if (hr->continuesHumongous()) {
1673       return false;
1674     }
1675     // We use a claim value of zero here because all regions
1676     // were claimed with value 1 in the FinalCount task.
1677     _g1->reset_gc_time_stamps(hr);
1678     double start = os::elapsedTime();
1679     _regions_claimed++;
1680     hr->note_end_of_marking();
1681     _max_live_bytes += hr->max_live_bytes();
1682     _g1->free_region_if_empty(hr,
1683                               &_freed_bytes,
1684                               _local_cleanup_list,
1685                               _old_proxy_set,
1686                               _humongous_proxy_set,
1687                               _hrrs_cleanup_task,
1688                               true /* par */);
1689     double region_time = (os::elapsedTime() - start);
1690     _claimed_region_time += region_time;
1691     if (region_time > _max_region_time) {
1692       _max_region_time = region_time;
1693     }
1694     return false;
1695   }
1696 
1697   size_t max_live_bytes() { return _max_live_bytes; }
1698   uint regions_claimed() { return _regions_claimed; }
1699   double claimed_region_time_sec() { return _claimed_region_time; }
1700   double max_region_time_sec() { return _max_region_time; }
1701 };
1702 
1703 class G1ParNoteEndTask: public AbstractGangTask {
1704   friend class G1NoteEndOfConcMarkClosure;
1705 
1706 protected:
1707   G1CollectedHeap* _g1h;
1708   size_t _max_live_bytes;
1709   size_t _freed_bytes;
1710   FreeRegionList* _cleanup_list;
1711 
1712 public:
1713   G1ParNoteEndTask(G1CollectedHeap* g1h,
1714                    FreeRegionList* cleanup_list) :
1715     AbstractGangTask("G1 note end"), _g1h(g1h),
1716     _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1717 
1718   void work(uint worker_id) {
1719     double start = os::elapsedTime();
1720     FreeRegionList local_cleanup_list("Local Cleanup List");
1721     OldRegionSet old_proxy_set("Local Cleanup Old Proxy Set");
1722     HumongousRegionSet humongous_proxy_set("Local Cleanup Humongous Proxy Set");
1723     HRRSCleanupTask hrrs_cleanup_task;
1724     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, worker_id, &local_cleanup_list,
1725                                            &old_proxy_set,
1726                                            &humongous_proxy_set,
1727                                            &hrrs_cleanup_task);
1728     if (G1CollectedHeap::use_parallel_gc_threads()) {
1729       _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1730                                             _g1h->workers()->active_workers(),
1731                                             HeapRegion::NoteEndClaimValue);
1732     } else {
1733       _g1h->heap_region_iterate(&g1_note_end);
1734     }
1735     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1736 
1737     // Now update the lists
1738     _g1h->update_sets_after_freeing_regions(g1_note_end.freed_bytes(),
1739                                             NULL /* free_list */,
1740                                             &old_proxy_set,
1741                                             &humongous_proxy_set,
1742                                             true /* par */);
1743     {
1744       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1745       _max_live_bytes += g1_note_end.max_live_bytes();
1746       _freed_bytes += g1_note_end.freed_bytes();
1747 
1748       // If we iterate over the global cleanup list at the end of
1749       // cleanup to do this printing we will not guarantee to only
1750       // generate output for the newly-reclaimed regions (the list
1751       // might not be empty at the beginning of cleanup; we might
1752       // still be working on its previous contents). So we do the
1753       // printing here, before we append the new regions to the global
1754       // cleanup list.
1755 
1756       G1HRPrinter* hr_printer = _g1h->hr_printer();
1757       if (hr_printer->is_active()) {
1758         HeapRegionLinkedListIterator iter(&local_cleanup_list);
1759         while (iter.more_available()) {
1760           HeapRegion* hr = iter.get_next();
1761           hr_printer->cleanup(hr);
1762         }
1763       }
1764 
1765       _cleanup_list->add_as_tail(&local_cleanup_list);
1766       assert(local_cleanup_list.is_empty(), "post-condition");
1767 
1768       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1769     }
1770   }
1771   size_t max_live_bytes() { return _max_live_bytes; }
1772   size_t freed_bytes() { return _freed_bytes; }
1773 };
1774 
1775 class G1ParScrubRemSetTask: public AbstractGangTask {
1776 protected:
1777   G1RemSet* _g1rs;
1778   BitMap* _region_bm;
1779   BitMap* _card_bm;
1780 public:
1781   G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1782                        BitMap* region_bm, BitMap* card_bm) :
1783     AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1784     _region_bm(region_bm), _card_bm(card_bm) { }
1785 
1786   void work(uint worker_id) {
1787     if (G1CollectedHeap::use_parallel_gc_threads()) {
1788       _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1789                        HeapRegion::ScrubRemSetClaimValue);
1790     } else {
1791       _g1rs->scrub(_region_bm, _card_bm);
1792     }
1793   }
1794 
1795 };
1796 
1797 void ConcurrentMark::cleanup() {
1798   // world is stopped at this checkpoint
1799   assert(SafepointSynchronize::is_at_safepoint(),
1800          "world should be stopped");
1801   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1802 
1803   // If a full collection has happened, we shouldn't do this.
1804   if (has_aborted()) {
1805     g1h->set_marking_complete(); // So bitmap clearing isn't confused
1806     return;
1807   }
1808 
1809   HRSPhaseSetter x(HRSPhaseCleanup);
1810   g1h->verify_region_sets_optional();
1811 
1812   if (VerifyDuringGC) {
1813     HandleMark hm;  // handle scope
1814     gclog_or_tty->print(" VerifyDuringGC:(before)");
1815     Universe::heap()->prepare_for_verify();
1816     Universe::verify(/* silent */ false,
1817                      /* option */ VerifyOption_G1UsePrevMarking);
1818   }
1819 
1820   G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1821   g1p->record_concurrent_mark_cleanup_start();
1822 
1823   double start = os::elapsedTime();
1824 
1825   HeapRegionRemSet::reset_for_cleanup_tasks();
1826 
1827   uint n_workers;
1828 
1829   // Do counting once more with the world stopped for good measure.
1830   G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1831 
1832   if (G1CollectedHeap::use_parallel_gc_threads()) {
1833    assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1834            "sanity check");
1835 
1836     g1h->set_par_threads();
1837     n_workers = g1h->n_par_threads();
1838     assert(g1h->n_par_threads() == n_workers,
1839            "Should not have been reset");
1840     g1h->workers()->run_task(&g1_par_count_task);
1841     // Done with the parallel phase so reset to 0.
1842     g1h->set_par_threads(0);
1843 
1844     assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
1845            "sanity check");
1846   } else {
1847     n_workers = 1;
1848     g1_par_count_task.work(0);
1849   }
1850 
1851   if (VerifyDuringGC) {
1852     // Verify that the counting data accumulated during marking matches
1853     // that calculated by walking the marking bitmap.
1854 
1855     // Bitmaps to hold expected values
1856     BitMap expected_region_bm(_region_bm.size(), false);
1857     BitMap expected_card_bm(_card_bm.size(), false);
1858 
1859     G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
1860                                                  &_region_bm,
1861                                                  &_card_bm,
1862                                                  &expected_region_bm,
1863                                                  &expected_card_bm);
1864 
1865     if (G1CollectedHeap::use_parallel_gc_threads()) {
1866       g1h->set_par_threads((int)n_workers);
1867       g1h->workers()->run_task(&g1_par_verify_task);
1868       // Done with the parallel phase so reset to 0.
1869       g1h->set_par_threads(0);
1870 
1871       assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
1872              "sanity check");
1873     } else {
1874       g1_par_verify_task.work(0);
1875     }
1876 
1877     guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
1878   }
1879 
1880   size_t start_used_bytes = g1h->used();
1881   g1h->set_marking_complete();
1882 
1883   double count_end = os::elapsedTime();
1884   double this_final_counting_time = (count_end - start);
1885   _total_counting_time += this_final_counting_time;
1886 
1887   if (G1PrintRegionLivenessInfo) {
1888     G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
1889     _g1h->heap_region_iterate(&cl);
1890   }
1891 
1892   // Install newly created mark bitMap as "prev".
1893   swapMarkBitMaps();
1894 
1895   g1h->reset_gc_time_stamp();
1896 
1897   // Note end of marking in all heap regions.
1898   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
1899   if (G1CollectedHeap::use_parallel_gc_threads()) {
1900     g1h->set_par_threads((int)n_workers);
1901     g1h->workers()->run_task(&g1_par_note_end_task);
1902     g1h->set_par_threads(0);
1903 
1904     assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
1905            "sanity check");
1906   } else {
1907     g1_par_note_end_task.work(0);
1908   }
1909   g1h->check_gc_time_stamps();
1910 
1911   if (!cleanup_list_is_empty()) {
1912     // The cleanup list is not empty, so we'll have to process it
1913     // concurrently. Notify anyone else that might be wanting free
1914     // regions that there will be more free regions coming soon.
1915     g1h->set_free_regions_coming();
1916   }
1917 
1918   // call below, since it affects the metric by which we sort the heap
1919   // regions.
1920   if (G1ScrubRemSets) {
1921     double rs_scrub_start = os::elapsedTime();
1922     G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
1923     if (G1CollectedHeap::use_parallel_gc_threads()) {
1924       g1h->set_par_threads((int)n_workers);
1925       g1h->workers()->run_task(&g1_par_scrub_rs_task);
1926       g1h->set_par_threads(0);
1927 
1928       assert(g1h->check_heap_region_claim_values(
1929                                             HeapRegion::ScrubRemSetClaimValue),
1930              "sanity check");
1931     } else {
1932       g1_par_scrub_rs_task.work(0);
1933     }
1934 
1935     double rs_scrub_end = os::elapsedTime();
1936     double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
1937     _total_rs_scrub_time += this_rs_scrub_time;
1938   }
1939 
1940   // this will also free any regions totally full of garbage objects,
1941   // and sort the regions.
1942   g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
1943 
1944   // Statistics.
1945   double end = os::elapsedTime();
1946   _cleanup_times.add((end - start) * 1000.0);
1947 
1948   if (G1Log::fine()) {
1949     g1h->print_size_transition(gclog_or_tty,
1950                                start_used_bytes,
1951                                g1h->used(),
1952                                g1h->capacity());
1953   }
1954 
1955   // Clean up will have freed any regions completely full of garbage.
1956   // Update the soft reference policy with the new heap occupancy.
1957   Universe::update_heap_info_at_gc();
1958 
1959   // We need to make this be a "collection" so any collection pause that
1960   // races with it goes around and waits for completeCleanup to finish.
1961   g1h->increment_total_collections();
1962 
1963   // We reclaimed old regions so we should calculate the sizes to make
1964   // sure we update the old gen/space data.
1965   g1h->g1mm()->update_sizes();
1966 
1967   if (VerifyDuringGC) {
1968     HandleMark hm;  // handle scope
1969     gclog_or_tty->print(" VerifyDuringGC:(after)");
1970     Universe::heap()->prepare_for_verify();
1971     Universe::verify(/* silent */ false,
1972                      /* option */ VerifyOption_G1UsePrevMarking);
1973   }
1974 
1975   g1h->verify_region_sets_optional();
1976 }
1977 
1978 void ConcurrentMark::completeCleanup() {
1979   if (has_aborted()) return;
1980 
1981   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1982 
1983   _cleanup_list.verify_optional();
1984   FreeRegionList tmp_free_list("Tmp Free List");
1985 
1986   if (G1ConcRegionFreeingVerbose) {
1987     gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
1988                            "cleanup list has %u entries",
1989                            _cleanup_list.length());
1990   }
1991 
1992   // Noone else should be accessing the _cleanup_list at this point,
1993   // so it's not necessary to take any locks
1994   while (!_cleanup_list.is_empty()) {
1995     HeapRegion* hr = _cleanup_list.remove_head();
1996     assert(hr != NULL, "the list was not empty");
1997     hr->par_clear();
1998     tmp_free_list.add_as_tail(hr);
1999 
2000     // Instead of adding one region at a time to the secondary_free_list,
2001     // we accumulate them in the local list and move them a few at a
2002     // time. This also cuts down on the number of notify_all() calls
2003     // we do during this process. We'll also append the local list when
2004     // _cleanup_list is empty (which means we just removed the last
2005     // region from the _cleanup_list).
