1 /*
   2  * Copyright (c) 2001, 2016, 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/metadataOnStackMark.hpp"
  27 #include "classfile/symbolTable.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc/g1/concurrentMarkThread.inline.hpp"
  30 #include "gc/g1/g1CollectedHeap.inline.hpp"
  31 #include "gc/g1/g1CollectorPolicy.hpp"
  32 #include "gc/g1/g1CollectorState.hpp"
  33 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  34 #include "gc/g1/g1HeapVerifier.hpp"
  35 #include "gc/g1/g1OopClosures.inline.hpp"
  36 #include "gc/g1/g1CardLiveData.inline.hpp"
  37 #include "gc/g1/g1StringDedup.hpp"
  38 #include "gc/g1/heapRegion.inline.hpp"
  39 #include "gc/g1/heapRegionRemSet.hpp"
  40 #include "gc/g1/heapRegionSet.inline.hpp"
  41 #include "gc/g1/suspendibleThreadSet.hpp"
  42 #include "gc/shared/gcId.hpp"
  43 #include "gc/shared/gcTimer.hpp"
  44 #include "gc/shared/gcTrace.hpp"
  45 #include "gc/shared/gcTraceTime.inline.hpp"
  46 #include "gc/shared/genOopClosures.inline.hpp"
  47 #include "gc/shared/referencePolicy.hpp"
  48 #include "gc/shared/strongRootsScope.hpp"
  49 #include "gc/shared/taskqueue.inline.hpp"
  50 #include "gc/shared/vmGCOperations.hpp"
  51 #include "logging/log.hpp"
  52 #include "memory/allocation.hpp"
  53 #include "memory/resourceArea.hpp"
  54 #include "oops/oop.inline.hpp"
  55 #include "runtime/atomic.inline.hpp"
  56 #include "runtime/handles.inline.hpp"
  57 #include "runtime/java.hpp"
  58 #include "runtime/prefetch.inline.hpp"
  59 #include "services/memTracker.hpp"
  60 
  61 // Concurrent marking bit map wrapper
  62 
  63 G1CMBitMapRO::G1CMBitMapRO(int shifter) :
  64   _bm(),
  65   _shifter(shifter) {
  66   _bmStartWord = 0;
  67   _bmWordSize = 0;
  68 }
  69 
  70 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
  71                                                  const HeapWord* limit) const {
  72   // First we must round addr *up* to a possible object boundary.
  73   addr = (HeapWord*)align_size_up((intptr_t)addr,
  74                                   HeapWordSize << _shifter);
  75   size_t addrOffset = heapWordToOffset(addr);
  76   assert(limit != NULL, "limit must not be NULL");
  77   size_t limitOffset = heapWordToOffset(limit);
  78   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  79   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  80   assert(nextAddr >= addr, "get_next_one postcondition");
  81   assert(nextAddr == limit || isMarked(nextAddr),
  82          "get_next_one postcondition");
  83   return nextAddr;
  84 }
  85 
  86 #ifndef PRODUCT
  87 bool G1CMBitMapRO::covers(MemRegion heap_rs) const {
  88   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  89   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
  90          "size inconsistency");
  91   return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
  92          _bmWordSize  == heap_rs.word_size();
  93 }
  94 #endif
  95 
  96 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
  97   _bm.print_on_error(st, prefix);
  98 }
  99 
 100 size_t G1CMBitMap::compute_size(size_t heap_size) {
 101   return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
 102 }
 103 
 104 size_t G1CMBitMap::mark_distance() {
 105   return MinObjAlignmentInBytes * BitsPerByte;
 106 }
 107 
 108 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
 109   _bmStartWord = heap.start();
 110   _bmWordSize = heap.word_size();
 111 
 112   _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
 113   _bm.set_size(_bmWordSize >> _shifter);
 114 
 115   storage->set_mapping_changed_listener(&_listener);
 116 }
 117 
 118 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
 119   if (zero_filled) {
 120     return;
 121   }
 122   // We need to clear the bitmap on commit, removing any existing information.
 123   MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
 124   _bm->clear_range(mr);
 125 }
 126 
 127 void G1CMBitMap::clear_range(MemRegion mr) {
 128   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 129   assert(!mr.is_empty(), "unexpected empty region");
 130   // convert address range into offset range
 131   _bm.at_put_range(heapWordToOffset(mr.start()),
 132                    heapWordToOffset(mr.end()), false);
 133 }
 134 
 135 G1CMMarkStack::G1CMMarkStack(G1ConcurrentMark* cm) :
 136   _base(NULL), _cm(cm)
 137 {}
 138 
 139 bool G1CMMarkStack::allocate(size_t capacity) {
 140   // allocate a stack of the requisite depth
 141   ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
 142   if (!rs.is_reserved()) {
 143     log_warning(gc)("ConcurrentMark MarkStack allocation failure");
 144     return false;
 145   }
 146   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 147   if (!_virtual_space.initialize(rs, rs.size())) {
 148     log_warning(gc)("ConcurrentMark MarkStack backing store failure");
 149     // Release the virtual memory reserved for the marking stack
 150     rs.release();
 151     return false;
 152   }
 153   assert(_virtual_space.committed_size() == rs.size(),
 154          "Didn't reserve backing store for all of G1ConcurrentMark stack?");
 155   _base = (oop*) _virtual_space.low();
 156   setEmpty();
 157   _capacity = (jint) capacity;
 158   _saved_index = -1;
 159   _should_expand = false;
 160   return true;
 161 }
 162 
 163 void G1CMMarkStack::expand() {
 164   // Called, during remark, if we've overflown the marking stack during marking.
 165   assert(isEmpty(), "stack should been emptied while handling overflow");
 166   assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
 167   // Clear expansion flag
 168   _should_expand = false;
 169   if (_capacity == (jint) MarkStackSizeMax) {
 170     log_trace(gc)("(benign) Can't expand marking stack capacity, at max size limit");
 171     return;
 172   }
 173   // Double capacity if possible
 174   jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
 175   // Do not give up existing stack until we have managed to
 176   // get the double capacity that we desired.
 177   ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
 178                                                            sizeof(oop)));
 179   if (rs.is_reserved()) {
 180     // Release the backing store associated with old stack
 181     _virtual_space.release();
 182     // Reinitialize virtual space for new stack
 183     if (!_virtual_space.initialize(rs, rs.size())) {
 184       fatal("Not enough swap for expanded marking stack capacity");
 185     }
 186     _base = (oop*)(_virtual_space.low());
 187     _index = 0;
 188     _capacity = new_capacity;
 189   } else {
 190     // Failed to double capacity, continue;
 191     log_trace(gc)("(benign) Failed to expand marking stack capacity from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
 192                   _capacity / K, new_capacity / K);
 193   }
 194 }
 195 
 196 void G1CMMarkStack::set_should_expand() {
 197   // If we're resetting the marking state because of an
 198   // marking stack overflow, record that we should, if
 199   // possible, expand the stack.
 200   _should_expand = _cm->has_overflown();
 201 }
 202 
 203 G1CMMarkStack::~G1CMMarkStack() {
 204   if (_base != NULL) {
 205     _base = NULL;
 206     _virtual_space.release();
 207   }
 208 }
 209 
 210 void G1CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 211   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 212   jint start = _index;
 213   jint next_index = start + n;
 214   if (next_index > _capacity) {
 215     _overflow = true;
 216     return;
 217   }
 218   // Otherwise.
 219   _index = next_index;
 220   for (int i = 0; i < n; i++) {
 221     int ind = start + i;
 222     assert(ind < _capacity, "By overflow test above.");
 223     _base[ind] = ptr_arr[i];
 224   }
 225 }
 226 
 227 bool G1CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 228   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 229   jint index = _index;
 230   if (index == 0) {
 231     *n = 0;
 232     return false;
 233   } else {
 234     int k = MIN2(max, index);
 235     jint  new_ind = index - k;
 236     for (int j = 0; j < k; j++) {
 237       ptr_arr[j] = _base[new_ind + j];
 238     }
 239     _index = new_ind;
 240     *n = k;
 241     return true;
 242   }
 243 }
 244 
 245 void G1CMMarkStack::note_start_of_gc() {
 246   assert(_saved_index == -1,
 247          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 248   _saved_index = _index;
 249 }
 250 
 251 void G1CMMarkStack::note_end_of_gc() {
 252   // This is intentionally a guarantee, instead of an assert. If we
 253   // accidentally add something to the mark stack during GC, it
 254   // will be a correctness issue so it's better if we crash. we'll
 255   // only check this once per GC anyway, so it won't be a performance
 256   // issue in any way.
 257   guarantee(_saved_index == _index,
 258             "saved index: %d index: %d", _saved_index, _index);
 259   _saved_index = -1;
 260 }
 261 
 262 G1CMRootRegions::G1CMRootRegions() :
 263   _young_list(NULL), _cm(NULL), _scan_in_progress(false),
 264   _should_abort(false),  _next_survivor(NULL) { }
 265 
 266 void G1CMRootRegions::init(G1CollectedHeap* g1h, G1ConcurrentMark* cm) {
 267   _young_list = g1h->young_list();
 268   _cm = cm;
 269 }
 270 
 271 void G1CMRootRegions::prepare_for_scan() {
 272   assert(!scan_in_progress(), "pre-condition");
 273 
 274   // Currently, only survivors can be root regions.
 275   assert(_next_survivor == NULL, "pre-condition");
 276   _next_survivor = _young_list->first_survivor_region();
 277   _scan_in_progress = (_next_survivor != NULL);
 278   _should_abort = false;
 279 }
 280 
 281 HeapRegion* G1CMRootRegions::claim_next() {
 282   if (_should_abort) {
 283     // If someone has set the should_abort flag, we return NULL to
 284     // force the caller to bail out of their loop.
 285     return NULL;
 286   }
 287 
 288   // Currently, only survivors can be root regions.
 289   HeapRegion* res = _next_survivor;
 290   if (res != NULL) {
 291     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 292     // Read it again in case it changed while we were waiting for the lock.
 293     res = _next_survivor;
 294     if (res != NULL) {
 295       if (res == _young_list->last_survivor_region()) {
 296         // We just claimed the last survivor so store NULL to indicate
 297         // that we're done.
 298         _next_survivor = NULL;
 299       } else {
 300         _next_survivor = res->get_next_young_region();
 301       }
 302     } else {
 303       // Someone else claimed the last survivor while we were trying
 304       // to take the lock so nothing else to do.
 305     }
 306   }
 307   assert(res == NULL || res->is_survivor(), "post-condition");
 308 
 309   return res;
 310 }
 311 
 312 void G1CMRootRegions::notify_scan_done() {
 313   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 314   _scan_in_progress = false;
 315   RootRegionScan_lock->notify_all();
 316 }
 317 
 318 void G1CMRootRegions::cancel_scan() {
 319   notify_scan_done();
 320 }
 321 
 322 void G1CMRootRegions::scan_finished() {
 323   assert(scan_in_progress(), "pre-condition");
 324 
 325   // Currently, only survivors can be root regions.
 326   if (!_should_abort) {
 327     assert(_next_survivor == NULL, "we should have claimed all survivors");
 328   }
 329   _next_survivor = NULL;
 330 
 331   notify_scan_done();
 332 }
 333 
 334 bool G1CMRootRegions::wait_until_scan_finished() {
 335   if (!scan_in_progress()) return false;
 336 
 337   {
 338     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 339     while (scan_in_progress()) {
 340       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 341     }
 342   }
 343   return true;
 344 }
 345 
 346 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 347   return MAX2((n_par_threads + 2) / 4, 1U);
 348 }
 349 
 350 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
 351   _g1h(g1h),
 352   _markBitMap1(),
 353   _markBitMap2(),
 354   _parallel_marking_threads(0),
 355   _max_parallel_marking_threads(0),
 356   _sleep_factor(0.0),
 357   _marking_task_overhead(1.0),
 358   _cleanup_list("Cleanup List"),
 359 
 360   _prevMarkBitMap(&_markBitMap1),
 361   _nextMarkBitMap(&_markBitMap2),
 362 
 363   _markStack(this),
 364   // _finger set in set_non_marking_state
 365 
 366   _max_worker_id(ParallelGCThreads),
 367   // _active_tasks set in set_non_marking_state
 368   // _tasks set inside the constructor
 369   _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)),
 370   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 371 
 372   _has_overflown(false),
 373   _concurrent(false),
 374   _has_aborted(false),
 375   _restart_for_overflow(false),
 376   _concurrent_marking_in_progress(false),
 377   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 378   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 379 
 380   // _verbose_level set below
 381 
 382   _init_times(),
 383   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 384   _cleanup_times(),
 385   _total_counting_time(0.0),
 386   _total_rs_scrub_time(0.0),
 387 
 388   _parallel_workers(NULL),
 389 
 390   _completed_initialization(false) {
 391 
 392   _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 393   _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
 394 
 395   // Create & start a ConcurrentMark thread.
 396   _cmThread = new ConcurrentMarkThread(this);
 397   assert(cmThread() != NULL, "CM Thread should have been created");
 398   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 399   if (_cmThread->osthread() == NULL) {
 400       vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 401   }
 402 
 403   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 404   assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
 405   assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
 406 
 407   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 408   satb_qs.set_buffer_size(G1SATBBufferSize);
 409 
 410   _root_regions.init(_g1h, this);
 411 
 412   if (ConcGCThreads > ParallelGCThreads) {
 413     log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).",
 414                     ConcGCThreads, ParallelGCThreads);
 415     return;
 416   }
 417   if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
 418     // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
 419     // if both are set
 420     _sleep_factor             = 0.0;
 421     _marking_task_overhead    = 1.0;
 422   } else if (G1MarkingOverheadPercent > 0) {
 423     // We will calculate the number of parallel marking threads based
 424     // on a target overhead with respect to the soft real-time goal
 425     double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 426     double overall_cm_overhead =
 427       (double) MaxGCPauseMillis * marking_overhead /
 428       (double) GCPauseIntervalMillis;
 429     double cpu_ratio = 1.0 / (double) os::processor_count();
 430     double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 431     double marking_task_overhead =
 432       overall_cm_overhead / marking_thread_num *
 433                                               (double) os::processor_count();
 434     double sleep_factor =
 435                        (1.0 - marking_task_overhead) / marking_task_overhead;
 436 
 437     FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num);
 438     _sleep_factor             = sleep_factor;
 439     _marking_task_overhead    = marking_task_overhead;
 440   } else {
 441     // Calculate the number of parallel marking threads by scaling
 442     // the number of parallel GC threads.
 443     uint marking_thread_num = scale_parallel_threads(ParallelGCThreads);
 444     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 445     _sleep_factor             = 0.0;
 446     _marking_task_overhead    = 1.0;
 447   }
 448 
 449   assert(ConcGCThreads > 0, "Should have been set");
 450   _parallel_marking_threads = ConcGCThreads;
 451   _max_parallel_marking_threads = _parallel_marking_threads;
 452 
 453   _parallel_workers = new WorkGang("G1 Marker",
 454        _max_parallel_marking_threads, false, true);
 455   if (_parallel_workers == NULL) {
 456     vm_exit_during_initialization("Failed necessary allocation.");
 457   } else {
 458     _parallel_workers->initialize_workers();
 459   }
 460 
 461   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 462     size_t mark_stack_size =
 463       MIN2(MarkStackSizeMax,
 464           MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 465     // Verify that the calculated value for MarkStackSize is in range.
 466     // It would be nice to use the private utility routine from Arguments.
 467     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 468       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 469                       "must be between 1 and " SIZE_FORMAT,
 470                       mark_stack_size, MarkStackSizeMax);
 471       return;
 472     }
 473     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 474   } else {
 475     // Verify MarkStackSize is in range.
 476     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 477       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 478         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 479           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 480                           "must be between 1 and " SIZE_FORMAT,
 481                           MarkStackSize, MarkStackSizeMax);
 482           return;
 483         }
 484       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 485         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 486           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 487                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 488                           MarkStackSize, MarkStackSizeMax);
 489           return;
 490         }
 491       }
 492     }
 493   }
 494 
 495   if (!_markStack.allocate(MarkStackSize)) {
 496     log_warning(gc)("Failed to allocate CM marking stack");
 497     return;
 498   }
 499 
 500   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC);
 501   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 502 
 503   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 504   _active_tasks = _max_worker_id;
 505 
 506   for (uint i = 0; i < _max_worker_id; ++i) {
 507     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 508     task_queue->initialize();
 509     _task_queues->register_queue(i, task_queue);
 510 
 511     _tasks[i] = new G1CMTask(i, this, task_queue, _task_queues);
 512 
 513     _accum_task_vtime[i] = 0.0;
 514   }
 515 
 516   // so that the call below can read a sensible value
 517   _heap_start = g1h->reserved_region().start();
 518   set_non_marking_state();
 519   _completed_initialization = true;
 520 }
 521 
 522 void G1ConcurrentMark::reset() {
 523   // Starting values for these two. This should be called in a STW
 524   // phase.
 525   MemRegion reserved = _g1h->g1_reserved();
 526   _heap_start = reserved.start();
 527   _heap_end   = reserved.end();
 528 
 529   // Separated the asserts so that we know which one fires.
 530   assert(_heap_start != NULL, "heap bounds should look ok");
 531   assert(_heap_end != NULL, "heap bounds should look ok");
 532   assert(_heap_start < _heap_end, "heap bounds should look ok");
 533 
 534   // Reset all the marking data structures and any necessary flags
 535   reset_marking_state();
 536 
 537   // We do reset all of them, since different phases will use
 538   // different number of active threads. So, it's easiest to have all
 539   // of them ready.
 540   for (uint i = 0; i < _max_worker_id; ++i) {
 541     _tasks[i]->reset(_nextMarkBitMap);
 542   }
 543 
 544   // we need this to make sure that the flag is on during the evac
 545   // pause with initial mark piggy-backed
 546   set_concurrent_marking_in_progress();
 547 }
 548 
 549 
 550 void G1ConcurrentMark::reset_marking_state(bool clear_overflow) {
 551   _markStack.set_should_expand();
 552   _markStack.setEmpty();        // Also clears the _markStack overflow flag
 553   if (clear_overflow) {
 554     clear_has_overflown();
 555   } else {
 556     assert(has_overflown(), "pre-condition");
 557   }
 558   _finger = _heap_start;
 559 
 560   for (uint i = 0; i < _max_worker_id; ++i) {
 561     G1CMTaskQueue* queue = _task_queues->queue(i);
 562     queue->set_empty();
 563   }
 564 }
 565 
 566 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 567   assert(active_tasks <= _max_worker_id, "we should not have more");
 568 
 569   _active_tasks = active_tasks;
 570   // Need to update the three data structures below according to the
 571   // number of active threads for this phase.
 572   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 573   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 574   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 575 }
 576 
 577 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 578   set_concurrency(active_tasks);
 579 
 580   _concurrent = concurrent;
 581   // We propagate this to all tasks, not just the active ones.
 582   for (uint i = 0; i < _max_worker_id; ++i)
 583     _tasks[i]->set_concurrent(concurrent);
 584 
 585   if (concurrent) {
 586     set_concurrent_marking_in_progress();
 587   } else {
 588     // We currently assume that the concurrent flag has been set to
 589     // false before we start remark. At this point we should also be
 590     // in a STW phase.
 591     assert(!concurrent_marking_in_progress(), "invariant");
 592     assert(out_of_regions(),
 593            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 594            p2i(_finger), p2i(_heap_end));
 595   }
 596 }
 597 
 598 void G1ConcurrentMark::set_non_marking_state() {
 599   // We set the global marking state to some default values when we're
 600   // not doing marking.
 601   reset_marking_state();
 602   _active_tasks = 0;
 603   clear_concurrent_marking_in_progress();
 604 }
 605 
 606 G1ConcurrentMark::~G1ConcurrentMark() {
 607   // The G1ConcurrentMark instance is never freed.
 608   ShouldNotReachHere();
 609 }
 610 
 611 class G1ClearBitMapTask : public AbstractGangTask {
 612   // Heap region closure used for clearing the given mark bitmap.
 613   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 614   private:
 615     G1CMBitMap* _bitmap;
 616     G1ConcurrentMark* _cm;
 617   public:
 618     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
 619     }
 620 
 621     virtual bool doHeapRegion(HeapRegion* r) {
 622       size_t const chunk_size_in_words = M / HeapWordSize;
 623 
 624       HeapWord* cur = r->bottom();
 625       HeapWord* const end = r->end();
 626 
 627       while (cur < end) {
 628         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 629         _bitmap->clear_range(mr);
 630 
 631         cur += chunk_size_in_words;
 632 
 633         // Abort iteration if after yielding the marking has been aborted.
 634         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 635           return true;
 636         }
 637         // Repeat the asserts from before the start of the closure. We will do them
 638         // as asserts here to minimize their overhead on the product. However, we
 639         // will have them as guarantees at the beginning / end of the bitmap
 640         // clearing to get some checking in the product.
 641         assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant");
 642         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
 643       }
 644       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 645 
 646       return false;
 647     }
 648   };
 649 
 650   G1ClearBitmapHRClosure _cl;
 651   HeapRegionClaimer _hr_claimer;
 652   bool _suspendible; // If the task is suspendible, workers must join the STS.
 653 
 654 public:
 655   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 656     AbstractGangTask("Parallel Clear Bitmap Task"),
 657     _cl(bitmap, suspendible ? cm : NULL),
 658     _hr_claimer(n_workers),
 659     _suspendible(suspendible)
 660   { }
 661 
 662   void work(uint worker_id) {
 663     SuspendibleThreadSetJoiner sts_join(_suspendible);
 664     G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer, true);
 665   }
 666 
 667   bool is_complete() {
 668     return _cl.complete();
 669   }
 670 };
 671 
 672 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 673   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 674 
 675   G1ClearBitMapTask task(bitmap, this, workers->active_workers(), may_yield);
 676   workers->run_task(&task);
 677   guarantee(!may_yield || task.is_complete(), "Must have completed iteration when not yielding.");
 678 }
 679 
 680 void G1ConcurrentMark::cleanup_for_next_mark() {
 681   // Make sure that the concurrent mark thread looks to still be in
 682   // the current cycle.
 683   guarantee(cmThread()->during_cycle(), "invariant");
 684 
 685   // We are finishing up the current cycle by clearing the next
 686   // marking bitmap and getting it ready for the next cycle. During
 687   // this time no other cycle can start. So, let's make sure that this
 688   // is the case.
 689   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 690 
 691   clear_bitmap(_nextMarkBitMap, _parallel_workers, true);
 692 
 693   // Clear the live count data. If the marking has been aborted, the abort()
 694   // call already did that.
 695   if (!has_aborted()) {
 696     clear_live_data(_parallel_workers);
 697     DEBUG_ONLY(verify_live_data_clear());
 698   }
 699 
 700   // Repeat the asserts from above.
 701   guarantee(cmThread()->during_cycle(), "invariant");
 702   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 703 }
 704 
 705 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 706   assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint.");
 707   clear_bitmap((G1CMBitMap*)_prevMarkBitMap, workers, false);
 708 }
 709 
 710 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 711   G1CMBitMap* _bitmap;
 712   bool _error;
 713  public:
 714   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 715   }
 716 
 717   virtual bool doHeapRegion(HeapRegion* r) {
 718     // This closure can be called concurrently to the mutator, so we must make sure
 719     // that the result of the getNextMarkedWordAddress() call is compared to the
 720     // value passed to it as limit to detect any found bits.
 721     // end never changes in G1.
 722     HeapWord* end = r->end();
 723     return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
 724   }
 725 };
 726 
 727 bool G1ConcurrentMark::nextMarkBitmapIsClear() {
 728   CheckBitmapClearHRClosure cl(_nextMarkBitMap);
 729   _g1h->heap_region_iterate(&cl);
 730   return cl.complete();
 731 }
 732 
 733 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 734 public:
 735   bool doHeapRegion(HeapRegion* r) {
 736     r->note_start_of_marking();
 737     return false;
 738   }
 739 };
 740 
 741 void G1ConcurrentMark::checkpointRootsInitialPre() {
 742   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 743   G1CollectorPolicy* g1p = g1h->g1_policy();
 744 
 745   _has_aborted = false;
 746 
 747   // Initialize marking structures. This has to be done in a STW phase.