2006     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2007         _cleanup_list.is_empty()) {
2008       if (G1ConcRegionFreeingVerbose) {
2009         gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2010                                "appending %u entries to the secondary_free_list, "
2011                                "cleanup list still has %u entries",
2012                                tmp_free_list.length(),
2013                                _cleanup_list.length());
2014       }
2015 
2016       {
2017         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2018         g1h->secondary_free_list_add_as_tail(&tmp_free_list);
2019         SecondaryFreeList_lock->notify_all();
2020       }
2021 
2022       if (G1StressConcRegionFreeing) {
2023         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2024           os::sleep(Thread::current(), (jlong) 1, false);
2025         }
2026       }
2027     }
2028   }
2029   assert(tmp_free_list.is_empty(), "post-condition");
2030 }
2031 
2032 // Support closures for reference procssing in G1
2033 
2034 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2035   HeapWord* addr = (HeapWord*)obj;
2036   return addr != NULL &&
2037          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2038 }
2039 
2040 class G1CMKeepAliveClosure: public OopClosure {
2041   G1CollectedHeap* _g1;
2042   ConcurrentMark*  _cm;
2043  public:
2044   G1CMKeepAliveClosure(G1CollectedHeap* g1, ConcurrentMark* cm) :
2045     _g1(g1), _cm(cm) {
2046     assert(Thread::current()->is_VM_thread(), "otherwise fix worker id");
2047   }
2048 
2049   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2050   virtual void do_oop(      oop* p) { do_oop_work(p); }
2051 
2052   template <class T> void do_oop_work(T* p) {
2053     oop obj = oopDesc::load_decode_heap_oop(p);
2054     HeapWord* addr = (HeapWord*)obj;
2055 
2056     if (_cm->verbose_high()) {
2057       gclog_or_tty->print_cr("\t[0] we're looking at location "
2058                              "*"PTR_FORMAT" = "PTR_FORMAT,
2059                              p, (void*) obj);
2060     }
2061 
2062     if (_g1->is_in_g1_reserved(addr) && _g1->is_obj_ill(obj)) {
2063       _cm->mark_and_count(obj);
2064       _cm->mark_stack_push(obj);
2065     }
2066   }
2067 };
2068 
2069 class G1CMDrainMarkingStackClosure: public VoidClosure {
2070   ConcurrentMark*               _cm;
2071   CMMarkStack*                  _markStack;
2072   G1CMKeepAliveClosure*         _oopClosure;
2073  public:
2074   G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMMarkStack* markStack,
2075                                G1CMKeepAliveClosure* oopClosure) :
2076     _cm(cm),
2077     _markStack(markStack),
2078     _oopClosure(oopClosure) { }
2079 
2080   void do_void() {
2081     _markStack->drain((OopClosure*)_oopClosure, _cm->nextMarkBitMap(), false);
2082   }
2083 };
2084 
2085 // 'Keep Alive' closure used by parallel reference processing.
2086 // An instance of this closure is used in the parallel reference processing
2087 // code rather than an instance of G1CMKeepAliveClosure. We could have used
2088 // the G1CMKeepAliveClosure as it is MT-safe. Also reference objects are
2089 // placed on to discovered ref lists once so we can mark and push with no
2090 // need to check whether the object has already been marked. Using the
2091 // G1CMKeepAliveClosure would mean, however, having all the worker threads
2092 // operating on the global mark stack. This means that an individual
2093 // worker would be doing lock-free pushes while it processes its own
2094 // discovered ref list followed by drain call. If the discovered ref lists
2095 // are unbalanced then this could cause interference with the other
2096 // workers. Using a CMTask (and its embedded local data structures)
2097 // avoids that potential interference.
2098 class G1CMParKeepAliveAndDrainClosure: public OopClosure {
2099   ConcurrentMark*  _cm;
2100   CMTask*          _task;
2101   int              _ref_counter_limit;
2102   int              _ref_counter;
2103  public:
2104   G1CMParKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task) :
2105     _cm(cm), _task(task),
2106     _ref_counter_limit(G1RefProcDrainInterval) {
2107     assert(_ref_counter_limit > 0, "sanity");
2108     _ref_counter = _ref_counter_limit;
2109   }
2110 
2111   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2112   virtual void do_oop(      oop* p) { do_oop_work(p); }
2113 
2114   template <class T> void do_oop_work(T* p) {
2115     if (!_cm->has_overflown()) {
2116       oop obj = oopDesc::load_decode_heap_oop(p);
2117       if (_cm->verbose_high()) {
2118         gclog_or_tty->print_cr("\t[%d] we're looking at location "
2119                                "*"PTR_FORMAT" = "PTR_FORMAT,
2120                                _task->task_id(), p, (void*) obj);
2121       }
2122 
2123       _task->deal_with_reference(obj);
2124       _ref_counter--;
2125 
2126       if (_ref_counter == 0) {
2127         // We have dealt with _ref_counter_limit references, pushing them and objects
2128         // reachable from them on to the local stack (and possibly the global stack).
2129         // Call do_marking_step() to process these entries. We call the routine in a
2130         // loop, which we'll exit if there's nothing more to do (i.e. we're done
2131         // with the entries that we've pushed as a result of the deal_with_reference
2132         // calls above) or we overflow.
2133         // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() flag
2134         // while there may still be some work to do. (See the comment at the
2135         // beginning of CMTask::do_marking_step() for those conditions - one of which
2136         // is reaching the specified time target.) It is only when
2137         // CMTask::do_marking_step() returns without setting the has_aborted() flag
2138         // that the marking has completed.
2139         do {
2140           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2141           _task->do_marking_step(mark_step_duration_ms,
2142                                  false /* do_stealing    */,
2143                                  false /* do_termination */);
2144         } while (_task->has_aborted() && !_cm->has_overflown());
2145         _ref_counter = _ref_counter_limit;
2146       }
2147     } else {
2148       if (_cm->verbose_high()) {
2149          gclog_or_tty->print_cr("\t[%d] CM Overflow", _task->task_id());
2150       }
2151     }
2152   }
2153 };
2154 
2155 class G1CMParDrainMarkingStackClosure: public VoidClosure {
2156   ConcurrentMark* _cm;
2157   CMTask* _task;
2158  public:
2159   G1CMParDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task) :
2160     _cm(cm), _task(task) { }
2161 
2162   void do_void() {
2163     do {
2164       if (_cm->verbose_high()) {
2165         gclog_or_tty->print_cr("\t[%d] Drain: Calling do marking_step",
2166                                _task->task_id());
2167       }
2168 
2169       // We call CMTask::do_marking_step() to completely drain the local and
2170       // global marking stacks. The routine is called in a loop, which we'll
2171       // exit if there's nothing more to do (i.e. we'completely drained the
2172       // entries that were pushed as a result of applying the
2173       // G1CMParKeepAliveAndDrainClosure to the entries on the discovered ref
2174       // lists above) or we overflow the global marking stack.
2175       // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() flag
2176       // while there may still be some work to do. (See the comment at the
2177       // beginning of CMTask::do_marking_step() for those conditions - one of which
2178       // is reaching the specified time target.) It is only when
2179       // CMTask::do_marking_step() returns without setting the has_aborted() flag
2180       // that the marking has completed.
2181 
2182       _task->do_marking_step(1000000000.0 /* something very large */,
2183                              true /* do_stealing    */,
2184                              true /* do_termination */);
2185     } while (_task->has_aborted() && !_cm->has_overflown());
2186   }
2187 };
2188 
2189 // Implementation of AbstractRefProcTaskExecutor for parallel
2190 // reference processing at the end of G1 concurrent marking
2191 
2192 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2193 private:
2194   G1CollectedHeap* _g1h;
2195   ConcurrentMark*  _cm;
2196   WorkGang*        _workers;
2197   int              _active_workers;
2198 
2199 public:
2200   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2201                         ConcurrentMark* cm,
2202                         WorkGang* workers,
2203                         int n_workers) :
2204     _g1h(g1h), _cm(cm),
2205     _workers(workers), _active_workers(n_workers) { }
2206 
2207   // Executes the given task using concurrent marking worker threads.
2208   virtual void execute(ProcessTask& task);
2209   virtual void execute(EnqueueTask& task);
2210 };
2211 
2212 class G1CMRefProcTaskProxy: public AbstractGangTask {
2213   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2214   ProcessTask&     _proc_task;
2215   G1CollectedHeap* _g1h;
2216   ConcurrentMark*  _cm;
2217 
2218 public:
2219   G1CMRefProcTaskProxy(ProcessTask& proc_task,
2220                      G1CollectedHeap* g1h,
2221                      ConcurrentMark* cm) :
2222     AbstractGangTask("Process reference objects in parallel"),
2223     _proc_task(proc_task), _g1h(g1h), _cm(cm) { }
2224 
2225   virtual void work(uint worker_id) {
2226     CMTask* marking_task = _cm->task(worker_id);
2227     G1CMIsAliveClosure g1_is_alive(_g1h);
2228     G1CMParKeepAliveAndDrainClosure g1_par_keep_alive(_cm, marking_task);
2229     G1CMParDrainMarkingStackClosure g1_par_drain(_cm, marking_task);
2230 
2231     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2232   }
2233 };
2234 
2235 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2236   assert(_workers != NULL, "Need parallel worker threads.");
2237 
2238   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2239 
2240   // We need to reset the phase for each task execution so that
2241   // the termination protocol of CMTask::do_marking_step works.
2242   _cm->set_phase(_active_workers, false /* concurrent */);
2243   _g1h->set_par_threads(_active_workers);
2244   _workers->run_task(&proc_task_proxy);
2245   _g1h->set_par_threads(0);
2246 }
2247 
2248 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2249   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2250   EnqueueTask& _enq_task;
2251 
2252 public:
2253   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2254     AbstractGangTask("Enqueue reference objects in parallel"),
2255     _enq_task(enq_task) { }
2256 
2257   virtual void work(uint worker_id) {
2258     _enq_task.work(worker_id);
2259   }
2260 };
2261 
2262 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2263   assert(_workers != NULL, "Need parallel worker threads.");
2264 
2265   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2266 
2267   _g1h->set_par_threads(_active_workers);
2268   _workers->run_task(&enq_task_proxy);
2269   _g1h->set_par_threads(0);
2270 }
2271 
2272 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2273   ResourceMark rm;
2274   HandleMark   hm;
2275 
2276   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2277 
2278   // Is alive closure.
2279   G1CMIsAliveClosure g1_is_alive(g1h);
2280 
2281   // Inner scope to exclude the cleaning of the string and symbol
2282   // tables from the displayed time.
2283   {
2284     if (G1Log::finer()) {
2285       gclog_or_tty->put(' ');
2286     }
2287     TraceTime t("GC ref-proc", G1Log::finer(), false, gclog_or_tty);
2288 
2289     ReferenceProcessor* rp = g1h->ref_processor_cm();
2290 
2291     // See the comment in G1CollectedHeap::ref_processing_init()
2292     // about how reference processing currently works in G1.
2293 
2294     // Process weak references.
2295     rp->setup_policy(clear_all_soft_refs);
2296     assert(_markStack.isEmpty(), "mark stack should be empty");
2297 
2298     G1CMKeepAliveClosure g1_keep_alive(g1h, this);
2299     G1CMDrainMarkingStackClosure
2300       g1_drain_mark_stack(this, &_markStack, &g1_keep_alive);
2301 
2302     // We use the work gang from the G1CollectedHeap and we utilize all
2303     // the worker threads.
2304     uint active_workers = g1h->workers() ? g1h->workers()->active_workers() : 1U;
2305     active_workers = MAX2(MIN2(active_workers, _max_task_num), 1U);
2306 
2307     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2308                                               g1h->workers(), active_workers);
2309 
2310     if (rp->processing_is_mt()) {
2311       // Set the degree of MT here.  If the discovery is done MT, there
2312       // may have been a different number of threads doing the discovery
2313       // and a different number of discovered lists may have Ref objects.
2314       // That is OK as long as the Reference lists are balanced (see
2315       // balance_all_queues() and balance_queues()).
2316       rp->set_active_mt_degree(active_workers);
2317 
2318       rp->process_discovered_references(&g1_is_alive,
2319                                       &g1_keep_alive,
2320                                       &g1_drain_mark_stack,
2321                                       &par_task_executor);
2322 
2323       // The work routines of the parallel keep_alive and drain_marking_stack
2324       // will set the has_overflown flag if we overflow the global marking
2325       // stack.
2326     } else {
2327       rp->process_discovered_references(&g1_is_alive,
2328                                         &g1_keep_alive,
2329                                         &g1_drain_mark_stack,
2330                                         NULL);
2331     }
2332 
2333     assert(_markStack.overflow() || _markStack.isEmpty(),
2334             "mark stack should be empty (unless it overflowed)");
2335     if (_markStack.overflow()) {
2336       // Should have been done already when we tried to push an
2337       // entry on to the global mark stack. But let's do it again.
2338       set_has_overflown();
2339     }
2340 
2341     if (rp->processing_is_mt()) {
2342       assert(rp->num_q() == active_workers, "why not");
2343       rp->enqueue_discovered_references(&par_task_executor);
2344     } else {
2345       rp->enqueue_discovered_references();
2346     }
2347 
2348     rp->verify_no_references_recorded();
2349     assert(!rp->discovery_enabled(), "Post condition");
2350   }
2351 
2352   // Now clean up stale oops in StringTable
2353   StringTable::unlink(&g1_is_alive);
2354   // Clean up unreferenced symbols in symbol table.
2355   SymbolTable::unlink();
2356 }
2357 
2358 void ConcurrentMark::swapMarkBitMaps() {
2359   CMBitMapRO* temp = _prevMarkBitMap;
2360   _prevMarkBitMap  = (CMBitMapRO*)_nextMarkBitMap;
2361   _nextMarkBitMap  = (CMBitMap*)  temp;
2362 }
2363 
2364 class CMRemarkTask: public AbstractGangTask {
2365 private:
2366   ConcurrentMark *_cm;
2367 
2368 public:
2369   void work(uint worker_id) {
2370     // Since all available tasks are actually started, we should
2371     // only proceed if we're supposed to be actived.