 748   reset();
 749 
 750   // For each region note start of marking.
 751   NoteStartOfMarkHRClosure startcl;
 752   g1h->heap_region_iterate(&startcl);
 753 }
 754 
 755 
 756 void G1ConcurrentMark::checkpointRootsInitialPost() {
 757   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 758 
 759   // Start Concurrent Marking weak-reference discovery.
 760   ReferenceProcessor* rp = g1h->ref_processor_cm();
 761   // enable ("weak") refs discovery
 762   rp->enable_discovery();
 763   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 764 
 765   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 766   // This is the start of  the marking cycle, we're expected all
 767   // threads to have SATB queues with active set to false.
 768   satb_mq_set.set_active_all_threads(true, /* new active value */
 769                                      false /* expected_active */);
 770 
 771   _root_regions.prepare_for_scan();
 772 
 773   // update_g1_committed() will be called at the end of an evac pause
 774   // when marking is on. So, it's also called at the end of the
 775   // initial-mark pause to update the heap end, if the heap expands
 776   // during it. No need to call it here.
 777 }
 778 
 779 /*
 780  * Notice that in the next two methods, we actually leave the STS
 781  * during the barrier sync and join it immediately afterwards. If we
 782  * do not do this, the following deadlock can occur: one thread could
 783  * be in the barrier sync code, waiting for the other thread to also
 784  * sync up, whereas another one could be trying to yield, while also
 785  * waiting for the other threads to sync up too.
 786  *
 787  * Note, however, that this code is also used during remark and in
 788  * this case we should not attempt to leave / enter the STS, otherwise
 789  * we'll either hit an assert (debug / fastdebug) or deadlock
 790  * (product). So we should only leave / enter the STS if we are
 791  * operating concurrently.
 792  *
 793  * Because the thread that does the sync barrier has left the STS, it
 794  * is possible to be suspended for a Full GC or an evacuation pause
 795  * could occur. This is actually safe, since the entering the sync
 796  * barrier is one of the last things do_marking_step() does, and it
 797  * doesn't manipulate any data structures afterwards.
 798  */
 799 
 800 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 801   bool barrier_aborted;
 802   {
 803     SuspendibleThreadSetLeaver sts_leave(concurrent());
 804     barrier_aborted = !_first_overflow_barrier_sync.enter();
 805   }
 806 
 807   // at this point everyone should have synced up and not be doing any
 808   // more work
 809 
 810   if (barrier_aborted) {
 811     // If the barrier aborted we ignore the overflow condition and
 812     // just abort the whole marking phase as quickly as possible.
 813     return;
 814   }
 815 
 816   // If we're executing the concurrent phase of marking, reset the marking
 817   // state; otherwise the marking state is reset after reference processing,
 818   // during the remark pause.
 819   // If we reset here as a result of an overflow during the remark we will
 820   // see assertion failures from any subsequent set_concurrency_and_phase()
 821   // calls.
 822   if (concurrent()) {
 823     // let the task associated with with worker 0 do this
 824     if (worker_id == 0) {
 825       // task 0 is responsible for clearing the global data structures
 826       // We should be here because of an overflow. During STW we should
 827       // not clear the overflow flag since we rely on it being true when
 828       // we exit this method to abort the pause and restart concurrent
 829       // marking.
 830       reset_marking_state(true /* clear_overflow */);
 831 
 832       log_info(gc, marking)("Concurrent Mark reset for overflow");
 833     }
 834   }
 835 
 836   // after this, each task should reset its own data structures then
 837   // then go into the second barrier
 838 }
 839 
 840 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 841   SuspendibleThreadSetLeaver sts_leave(concurrent());
 842   _second_overflow_barrier_sync.enter();
 843 
 844   // at this point everything should be re-initialized and ready to go
 845 }
 846 
 847 class G1CMConcurrentMarkingTask: public AbstractGangTask {
 848 private:
 849   G1ConcurrentMark*     _cm;
 850   ConcurrentMarkThread* _cmt;
 851 
 852 public:
 853   void work(uint worker_id) {
 854     assert(Thread::current()->is_ConcurrentGC_thread(),
 855            "this should only be done by a conc GC thread");
 856     ResourceMark rm;
 857 
 858     double start_vtime = os::elapsedVTime();
 859 
 860     {
 861       SuspendibleThreadSetJoiner sts_join;
 862 
 863       assert(worker_id < _cm->active_tasks(), "invariant");
 864       G1CMTask* the_task = _cm->task(worker_id);
 865       the_task->record_start_time();
 866       if (!_cm->has_aborted()) {
 867         do {
 868           double start_vtime_sec = os::elapsedVTime();
 869           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
 870 
 871           the_task->do_marking_step(mark_step_duration_ms,
 872                                     true  /* do_termination */,
 873                                     false /* is_serial*/);
 874 
 875           double end_vtime_sec = os::elapsedVTime();
 876           double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
 877           _cm->clear_has_overflown();
 878 
 879           _cm->do_yield_check();
 880 
 881           jlong sleep_time_ms;
 882           if (!_cm->has_aborted() && the_task->has_aborted()) {
 883             sleep_time_ms =
 884               (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
 885             {
 886               SuspendibleThreadSetLeaver sts_leave;
 887               os::sleep(Thread::current(), sleep_time_ms, false);
 888             }
 889           }
 890         } while (!_cm->has_aborted() && the_task->has_aborted());
 891       }
 892       the_task->record_end_time();
 893       guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
 894     }
 895 
 896     double end_vtime = os::elapsedVTime();
 897     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 898   }
 899 
 900   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm,
 901                             ConcurrentMarkThread* cmt) :
 902       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
 903 
 904   ~G1CMConcurrentMarkingTask() { }
 905 };
 906 
 907 // Calculates the number of active workers for a concurrent
 908 // phase.
 909 uint G1ConcurrentMark::calc_parallel_marking_threads() {
 910   uint n_conc_workers = 0;
 911   if (!UseDynamicNumberOfGCThreads ||
 912       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 913        !ForceDynamicNumberOfGCThreads)) {
 914     n_conc_workers = max_parallel_marking_threads();
 915   } else {
 916     n_conc_workers =
 917       AdaptiveSizePolicy::calc_default_active_workers(
 918                                    max_parallel_marking_threads(),
 919                                    1, /* Minimum workers */
 920                                    parallel_marking_threads(),
 921                                    Threads::number_of_non_daemon_threads());
 922     // Don't scale down "n_conc_workers" by scale_parallel_threads() because
 923     // that scaling has already gone into "_max_parallel_marking_threads".
 924   }
 925   assert(n_conc_workers > 0, "Always need at least 1");
 926   return n_conc_workers;
 927 }
 928 
 929 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr) {
 930   // Currently, only survivors can be root regions.
 931   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 932   G1RootRegionScanClosure cl(_g1h, this);
 933 
 934   const uintx interval = PrefetchScanIntervalInBytes;
 935   HeapWord* curr = hr->bottom();
 936   const HeapWord* end = hr->top();
 937   while (curr < end) {
 938     Prefetch::read(curr, interval);
 939     oop obj = oop(curr);
 940     int size = obj->oop_iterate_size(&cl);
 941     assert(size == obj->size(), "sanity");
 942     curr += size;
 943   }
 944 }
 945 
 946 class G1CMRootRegionScanTask : public AbstractGangTask {
 947 private:
 948   G1ConcurrentMark* _cm;
 949 
 950 public:
 951   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 952     AbstractGangTask("Root Region Scan"), _cm(cm) { }
 953 
 954   void work(uint worker_id) {
 955     assert(Thread::current()->is_ConcurrentGC_thread(),
 956            "this should only be done by a conc GC thread");
 957 
 958     G1CMRootRegions* root_regions = _cm->root_regions();
 959     HeapRegion* hr = root_regions->claim_next();
 960     while (hr != NULL) {
 961       _cm->scanRootRegion(hr);
 962       hr = root_regions->claim_next();
 963     }
 964   }
 965 };
 966 
 967 void G1ConcurrentMark::scan_root_regions() {
 968   // scan_in_progress() will have been set to true only if there was
 969   // at least one root region to scan. So, if it's false, we
 970   // should not attempt to do any further work.
 971   if (root_regions()->scan_in_progress()) {
 972     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 973 
 974     _parallel_marking_threads = calc_parallel_marking_threads();
 975     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
 976            "Maximum number of marking threads exceeded");
 977     uint active_workers = MAX2(1U, parallel_marking_threads());
 978 
 979     G1CMRootRegionScanTask task(this);
 980     _parallel_workers->set_active_workers(active_workers);
 981     _parallel_workers->run_task(&task);
 982 
 983     // It's possible that has_aborted() is true here without actually
 984     // aborting the survivor scan earlier. This is OK as it's
 985     // mainly used for sanity checking.
 986     root_regions()->scan_finished();
 987   }
 988 }
 989 
 990 void G1ConcurrentMark::concurrent_cycle_start() {
 991   _gc_timer_cm->register_gc_start();
 992 
 993   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
 994 
 995   _g1h->trace_heap_before_gc(_gc_tracer_cm);
 996 }
 997 
 998 void G1ConcurrentMark::concurrent_cycle_end() {
 999   _g1h->trace_heap_after_gc(_gc_tracer_cm);
1000 
1001   if (has_aborted()) {
1002     _gc_tracer_cm->report_concurrent_mode_failure();
1003   }
1004 
1005   _gc_timer_cm->register_gc_end();
1006 
1007   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1008 }
1009 
1010 void G1ConcurrentMark::mark_from_roots() {
1011   // we might be tempted to assert that:
1012   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1013   //        "inconsistent argument?");
1014   // However that wouldn't be right, because it's possible that
1015   // a safepoint is indeed in progress as a younger generation
1016   // stop-the-world GC happens even as we mark in this generation.
1017 
1018   _restart_for_overflow = false;
1019 
1020   // _g1h has _n_par_threads
1021   _parallel_marking_threads = calc_parallel_marking_threads();
1022   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1023     "Maximum number of marking threads exceeded");
1024 
1025   uint active_workers = MAX2(1U, parallel_marking_threads());
1026   assert(active_workers > 0, "Should have been set");
1027 
1028   // Parallel task terminator is set in "set_concurrency_and_phase()"
1029   set_concurrency_and_phase(active_workers, true /* concurrent */);
1030 
1031   G1CMConcurrentMarkingTask markingTask(this, cmThread());
1032   _parallel_workers->set_active_workers(active_workers);
1033   _parallel_workers->run_task(&markingTask);
1034   print_stats();
1035 }
1036 
1037 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1038   // world is stopped at this checkpoint
1039   assert(SafepointSynchronize::is_at_safepoint(),
1040          "world should be stopped");
1041 
1042   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1043 
1044   // If a full collection has happened, we shouldn't do this.
1045   if (has_aborted()) {
1046     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1047     return;
1048   }
1049 
1050   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1051 
1052   if (VerifyDuringGC) {
1053     HandleMark hm;  // handle scope
1054     g1h->prepare_for_verify();
1055     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1056   }
1057   g1h->verifier()->check_bitmaps("Remark Start");
1058 
1059   G1CollectorPolicy* g1p = g1h->g1_policy();
1060   g1p->record_concurrent_mark_remark_start();
1061 
1062   double start = os::elapsedTime();
1063 
1064   checkpointRootsFinalWork();
1065 
1066   double mark_work_end = os::elapsedTime();
1067 
1068   weakRefsWork(clear_all_soft_refs);
1069 
1070   if (has_overflown()) {
1071     // Oops.  We overflowed.  Restart concurrent marking.
1072     _restart_for_overflow = true;
1073 
1074     // Verify the heap w.r.t. the previous marking bitmap.
1075     if (VerifyDuringGC) {
1076       HandleMark hm;  // handle scope
1077       g1h->prepare_for_verify();
1078       Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)");
1079     }
1080 
1081     // Clear the marking state because we will be restarting
1082     // marking due to overflowing the global mark stack.
1083     reset_marking_state();
1084   } else {
1085     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1086     // We're done with marking.
1087     // This is the end of  the marking cycle, we're expected all
1088     // threads to have SATB queues with active set to true.
1089     satb_mq_set.set_active_all_threads(false, /* new active value */
1090                                        true /* expected_active */);
1091 
1092     if (VerifyDuringGC) {
1093       HandleMark hm;  // handle scope
1094       g1h->prepare_for_verify();
1095       Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)");
1096     }
1097     g1h->verifier()->check_bitmaps("Remark End");
1098     assert(!restart_for_overflow(), "sanity");
1099     // Completely reset the marking state since marking completed
1100     set_non_marking_state();
1101   }
1102 
1103   // Expand the marking stack, if we have to and if we can.
1104   if (_markStack.should_expand()) {
1105     _markStack.expand();
1106   }
1107 
1108   // Statistics
1109   double now = os::elapsedTime();
1110   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1111   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1112   _remark_times.add((now - start) * 1000.0);
1113 
1114   g1p->record_concurrent_mark_remark_end();
1115 
1116   G1CMIsAliveClosure is_alive(g1h);
1117   _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1118 }
1119 
1120 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1121   G1CollectedHeap* _g1;
1122   size_t _freed_bytes;
1123   FreeRegionList* _local_cleanup_list;
1124   uint _old_regions_removed;
1125   uint _humongous_regions_removed;
1126   HRRSCleanupTask* _hrrs_cleanup_task;
1127 
1128 public:
1129   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1130                              FreeRegionList* local_cleanup_list,
1131                              HRRSCleanupTask* hrrs_cleanup_task) :
1132     _g1(g1),
1133     _freed_bytes(0),
1134     _local_cleanup_list(local_cleanup_list),
1135     _old_regions_removed(0),
1136     _humongous_regions_removed(0),
1137     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1138 
1139   size_t freed_bytes() { return _freed_bytes; }
1140   const uint old_regions_removed() { return _old_regions_removed; }
1141   const uint humongous_regions_removed() { return _humongous_regions_removed; }
1142 
1143   bool doHeapRegion(HeapRegion *hr) {
1144     if (hr->is_archive()) {
1145       return false;
1146     }
1147     _g1->reset_gc_time_stamps(hr);
1148     hr->note_end_of_marking();
1149 
1150     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1151       _freed_bytes += hr->used();
1152       hr->set_containing_set(NULL);
1153       if (hr->is_humongous()) {
1154         _humongous_regions_removed++;
1155         _g1->free_humongous_region(hr, _local_cleanup_list, true);
1156       } else {
1157         _old_regions_removed++;
1158         _g1->free_region(hr, _local_cleanup_list, true);
1159       }
1160     } else {
1161       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1162     }
1163 
1164     return false;
1165   }
1166 };
1167 
1168 class G1ParNoteEndTask: public AbstractGangTask {
1169   friend class G1NoteEndOfConcMarkClosure;
1170 
1171 protected:
1172   G1CollectedHeap* _g1h;
1173   FreeRegionList* _cleanup_list;
1174   HeapRegionClaimer _hrclaimer;
1175 
1176 public:
1177   G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1178       AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1179   }
1180 
1181   void work(uint worker_id) {
1182     FreeRegionList local_cleanup_list("Local Cleanup List");
1183     HRRSCleanupTask hrrs_cleanup_task;
1184     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1185                                            &hrrs_cleanup_task);
1186     _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1187     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1188 
1189     // Now update the lists
1190     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1191     {
1192       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1193       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1194 
1195       // If we iterate over the global cleanup list at the end of
1196       // cleanup to do this printing we will not guarantee to only
1197       // generate output for the newly-reclaimed regions (the list
1198       // might not be empty at the beginning of cleanup; we might
1199       // still be working on its previous contents). So we do the
1200       // printing here, before we append the new regions to the global
1201       // cleanup list.
1202 
1203       G1HRPrinter* hr_printer = _g1h->hr_printer();
1204       if (hr_printer->is_active()) {
1205         FreeRegionListIterator iter(&local_cleanup_list);
1206         while (iter.more_available()) {
1207           HeapRegion* hr = iter.get_next();
1208           hr_printer->cleanup(hr);
1209         }
1210       }
1211 
1212       _cleanup_list->add_ordered(&local_cleanup_list);
1213       assert(local_cleanup_list.is_empty(), "post-condition");
1214 
1215       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1216     }
1217   }
1218 };
1219 
1220 void G1ConcurrentMark::cleanup() {
1221   // world is stopped at this checkpoint
1222   assert(SafepointSynchronize::is_at_safepoint(),
1223          "world should be stopped");
1224   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1225 
1226   // If a full collection has happened, we shouldn't do this.
1227   if (has_aborted()) {
1228     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1229     return;
1230   }
1231 
1232   g1h->verifier()->verify_region_sets_optional();
1233 
1234   if (VerifyDuringGC) {
1235     HandleMark hm;  // handle scope
1236     g1h->prepare_for_verify();
1237     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1238   }
1239   g1h->verifier()->check_bitmaps("Cleanup Start");
1240 
1241   G1CollectorPolicy* g1p = g1h->g1_policy();
1242   g1p->record_concurrent_mark_cleanup_start();
1243 
1244   double start = os::elapsedTime();
1245 
1246   HeapRegionRemSet::reset_for_cleanup_tasks();
1247 
1248   {
1249     GCTraceTime(Debug, gc)("Finalize Live Data");
1250     finalize_live_data();
1251   }
1252 
1253   if (VerifyDuringGC) {
1254     GCTraceTime(Debug, gc)("Verify Live Data");
1255     verify_live_data();
1256   }
1257 
1258   g1h->collector_state()->set_mark_in_progress(false);
1259 
1260   double count_end = os::elapsedTime();
1261   double this_final_counting_time = (count_end - start);
1262   _total_counting_time += this_final_counting_time;
1263 
1264   if (log_is_enabled(Trace, gc, liveness)) {
1265     G1PrintRegionLivenessInfoClosure cl("Post-Marking");
1266     _g1h->heap_region_iterate(&cl);
1267   }
1268 
1269   // Install newly created mark bitMap as "prev".
1270   swapMarkBitMaps();
1271 
1272   g1h->reset_gc_time_stamp();
1273 
1274   uint n_workers = _g1h->workers()->active_workers();
1275 
1276   // Note end of marking in all heap regions.
1277   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1278   g1h->workers()->run_task(&g1_par_note_end_task);
1279   g1h->check_gc_time_stamps();
1280 
1281   if (!cleanup_list_is_empty()) {
1282     // The cleanup list is not empty, so we'll have to process it
1283     // concurrently. Notify anyone else that might be wanting free
1284     // regions that there will be more free regions coming soon.
1285     g1h->set_free_regions_coming();
1286   }
1287 
1288   // call below, since it affects the metric by which we sort the heap
1289   // regions.
1290   if (G1ScrubRemSets) {
1291     double rs_scrub_start = os::elapsedTime();
1292     g1h->scrub_rem_set();
1293     _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start);
1294   }
1295 
1296   // this will also free any regions totally full of garbage objects,
1297   // and sort the regions.
1298   g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1299 
1300   // Statistics.
1301   double end = os::elapsedTime();
1302   _cleanup_times.add((end - start) * 1000.0);
1303 
1304   // Clean up will have freed any regions completely full of garbage.
1305   // Update the soft reference policy with the new heap occupancy.
1306   Universe::update_heap_info_at_gc();
1307 
1308   if (VerifyDuringGC) {
1309     HandleMark hm;  // handle scope
1310     g1h->prepare_for_verify();
1311     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)");
1312   }
1313 
1314   g1h->verifier()->check_bitmaps("Cleanup End");
1315 
1316   g1h->verifier()->verify_region_sets_optional();
1317 
1318   // We need to make this be a "collection" so any collection pause that
1319   // races with it goes around and waits for completeCleanup to finish.
1320   g1h->increment_total_collections();
1321 
1322   // Clean out dead classes and update Metaspace sizes.
1323   if (ClassUnloadingWithConcurrentMark) {
1324     ClassLoaderDataGraph::purge();
1325   }
1326   MetaspaceGC::compute_new_size();
1327 
1328   // We reclaimed old regions so we should calculate the sizes to make
1329   // sure we update the old gen/space data.
1330   g1h->g1mm()->update_sizes();
1331   g1h->allocation_context_stats().update_after_mark();
1332 }
1333 
1334 void G1ConcurrentMark::complete_cleanup() {
1335   if (has_aborted()) return;
1336 
1337   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1338 
1339   _cleanup_list.verify_optional();
1340   FreeRegionList tmp_free_list("Tmp Free List");
1341 
1342   log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1343                                   "cleanup list has %u entries",
1344                                   _cleanup_list.length());
1345 
1346   // No one else should be accessing the _cleanup_list at this point,
1347   // so it is not necessary to take any locks
1348   while (!_cleanup_list.is_empty()) {
1349     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
1350     assert(hr != NULL, "Got NULL from a non-empty list");
1351     hr->par_clear();
1352     tmp_free_list.add_ordered(hr);
1353 
1354     // Instead of adding one region at a time to the secondary_free_list,
1355     // we accumulate them in the local list and move them a few at a
1356     // time. This also cuts down on the number of notify_all() calls
1357     // we do during this process. We'll also append the local list when
1358     // _cleanup_list is empty (which means we just removed the last
1359     // region from the _cleanup_list).
1360     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
1361         _cleanup_list.is_empty()) {
1362       log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1363                                       "appending %u entries to the secondary_free_list, "
1364                                       "cleanup list still has %u entries",
1365                                       tmp_free_list.length(),
1366                                       _cleanup_list.length());
1367 
1368       {
1369         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
1370         g1h->secondary_free_list_add(&tmp_free_list);
1371         SecondaryFreeList_lock->notify_all();
1372       }
1373 #ifndef PRODUCT
1374       if (G1StressConcRegionFreeing) {
1375         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
1376           os::sleep(Thread::current(), (jlong) 1, false);
1377         }
1378       }
1379 #endif
1380     }
1381   }
1382   assert(tmp_free_list.is_empty(), "post-condition");
1383 }
1384 
1385 // Supporting Object and Oop closures for reference discovery
1386 // and processing in during marking
1387 
1388 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1389   HeapWord* addr = (HeapWord*)obj;
1390   return addr != NULL &&
1391          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
1392 }
1393 
1394 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1395 // Uses the G1CMTask associated with a worker thread (for serial reference
1396 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1397 // trace referent objects.
1398 //
1399 // Using the G1CMTask and embedded local queues avoids having the worker
1400 // threads operating on the global mark stack. This reduces the risk
1401 // of overflowing the stack - which we would rather avoid at this late
1402 // state. Also using the tasks' local queues removes the potential
1403 // of the workers interfering with each other that could occur if
1404 // operating on the global stack.
1405 
1406 class G1CMKeepAliveAndDrainClosure: public OopClosure {
1407   G1ConcurrentMark* _cm;
1408   G1CMTask*         _task;
1409   int               _ref_counter_limit;
1410   int               _ref_counter;
1411   bool              _is_serial;
1412  public:
1413   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1414     _cm(cm), _task(task), _is_serial(is_serial),
1415     _ref_counter_limit(G1RefProcDrainInterval) {
1416     assert(_ref_counter_limit > 0, "sanity");
1417     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1418     _ref_counter = _ref_counter_limit;
1419   }
1420 
1421   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1422   virtual void do_oop(      oop* p) { do_oop_work(p); }
1423 
1424   template <class T> void do_oop_work(T* p) {
1425     if (!_cm->has_overflown()) {
1426       oop obj = oopDesc::load_decode_heap_oop(p);
1427       _task->deal_with_reference(obj);
1428       _ref_counter--;
1429 
1430       if (_ref_counter == 0) {
1431         // We have dealt with _ref_counter_limit references, pushing them
1432         // and objects reachable from them on to the local stack (and
1433         // possibly the global stack). Call G1CMTask::do_marking_step() to
1434         // process these entries.