2372     if (worker_id < _cm->active_tasks()) {
2373       CMTask* task = _cm->task(worker_id);
2374       task->record_start_time();
2375       do {
2376         task->do_marking_step(1000000000.0 /* something very large */,
2377                               true /* do_stealing    */,
2378                               true /* do_termination */);
2379       } while (task->has_aborted() && !_cm->has_overflown());
2380       // If we overflow, then we do not want to restart. We instead
2381       // want to abort remark and do concurrent marking again.
2382       task->record_end_time();
2383     }
2384   }
2385 
2386   CMRemarkTask(ConcurrentMark* cm, int active_workers) :
2387     AbstractGangTask("Par Remark"), _cm(cm) {
2388     _cm->terminator()->reset_for_reuse(active_workers);
2389   }
2390 };
2391 
2392 void ConcurrentMark::checkpointRootsFinalWork() {
2393   ResourceMark rm;
2394   HandleMark   hm;
2395   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2396 
2397   g1h->ensure_parsability(false);
2398 
2399   if (G1CollectedHeap::use_parallel_gc_threads()) {
2400     G1CollectedHeap::StrongRootsScope srs(g1h);
2401     // this is remark, so we'll use up all active threads
2402     uint active_workers = g1h->workers()->active_workers();
2403     if (active_workers == 0) {
2404       assert(active_workers > 0, "Should have been set earlier");
2405       active_workers = (uint) ParallelGCThreads;
2406       g1h->workers()->set_active_workers(active_workers);
2407     }
2408     set_phase(active_workers, false /* concurrent */);
2409     // Leave _parallel_marking_threads at it's
2410     // value originally calculated in the ConcurrentMark
2411     // constructor and pass values of the active workers
2412     // through the gang in the task.
2413 
2414     CMRemarkTask remarkTask(this, active_workers);
2415     g1h->set_par_threads(active_workers);
2416     g1h->workers()->run_task(&remarkTask);
2417     g1h->set_par_threads(0);
2418   } else {
2419     G1CollectedHeap::StrongRootsScope srs(g1h);
2420     // this is remark, so we'll use up all available threads
2421     uint active_workers = 1;
2422     set_phase(active_workers, false /* concurrent */);
2423 
2424     CMRemarkTask remarkTask(this, active_workers);
2425     // We will start all available threads, even if we decide that the
2426     // active_workers will be fewer. The extra ones will just bail out
2427     // immediately.
2428     remarkTask.work(0);
2429   }
2430   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2431   guarantee(satb_mq_set.completed_buffers_num() == 0, "invariant");
2432 
2433   print_stats();
2434 
2435 #if VERIFY_OBJS_PROCESSED
2436   if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) {
2437     gclog_or_tty->print_cr("Processed = %d, enqueued = %d.",
2438                            _scan_obj_cl.objs_processed,
2439                            ThreadLocalObjQueue::objs_enqueued);
2440     guarantee(_scan_obj_cl.objs_processed ==
2441               ThreadLocalObjQueue::objs_enqueued,
2442               "Different number of objs processed and enqueued.");
2443   }
2444 #endif
2445 }
2446 
2447 #ifndef PRODUCT
2448 
2449 class PrintReachableOopClosure: public OopClosure {
2450 private:
2451   G1CollectedHeap* _g1h;
2452   outputStream*    _out;
2453   VerifyOption     _vo;
2454   bool             _all;
2455 
2456 public:
2457   PrintReachableOopClosure(outputStream* out,
2458                            VerifyOption  vo,
2459                            bool          all) :
2460     _g1h(G1CollectedHeap::heap()),
2461     _out(out), _vo(vo), _all(all) { }
2462 
2463   void do_oop(narrowOop* p) { do_oop_work(p); }
2464   void do_oop(      oop* p) { do_oop_work(p); }
2465 
2466   template <class T> void do_oop_work(T* p) {
2467     oop         obj = oopDesc::load_decode_heap_oop(p);
2468     const char* str = NULL;
2469     const char* str2 = "";
2470 
2471     if (obj == NULL) {
2472       str = "";
2473     } else if (!_g1h->is_in_g1_reserved(obj)) {
2474       str = " O";
2475     } else {
2476       HeapRegion* hr  = _g1h->heap_region_containing(obj);
2477       guarantee(hr != NULL, "invariant");
2478       bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2479       bool marked = _g1h->is_marked(obj, _vo);
2480 
2481       if (over_tams) {
2482         str = " >";
2483         if (marked) {
2484           str2 = " AND MARKED";
2485         }
2486       } else if (marked) {
2487         str = " M";
2488       } else {
2489         str = " NOT";
2490       }
2491     }
2492 
2493     _out->print_cr("  "PTR_FORMAT": "PTR_FORMAT"%s%s",
2494                    p, (void*) obj, str, str2);
2495   }
2496 };
2497 
2498 class PrintReachableObjectClosure : public ObjectClosure {
2499 private:
2500   G1CollectedHeap* _g1h;
2501   outputStream*    _out;
2502   VerifyOption     _vo;
2503   bool             _all;
2504   HeapRegion*      _hr;
2505 
2506 public:
2507   PrintReachableObjectClosure(outputStream* out,
2508                               VerifyOption  vo,
2509                               bool          all,
2510                               HeapRegion*   hr) :
2511     _g1h(G1CollectedHeap::heap()),
2512     _out(out), _vo(vo), _all(all), _hr(hr) { }
2513 
2514   void do_object(oop o) {
2515     bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2516     bool marked = _g1h->is_marked(o, _vo);
2517     bool print_it = _all || over_tams || marked;
2518 
2519     if (print_it) {
2520       _out->print_cr(" "PTR_FORMAT"%s",
2521                      o, (over_tams) ? " >" : (marked) ? " M" : "");
2522       PrintReachableOopClosure oopCl(_out, _vo, _all);
2523       o->oop_iterate(&oopCl);
2524     }
2525   }
2526 };
2527 
2528 class PrintReachableRegionClosure : public HeapRegionClosure {
2529 private:
2530   G1CollectedHeap* _g1h;
2531   outputStream*    _out;
2532   VerifyOption     _vo;
2533   bool             _all;
2534 
2535 public:
2536   bool doHeapRegion(HeapRegion* hr) {
2537     HeapWord* b = hr->bottom();
2538     HeapWord* e = hr->end();
2539     HeapWord* t = hr->top();
2540     HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2541     _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2542                    "TAMS: "PTR_FORMAT, b, e, t, p);
2543     _out->cr();
2544 
2545     HeapWord* from = b;
2546     HeapWord* to   = t;
2547 
2548     if (to > from) {
2549       _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to);
2550       _out->cr();
2551       PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2552       hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2553       _out->cr();
2554     }
2555 
2556     return false;
2557   }
2558 
2559   PrintReachableRegionClosure(outputStream* out,
2560                               VerifyOption  vo,
2561                               bool          all) :
2562     _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2563 };
2564 
2565 void ConcurrentMark::print_reachable(const char* str,
2566                                      VerifyOption vo,
2567                                      bool all) {
2568   gclog_or_tty->cr();
2569   gclog_or_tty->print_cr("== Doing heap dump... ");
2570 
2571   if (G1PrintReachableBaseFile == NULL) {
2572     gclog_or_tty->print_cr("  #### error: no base file defined");
2573     return;
2574   }
2575 
2576   if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2577       (JVM_MAXPATHLEN - 1)) {
2578     gclog_or_tty->print_cr("  #### error: file name too long");
2579     return;
2580   }
2581 
2582   char file_name[JVM_MAXPATHLEN];
2583   sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2584   gclog_or_tty->print_cr("  dumping to file %s", file_name);
2585 
2586   fileStream fout(file_name);
2587   if (!fout.is_open()) {
2588     gclog_or_tty->print_cr("  #### error: could not open file");
2589     return;
2590   }
2591 
2592   outputStream* out = &fout;
2593   out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2594   out->cr();
2595 
2596   out->print_cr("--- ITERATING OVER REGIONS");
2597   out->cr();
2598   PrintReachableRegionClosure rcl(out, vo, all);
2599   _g1h->heap_region_iterate(&rcl);
2600   out->cr();
2601 
2602   gclog_or_tty->print_cr("  done");
2603   gclog_or_tty->flush();
2604 }
2605 
2606 #endif // PRODUCT
2607 
2608 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2609   // Note we are overriding the read-only view of the prev map here, via
2610   // the cast.
2611   ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2612 }
2613 
2614 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2615   _nextMarkBitMap->clearRange(mr);
2616 }
2617 
2618 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2619   clearRangePrevBitmap(mr);
2620   clearRangeNextBitmap(mr);
2621 }
2622 
2623 HeapRegion*
2624 ConcurrentMark::claim_region(int task_num) {
2625   // "checkpoint" the finger
2626   HeapWord* finger = _finger;
2627 
2628   // _heap_end will not change underneath our feet; it only changes at
2629   // yield points.
2630   while (finger < _heap_end) {
2631     assert(_g1h->is_in_g1_reserved(finger), "invariant");
2632 
2633     // Note on how this code handles humongous regions. In the
2634     // normal case the finger will reach the start of a "starts
2635     // humongous" (SH) region. Its end will either be the end of the
2636     // last "continues humongous" (CH) region in the sequence, or the
2637     // standard end of the SH region (if the SH is the only region in
2638     // the sequence). That way claim_region() will skip over the CH
2639     // regions. However, there is a subtle race between a CM thread
2640     // executing this method and a mutator thread doing a humongous
2641     // object allocation. The two are not mutually exclusive as the CM
2642     // thread does not need to hold the Heap_lock when it gets
2643     // here. So there is a chance that claim_region() will come across
2644     // a free region that's in the progress of becoming a SH or a CH
2645     // region. In the former case, it will either
2646     //   a) Miss the update to the region's end, in which case it will
2647     //      visit every subsequent CH region, will find their bitmaps
2648     //      empty, and do nothing, or
2649     //   b) Will observe the update of the region's end (in which case
2650     //      it will skip the subsequent CH regions).
2651     // If it comes across a region that suddenly becomes CH, the
2652     // scenario will be similar to b). So, the race between
2653     // claim_region() and a humongous object allocation might force us
2654     // to do a bit of unnecessary work (due to some unnecessary bitmap
2655     // iterations) but it should not introduce and correctness issues.
2656     HeapRegion* curr_region   = _g1h->heap_region_containing_raw(finger);
2657     HeapWord*   bottom        = curr_region->bottom();
2658     HeapWord*   end           = curr_region->end();
2659     HeapWord*   limit         = curr_region->next_top_at_mark_start();
2660 
2661     if (verbose_low()) {
2662       gclog_or_tty->print_cr("[%d] curr_region = "PTR_FORMAT" "
2663                              "["PTR_FORMAT", "PTR_FORMAT"), "
2664                              "limit = "PTR_FORMAT,
2665                              task_num, curr_region, bottom, end, limit);
2666     }
2667 
2668     // Is the gap between reading the finger and doing the CAS too long?
2669     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2670     if (res == finger) {
2671       // we succeeded
2672 
2673       // notice that _finger == end cannot be guaranteed here since,
2674       // someone else might have moved the finger even further
2675       assert(_finger >= end, "the finger should have moved forward");
2676 
2677       if (verbose_low()) {
2678         gclog_or_tty->print_cr("[%d] we were successful with region = "
2679                                PTR_FORMAT, task_num, curr_region);
2680       }
2681 
2682       if (limit > bottom) {
2683         if (verbose_low()) {
2684           gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is not empty, "
2685                                  "returning it ", task_num, curr_region);
2686         }
2687         return curr_region;
2688       } else {
2689         assert(limit == bottom,
2690                "the region limit should be at bottom");
2691         if (verbose_low()) {
2692           gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is empty, "
2693                                  "returning NULL", task_num, curr_region);
2694         }
2695         // we return NULL and the caller should try calling
2696         // claim_region() again.
2697         return NULL;
2698       }
2699     } else {
2700       assert(_finger > finger, "the finger should have moved forward");
2701       if (verbose_low()) {
2702         gclog_or_tty->print_cr("[%d] somebody else moved the finger, "
2703                                "global finger = "PTR_FORMAT", "
2704                                "our finger = "PTR_FORMAT,
2705                                task_num, _finger, finger);
2706       }
2707 
2708       // read it again
2709       finger = _finger;
2710     }
2711   }
2712 
2713   return NULL;
2714 }
2715 
2716 #ifndef PRODUCT
2717 enum VerifyNoCSetOopsPhase {
2718   VerifyNoCSetOopsStack,
2719   VerifyNoCSetOopsQueues,
2720   VerifyNoCSetOopsSATBCompleted,
2721   VerifyNoCSetOopsSATBThread
2722 };
2723 
2724 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure  {
2725 private:
2726   G1CollectedHeap* _g1h;
2727   VerifyNoCSetOopsPhase _phase;
2728   int _info;
2729 
2730   const char* phase_str() {
2731     switch (_phase) {
2732     case VerifyNoCSetOopsStack:         return "Stack";
2733     case VerifyNoCSetOopsQueues:        return "Queue";
2734     case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2735     case VerifyNoCSetOopsSATBThread:    return "Thread SATB Buffers";
2736     default:                            ShouldNotReachHere();
2737     }
2738     return NULL;
2739   }
2740 
2741   void do_object_work(oop obj) {
2742     guarantee(!_g1h->obj_in_cs(obj),
2743               err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2744                       (void*) obj, phase_str(), _info));
2745   }
2746 
2747 public:
2748   VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2749 
2750   void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2751     _phase = phase;
2752     _info = info;
2753   }
2754 
2755   virtual void do_oop(oop* p) {
2756     oop obj = oopDesc::load_decode_heap_oop(p);
2757     do_object_work(obj);
2758   }
2759 
2760   virtual void do_oop(narrowOop* p) {
2761     // We should not come across narrow oops while scanning marking
2762     // stacks and SATB buffers.