1435         //
1436         // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1437         // there's nothing more to do (i.e. we're done with the entries that
1438         // were pushed as a result of the G1CMTask::deal_with_reference() calls
1439         // above) or we overflow.
1440         //
1441         // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1442         // flag while there may still be some work to do. (See the comment at
1443         // the beginning of G1CMTask::do_marking_step() for those conditions -
1444         // one of which is reaching the specified time target.) It is only
1445         // when G1CMTask::do_marking_step() returns without setting the
1446         // has_aborted() flag that the marking step has completed.
1447         do {
1448           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1449           _task->do_marking_step(mark_step_duration_ms,
1450                                  false      /* do_termination */,
1451                                  _is_serial);
1452         } while (_task->has_aborted() && !_cm->has_overflown());
1453         _ref_counter = _ref_counter_limit;
1454       }
1455     }
1456   }
1457 };
1458 
1459 // 'Drain' oop closure used by both serial and parallel reference processing.
1460 // Uses the G1CMTask associated with a given worker thread (for serial
1461 // reference processing the G1CMtask for worker 0 is used). Calls the
1462 // do_marking_step routine, with an unbelievably large timeout value,
1463 // to drain the marking data structures of the remaining entries
1464 // added by the 'keep alive' oop closure above.
1465 
1466 class G1CMDrainMarkingStackClosure: public VoidClosure {
1467   G1ConcurrentMark* _cm;
1468   G1CMTask*         _task;
1469   bool              _is_serial;
1470  public:
1471   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1472     _cm(cm), _task(task), _is_serial(is_serial) {
1473     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1474   }
1475 
1476   void do_void() {
1477     do {
1478       // We call G1CMTask::do_marking_step() to completely drain the local
1479       // and global marking stacks of entries pushed by the 'keep alive'
1480       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1481       //
1482       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1483       // if there's nothing more to do (i.e. we've completely drained the
1484       // entries that were pushed as a a result of applying the 'keep alive'
1485       // closure to the entries on the discovered ref lists) or we overflow
1486       // the global marking stack.
1487       //
1488       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1489       // flag while there may still be some work to do. (See the comment at
1490       // the beginning of G1CMTask::do_marking_step() for those conditions -
1491       // one of which is reaching the specified time target.) It is only
1492       // when G1CMTask::do_marking_step() returns without setting the
1493       // has_aborted() flag that the marking step has completed.
1494 
1495       _task->do_marking_step(1000000000.0 /* something very large */,
1496                              true         /* do_termination */,
1497                              _is_serial);
1498     } while (_task->has_aborted() && !_cm->has_overflown());
1499   }
1500 };
1501 
1502 // Implementation of AbstractRefProcTaskExecutor for parallel
1503 // reference processing at the end of G1 concurrent marking
1504 
1505 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
1506 private:
1507   G1CollectedHeap*  _g1h;
1508   G1ConcurrentMark* _cm;
1509   WorkGang*         _workers;
1510   uint              _active_workers;
1511 
1512 public:
1513   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1514                           G1ConcurrentMark* cm,
1515                           WorkGang* workers,
1516                           uint n_workers) :
1517     _g1h(g1h), _cm(cm),
1518     _workers(workers), _active_workers(n_workers) { }
1519 
1520   // Executes the given task using concurrent marking worker threads.
1521   virtual void execute(ProcessTask& task);
1522   virtual void execute(EnqueueTask& task);
1523 };
1524 
1525 class G1CMRefProcTaskProxy: public AbstractGangTask {
1526   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1527   ProcessTask&      _proc_task;
1528   G1CollectedHeap*  _g1h;
1529   G1ConcurrentMark* _cm;
1530 
1531 public:
1532   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1533                        G1CollectedHeap* g1h,
1534                        G1ConcurrentMark* cm) :
1535     AbstractGangTask("Process reference objects in parallel"),
1536     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1537     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1538     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1539   }
1540 
1541   virtual void work(uint worker_id) {
1542     ResourceMark rm;
1543     HandleMark hm;
1544     G1CMTask* task = _cm->task(worker_id);
1545     G1CMIsAliveClosure g1_is_alive(_g1h);
1546     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1547     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1548 
1549     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1550   }
1551 };
1552 
1553 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1554   assert(_workers != NULL, "Need parallel worker threads.");
1555   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1556 
1557   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1558 
1559   // We need to reset the concurrency level before each
1560   // proxy task execution, so that the termination protocol
1561   // and overflow handling in G1CMTask::do_marking_step() knows
1562   // how many workers to wait for.
1563   _cm->set_concurrency(_active_workers);
1564   _workers->run_task(&proc_task_proxy);
1565 }
1566 
1567 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
1568   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1569   EnqueueTask& _enq_task;
1570 
1571 public:
1572   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1573     AbstractGangTask("Enqueue reference objects in parallel"),
1574     _enq_task(enq_task) { }
1575 
1576   virtual void work(uint worker_id) {
1577     _enq_task.work(worker_id);
1578   }
1579 };
1580 
1581 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1582   assert(_workers != NULL, "Need parallel worker threads.");
1583   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1584 
1585   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1586 
1587   // Not strictly necessary but...
1588   //
1589   // We need to reset the concurrency level before each
1590   // proxy task execution, so that the termination protocol
1591   // and overflow handling in G1CMTask::do_marking_step() knows
1592   // how many workers to wait for.
1593   _cm->set_concurrency(_active_workers);
1594   _workers->run_task(&enq_task_proxy);
1595 }
1596 
1597 void G1ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
1598   G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
1599 }
1600 
1601 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
1602   if (has_overflown()) {
1603     // Skip processing the discovered references if we have
1604     // overflown the global marking stack. Reference objects
1605     // only get discovered once so it is OK to not
1606     // de-populate the discovered reference lists. We could have,
1607     // but the only benefit would be that, when marking restarts,
1608     // less reference objects are discovered.
1609     return;
1610   }
1611 
1612   ResourceMark rm;
1613   HandleMark   hm;
1614 
1615   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1616 
1617   // Is alive closure.
1618   G1CMIsAliveClosure g1_is_alive(g1h);
1619 
1620   // Inner scope to exclude the cleaning of the string and symbol
1621   // tables from the displayed time.
1622   {
1623     GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm);
1624 
1625     ReferenceProcessor* rp = g1h->ref_processor_cm();
1626 
1627     // See the comment in G1CollectedHeap::ref_processing_init()
1628     // about how reference processing currently works in G1.
1629 
1630     // Set the soft reference policy
1631     rp->setup_policy(clear_all_soft_refs);
1632     assert(_markStack.isEmpty(), "mark stack should be empty");
1633 
1634     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1635     // in serial reference processing. Note these closures are also
1636     // used for serially processing (by the the current thread) the
1637     // JNI references during parallel reference processing.
1638     //
1639     // These closures do not need to synchronize with the worker
1640     // threads involved in parallel reference processing as these
1641     // instances are executed serially by the current thread (e.g.
1642     // reference processing is not multi-threaded and is thus
1643     // performed by the current thread instead of a gang worker).
1644     //
1645     // The gang tasks involved in parallel reference processing create
1646     // their own instances of these closures, which do their own
1647     // synchronization among themselves.
1648     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1649     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1650 
1651     // We need at least one active thread. If reference processing
1652     // is not multi-threaded we use the current (VMThread) thread,
1653     // otherwise we use the work gang from the G1CollectedHeap and
1654     // we utilize all the worker threads we can.
1655     bool processing_is_mt = rp->processing_is_mt();
1656     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
1657     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
1658 
1659     // Parallel processing task executor.
1660     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
1661                                               g1h->workers(), active_workers);
1662     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1663 
1664     // Set the concurrency level. The phase was already set prior to
1665     // executing the remark task.
1666     set_concurrency(active_workers);
1667 
1668     // Set the degree of MT processing here.  If the discovery was done MT,
1669     // the number of threads involved during discovery could differ from
1670     // the number of active workers.  This is OK as long as the discovered
1671     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1672     rp->set_active_mt_degree(active_workers);
1673 
1674     // Process the weak references.
1675     const ReferenceProcessorStats& stats =
1676         rp->process_discovered_references(&g1_is_alive,
1677                                           &g1_keep_alive,
1678                                           &g1_drain_mark_stack,
1679                                           executor,
1680                                           _gc_timer_cm);
1681     _gc_tracer_cm->report_gc_reference_stats(stats);
1682 
1683     // The do_oop work routines of the keep_alive and drain_marking_stack
1684     // oop closures will set the has_overflown flag if we overflow the
1685     // global marking stack.
1686 
1687     assert(_markStack.overflow() || _markStack.isEmpty(),
1688             "mark stack should be empty (unless it overflowed)");
1689 
1690     if (_markStack.overflow()) {
1691       // This should have been done already when we tried to push an
1692       // entry on to the global mark stack. But let's do it again.
1693       set_has_overflown();
1694     }
1695 
1696     assert(rp->num_q() == active_workers, "why not");
1697 
1698     rp->enqueue_discovered_references(executor);
1699 
1700     rp->verify_no_references_recorded();
1701     assert(!rp->discovery_enabled(), "Post condition");
1702   }
1703 
1704   if (has_overflown()) {
1705     // We can not trust g1_is_alive if the marking stack overflowed
1706     return;
1707   }
1708 
1709   assert(_markStack.isEmpty(), "Marking should have completed");
1710 
1711   // Unload Klasses, String, Symbols, Code Cache, etc.
1712   if (ClassUnloadingWithConcurrentMark) {
1713     bool purged_classes;
1714 
1715     {
1716       GCTraceTime(Debug, gc, phases) trace("System Dictionary Unloading", _gc_timer_cm);
1717       purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
1718     }
1719 
1720     {
1721       GCTraceTime(Debug, gc, phases) trace("Parallel Unloading", _gc_timer_cm);
1722       weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
1723     }
1724   }
1725 
1726   if (G1StringDedup::is_enabled()) {
1727     GCTraceTime(Debug, gc, phases) trace("String Deduplication Unlink", _gc_timer_cm);
1728     G1StringDedup::unlink(&g1_is_alive);
1729   }
1730 }
1731 
1732 void G1ConcurrentMark::swapMarkBitMaps() {
1733   G1CMBitMapRO* temp = _prevMarkBitMap;
1734   _prevMarkBitMap    = (G1CMBitMapRO*)_nextMarkBitMap;
1735   _nextMarkBitMap    = (G1CMBitMap*)  temp;
1736 }
1737 
1738 // Closure for marking entries in SATB buffers.
1739 class G1CMSATBBufferClosure : public SATBBufferClosure {
1740 private:
1741   G1CMTask* _task;
1742   G1CollectedHeap* _g1h;
1743 
1744   // This is very similar to G1CMTask::deal_with_reference, but with
1745   // more relaxed requirements for the argument, so this must be more
1746   // circumspect about treating the argument as an object.
1747   void do_entry(void* entry) const {
1748     _task->increment_refs_reached();
1749     HeapRegion* hr = _g1h->heap_region_containing(entry);
1750     if (entry < hr->next_top_at_mark_start()) {
1751       // Until we get here, we don't know whether entry refers to a valid
1752       // object; it could instead have been a stale reference.
1753       oop obj = static_cast<oop>(entry);
1754       assert(obj->is_oop(true /* ignore mark word */),
1755              "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj));
1756       _task->make_reference_grey(obj);
1757     }
1758   }
1759 
1760 public:
1761   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1762     : _task(task), _g1h(g1h) { }
1763 
1764   virtual void do_buffer(void** buffer, size_t size) {
1765     for (size_t i = 0; i < size; ++i) {
1766       do_entry(buffer[i]);
1767     }
1768   }
1769 };
1770 
1771 class G1RemarkThreadsClosure : public ThreadClosure {
1772   G1CMSATBBufferClosure _cm_satb_cl;
1773   G1CMOopClosure _cm_cl;
1774   MarkingCodeBlobClosure _code_cl;
1775   int _thread_parity;
1776 
1777  public:
1778   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1779     _cm_satb_cl(task, g1h),
1780     _cm_cl(g1h, g1h->concurrent_mark(), task),
1781     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1782     _thread_parity(Threads::thread_claim_parity()) {}
1783 
1784   void do_thread(Thread* thread) {
1785     if (thread->is_Java_thread()) {
1786       if (thread->claim_oops_do(true, _thread_parity)) {
1787         JavaThread* jt = (JavaThread*)thread;
1788 
1789         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1790         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1791         // * Alive if on the stack of an executing method
1792         // * Weakly reachable otherwise
1793         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1794         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1795         jt->nmethods_do(&_code_cl);
1796 
1797         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
1798       }
1799     } else if (thread->is_VM_thread()) {
1800       if (thread->claim_oops_do(true, _thread_parity)) {
1801         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1802       }
1803     }
1804   }
1805 };
1806 
1807 class G1CMRemarkTask: public AbstractGangTask {
1808 private:
1809   G1ConcurrentMark* _cm;
1810 public:
1811   void work(uint worker_id) {
1812     // Since all available tasks are actually started, we should
1813     // only proceed if we're supposed to be active.
1814     if (worker_id < _cm->active_tasks()) {
1815       G1CMTask* task = _cm->task(worker_id);
1816       task->record_start_time();
1817       {
1818         ResourceMark rm;
1819         HandleMark hm;
1820 
1821         G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1822         Threads::threads_do(&threads_f);
1823       }
1824 
1825       do {
1826         task->do_marking_step(1000000000.0 /* something very large */,
1827                               true         /* do_termination       */,
1828                               false        /* is_serial            */);
1829       } while (task->has_aborted() && !_cm->has_overflown());
1830       // If we overflow, then we do not want to restart. We instead
1831       // want to abort remark and do concurrent marking again.
1832       task->record_end_time();
1833     }
1834   }
1835 
1836   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1837     AbstractGangTask("Par Remark"), _cm(cm) {
1838     _cm->terminator()->reset_for_reuse(active_workers);
1839   }
1840 };
1841 
1842 void G1ConcurrentMark::checkpointRootsFinalWork() {
1843   ResourceMark rm;
1844   HandleMark   hm;
1845   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1846 
1847   GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm);
1848 
1849   g1h->ensure_parsability(false);
1850 
1851   // this is remark, so we'll use up all active threads
1852   uint active_workers = g1h->workers()->active_workers();
1853   set_concurrency_and_phase(active_workers, false /* concurrent */);
1854   // Leave _parallel_marking_threads at it's
1855   // value originally calculated in the G1ConcurrentMark
1856   // constructor and pass values of the active workers
1857   // through the gang in the task.
1858 
1859   {
1860     StrongRootsScope srs(active_workers);
1861 
1862     G1CMRemarkTask remarkTask(this, active_workers);
1863     // We will start all available threads, even if we decide that the
1864     // active_workers will be fewer. The extra ones will just bail out
1865     // immediately.
1866     g1h->workers()->run_task(&remarkTask);
1867   }
1868 
1869   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1870   guarantee(has_overflown() ||
1871             satb_mq_set.completed_buffers_num() == 0,
1872             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1873             BOOL_TO_STR(has_overflown()),
1874             satb_mq_set.completed_buffers_num());
1875 
1876   print_stats();
1877 }
1878 
1879 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
1880   // Note we are overriding the read-only view of the prev map here, via
1881   // the cast.
1882   ((G1CMBitMap*)_prevMarkBitMap)->clear_range(mr);
1883 }
1884 
1885 HeapRegion*
1886 G1ConcurrentMark::claim_region(uint worker_id) {
1887   // "checkpoint" the finger
1888   HeapWord* finger = _finger;
1889 
1890   // _heap_end will not change underneath our feet; it only changes at
1891   // yield points.
1892   while (finger < _heap_end) {
1893     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1894 
1895     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1896 
1897     // Above heap_region_containing may return NULL as we always scan claim
1898     // until the end of the heap. In this case, just jump to the next region.
1899     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1900 
1901     // Is the gap between reading the finger and doing the CAS too long?
1902     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
1903     if (res == finger && curr_region != NULL) {
1904       // we succeeded
1905       HeapWord*   bottom        = curr_region->bottom();
1906       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1907 
1908       // notice that _finger == end cannot be guaranteed here since,
1909       // someone else might have moved the finger even further
1910       assert(_finger >= end, "the finger should have moved forward");
1911 
1912       if (limit > bottom) {
1913         return curr_region;
1914       } else {
1915         assert(limit == bottom,
1916                "the region limit should be at bottom");
1917         // we return NULL and the caller should try calling
1918         // claim_region() again.
1919         return NULL;
1920       }
1921     } else {
1922       assert(_finger > finger, "the finger should have moved forward");
1923       // read it again
1924       finger = _finger;
1925     }
1926   }
1927 
1928   return NULL;
1929 }
1930 
1931 #ifndef PRODUCT
1932 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
1933 private:
1934   G1CollectedHeap* _g1h;
1935   const char* _phase;
1936   int _info;
1937 
1938 public:
1939   VerifyNoCSetOops(const char* phase, int info = -1) :
1940     _g1h(G1CollectedHeap::heap()),
1941     _phase(phase),
1942     _info(info)
1943   { }
1944 
1945   void operator()(oop obj) const {
1946     guarantee(obj->is_oop(),
1947               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1948               p2i(obj), _phase, _info);
1949     guarantee(!_g1h->obj_in_cs(obj),
1950               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1951               p2i(obj), _phase, _info);
1952   }
1953 };
1954 
1955 void G1ConcurrentMark::verify_no_cset_oops() {
1956   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1957   if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
1958     return;
1959   }
1960 
1961   // Verify entries on the global mark stack
1962   _markStack.iterate(VerifyNoCSetOops("Stack"));
1963 
1964   // Verify entries on the task queues
1965   for (uint i = 0; i < _max_worker_id; ++i) {
1966     G1CMTaskQueue* queue = _task_queues->queue(i);
1967     queue->iterate(VerifyNoCSetOops("Queue", i));
1968   }
1969 
1970   // Verify the global finger
1971   HeapWord* global_finger = finger();
1972   if (global_finger != NULL && global_finger < _heap_end) {
1973     // Since we always iterate over all regions, we might get a NULL HeapRegion
1974     // here.
1975     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1976     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1977               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1978               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1979   }
1980 
1981   // Verify the task fingers
1982   assert(parallel_marking_threads() <= _max_worker_id, "sanity");
1983   for (uint i = 0; i < parallel_marking_threads(); ++i) {
1984     G1CMTask* task = _tasks[i];
1985     HeapWord* task_finger = task->finger();
1986     if (task_finger != NULL && task_finger < _heap_end) {
1987       // See above note on the global finger verification.
1988       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
1989       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
1990                 !task_hr->in_collection_set(),
1991                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
1992                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
1993     }
1994   }
1995 }
1996 #endif // PRODUCT
1997 void G1ConcurrentMark::create_live_data() {
1998   _g1h->g1_rem_set()->create_card_live_data(_parallel_workers, _nextMarkBitMap);
1999 }
2000 
2001 void G1ConcurrentMark::finalize_live_data() {
2002   _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _nextMarkBitMap);
2003 }
2004 
2005 void G1ConcurrentMark::verify_live_data() {
2006   _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _nextMarkBitMap);
2007 }
2008 
2009 void G1ConcurrentMark::clear_live_data(WorkGang* workers) {
2010   _g1h->g1_rem_set()->clear_card_live_data(workers);
2011 }
2012 
2013 #ifdef ASSERT
2014 void G1ConcurrentMark::verify_live_data_clear() {
2015   _g1h->g1_rem_set()->verify_card_live_data_is_clear();
2016 }
2017 #endif
2018 
2019 void G1ConcurrentMark::print_stats() {
2020   if (!log_is_enabled(Debug, gc, stats)) {
2021     return;
2022   }
2023   log_debug(gc, stats)("---------------------------------------------------------------------");
2024   for (size_t i = 0; i < _active_tasks; ++i) {
2025     _tasks[i]->print_stats();
2026     log_debug(gc, stats)("---------------------------------------------------------------------");
2027   }
2028 }
2029 
2030 void G1ConcurrentMark::abort() {
2031   if (!cmThread()->during_cycle() || _has_aborted) {
2032     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2033     return;
2034   }
2035 
2036   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2037   // concurrent bitmap clearing.
2038   {
2039     GCTraceTime(Debug, gc)("Clear Next Bitmap");
2040     clear_bitmap(_nextMarkBitMap, _g1h->workers(), false);
2041   }
2042   // Note we cannot clear the previous marking bitmap here
2043   // since VerifyDuringGC verifies the objects marked during
2044   // a full GC against the previous bitmap.
2045 
2046   {
2047     GCTraceTime(Debug, gc)("Clear Live Data");
2048     clear_live_data(_g1h->workers());
2049   }
2050   DEBUG_ONLY({
2051     GCTraceTime(Debug, gc)("Verify Live Data Clear");
2052     verify_live_data_clear();
2053   })
2054   // Empty mark stack
2055   reset_marking_state();
2056   for (uint i = 0; i < _max_worker_id; ++i) {
2057     _tasks[i]->clear_region_fields();
2058   }
2059   _first_overflow_barrier_sync.abort();
2060   _second_overflow_barrier_sync.abort();
2061   _has_aborted = true;
2062 
2063   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2064   satb_mq_set.abandon_partial_marking();
2065   // This can be called either during or outside marking, we'll read
2066   // the expected_active value from the SATB queue set.
2067   satb_mq_set.set_active_all_threads(
2068                                  false, /* new active value */
2069                                  satb_mq_set.is_active() /* expected_active */);
2070 }
2071 
2072 static void print_ms_time_info(const char* prefix, const char* name,
2073                                NumberSeq& ns) {
2074   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2075                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2076   if (ns.num() > 0) {
2077     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2078                            prefix, ns.sd(), ns.maximum());
2079   }
2080 }
2081 
2082 void G1ConcurrentMark::print_summary_info() {
2083   Log(gc, marking) log;
2084   if (!log.is_trace()) {
2085     return;
2086   }
2087 
2088   log.trace(" Concurrent marking:");
2089   print_ms_time_info("  ", "init marks", _init_times);
2090   print_ms_time_info("  ", "remarks", _remark_times);
2091   {
2092     print_ms_time_info("     ", "final marks", _remark_mark_times);
2093     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2094 
2095   }
2096   print_ms_time_info("  ", "cleanups", _cleanup_times);
2097   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2098             _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2099   if (G1ScrubRemSets) {
2100     log.trace("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
2101               _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2102   }
2103   log.trace("  Total stop_world time = %8.2f s.",
2104             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2105   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2106             cmThread()->vtime_accum(), cmThread()->vtime_mark_accum());
2107 }
2108 
2109 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2110   _parallel_workers->print_worker_threads_on(st);
2111 }
2112 
2113 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2114   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2115       p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
2116   _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
2117   _nextMarkBitMap->print_on_error(st, " Next Bits: ");
2118 }
2119 
2120 // Closure for iteration over bitmaps
2121 class G1CMBitMapClosure : public BitMapClosure {
2122 private:
2123   // the bitmap that is being iterated over
2124   G1CMBitMap*                 _nextMarkBitMap;
2125   G1ConcurrentMark*           _cm;
2126   G1CMTask*                   _task;
2127 
2128 public:
2129   G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm, G1CMBitMap* nextMarkBitMap) :
2130     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
2131 
2132   bool do_bit(size_t offset) {
2133     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
2134     assert(_nextMarkBitMap->isMarked(addr), "invariant");
2135     assert( addr < _cm->finger(), "invariant");
2136     assert(addr >= _task->finger(), "invariant");
2137 
2138     // We move that task's local finger along.