2763     ShouldNotReachHere();
2764   }
2765 
2766   virtual void do_object(oop obj) {
2767     do_object_work(obj);
2768   }
2769 };
2770 
2771 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
2772                                          bool verify_enqueued_buffers,
2773                                          bool verify_thread_buffers,
2774                                          bool verify_fingers) {
2775   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2776   if (!G1CollectedHeap::heap()->mark_in_progress()) {
2777     return;
2778   }
2779 
2780   VerifyNoCSetOopsClosure cl;
2781 
2782   if (verify_stacks) {
2783     // Verify entries on the global mark stack
2784     cl.set_phase(VerifyNoCSetOopsStack);
2785     _markStack.oops_do(&cl);
2786 
2787     // Verify entries on the task queues
2788     for (int i = 0; i < (int) _max_task_num; i += 1) {
2789       cl.set_phase(VerifyNoCSetOopsQueues, i);
2790       OopTaskQueue* queue = _task_queues->queue(i);
2791       queue->oops_do(&cl);
2792     }
2793   }
2794 
2795   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
2796 
2797   // Verify entries on the enqueued SATB buffers
2798   if (verify_enqueued_buffers) {
2799     cl.set_phase(VerifyNoCSetOopsSATBCompleted);
2800     satb_qs.iterate_completed_buffers_read_only(&cl);
2801   }
2802 
2803   // Verify entries on the per-thread SATB buffers
2804   if (verify_thread_buffers) {
2805     cl.set_phase(VerifyNoCSetOopsSATBThread);
2806     satb_qs.iterate_thread_buffers_read_only(&cl);
2807   }
2808 
2809   if (verify_fingers) {
2810     // Verify the global finger
2811     HeapWord* global_finger = finger();
2812     if (global_finger != NULL && global_finger < _heap_end) {
2813       // The global finger always points to a heap region boundary. We
2814       // use heap_region_containing_raw() to get the containing region
2815       // given that the global finger could be pointing to a free region
2816       // which subsequently becomes continues humongous. If that
2817       // happens, heap_region_containing() will return the bottom of the
2818       // corresponding starts humongous region and the check below will
2819       // not hold any more.
2820       HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
2821       guarantee(global_finger == global_hr->bottom(),
2822                 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
2823                         global_finger, HR_FORMAT_PARAMS(global_hr)));
2824     }
2825 
2826     // Verify the task fingers
2827     assert(parallel_marking_threads() <= _max_task_num, "sanity");
2828     for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
2829       CMTask* task = _tasks[i];
2830       HeapWord* task_finger = task->finger();
2831       if (task_finger != NULL && task_finger < _heap_end) {
2832         // See above note on the global finger verification.
2833         HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
2834         guarantee(task_finger == task_hr->bottom() ||
2835                   !task_hr->in_collection_set(),
2836                   err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
2837                           task_finger, HR_FORMAT_PARAMS(task_hr)));
2838       }
2839     }
2840   }
2841 }
2842 #endif // PRODUCT
2843 
2844 void ConcurrentMark::clear_marking_state(bool clear_overflow) {
2845   _markStack.setEmpty();
2846   _markStack.clear_overflow();
2847   if (clear_overflow) {
2848     clear_has_overflown();
2849   } else {
2850     assert(has_overflown(), "pre-condition");
2851   }
2852   _finger = _heap_start;
2853 
2854   for (int i = 0; i < (int)_max_task_num; ++i) {
2855     OopTaskQueue* queue = _task_queues->queue(i);
2856     queue->set_empty();
2857   }
2858 }
2859 
2860 // Aggregate the counting data that was constructed concurrently
2861 // with marking.
2862 class AggregateCountDataHRClosure: public HeapRegionClosure {
2863   G1CollectedHeap* _g1h;
2864   ConcurrentMark* _cm;
2865   CardTableModRefBS* _ct_bs;
2866   BitMap* _cm_card_bm;
2867   size_t _max_task_num;
2868 
2869  public:
2870   AggregateCountDataHRClosure(G1CollectedHeap* g1h,
2871                               BitMap* cm_card_bm,
2872                               size_t max_task_num) :
2873     _g1h(g1h), _cm(g1h->concurrent_mark()),
2874     _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
2875     _cm_card_bm(cm_card_bm), _max_task_num(max_task_num) { }
2876 
2877   bool doHeapRegion(HeapRegion* hr) {
2878     if (hr->continuesHumongous()) {
2879       // We will ignore these here and process them when their
2880       // associated "starts humongous" region is processed.
2881       // Note that we cannot rely on their associated
2882       // "starts humongous" region to have their bit set to 1
2883       // since, due to the region chunking in the parallel region
2884       // iteration, a "continues humongous" region might be visited
2885       // before its associated "starts humongous".
2886       return false;
2887     }
2888 
2889     HeapWord* start = hr->bottom();
2890     HeapWord* limit = hr->next_top_at_mark_start();
2891     HeapWord* end = hr->end();
2892 
2893     assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
2894            err_msg("Preconditions not met - "
2895                    "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
2896                    "top: "PTR_FORMAT", end: "PTR_FORMAT,
2897                    start, limit, hr->top(), hr->end()));
2898 
2899     assert(hr->next_marked_bytes() == 0, "Precondition");
2900 
2901     if (start == limit) {
2902       // NTAMS of this region has not been set so nothing to do.
2903       return false;
2904     }
2905 
2906     // 'start' should be in the heap.
2907     assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
2908     // 'end' *may* be just beyone the end of the heap (if hr is the last region)
2909     assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
2910 
2911     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
2912     BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
2913     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
2914 
2915     // If ntams is not card aligned then we bump card bitmap index
2916     // for limit so that we get the all the cards spanned by
2917     // the object ending at ntams.
2918     // Note: if this is the last region in the heap then ntams
2919     // could be actually just beyond the end of the the heap;
2920     // limit_idx will then  correspond to a (non-existent) card
2921     // that is also outside the heap.
2922     if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
2923       limit_idx += 1;
2924     }
2925 
2926     assert(limit_idx <= end_idx, "or else use atomics");
2927 
2928     // Aggregate the "stripe" in the count data associated with hr.
2929     uint hrs_index = hr->hrs_index();
2930     size_t marked_bytes = 0;
2931 
2932     for (int i = 0; (size_t)i < _max_task_num; i += 1) {
2933       size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
2934       BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
2935 
2936       // Fetch the marked_bytes in this region for task i and
2937       // add it to the running total for this region.
2938       marked_bytes += marked_bytes_array[hrs_index];
2939 
2940       // Now union the bitmaps[0,max_task_num)[start_idx..limit_idx)
2941       // into the global card bitmap.
2942       BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
2943 
2944       while (scan_idx < limit_idx) {
2945         assert(task_card_bm->at(scan_idx) == true, "should be");
2946         _cm_card_bm->set_bit(scan_idx);
2947         assert(_cm_card_bm->at(scan_idx) == true, "should be");
2948 
2949         // BitMap::get_next_one_offset() can handle the case when
2950         // its left_offset parameter is greater than its right_offset
2951         // parameter. It does, however, have an early exit if
2952         // left_offset == right_offset. So let's limit the value
2953         // passed in for left offset here.
2954         BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
2955         scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
2956       }
2957     }
2958 
2959     // Update the marked bytes for this region.
2960     hr->add_to_marked_bytes(marked_bytes);
2961 
2962     // Next heap region
2963     return false;
2964   }
2965 };
2966 
2967 class G1AggregateCountDataTask: public AbstractGangTask {
2968 protected:
2969   G1CollectedHeap* _g1h;
2970   ConcurrentMark* _cm;
2971   BitMap* _cm_card_bm;
2972   size_t _max_task_num;
2973   int _active_workers;
2974 
2975 public:
2976   G1AggregateCountDataTask(G1CollectedHeap* g1h,
2977                            ConcurrentMark* cm,
2978                            BitMap* cm_card_bm,
2979                            size_t max_task_num,
2980                            int n_workers) :
2981     AbstractGangTask("Count Aggregation"),
2982     _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
2983     _max_task_num(max_task_num),
2984     _active_workers(n_workers) { }
2985 
2986   void work(uint worker_id) {
2987     AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_task_num);
2988 
2989     if (G1CollectedHeap::use_parallel_gc_threads()) {
2990       _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
2991                                             _active_workers,
2992                                             HeapRegion::AggregateCountClaimValue);
2993     } else {
2994       _g1h->heap_region_iterate(&cl);
2995     }
2996   }
2997 };
2998 
2999 
3000 void ConcurrentMark::aggregate_count_data() {
3001   int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3002                         _g1h->workers()->active_workers() :
3003                         1);
3004 
3005   G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3006                                            _max_task_num, n_workers);
3007 
3008   if (G1CollectedHeap::use_parallel_gc_threads()) {
3009     assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3010            "sanity check");
3011     _g1h->set_par_threads(n_workers);
3012     _g1h->workers()->run_task(&g1_par_agg_task);
3013     _g1h->set_par_threads(0);
3014 
3015     assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3016            "sanity check");
3017     _g1h->reset_heap_region_claim_values();
3018   } else {
3019     g1_par_agg_task.work(0);
3020   }
3021 }
3022 
3023 // Clear the per-worker arrays used to store the per-region counting data
3024 void ConcurrentMark::clear_all_count_data() {
3025   // Clear the global card bitmap - it will be filled during
3026   // liveness count aggregation (during remark) and the
3027   // final counting task.
3028   _card_bm.clear();
3029 
3030   // Clear the global region bitmap - it will be filled as part
3031   // of the final counting task.
3032   _region_bm.clear();
3033 
3034   uint max_regions = _g1h->max_regions();
3035   assert(_max_task_num != 0, "unitialized");
3036 
3037   for (int i = 0; (size_t) i < _max_task_num; i += 1) {
3038     BitMap* task_card_bm = count_card_bitmap_for(i);
3039     size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3040 
3041     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3042     assert(marked_bytes_array != NULL, "uninitialized");
3043 
3044     memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3045     task_card_bm->clear();
3046   }
3047 }
3048 
3049 void ConcurrentMark::print_stats() {
3050   if (verbose_stats()) {
3051     gclog_or_tty->print_cr("---------------------------------------------------------------------");
3052     for (size_t i = 0; i < _active_tasks; ++i) {
3053       _tasks[i]->print_stats();
3054       gclog_or_tty->print_cr("---------------------------------------------------------------------");
3055     }
3056   }
3057 }
3058 
3059 // abandon current marking iteration due to a Full GC
3060 void ConcurrentMark::abort() {
3061   // Clear all marks to force marking thread to do nothing
3062   _nextMarkBitMap->clearAll();
3063   // Clear the liveness counting data
3064   clear_all_count_data();
3065   // Empty mark stack
3066   clear_marking_state();
3067   for (int i = 0; i < (int)_max_task_num; ++i) {
3068     _tasks[i]->clear_region_fields();
3069   }
3070   _has_aborted = true;
3071 
3072   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3073   satb_mq_set.abandon_partial_marking();
3074   // This can be called either during or outside marking, we'll read
3075   // the expected_active value from the SATB queue set.
3076   satb_mq_set.set_active_all_threads(
3077                                  false, /* new active value */
3078                                  satb_mq_set.is_active() /* expected_active */);
3079 }
3080 
3081 static void print_ms_time_info(const char* prefix, const char* name,
3082                                NumberSeq& ns) {
3083   gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3084                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3085   if (ns.num() > 0) {
3086     gclog_or_tty->print_cr("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
3087                            prefix, ns.sd(), ns.maximum());
3088   }
3089 }
3090 
3091 void ConcurrentMark::print_summary_info() {
3092   gclog_or_tty->print_cr(" Concurrent marking:");
3093   print_ms_time_info("  ", "init marks", _init_times);
3094   print_ms_time_info("  ", "remarks", _remark_times);
3095   {
3096     print_ms_time_info("     ", "final marks", _remark_mark_times);
3097     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
3098 
3099   }
3100   print_ms_time_info("  ", "cleanups", _cleanup_times);
3101   gclog_or_tty->print_cr("    Final counting total time = %8.2f s (avg = %8.2f ms).",
3102                          _total_counting_time,
3103                          (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3104                           (double)_cleanup_times.num()
3105                          : 0.0));
3106   if (G1ScrubRemSets) {
3107     gclog_or_tty->print_cr("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
3108                            _total_rs_scrub_time,
3109                            (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3110                             (double)_cleanup_times.num()
3111                            : 0.0));
3112   }
3113   gclog_or_tty->print_cr("  Total stop_world time = %8.2f s.",
3114                          (_init_times.sum() + _remark_times.sum() +
3115                           _cleanup_times.sum())/1000.0);
3116   gclog_or_tty->print_cr("  Total concurrent time = %8.2f s "
3117                 "(%8.2f s marking).",
3118                 cmThread()->vtime_accum(),
3119                 cmThread()->vtime_mark_accum());
3120 }
3121 
3122 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3123   _parallel_workers->print_worker_threads_on(st);
3124 }
3125 
3126 // We take a break if someone is trying to stop the world.
3127 bool ConcurrentMark::do_yield_check(uint worker_id) {
3128   if (should_yield()) {
3129     if (worker_id == 0) {
3130       _g1h->g1_policy()->record_concurrent_pause();
3131     }
3132     cmThread()->yield();
3133     return true;
3134   } else {
3135     return false;
3136   }
3137 }
3138 
3139 bool ConcurrentMark::should_yield() {
3140   return cmThread()->should_yield();
3141 }
3142 
3143 bool ConcurrentMark::containing_card_is_marked(void* p) {
3144   size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3145   return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3146 }
3147 
3148 bool ConcurrentMark::containing_cards_are_marked(void* start,
3149                                                  void* last) {
3150   return containing_card_is_marked(start) &&
3151          containing_card_is_marked(last);
3152 }
3153 
3154 #ifndef PRODUCT
3155 // for debugging purposes
3156 void ConcurrentMark::print_finger() {
3157   gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3158                          _heap_start, _heap_end, _finger);
3159   for (int i = 0; i < (int) _max_task_num; ++i) {
3160     gclog_or_tty->print("   %d: "PTR_FORMAT, i, _tasks[i]->finger());
3161   }
3162   gclog_or_tty->print_cr("");
3163 }
3164 #endif
3165 
3166 void CMTask::scan_object(oop obj) {
3167   assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3168 
3169   if (_cm->verbose_high()) {
3170     gclog_or_tty->print_cr("[%d] we're scanning object "PTR_FORMAT,
3171                            _task_id, (void*) obj);
3172   }
3173 
3174   size_t obj_size = obj->size();
3175   _words_scanned += obj_size;
3176 
3177   obj->oop_iterate(_cm_oop_closure);
3178   statsOnly( ++_objs_scanned );
3179   check_limits();
3180 }
3181 
3182 // Closure for iteration over bitmaps
3183 class CMBitMapClosure : public BitMapClosure {
3184 private:
3185   // the bitmap that is being iterated over
3186   CMBitMap*                   _nextMarkBitMap;
3187   ConcurrentMark*             _cm;
3188   CMTask*                     _task;
3189 
3190 public:
3191   CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3192     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3193 
3194   bool do_bit(size_t offset) {
3195     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3196     assert(_nextMarkBitMap->isMarked(addr), "invariant");
3197     assert( addr < _cm->finger(), "invariant");
3198 
3199     statsOnly( _task->increase_objs_found_on_bitmap() );
3200     assert(addr >= _task->finger(), "invariant");
3201 
3202     // We move that task's local finger along.
3203     _task->move_finger_to(addr);
3204 
3205     _task->scan_object(oop(addr));
3206     // we only partially drain the local queue and global stack
3207     _task->drain_local_queue(true);
3208     _task->drain_global_stack(true);
3209 
3210     // if the has_aborted flag has been raised, we need to bail out of
3211     // the iteration
3212     return !_task->has_aborted();
3213   }
3214 };
3215 
3216 // Closure for iterating over objects, currently only used for
3217 // processing SATB buffers.
3218 class CMObjectClosure : public ObjectClosure {
3219 private:
3220   CMTask* _task;
3221 
3222 public:
3223   void do_object(oop obj) {
3224     _task->deal_with_reference(obj);
3225   }
3226 
3227   CMObjectClosure(CMTask* task) : _task(task) { }
3228 };
3229 
3230 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3231                                ConcurrentMark* cm,
3232                                CMTask* task)
3233   : _g1h(g1h), _cm(cm), _task(task) {
3234   assert(_ref_processor == NULL, "should be initialized to NULL");
3235 
3236   if (G1UseConcMarkReferenceProcessing) {
3237     _ref_processor = g1h->ref_processor_cm();
3238     assert(_ref_processor != NULL, "should not be NULL");
3239   }
3240 }
3241 
3242 void CMTask::setup_for_region(HeapRegion* hr) {
3243   // Separated the asserts so that we know which one fires.
3244   assert(hr != NULL,
3245         "claim_region() should have filtered out continues humongous regions");
3246   assert(!hr->continuesHumongous(),
3247         "claim_region() should have filtered out continues humongous regions");
3248 
3249   if (_cm->verbose_low()) {
3250     gclog_or_tty->print_cr("[%d] setting up for region "PTR_FORMAT,
3251                            _task_id, hr);
3252   }
3253 
3254   _curr_region  = hr;
3255   _finger       = hr->bottom();
3256   update_region_limit();
3257 }
3258 
3259 void CMTask::update_region_limit() {
3260   HeapRegion* hr            = _curr_region;
3261   HeapWord* bottom          = hr->bottom();
3262   HeapWord* limit           = hr->next_top_at_mark_start();
3263 
3264   if (limit == bottom) {
3265     if (_cm->verbose_low()) {
3266       gclog_or_tty->print_cr("[%d] found an empty region "
3267                              "["PTR_FORMAT", "PTR_FORMAT")",
3268                              _task_id, bottom, limit);
3269     }
3270     // The region was collected underneath our feet.
3271     // We set the finger to bottom to ensure that the bitmap
3272     // iteration that will follow this will not do anything.
3273     // (this is not a condition that holds when we set the region up,
3274     // as the region is not supposed to be empty in the first place)
3275     _finger = bottom;
3276   } else if (limit >= _region_limit) {
3277     assert(limit >= _finger, "peace of mind");
3278   } else {
3279     assert(limit < _region_limit, "only way to get here");
3280     // This can happen under some pretty unusual circumstances.  An
3281     // evacuation pause empties the region underneath our feet (NTAMS
3282     // at bottom). We then do some allocation in the region (NTAMS
3283     // stays at bottom), followed by the region being used as a GC
3284     // alloc region (NTAMS will move to top() and the objects
3285     // originally below it will be grayed). All objects now marked in
3286     // the region are explicitly grayed, if below the global finger,
3287     // and we do not need in fact to scan anything else. So, we simply
3288     // set _finger to be limit to ensure that the bitmap iteration
3289     // doesn't do anything.
3290     _finger = limit;
3291   }
3292 
3293   _region_limit = limit;
3294 }
3295 
3296 void CMTask::giveup_current_region() {
3297   assert(_curr_region != NULL, "invariant");
3298   if (_cm->verbose_low()) {
3299     gclog_or_tty->print_cr("[%d] giving up region "PTR_FORMAT,
3300                            _task_id, _curr_region);
3301   }
3302   clear_region_fields();
3303 }
3304 
3305 void CMTask::clear_region_fields() {
3306   // Values for these three fields that indicate that we're not
3307   // holding on to a region.
3308   _curr_region   = NULL;
3309   _finger        = NULL;
3310   _region_limit  = NULL;
3311 }
3312 
3313 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3314   if (cm_oop_closure == NULL) {
3315     assert(_cm_oop_closure != NULL, "invariant");
3316   } else {
3317     assert(_cm_oop_closure == NULL, "invariant");
3318   }
3319   _cm_oop_closure = cm_oop_closure;
3320 }
3321 
3322 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3323   guarantee(nextMarkBitMap != NULL, "invariant");
3324 
3325   if (_cm->verbose_low()) {
3326     gclog_or_tty->print_cr("[%d] resetting", _task_id);
3327   }
3328 
3329   _nextMarkBitMap                = nextMarkBitMap;
3330   clear_region_fields();
3331 
3332   _calls                         = 0;
3333   _elapsed_time_ms               = 0.0;
3334   _termination_time_ms           = 0.0;
3335   _termination_start_time_ms     = 0.0;
3336 
3337 #if _MARKING_STATS_
3338   _local_pushes                  = 0;
3339   _local_pops                    = 0;
3340   _local_max_size                = 0;
3341   _objs_scanned                  = 0;
3342   _global_pushes                 = 0;
3343   _global_pops                   = 0;
3344   _global_max_size               = 0;
3345   _global_transfers_to           = 0;
3346   _global_transfers_from         = 0;
3347   _regions_claimed               = 0;
3348   _objs_found_on_bitmap          = 0;
3349   _satb_buffers_processed        = 0;
3350   _steal_attempts                = 0;
3351   _steals                        = 0;
3352   _aborted                       = 0;
3353   _aborted_overflow              = 0;
3354   _aborted_cm_aborted            = 0;
3355   _aborted_yield                 = 0;
3356   _aborted_timed_out             = 0;
3357   _aborted_satb                  = 0;
3358   _aborted_termination           = 0;
3359 #endif // _MARKING_STATS_
3360 }
3361 
3362 bool CMTask::should_exit_termination() {
3363   regular_clock_call();
3364   // This is called when we are in the termination protocol. We should
3365   // quit if, for some reason, this task wants to abort or the global
3366   // stack is not empty (this means that we can get work from it).
3367   return !_cm->mark_stack_empty() || has_aborted();
3368 }
3369 
3370 void CMTask::reached_limit() {
3371   assert(_words_scanned >= _words_scanned_limit ||
3372          _refs_reached >= _refs_reached_limit ,
3373          "shouldn't have been called otherwise");
3374   regular_clock_call();
3375 }
3376 
3377 void CMTask::regular_clock_call() {
3378   if (has_aborted()) return;
3379 
3380   // First, we need to recalculate the words scanned and refs reached
3381   // limits for the next clock call.
3382   recalculate_limits();
3383 
3384   // During the regular clock call we do the following
3385 
3386   // (1) If an overflow has been flagged, then we abort.
3387   if (_cm->has_overflown()) {
3388     set_has_aborted();
3389     return;
3390   }
3391 
3392   // If we are not concurrent (i.e. we're doing remark) we don't need
3393   // to check anything else. The other steps are only needed during
3394   // the concurrent marking phase.
3395   if (!concurrent()) return;
3396 
3397   // (2) If marking has been aborted for Full GC, then we also abort.
3398   if (_cm->has_aborted()) {
3399     set_has_aborted();
3400     statsOnly( ++_aborted_cm_aborted );
3401     return;
3402   }
3403 
3404   double curr_time_ms = os::elapsedVTime() * 1000.0;
3405 
3406   // (3) If marking stats are enabled, then we update the step history.
3407 #if _MARKING_STATS_
3408   if (_words_scanned >= _words_scanned_limit) {
3409     ++_clock_due_to_scanning;
3410   }
3411   if (_refs_reached >= _refs_reached_limit) {
3412     ++_clock_due_to_marking;
3413   }
3414 
3415   double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3416   _interval_start_time_ms = curr_time_ms;
3417   _all_clock_intervals_ms.add(last_interval_ms);
3418 
3419   if (_cm->verbose_medium()) {
3420       gclog_or_tty->print_cr("[%d] regular clock, interval = %1.2lfms, "
3421                         "scanned = %d%s, refs reached = %d%s",
3422                         _task_id, last_interval_ms,
3423                         _words_scanned,
3424                         (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3425                         _refs_reached,
3426                         (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3427   }
3428 #endif // _MARKING_STATS_
3429 
3430   // (4) We check whether we should yield. If we have to, then we abort.
3431   if (_cm->should_yield()) {
3432     // We should yield. To do this we abort the task. The caller is
3433     // responsible for yielding.
3434     set_has_aborted();
3435     statsOnly( ++_aborted_yield );
3436     return;
3437   }
3438 
3439   // (5) We check whether we've reached our time quota. If we have,
3440   // then we abort.
3441   double elapsed_time_ms = curr_time_ms - _start_time_ms;
3442   if (elapsed_time_ms > _time_target_ms) {
3443     set_has_aborted();
3444     _has_timed_out = true;
3445     statsOnly( ++_aborted_timed_out );
3446     return;
3447   }
3448 
3449   // (6) Finally, we check whether there are enough completed STAB
3450   // buffers available for processing. If there are, we abort.
3451   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3452   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3453     if (_cm->verbose_low()) {
3454       gclog_or_tty->print_cr("[%d] aborting to deal with pending SATB buffers",
3455                              _task_id);
3456     }
3457     // we do need to process SATB buffers, we'll abort and restart
3458     // the marking task to do so
3459     set_has_aborted();
3460     statsOnly( ++_aborted_satb );
3461     return;
3462   }
3463 }
3464 
3465 void CMTask::recalculate_limits() {
3466   _real_words_scanned_limit = _words_scanned + words_scanned_period;
3467   _words_scanned_limit      = _real_words_scanned_limit;
3468 
3469   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
3470   _refs_reached_limit       = _real_refs_reached_limit;
3471 }
3472 
3473 void CMTask::decrease_limits() {
3474   // This is called when we believe that we're going to do an infrequent
3475   // operation which will increase the per byte scanned cost (i.e. move
3476   // entries to/from the global stack). It basically tries to decrease the
3477   // scanning limit so that the clock is called earlier.