2139     _task->move_finger_to(addr);
2140 
2141     _task->scan_object(oop(addr));
2142     // we only partially drain the local queue and global stack
2143     _task->drain_local_queue(true);
2144     _task->drain_global_stack(true);
2145 
2146     // if the has_aborted flag has been raised, we need to bail out of
2147     // the iteration
2148     return !_task->has_aborted();
2149   }
2150 };
2151 
2152 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2153   ReferenceProcessor* result = g1h->ref_processor_cm();
2154   assert(result != NULL, "CM reference processor should not be NULL");
2155   return result;
2156 }
2157 
2158 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2159                                G1ConcurrentMark* cm,
2160                                G1CMTask* task)
2161   : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2162     _g1h(g1h), _cm(cm), _task(task)
2163 { }
2164 
2165 void G1CMTask::setup_for_region(HeapRegion* hr) {
2166   assert(hr != NULL,
2167         "claim_region() should have filtered out NULL regions");
2168   _curr_region  = hr;
2169   _finger       = hr->bottom();
2170   update_region_limit();
2171 }
2172 
2173 void G1CMTask::update_region_limit() {
2174   HeapRegion* hr            = _curr_region;
2175   HeapWord* bottom          = hr->bottom();
2176   HeapWord* limit           = hr->next_top_at_mark_start();
2177 
2178   if (limit == bottom) {
2179     // The region was collected underneath our feet.
2180     // We set the finger to bottom to ensure that the bitmap
2181     // iteration that will follow this will not do anything.
2182     // (this is not a condition that holds when we set the region up,
2183     // as the region is not supposed to be empty in the first place)
2184     _finger = bottom;
2185   } else if (limit >= _region_limit) {
2186     assert(limit >= _finger, "peace of mind");
2187   } else {
2188     assert(limit < _region_limit, "only way to get here");
2189     // This can happen under some pretty unusual circumstances.  An
2190     // evacuation pause empties the region underneath our feet (NTAMS
2191     // at bottom). We then do some allocation in the region (NTAMS
2192     // stays at bottom), followed by the region being used as a GC
2193     // alloc region (NTAMS will move to top() and the objects
2194     // originally below it will be grayed). All objects now marked in
2195     // the region are explicitly grayed, if below the global finger,
2196     // and we do not need in fact to scan anything else. So, we simply
2197     // set _finger to be limit to ensure that the bitmap iteration
2198     // doesn't do anything.
2199     _finger = limit;
2200   }
2201 
2202   _region_limit = limit;
2203 }
2204 
2205 void G1CMTask::giveup_current_region() {
2206   assert(_curr_region != NULL, "invariant");
2207   clear_region_fields();
2208 }
2209 
2210 void G1CMTask::clear_region_fields() {
2211   // Values for these three fields that indicate that we're not
2212   // holding on to a region.
2213   _curr_region   = NULL;
2214   _finger        = NULL;
2215   _region_limit  = NULL;
2216 }
2217 
2218 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2219   if (cm_oop_closure == NULL) {
2220     assert(_cm_oop_closure != NULL, "invariant");
2221   } else {
2222     assert(_cm_oop_closure == NULL, "invariant");
2223   }
2224   _cm_oop_closure = cm_oop_closure;
2225 }
2226 
2227 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) {
2228   guarantee(nextMarkBitMap != NULL, "invariant");
2229   _nextMarkBitMap                = nextMarkBitMap;
2230   clear_region_fields();
2231 
2232   _calls                         = 0;
2233   _elapsed_time_ms               = 0.0;
2234   _termination_time_ms           = 0.0;
2235   _termination_start_time_ms     = 0.0;
2236 }
2237 
2238 bool G1CMTask::should_exit_termination() {
2239   regular_clock_call();
2240   // This is called when we are in the termination protocol. We should
2241   // quit if, for some reason, this task wants to abort or the global
2242   // stack is not empty (this means that we can get work from it).
2243   return !_cm->mark_stack_empty() || has_aborted();
2244 }
2245 
2246 void G1CMTask::reached_limit() {
2247   assert(_words_scanned >= _words_scanned_limit ||
2248          _refs_reached >= _refs_reached_limit ,
2249          "shouldn't have been called otherwise");
2250   regular_clock_call();
2251 }
2252 
2253 void G1CMTask::regular_clock_call() {
2254   if (has_aborted()) return;
2255 
2256   // First, we need to recalculate the words scanned and refs reached
2257   // limits for the next clock call.
2258   recalculate_limits();
2259 
2260   // During the regular clock call we do the following
2261 
2262   // (1) If an overflow has been flagged, then we abort.
2263   if (_cm->has_overflown()) {
2264     set_has_aborted();
2265     return;
2266   }
2267 
2268   // If we are not concurrent (i.e. we're doing remark) we don't need
2269   // to check anything else. The other steps are only needed during
2270   // the concurrent marking phase.
2271   if (!concurrent()) return;
2272 
2273   // (2) If marking has been aborted for Full GC, then we also abort.
2274   if (_cm->has_aborted()) {
2275     set_has_aborted();
2276     return;
2277   }
2278 
2279   double curr_time_ms = os::elapsedVTime() * 1000.0;
2280 
2281   // (4) We check whether we should yield. If we have to, then we abort.
2282   if (SuspendibleThreadSet::should_yield()) {
2283     // We should yield. To do this we abort the task. The caller is
2284     // responsible for yielding.
2285     set_has_aborted();
2286     return;
2287   }
2288 
2289   // (5) We check whether we've reached our time quota. If we have,
2290   // then we abort.
2291   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2292   if (elapsed_time_ms > _time_target_ms) {
2293     set_has_aborted();
2294     _has_timed_out = true;
2295     return;
2296   }
2297 
2298   // (6) Finally, we check whether there are enough completed STAB
2299   // buffers available for processing. If there are, we abort.
2300   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2301   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2302     // we do need to process SATB buffers, we'll abort and restart
2303     // the marking task to do so
2304     set_has_aborted();
2305     return;
2306   }
2307 }
2308 
2309 void G1CMTask::recalculate_limits() {
2310   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2311   _words_scanned_limit      = _real_words_scanned_limit;
2312 
2313   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2314   _refs_reached_limit       = _real_refs_reached_limit;
2315 }
2316 
2317 void G1CMTask::decrease_limits() {
2318   // This is called when we believe that we're going to do an infrequent
2319   // operation which will increase the per byte scanned cost (i.e. move
2320   // entries to/from the global stack). It basically tries to decrease the
2321   // scanning limit so that the clock is called earlier.
2322 
2323   _words_scanned_limit = _real_words_scanned_limit -
2324     3 * words_scanned_period / 4;
2325   _refs_reached_limit  = _real_refs_reached_limit -
2326     3 * refs_reached_period / 4;
2327 }
2328 
2329 void G1CMTask::move_entries_to_global_stack() {
2330   // local array where we'll store the entries that will be popped
2331   // from the local queue
2332   oop buffer[global_stack_transfer_size];
2333 
2334   int n = 0;
2335   oop obj;
2336   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
2337     buffer[n] = obj;
2338     ++n;
2339   }
2340 
2341   if (n > 0) {
2342     // we popped at least one entry from the local queue
2343 
2344     if (!_cm->mark_stack_push(buffer, n)) {
2345       set_has_aborted();
2346     }
2347   }
2348 
2349   // this operation was quite expensive, so decrease the limits
2350   decrease_limits();
2351 }
2352 
2353 void G1CMTask::get_entries_from_global_stack() {
2354   // local array where we'll store the entries that will be popped
2355   // from the global stack.
2356   oop buffer[global_stack_transfer_size];
2357   int n;
2358   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
2359   assert(n <= global_stack_transfer_size,
2360          "we should not pop more than the given limit");
2361   if (n > 0) {
2362     // yes, we did actually pop at least one entry
2363     for (int i = 0; i < n; ++i) {
2364       bool success = _task_queue->push(buffer[i]);
2365       // We only call this when the local queue is empty or under a
2366       // given target limit. So, we do not expect this push to fail.
2367       assert(success, "invariant");
2368     }
2369   }
2370 
2371   // this operation was quite expensive, so decrease the limits
2372   decrease_limits();
2373 }
2374 
2375 void G1CMTask::drain_local_queue(bool partially) {
2376   if (has_aborted()) return;
2377 
2378   // Decide what the target size is, depending whether we're going to
2379   // drain it partially (so that other tasks can steal if they run out
2380   // of things to do) or totally (at the very end).
2381   size_t target_size;
2382   if (partially) {
2383     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2384   } else {
2385     target_size = 0;
2386   }
2387 
2388   if (_task_queue->size() > target_size) {
2389     oop obj;
2390     bool ret = _task_queue->pop_local(obj);
2391     while (ret) {
2392       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
2393       assert(!_g1h->is_on_master_free_list(
2394                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
2395 
2396       scan_object(obj);
2397 
2398       if (_task_queue->size() <= target_size || has_aborted()) {
2399         ret = false;
2400       } else {
2401         ret = _task_queue->pop_local(obj);
2402       }
2403     }
2404   }
2405 }
2406 
2407 void G1CMTask::drain_global_stack(bool partially) {
2408   if (has_aborted()) return;
2409 
2410   // We have a policy to drain the local queue before we attempt to
2411   // drain the global stack.
2412   assert(partially || _task_queue->size() == 0, "invariant");
2413 
2414   // Decide what the target size is, depending whether we're going to
2415   // drain it partially (so that other tasks can steal if they run out
2416   // of things to do) or totally (at the very end).  Notice that,
2417   // because we move entries from the global stack in chunks or
2418   // because another task might be doing the same, we might in fact
2419   // drop below the target. But, this is not a problem.
2420   size_t target_size;
2421   if (partially) {
2422     target_size = _cm->partial_mark_stack_size_target();
2423   } else {
2424     target_size = 0;
2425   }
2426 
2427   if (_cm->mark_stack_size() > target_size) {
2428     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2429       get_entries_from_global_stack();
2430       drain_local_queue(partially);
2431     }
2432   }
2433 }
2434 
2435 // SATB Queue has several assumptions on whether to call the par or
2436 // non-par versions of the methods. this is why some of the code is
2437 // replicated. We should really get rid of the single-threaded version
2438 // of the code to simplify things.
2439 void G1CMTask::drain_satb_buffers() {
2440   if (has_aborted()) return;
2441 
2442   // We set this so that the regular clock knows that we're in the
2443   // middle of draining buffers and doesn't set the abort flag when it
2444   // notices that SATB buffers are available for draining. It'd be
2445   // very counter productive if it did that. :-)
2446   _draining_satb_buffers = true;
2447 
2448   G1CMSATBBufferClosure satb_cl(this, _g1h);
2449   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2450 
2451   // This keeps claiming and applying the closure to completed buffers
2452   // until we run out of buffers or we need to abort.
2453   while (!has_aborted() &&
2454          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2455     regular_clock_call();
2456   }
2457 
2458   _draining_satb_buffers = false;
2459 
2460   assert(has_aborted() ||
2461          concurrent() ||
2462          satb_mq_set.completed_buffers_num() == 0, "invariant");
2463 
2464   // again, this was a potentially expensive operation, decrease the
2465   // limits to get the regular clock call early
2466   decrease_limits();
2467 }
2468 
2469 void G1CMTask::print_stats() {
2470   log_debug(gc, stats)("Marking Stats, task = %u, calls = %d",
2471                        _worker_id, _calls);
2472   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2473                        _elapsed_time_ms, _termination_time_ms);
2474   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
2475                        _step_times_ms.num(), _step_times_ms.avg(),
2476                        _step_times_ms.sd());
2477   log_debug(gc, stats)("                    max = %1.2lfms, total = %1.2lfms",
2478                        _step_times_ms.maximum(), _step_times_ms.sum());
2479 }
2480 
2481 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) {
2482   return _task_queues->steal(worker_id, hash_seed, obj);
2483 }
2484 
2485 /*****************************************************************************
2486 
2487     The do_marking_step(time_target_ms, ...) method is the building
2488     block of the parallel marking framework. It can be called in parallel
2489     with other invocations of do_marking_step() on different tasks
2490     (but only one per task, obviously) and concurrently with the
2491     mutator threads, or during remark, hence it eliminates the need
2492     for two versions of the code. When called during remark, it will
2493     pick up from where the task left off during the concurrent marking
2494     phase. Interestingly, tasks are also claimable during evacuation
2495     pauses too, since do_marking_step() ensures that it aborts before
2496     it needs to yield.