3478 
3479   if (_cm->verbose_medium()) {
3480     gclog_or_tty->print_cr("[%d] decreasing limits", _task_id);
3481   }
3482 
3483   _words_scanned_limit = _real_words_scanned_limit -
3484     3 * words_scanned_period / 4;
3485   _refs_reached_limit  = _real_refs_reached_limit -
3486     3 * refs_reached_period / 4;
3487 }
3488 
3489 void CMTask::move_entries_to_global_stack() {
3490   // local array where we'll store the entries that will be popped
3491   // from the local queue
3492   oop buffer[global_stack_transfer_size];
3493 
3494   int n = 0;
3495   oop obj;
3496   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3497     buffer[n] = obj;
3498     ++n;
3499   }
3500 
3501   if (n > 0) {
3502     // we popped at least one entry from the local queue
3503 
3504     statsOnly( ++_global_transfers_to; _local_pops += n );
3505 
3506     if (!_cm->mark_stack_push(buffer, n)) {
3507       if (_cm->verbose_low()) {
3508         gclog_or_tty->print_cr("[%d] aborting due to global stack overflow",
3509                                _task_id);
3510       }
3511       set_has_aborted();
3512     } else {
3513       // the transfer was successful
3514 
3515       if (_cm->verbose_medium()) {
3516         gclog_or_tty->print_cr("[%d] pushed %d entries to the global stack",
3517                                _task_id, n);
3518       }
3519       statsOnly( int tmp_size = _cm->mark_stack_size();
3520                  if (tmp_size > _global_max_size) {
3521                    _global_max_size = tmp_size;
3522                  }
3523                  _global_pushes += n );
3524     }
3525   }
3526 
3527   // this operation was quite expensive, so decrease the limits
3528   decrease_limits();
3529 }
3530 
3531 void CMTask::get_entries_from_global_stack() {
3532   // local array where we'll store the entries that will be popped
3533   // from the global stack.
3534   oop buffer[global_stack_transfer_size];
3535   int n;
3536   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3537   assert(n <= global_stack_transfer_size,
3538          "we should not pop more than the given limit");
3539   if (n > 0) {
3540     // yes, we did actually pop at least one entry
3541 
3542     statsOnly( ++_global_transfers_from; _global_pops += n );
3543     if (_cm->verbose_medium()) {
3544       gclog_or_tty->print_cr("[%d] popped %d entries from the global stack",
3545                              _task_id, n);
3546     }
3547     for (int i = 0; i < n; ++i) {
3548       bool success = _task_queue->push(buffer[i]);
3549       // We only call this when the local queue is empty or under a
3550       // given target limit. So, we do not expect this push to fail.
3551       assert(success, "invariant");
3552     }
3553 
3554     statsOnly( int tmp_size = _task_queue->size();
3555                if (tmp_size > _local_max_size) {
3556                  _local_max_size = tmp_size;
3557                }
3558                _local_pushes += n );
3559   }
3560 
3561   // this operation was quite expensive, so decrease the limits
3562   decrease_limits();
3563 }
3564 
3565 void CMTask::drain_local_queue(bool partially) {
3566   if (has_aborted()) return;
3567 
3568   // Decide what the target size is, depending whether we're going to
3569   // drain it partially (so that other tasks can steal if they run out
3570   // of things to do) or totally (at the very end).
3571   size_t target_size;
3572   if (partially) {
3573     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3574   } else {
3575     target_size = 0;
3576   }
3577 
3578   if (_task_queue->size() > target_size) {
3579     if (_cm->verbose_high()) {
3580       gclog_or_tty->print_cr("[%d] draining local queue, target size = %d",
3581                              _task_id, target_size);
3582     }
3583 
3584     oop obj;
3585     bool ret = _task_queue->pop_local(obj);
3586     while (ret) {
3587       statsOnly( ++_local_pops );
3588 
3589       if (_cm->verbose_high()) {
3590         gclog_or_tty->print_cr("[%d] popped "PTR_FORMAT, _task_id,
3591                                (void*) obj);
3592       }
3593 
3594       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3595       assert(!_g1h->is_on_master_free_list(
3596                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3597 
3598       scan_object(obj);
3599 
3600       if (_task_queue->size() <= target_size || has_aborted()) {
3601         ret = false;
3602       } else {
3603         ret = _task_queue->pop_local(obj);
3604       }
3605     }
3606 
3607     if (_cm->verbose_high()) {
3608       gclog_or_tty->print_cr("[%d] drained local queue, size = %d",
3609                              _task_id, _task_queue->size());
3610     }
3611   }
3612 }
3613 
3614 void CMTask::drain_global_stack(bool partially) {
3615   if (has_aborted()) return;
3616 
3617   // We have a policy to drain the local queue before we attempt to
3618   // drain the global stack.
3619   assert(partially || _task_queue->size() == 0, "invariant");
3620 
3621   // Decide what the target size is, depending whether we're going to
3622   // drain it partially (so that other tasks can steal if they run out
3623   // of things to do) or totally (at the very end).  Notice that,
3624   // because we move entries from the global stack in chunks or
3625   // because another task might be doing the same, we might in fact
3626   // drop below the target. But, this is not a problem.
3627   size_t target_size;
3628   if (partially) {
3629     target_size = _cm->partial_mark_stack_size_target();
3630   } else {
3631     target_size = 0;
3632   }
3633 
3634   if (_cm->mark_stack_size() > target_size) {
3635     if (_cm->verbose_low()) {
3636       gclog_or_tty->print_cr("[%d] draining global_stack, target size %d",
3637                              _task_id, target_size);
3638     }
3639 
3640     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3641       get_entries_from_global_stack();
3642       drain_local_queue(partially);
3643     }
3644 
3645     if (_cm->verbose_low()) {
3646       gclog_or_tty->print_cr("[%d] drained global stack, size = %d",
3647                              _task_id, _cm->mark_stack_size());
3648     }
3649   }
3650 }
3651 
3652 // SATB Queue has several assumptions on whether to call the par or
3653 // non-par versions of the methods. this is why some of the code is
3654 // replicated. We should really get rid of the single-threaded version
3655 // of the code to simplify things.
3656 void CMTask::drain_satb_buffers() {
3657   if (has_aborted()) return;
3658 
3659   // We set this so that the regular clock knows that we're in the
3660   // middle of draining buffers and doesn't set the abort flag when it
3661   // notices that SATB buffers are available for draining. It'd be
3662   // very counter productive if it did that. :-)
3663   _draining_satb_buffers = true;
3664 
3665   CMObjectClosure oc(this);
3666   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3667   if (G1CollectedHeap::use_parallel_gc_threads()) {
3668     satb_mq_set.set_par_closure(_task_id, &oc);
3669   } else {
3670     satb_mq_set.set_closure(&oc);
3671   }
3672 
3673   // This keeps claiming and applying the closure to completed buffers
3674   // until we run out of buffers or we need to abort.
3675   if (G1CollectedHeap::use_parallel_gc_threads()) {
3676     while (!has_aborted() &&
3677            satb_mq_set.par_apply_closure_to_completed_buffer(_task_id)) {
3678       if (_cm->verbose_medium()) {
3679         gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id);
3680       }
3681       statsOnly( ++_satb_buffers_processed );
3682       regular_clock_call();
3683     }
3684   } else {
3685     while (!has_aborted() &&
3686            satb_mq_set.apply_closure_to_completed_buffer()) {
3687       if (_cm->verbose_medium()) {
3688         gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id);
3689       }
3690       statsOnly( ++_satb_buffers_processed );
3691       regular_clock_call();
3692     }
3693   }
3694 
3695   if (!concurrent() && !has_aborted()) {
3696     // We should only do this during remark.
3697     if (G1CollectedHeap::use_parallel_gc_threads()) {
3698       satb_mq_set.par_iterate_closure_all_threads(_task_id);
3699     } else {
3700       satb_mq_set.iterate_closure_all_threads();
3701     }
3702   }
3703 
3704   _draining_satb_buffers = false;
3705 
3706   assert(has_aborted() ||
3707          concurrent() ||
3708          satb_mq_set.completed_buffers_num() == 0, "invariant");
3709 
3710   if (G1CollectedHeap::use_parallel_gc_threads()) {
3711     satb_mq_set.set_par_closure(_task_id, NULL);
3712   } else {
3713     satb_mq_set.set_closure(NULL);
3714   }
3715 
3716   // again, this was a potentially expensive operation, decrease the
3717   // limits to get the regular clock call early
3718   decrease_limits();
3719 }
3720 
3721 void CMTask::print_stats() {
3722   gclog_or_tty->print_cr("Marking Stats, task = %d, calls = %d",
3723                          _task_id, _calls);
3724   gclog_or_tty->print_cr("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3725                          _elapsed_time_ms, _termination_time_ms);
3726   gclog_or_tty->print_cr("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3727                          _step_times_ms.num(), _step_times_ms.avg(),
3728                          _step_times_ms.sd());
3729   gclog_or_tty->print_cr("                    max = %1.2lfms, total = %1.2lfms",
3730                          _step_times_ms.maximum(), _step_times_ms.sum());
3731 
3732 #if _MARKING_STATS_
3733   gclog_or_tty->print_cr("  Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3734                          _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3735                          _all_clock_intervals_ms.sd());
3736   gclog_or_tty->print_cr("                         max = %1.2lfms, total = %1.2lfms",
3737                          _all_clock_intervals_ms.maximum(),
3738                          _all_clock_intervals_ms.sum());
3739   gclog_or_tty->print_cr("  Clock Causes (cum): scanning = %d, marking = %d",
3740                          _clock_due_to_scanning, _clock_due_to_marking);
3741   gclog_or_tty->print_cr("  Objects: scanned = %d, found on the bitmap = %d",
3742                          _objs_scanned, _objs_found_on_bitmap);
3743   gclog_or_tty->print_cr("  Local Queue:  pushes = %d, pops = %d, max size = %d",
3744                          _local_pushes, _local_pops, _local_max_size);
3745   gclog_or_tty->print_cr("  Global Stack: pushes = %d, pops = %d, max size = %d",
3746                          _global_pushes, _global_pops, _global_max_size);
3747   gclog_or_tty->print_cr("                transfers to = %d, transfers from = %d",
3748                          _global_transfers_to,_global_transfers_from);
3749   gclog_or_tty->print_cr("  Regions: claimed = %d", _regions_claimed);
3750   gclog_or_tty->print_cr("  SATB buffers: processed = %d", _satb_buffers_processed);
3751   gclog_or_tty->print_cr("  Steals: attempts = %d, successes = %d",
3752                          _steal_attempts, _steals);
3753   gclog_or_tty->print_cr("  Aborted: %d, due to", _aborted);
3754   gclog_or_tty->print_cr("    overflow: %d, global abort: %d, yield: %d",
3755                          _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3756   gclog_or_tty->print_cr("    time out: %d, SATB: %d, termination: %d",
3757                          _aborted_timed_out, _aborted_satb, _aborted_termination);
3758 #endif // _MARKING_STATS_
3759 }
3760 
3761 /*****************************************************************************
3762 
3763     The do_marking_step(time_target_ms) method is the building block
3764     of the parallel marking framework. It can be called in parallel
3765     with other invocations of do_marking_step() on different tasks
3766     (but only one per task, obviously) and concurrently with the
3767     mutator threads, or during remark, hence it eliminates the need
3768     for two versions of the code. When called during remark, it will
3769     pick up from where the task left off during the concurrent marking
3770     phase. Interestingly, tasks are also claimable during evacuation
3771     pauses too, since do_marking_step() ensures that it aborts before
3772     it needs to yield.
3773 
3774     The data structures that is uses to do marking work are the
3775     following:
3776 
3777       (1) Marking Bitmap. If there are gray objects that appear only
3778       on the bitmap (this happens either when dealing with an overflow
3779       or when the initial marking phase has simply marked the roots
3780       and didn't push them on the stack), then tasks claim heap
3781       regions whose bitmap they then scan to find gray objects. A
3782       global finger indicates where the end of the last claimed region
3783       is. A local finger indicates how far into the region a task has
3784       scanned. The two fingers are used to determine how to gray an
3785       object (i.e. whether simply marking it is OK, as it will be
3786       visited by a task in the future, or whether it needs to be also
3787       pushed on a stack).
3788 
3789       (2) Local Queue. The local queue of the task which is accessed
3790       reasonably efficiently by the task. Other tasks can steal from
3791       it when they run out of work. Throughout the marking phase, a
3792       task attempts to keep its local queue short but not totally
3793       empty, so that entries are available for stealing by other
3794       tasks. Only when there is no more work, a task will totally
3795       drain its local queue.
3796 
3797       (3) Global Mark Stack. This handles local queue overflow. During
3798       marking only sets of entries are moved between it and the local
3799       queues, as access to it requires a mutex and more fine-grain
3800       interaction with it which might cause contention. If it
3801       overflows, then the marking phase should restart and iterate
3802       over the bitmap to identify gray objects. Throughout the marking
3803       phase, tasks attempt to keep the global mark stack at a small
3804       length but not totally empty, so that entries are available for
3805       popping by other tasks. Only when there is no more work, tasks
3806       will totally drain the global mark stack.
3807 
3808       (4) SATB Buffer Queue. This is where completed SATB buffers are
3809       made available. Buffers are regularly removed from this queue
3810       and scanned for roots, so that the queue doesn't get too
3811       long. During remark, all completed buffers are processed, as
3812       well as the filled in parts of any uncompleted buffers.
3813 
3814     The do_marking_step() method tries to abort when the time target
3815     has been reached. There are a few other cases when the
3816     do_marking_step() method also aborts:
3817 
3818       (1) When the marking phase has been aborted (after a Full GC).
3819 
3820       (2) When a global overflow (on the global stack) has been
3821       triggered. Before the task aborts, it will actually sync up with
3822       the other tasks to ensure that all the marking data structures
3823       (local queues, stacks, fingers etc.)  are re-initialised so that
3824       when do_marking_step() completes, the marking phase can
3825       immediately restart.
3826 
3827       (3) When enough completed SATB buffers are available. The
3828       do_marking_step() method only tries to drain SATB buffers right
3829       at the beginning. So, if enough buffers are available, the
3830       marking step aborts and the SATB buffers are processed at
3831       the beginning of the next invocation.
3832 
3833       (4) To yield. when we have to yield then we abort and yield
3834       right at the end of do_marking_step(). This saves us from a lot
3835       of hassle as, by yielding we might allow a Full GC. If this
3836       happens then objects will be compacted underneath our feet, the
3837       heap might shrink, etc. We save checking for this by just
3838       aborting and doing the yield right at the end.
3839 
3840     From the above it follows that the do_marking_step() method should
3841     be called in a loop (or, otherwise, regularly) until it completes.
3842 
3843     If a marking step completes without its has_aborted() flag being
3844     true, it means it has completed the current marking phase (and
3845     also all other marking tasks have done so and have all synced up).
3846 
3847     A method called regular_clock_call() is invoked "regularly" (in
3848     sub ms intervals) throughout marking. It is this clock method that
3849     checks all the abort conditions which were mentioned above and
3850     decides when the task should abort. A work-based scheme is used to
3851     trigger this clock method: when the number of object words the
3852     marking phase has scanned or the number of references the marking
3853     phase has visited reach a given limit. Additional invocations to
3854     the method clock have been planted in a few other strategic places
3855     too. The initial reason for the clock method was to avoid calling
3856     vtime too regularly, as it is quite expensive. So, once it was in
3857     place, it was natural to piggy-back all the other conditions on it
3858     too and not constantly check them throughout the code.
3859 
3860  *****************************************************************************/
3861 
3862 void CMTask::do_marking_step(double time_target_ms,
3863                              bool do_stealing,
3864                              bool do_termination) {
3865   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
3866   assert(concurrent() == _cm->concurrent(), "they should be the same");
3867 
3868   G1CollectorPolicy* g1_policy = _g1h->g1_policy();
3869   assert(_task_queues != NULL, "invariant");
3870   assert(_task_queue != NULL, "invariant");
3871   assert(_task_queues->queue(_task_id) == _task_queue, "invariant");
3872 
3873   assert(!_claimed,
3874          "only one thread should claim this task at any one time");
3875 
3876   // OK, this doesn't safeguard again all possible scenarios, as it is
3877   // possible for two threads to set the _claimed flag at the same
3878   // time. But it is only for debugging purposes anyway and it will
3879   // catch most problems.
3880   _claimed = true;
3881 
3882   _start_time_ms = os::elapsedVTime() * 1000.0;
3883   statsOnly( _interval_start_time_ms = _start_time_ms );
3884 
3885   double diff_prediction_ms =
3886     g1_policy->get_new_prediction(&_marking_step_diffs_ms);
3887   _time_target_ms = time_target_ms - diff_prediction_ms;
3888 
3889   // set up the variables that are used in the work-based scheme to
3890   // call the regular clock method
3891   _words_scanned = 0;
3892   _refs_reached  = 0;
3893   recalculate_limits();
3894 
3895   // clear all flags
3896   clear_has_aborted();
3897   _has_timed_out = false;
3898   _draining_satb_buffers = false;
3899 
3900   ++_calls;
3901 
3902   if (_cm->verbose_low()) {
3903     gclog_or_tty->print_cr("[%d] >>>>>>>>>> START, call = %d, "
3904                            "target = %1.2lfms >>>>>>>>>>",
3905                            _task_id, _calls, _time_target_ms);
3906   }
3907 
3908   // Set up the bitmap and oop closures. Anything that uses them is
3909   // eventually called from this method, so it is OK to allocate these
3910   // statically.
3911   CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
3912   G1CMOopClosure  cm_oop_closure(_g1h, _cm, this);
3913   set_cm_oop_closure(&cm_oop_closure);
3914 
3915   if (_cm->has_overflown()) {
3916     // This can happen if the mark stack overflows during a GC pause
3917     // and this task, after a yield point, restarts. We have to abort
3918     // as we need to get into the overflow protocol which happens
3919     // right at the end of this task.
3920     set_has_aborted();
3921   }
3922 
3923   // First drain any available SATB buffers. After this, we will not
3924   // look at SATB buffers before the next invocation of this method.
3925   // If enough completed SATB buffers are queued up, the regular clock
3926   // will abort this task so that it restarts.
3927   drain_satb_buffers();
3928   // ...then partially drain the local queue and the global stack
3929   drain_local_queue(true);
3930   drain_global_stack(true);
3931 
3932   do {
3933     if (!has_aborted() && _curr_region != NULL) {
3934       // This means that we're already holding on to a region.
3935       assert(_finger != NULL, "if region is not NULL, then the finger "
3936              "should not be NULL either");
3937 
3938       // We might have restarted this task after an evacuation pause
3939       // which might have evacuated the region we're holding on to
3940       // underneath our feet. Let's read its limit again to make sure
3941       // that we do not iterate over a region of the heap that
3942       // contains garbage (update_region_limit() will also move
3943       // _finger to the start of the region if it is found empty).
3944       update_region_limit();
3945       // We will start from _finger not from the start of the region,
3946       // as we might be restarting this task after aborting half-way
3947       // through scanning this region. In this case, _finger points to
3948       // the address where we last found a marked object. If this is a
3949       // fresh region, _finger points to start().
3950       MemRegion mr = MemRegion(_finger, _region_limit);
3951 
3952       if (_cm->verbose_low()) {
3953         gclog_or_tty->print_cr("[%d] we're scanning part "
3954                                "["PTR_FORMAT", "PTR_FORMAT") "
3955                                "of region "PTR_FORMAT,
3956                                _task_id, _finger, _region_limit, _curr_region);
3957       }
3958 
3959       // Let's iterate over the bitmap of the part of the
3960       // region that is left.
3961       if (mr.is_empty() || _nextMarkBitMap->iterate(&bitmap_closure, mr)) {
3962         // We successfully completed iterating over the region. Now,
3963         // let's give up the region.
3964         giveup_current_region();
3965         regular_clock_call();
3966       } else {
3967         assert(has_aborted(), "currently the only way to do so");
3968         // The only way to abort the bitmap iteration is to return
3969         // false from the do_bit() method. However, inside the
3970         // do_bit() method we move the _finger to point to the
3971         // object currently being looked at. So, if we bail out, we
3972         // have definitely set _finger to something non-null.
3973         assert(_finger != NULL, "invariant");
3974 
3975         // Region iteration was actually aborted. So now _finger
3976         // points to the address of the object we last scanned. If we
3977         // leave it there, when we restart this task, we will rescan
3978         // the object. It is easy to avoid this. We move the finger by
3979         // enough to point to the next possible object header (the
3980         // bitmap knows by how much we need to move it as it knows its
3981         // granularity).
3982         assert(_finger < _region_limit, "invariant");
3983         HeapWord* new_finger = _nextMarkBitMap->nextWord(_finger);
3984         // Check if bitmap iteration was aborted while scanning the last object
3985         if (new_finger >= _region_limit) {
3986           giveup_current_region();
3987         } else {
3988           move_finger_to(new_finger);
3989         }
3990       }
3991     }
3992     // At this point we have either completed iterating over the
3993     // region we were holding on to, or we have aborted.
3994 
3995     // We then partially drain the local queue and the global stack.
3996     // (Do we really need this?)
3997     drain_local_queue(true);
3998     drain_global_stack(true);
3999 
4000     // Read the note on the claim_region() method on why it might
4001     // return NULL with potentially more regions available for
4002     // claiming and why we have to check out_of_regions() to determine
4003     // whether we're done or not.
4004     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4005       // We are going to try to claim a new region. We should have
4006       // given up on the previous one.
4007       // Separated the asserts so that we know which one fires.
4008       assert(_curr_region  == NULL, "invariant");
4009       assert(_finger       == NULL, "invariant");
4010       assert(_region_limit == NULL, "invariant");
4011       if (_cm->verbose_low()) {
4012         gclog_or_tty->print_cr("[%d] trying to claim a new region", _task_id);
4013       }
4014       HeapRegion* claimed_region = _cm->claim_region(_task_id);
4015       if (claimed_region != NULL) {
4016         // Yes, we managed to claim one
4017         statsOnly( ++_regions_claimed );
4018 
4019         if (_cm->verbose_low()) {
4020           gclog_or_tty->print_cr("[%d] we successfully claimed "
4021                                  "region "PTR_FORMAT,
4022                                  _task_id, claimed_region);
4023         }
4024 
4025         setup_for_region(claimed_region);
4026         assert(_curr_region == claimed_region, "invariant");
4027       }
4028       // It is important to call the regular clock here. It might take
4029       // a while to claim a region if, for example, we hit a large
4030       // block of empty regions. So we need to call the regular clock
4031       // method once round the loop to make sure it's called
4032       // frequently enough.
4033       regular_clock_call();
4034     }
4035 
4036     if (!has_aborted() && _curr_region == NULL) {
4037       assert(_cm->out_of_regions(),
4038              "at this point we should be out of regions");
4039     }
4040   } while ( _curr_region != NULL && !has_aborted());
4041 
4042   if (!has_aborted()) {
4043     // We cannot check whether the global stack is empty, since other
4044     // tasks might be pushing objects to it concurrently.
4045     assert(_cm->out_of_regions(),
4046            "at this point we should be out of regions");
4047 
4048     if (_cm->verbose_low()) {
4049       gclog_or_tty->print_cr("[%d] all regions claimed", _task_id);
4050     }
4051 
4052     // Try to reduce the number of available SATB buffers so that
4053     // remark has less work to do.
4054     drain_satb_buffers();
4055   }
4056 
4057   // Since we've done everything else, we can now totally drain the
4058   // local queue and global stack.
4059   drain_local_queue(false);
4060   drain_global_stack(false);
4061 
4062   // Attempt at work stealing from other task's queues.
4063   if (do_stealing && !has_aborted()) {
4064     // We have not aborted. This means that we have finished all that
4065     // we could. Let's try to do some stealing...
4066 
4067     // We cannot check whether the global stack is empty, since other
4068     // tasks might be pushing objects to it concurrently.
4069     assert(_cm->out_of_regions() && _task_queue->size() == 0,
4070            "only way to reach here");
4071 
4072     if (_cm->verbose_low()) {
4073       gclog_or_tty->print_cr("[%d] starting to steal", _task_id);
4074     }
4075 
4076     while (!has_aborted()) {
4077       oop obj;
4078       statsOnly( ++_steal_attempts );
4079 
4080       if (_cm->try_stealing(_task_id, &_hash_seed, obj)) {
4081         if (_cm->verbose_medium()) {
4082           gclog_or_tty->print_cr("[%d] stolen "PTR_FORMAT" successfully",
4083                                  _task_id, (void*) obj);
4084         }
4085 
4086         statsOnly( ++_steals );
4087 
4088         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4089                "any stolen object should be marked");
4090         scan_object(obj);
4091 
4092         // And since we're towards the end, let's totally drain the
4093         // local queue and global stack.
4094         drain_local_queue(false);
4095         drain_global_stack(false);
4096       } else {
4097         break;
4098       }
4099     }
4100   }
4101 
4102   // If we are about to wrap up and go into termination, check if we
4103   // should raise the overflow flag.
4104   if (do_termination && !has_aborted()) {
4105     if (_cm->force_overflow()->should_force()) {
4106       _cm->set_has_overflown();
4107       regular_clock_call();
4108     }
4109   }
4110 
4111   // We still haven't aborted. Now, let's try to get into the
4112   // termination protocol.
4113   if (do_termination && !has_aborted()) {
4114     // We cannot check whether the global stack is empty, since other
4115     // tasks might be concurrently pushing objects on it.
4116     // Separated the asserts so that we know which one fires.
4117     assert(_cm->out_of_regions(), "only way to reach here");
4118     assert(_task_queue->size() == 0, "only way to reach here");
4119 
4120     if (_cm->verbose_low()) {
4121       gclog_or_tty->print_cr("[%d] starting termination protocol", _task_id);
4122     }
4123 
4124     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4125     // The CMTask class also extends the TerminatorTerminator class,
4126     // hence its should_exit_termination() method will also decide
4127     // whether to exit the termination protocol or not.
4128     bool finished = _cm->terminator()->offer_termination(this);
4129     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4130     _termination_time_ms +=
4131       termination_end_time_ms - _termination_start_time_ms;
4132 
4133     if (finished) {
4134       // We're all done.
4135 
4136       if (_task_id == 0) {
4137         // let's allow task 0 to do this
4138         if (concurrent()) {
4139           assert(_cm->concurrent_marking_in_progress(), "invariant");
4140           // we need to set this to false before the next
4141           // safepoint. This way we ensure that the marking phase
4142           // doesn't observe any more heap expansions.
4143           _cm->clear_concurrent_marking_in_progress();
4144         }
4145       }
4146 
4147       // We can now guarantee that the global stack is empty, since
4148       // all other tasks have finished. We separated the guarantees so
4149       // that, if a condition is false, we can immediately find out
4150       // which one.
4151       guarantee(_cm->out_of_regions(), "only way to reach here");
4152       guarantee(_cm->mark_stack_empty(), "only way to reach here");
4153       guarantee(_task_queue->size() == 0, "only way to reach here");
4154       guarantee(!_cm->has_overflown(), "only way to reach here");
4155       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4156 
4157       if (_cm->verbose_low()) {
4158         gclog_or_tty->print_cr("[%d] all tasks terminated", _task_id);
4159       }
4160     } else {
4161       // Apparently there's more work to do. Let's abort this task. It
4162       // will restart it and we can hopefully find more things to do.
4163 
4164       if (_cm->verbose_low()) {
4165         gclog_or_tty->print_cr("[%d] apparently there is more work to do",
4166                                _task_id);
4167       }
4168 
4169       set_has_aborted();
4170       statsOnly( ++_aborted_termination );
4171     }
4172   }
4173 
4174   // Mainly for debugging purposes to make sure that a pointer to the
4175   // closure which was statically allocated in this frame doesn't
4176   // escape it by accident.
4177   set_cm_oop_closure(NULL);
4178   double end_time_ms = os::elapsedVTime() * 1000.0;
4179   double elapsed_time_ms = end_time_ms - _start_time_ms;
4180   // Update the step history.
4181   _step_times_ms.add(elapsed_time_ms);
4182 
4183   if (has_aborted()) {
4184     // The task was aborted for some reason.
4185 
4186     statsOnly( ++_aborted );
4187 
4188     if (_has_timed_out) {
4189       double diff_ms = elapsed_time_ms - _time_target_ms;
4190       // Keep statistics of how well we did with respect to hitting
4191       // our target only if we actually timed out (if we aborted for
4192       // other reasons, then the results might get skewed).
4193       _marking_step_diffs_ms.add(diff_ms);
4194     }
4195 
4196     if (_cm->has_overflown()) {
4197       // This is the interesting one. We aborted because a global
4198       // overflow was raised. This means we have to restart the
4199       // marking phase and start iterating over regions. However, in
4200       // order to do this we have to make sure that all tasks stop
4201       // what they are doing and re-initialise in a safe manner. We
4202       // will achieve this with the use of two barrier sync points.
4203 
4204       if (_cm->verbose_low()) {
4205         gclog_or_tty->print_cr("[%d] detected overflow", _task_id);
4206       }
4207 
4208       _cm->enter_first_sync_barrier(_task_id);
4209       // When we exit this sync barrier we know that all tasks have
4210       // stopped doing marking work. So, it's now safe to
4211       // re-initialise our data structures. At the end of this method,
4212       // task 0 will clear the global data structures.
4213 
4214       statsOnly( ++_aborted_overflow );
4215 
4216       // We clear the local state of this task...
4217       clear_region_fields();
4218 
4219       // ...and enter the second barrier.
4220       _cm->enter_second_sync_barrier(_task_id);
4221       // At this point everything has bee re-initialised and we're
4222       // ready to restart.
4223     }
4224 
4225     if (_cm->verbose_low()) {
4226       gclog_or_tty->print_cr("[%d] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4227                              "elapsed = %1.2lfms <<<<<<<<<<",
4228                              _task_id, _time_target_ms, elapsed_time_ms);
4229       if (_cm->has_aborted()) {
4230         gclog_or_tty->print_cr("[%d] ========== MARKING ABORTED ==========",
4231                                _task_id);
4232       }
4233     }
4234   } else {
4235     if (_cm->verbose_low()) {
4236       gclog_or_tty->print_cr("[%d] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4237                              "elapsed = %1.2lfms <<<<<<<<<<",
4238                              _task_id, _time_target_ms, elapsed_time_ms);
4239     }
4240   }
4241 
4242   _claimed = false;
4243 }
4244 
4245 CMTask::CMTask(int task_id,
4246                ConcurrentMark* cm,
4247                size_t* marked_bytes,
4248                BitMap* card_bm,
4249                CMTaskQueue* task_queue,
4250                CMTaskQueueSet* task_queues)
4251   : _g1h(G1CollectedHeap::heap()),
4252     _task_id(task_id), _cm(cm),
4253     _claimed(false),
4254     _nextMarkBitMap(NULL), _hash_seed(17),
4255     _task_queue(task_queue),
4256     _task_queues(task_queues),
4257     _cm_oop_closure(NULL),
4258     _marked_bytes_array(marked_bytes),
4259     _card_bm(card_bm) {
4260   guarantee(task_queue != NULL, "invariant");
4261   guarantee(task_queues != NULL, "invariant");
4262 
4263   statsOnly( _clock_due_to_scanning = 0;
4264              _clock_due_to_marking  = 0 );
4265 
4266   _marking_step_diffs_ms.add(0.5);
4267 }
4268 
4269 // These are formatting macros that are used below to ensure
4270 // consistent formatting. The *_H_* versions are used to format the
4271 // header for a particular value and they should be kept consistent
4272 // with the corresponding macro. Also note that most of the macros add
4273 // the necessary white space (as a prefix) which makes them a bit
4274 // easier to compose.
4275 
4276 // All the output lines are prefixed with this string to be able to
4277 // identify them easily in a large log file.
4278 #define G1PPRL_LINE_PREFIX            "###"
4279 
4280 #define G1PPRL_ADDR_BASE_FORMAT    " "PTR_FORMAT"-"PTR_FORMAT
4281 #ifdef _LP64
4282 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
4283 #else // _LP64
4284 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
4285 #endif // _LP64
4286 
4287 // For per-region info
4288 #define G1PPRL_TYPE_FORMAT            "   %-4s"
4289 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
4290 #define G1PPRL_BYTE_FORMAT            "  "SIZE_FORMAT_W(9)
4291 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
4292 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
4293 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
4294 
4295 // For summary info
4296 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  "tag":"G1PPRL_ADDR_BASE_FORMAT
4297 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  "tag": "SIZE_FORMAT
4298 #define G1PPRL_SUM_MB_FORMAT(tag)      "  "tag": %1.2f MB"
4299 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4300 
4301 G1PrintRegionLivenessInfoClosure::
4302 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4303   : _out(out),
4304     _total_used_bytes(0), _total_capacity_bytes(0),
4305     _total_prev_live_bytes(0), _total_next_live_bytes(0),
4306     _hum_used_bytes(0), _hum_capacity_bytes(0),
4307     _hum_prev_live_bytes(0), _hum_next_live_bytes(0) {
4308   G1CollectedHeap* g1h = G1CollectedHeap::heap();
4309   MemRegion g1_committed = g1h->g1_committed();
4310   MemRegion g1_reserved = g1h->g1_reserved();
4311   double now = os::elapsedTime();
4312 
4313   // Print the header of the output.
4314   _out->cr();
4315   _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4316   _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4317                  G1PPRL_SUM_ADDR_FORMAT("committed")
4318                  G1PPRL_SUM_ADDR_FORMAT("reserved")
4319                  G1PPRL_SUM_BYTE_FORMAT("region-size"),
4320                  g1_committed.start(), g1_committed.end(),
4321                  g1_reserved.start(), g1_reserved.end(),
4322                  HeapRegion::GrainBytes);
4323   _out->print_cr(G1PPRL_LINE_PREFIX);
4324   _out->print_cr(G1PPRL_LINE_PREFIX
4325                  G1PPRL_TYPE_H_FORMAT
4326                  G1PPRL_ADDR_BASE_H_FORMAT
4327                  G1PPRL_BYTE_H_FORMAT
4328                  G1PPRL_BYTE_H_FORMAT
4329                  G1PPRL_BYTE_H_FORMAT
4330                  G1PPRL_DOUBLE_H_FORMAT,
4331                  "type", "address-range",
4332                  "used", "prev-live", "next-live", "gc-eff");
4333   _out->print_cr(G1PPRL_LINE_PREFIX
4334                  G1PPRL_TYPE_H_FORMAT
4335                  G1PPRL_ADDR_BASE_H_FORMAT
4336                  G1PPRL_BYTE_H_FORMAT
4337                  G1PPRL_BYTE_H_FORMAT
4338                  G1PPRL_BYTE_H_FORMAT
4339                  G1PPRL_DOUBLE_H_FORMAT,
4340                  "", "",
4341                  "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)");
4342 }
4343 
4344 // It takes as a parameter a reference to one of the _hum_* fields, it
4345 // deduces the corresponding value for a region in a humongous region
4346 // series (either the region size, or what's left if the _hum_* field
4347 // is < the region size), and updates the _hum_* field accordingly.
4348 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4349   size_t bytes = 0;
4350   // The > 0 check is to deal with the prev and next live bytes which
4351   // could be 0.
4352   if (*hum_bytes > 0) {
4353     bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4354     *hum_bytes -= bytes;
4355   }
4356   return bytes;
4357 }
4358 
4359 // It deduces the values for a region in a humongous region series
4360 // from the _hum_* fields and updates those accordingly. It assumes
4361 // that that _hum_* fields have already been set up from the "starts
4362 // humongous" region and we visit the regions in address order.
4363 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4364                                                      size_t* capacity_bytes,
4365                                                      size_t* prev_live_bytes,
4366                                                      size_t* next_live_bytes) {
4367   assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4368   *used_bytes      = get_hum_bytes(&_hum_used_bytes);
4369   *capacity_bytes  = get_hum_bytes(&_hum_capacity_bytes);
4370   *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4371   *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4372 }
4373 
4374 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4375   const char* type = "";
4376   HeapWord* bottom       = r->bottom();
4377   HeapWord* end          = r->end();
4378   size_t capacity_bytes  = r->capacity();
4379   size_t used_bytes      = r->used();
4380   size_t prev_live_bytes = r->live_bytes();
4381   size_t next_live_bytes = r->next_live_bytes();
4382   double gc_eff          = r->gc_efficiency();
4383   if (r->used() == 0) {
4384     type = "FREE";
4385   } else if (r->is_survivor()) {
4386     type = "SURV";
4387   } else if (r->is_young()) {
4388     type = "EDEN";
4389   } else if (r->startsHumongous()) {
4390     type = "HUMS";
4391 
4392     assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4393            _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4394            "they should have been zeroed after the last time we used them");
4395     // Set up the _hum_* fields.
4396     _hum_capacity_bytes  = capacity_bytes;
4397     _hum_used_bytes      = used_bytes;
4398     _hum_prev_live_bytes = prev_live_bytes;
4399     _hum_next_live_bytes = next_live_bytes;
4400     get_hum_bytes(&used_bytes, &capacity_bytes,
4401                   &prev_live_bytes, &next_live_bytes);
4402     end = bottom + HeapRegion::GrainWords;
4403   } else if (r->continuesHumongous()) {
4404     type = "HUMC";
4405     get_hum_bytes(&used_bytes, &capacity_bytes,
4406                   &prev_live_bytes, &next_live_bytes);
4407     assert(end == bottom + HeapRegion::GrainWords, "invariant");
4408   } else {
4409     type = "OLD";
4410   }
4411 
4412   _total_used_bytes      += used_bytes;
4413   _total_capacity_bytes  += capacity_bytes;
4414   _total_prev_live_bytes += prev_live_bytes;
4415   _total_next_live_bytes += next_live_bytes;
4416 
4417   // Print a line for this particular region.
4418   _out->print_cr(G1PPRL_LINE_PREFIX
4419                  G1PPRL_TYPE_FORMAT
4420                  G1PPRL_ADDR_BASE_FORMAT
4421                  G1PPRL_BYTE_FORMAT
4422                  G1PPRL_BYTE_FORMAT
4423                  G1PPRL_BYTE_FORMAT
4424                  G1PPRL_DOUBLE_FORMAT,
4425                  type, bottom, end,
4426                  used_bytes, prev_live_bytes, next_live_bytes, gc_eff);
4427 
4428   return false;
4429 }
4430 
4431 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4432   // Print the footer of the output.
4433   _out->print_cr(G1PPRL_LINE_PREFIX);
4434   _out->print_cr(G1PPRL_LINE_PREFIX
4435                  " SUMMARY"
4436                  G1PPRL_SUM_MB_FORMAT("capacity")
4437                  G1PPRL_SUM_MB_PERC_FORMAT("used")
4438                  G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4439                  G1PPRL_SUM_MB_PERC_FORMAT("next-live"),
4440                  bytes_to_mb(_total_capacity_bytes),
4441                  bytes_to_mb(_total_used_bytes),
4442                  perc(_total_used_bytes, _total_capacity_bytes),
4443                  bytes_to_mb(_total_prev_live_bytes),
4444                  perc(_total_prev_live_bytes, _total_capacity_bytes),
4445                  bytes_to_mb(_total_next_live_bytes),
4446                  perc(_total_next_live_bytes, _total_capacity_bytes));
4447   _out->cr();
4448 }