2497 
2498     The data structures that it uses to do marking work are the
2499     following:
2500 
2501       (1) Marking Bitmap. If there are gray objects that appear only
2502       on the bitmap (this happens either when dealing with an overflow
2503       or when the initial marking phase has simply marked the roots
2504       and didn't push them on the stack), then tasks claim heap
2505       regions whose bitmap they then scan to find gray objects. A
2506       global finger indicates where the end of the last claimed region
2507       is. A local finger indicates how far into the region a task has
2508       scanned. The two fingers are used to determine how to gray an
2509       object (i.e. whether simply marking it is OK, as it will be
2510       visited by a task in the future, or whether it needs to be also
2511       pushed on a stack).
2512 
2513       (2) Local Queue. The local queue of the task which is accessed
2514       reasonably efficiently by the task. Other tasks can steal from
2515       it when they run out of work. Throughout the marking phase, a
2516       task attempts to keep its local queue short but not totally
2517       empty, so that entries are available for stealing by other
2518       tasks. Only when there is no more work, a task will totally
2519       drain its local queue.
2520 
2521       (3) Global Mark Stack. This handles local queue overflow. During
2522       marking only sets of entries are moved between it and the local
2523       queues, as access to it requires a mutex and more fine-grain
2524       interaction with it which might cause contention. If it
2525       overflows, then the marking phase should restart and iterate
2526       over the bitmap to identify gray objects. Throughout the marking
2527       phase, tasks attempt to keep the global mark stack at a small
2528       length but not totally empty, so that entries are available for
2529       popping by other tasks. Only when there is no more work, tasks
2530       will totally drain the global mark stack.
2531 
2532       (4) SATB Buffer Queue. This is where completed SATB buffers are
2533       made available. Buffers are regularly removed from this queue
2534       and scanned for roots, so that the queue doesn't get too
2535       long. During remark, all completed buffers are processed, as
2536       well as the filled in parts of any uncompleted buffers.
2537 
2538     The do_marking_step() method tries to abort when the time target
2539     has been reached. There are a few other cases when the
2540     do_marking_step() method also aborts:
2541 
2542       (1) When the marking phase has been aborted (after a Full GC).
2543 
2544       (2) When a global overflow (on the global stack) has been
2545       triggered. Before the task aborts, it will actually sync up with
2546       the other tasks to ensure that all the marking data structures
2547       (local queues, stacks, fingers etc.)  are re-initialized so that
2548       when do_marking_step() completes, the marking phase can
2549       immediately restart.
2550 
2551       (3) When enough completed SATB buffers are available. The
2552       do_marking_step() method only tries to drain SATB buffers right
2553       at the beginning. So, if enough buffers are available, the
2554       marking step aborts and the SATB buffers are processed at
2555       the beginning of the next invocation.
2556 
2557       (4) To yield. when we have to yield then we abort and yield
2558       right at the end of do_marking_step(). This saves us from a lot
2559       of hassle as, by yielding we might allow a Full GC. If this
2560       happens then objects will be compacted underneath our feet, the
2561       heap might shrink, etc. We save checking for this by just
2562       aborting and doing the yield right at the end.
2563 
2564     From the above it follows that the do_marking_step() method should
2565     be called in a loop (or, otherwise, regularly) until it completes.
2566 
2567     If a marking step completes without its has_aborted() flag being
2568     true, it means it has completed the current marking phase (and
2569     also all other marking tasks have done so and have all synced up).
2570 
2571     A method called regular_clock_call() is invoked "regularly" (in
2572     sub ms intervals) throughout marking. It is this clock method that
2573     checks all the abort conditions which were mentioned above and
2574     decides when the task should abort. A work-based scheme is used to
2575     trigger this clock method: when the number of object words the
2576     marking phase has scanned or the number of references the marking
2577     phase has visited reach a given limit. Additional invocations to
2578     the method clock have been planted in a few other strategic places
2579     too. The initial reason for the clock method was to avoid calling
2580     vtime too regularly, as it is quite expensive. So, once it was in
2581     place, it was natural to piggy-back all the other conditions on it
2582     too and not constantly check them throughout the code.
2583 
2584     If do_termination is true then do_marking_step will enter its
2585     termination protocol.
2586 
2587     The value of is_serial must be true when do_marking_step is being
2588     called serially (i.e. by the VMThread) and do_marking_step should
2589     skip any synchronization in the termination and overflow code.
2590     Examples include the serial remark code and the serial reference
2591     processing closures.
2592 
2593     The value of is_serial must be false when do_marking_step is
2594     being called by any of the worker threads in a work gang.
2595     Examples include the concurrent marking code (CMMarkingTask),
2596     the MT remark code, and the MT reference processing closures.
2597 
2598  *****************************************************************************/
2599 
2600 void G1CMTask::do_marking_step(double time_target_ms,
2601                                bool do_termination,
2602                                bool is_serial) {
2603   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2604   assert(concurrent() == _cm->concurrent(), "they should be the same");
2605 
2606   G1CollectorPolicy* g1_policy = _g1h->g1_policy();
2607   assert(_task_queues != NULL, "invariant");
2608   assert(_task_queue != NULL, "invariant");
2609   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
2610 
2611   assert(!_claimed,
2612          "only one thread should claim this task at any one time");
2613 
2614   // OK, this doesn't safeguard again all possible scenarios, as it is
2615   // possible for two threads to set the _claimed flag at the same
2616   // time. But it is only for debugging purposes anyway and it will
2617   // catch most problems.
2618   _claimed = true;
2619 
2620   _start_time_ms = os::elapsedVTime() * 1000.0;
2621 
2622   // If do_stealing is true then do_marking_step will attempt to
2623   // steal work from the other G1CMTasks. It only makes sense to
2624   // enable stealing when the termination protocol is enabled
2625   // and do_marking_step() is not being called serially.
2626   bool do_stealing = do_termination && !is_serial;
2627 
2628   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2629   _time_target_ms = time_target_ms - diff_prediction_ms;
2630 
2631   // set up the variables that are used in the work-based scheme to
2632   // call the regular clock method
2633   _words_scanned = 0;
2634   _refs_reached  = 0;
2635   recalculate_limits();
2636 
2637   // clear all flags
2638   clear_has_aborted();
2639   _has_timed_out = false;
2640   _draining_satb_buffers = false;
2641 
2642   ++_calls;
2643 
2644   // Set up the bitmap and oop closures. Anything that uses them is
2645   // eventually called from this method, so it is OK to allocate these
2646   // statically.
2647   G1CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
2648   G1CMOopClosure    cm_oop_closure(_g1h, _cm, this);
2649   set_cm_oop_closure(&cm_oop_closure);
2650 
2651   if (_cm->has_overflown()) {
2652     // This can happen if the mark stack overflows during a GC pause
2653     // and this task, after a yield point, restarts. We have to abort
2654     // as we need to get into the overflow protocol which happens
2655     // right at the end of this task.
2656     set_has_aborted();
2657   }
2658 
2659   // First drain any available SATB buffers. After this, we will not
2660   // look at SATB buffers before the next invocation of this method.
2661   // If enough completed SATB buffers are queued up, the regular clock
2662   // will abort this task so that it restarts.
2663   drain_satb_buffers();
2664   // ...then partially drain the local queue and the global stack
2665   drain_local_queue(true);
2666   drain_global_stack(true);
2667 
2668   do {
2669     if (!has_aborted() && _curr_region != NULL) {
2670       // This means that we're already holding on to a region.
2671       assert(_finger != NULL, "if region is not NULL, then the finger "
2672              "should not be NULL either");
2673 
2674       // We might have restarted this task after an evacuation pause
2675       // which might have evacuated the region we're holding on to
2676       // underneath our feet. Let's read its limit again to make sure
2677       // that we do not iterate over a region of the heap that
2678       // contains garbage (update_region_limit() will also move
2679       // _finger to the start of the region if it is found empty).
2680       update_region_limit();
2681       // We will start from _finger not from the start of the region,
2682       // as we might be restarting this task after aborting half-way
2683       // through scanning this region. In this case, _finger points to
2684       // the address where we last found a marked object. If this is a
2685       // fresh region, _finger points to start().
2686       MemRegion mr = MemRegion(_finger, _region_limit);
2687 
2688       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2689              "humongous regions should go around loop once only");
2690 
2691       // Some special cases:
2692       // If the memory region is empty, we can just give up the region.
2693       // If the current region is humongous then we only need to check
2694       // the bitmap for the bit associated with the start of the object,
2695       // scan the object if it's live, and give up the region.
2696       // Otherwise, let's iterate over the bitmap of the part of the region
2697       // that is left.
2698       // If the iteration is successful, give up the region.
2699       if (mr.is_empty()) {
2700         giveup_current_region();
2701         regular_clock_call();
2702       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2703         if (_nextMarkBitMap->isMarked(mr.start())) {
2704           // The object is marked - apply the closure
2705           BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
2706           bitmap_closure.do_bit(offset);
2707         }
2708         // Even if this task aborted while scanning the humongous object
2709         // we can (and should) give up the current region.
2710         giveup_current_region();
2711         regular_clock_call();
2712       } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
2713         giveup_current_region();
2714         regular_clock_call();
2715       } else {
2716         assert(has_aborted(), "currently the only way to do so");
2717         // The only way to abort the bitmap iteration is to return
2718         // false from the do_bit() method. However, inside the
2719         // do_bit() method we move the _finger to point to the
2720         // object currently being looked at. So, if we bail out, we
2721         // have definitely set _finger to something non-null.
2722         assert(_finger != NULL, "invariant");
2723 
2724         // Region iteration was actually aborted. So now _finger
2725         // points to the address of the object we last scanned. If we
2726         // leave it there, when we restart this task, we will rescan
2727         // the object. It is easy to avoid this. We move the finger by
2728         // enough to point to the next possible object header (the
2729         // bitmap knows by how much we need to move it as it knows its
2730         // granularity).
2731         assert(_finger < _region_limit, "invariant");
2732         HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
2733         // Check if bitmap iteration was aborted while scanning the last object
2734         if (new_finger >= _region_limit) {
2735           giveup_current_region();
2736         } else {
2737           move_finger_to(new_finger);
2738         }
2739       }
2740     }
2741     // At this point we have either completed iterating over the
2742     // region we were holding on to, or we have aborted.
2743 
2744     // We then partially drain the local queue and the global stack.
2745     // (Do we really need this?)
2746     drain_local_queue(true);
2747     drain_global_stack(true);
2748 
2749     // Read the note on the claim_region() method on why it might
2750     // return NULL with potentially more regions available for
2751     // claiming and why we have to check out_of_regions() to determine
2752     // whether we're done or not.
2753     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2754       // We are going to try to claim a new region. We should have
2755       // given up on the previous one.
2756       // Separated the asserts so that we know which one fires.
2757       assert(_curr_region  == NULL, "invariant");
2758       assert(_finger       == NULL, "invariant");
2759       assert(_region_limit == NULL, "invariant");
2760       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2761       if (claimed_region != NULL) {
2762         // Yes, we managed to claim one
2763         setup_for_region(claimed_region);
2764         assert(_curr_region == claimed_region, "invariant");
2765       }
2766       // It is important to call the regular clock here. It might take
2767       // a while to claim a region if, for example, we hit a large
2768       // block of empty regions. So we need to call the regular clock
2769       // method once round the loop to make sure it's called
2770       // frequently enough.
2771       regular_clock_call();
2772     }
2773 
2774     if (!has_aborted() && _curr_region == NULL) {
2775       assert(_cm->out_of_regions(),
2776              "at this point we should be out of regions");
2777     }
2778   } while ( _curr_region != NULL && !has_aborted());
2779 
2780   if (!has_aborted()) {
2781     // We cannot check whether the global stack is empty, since other
2782     // tasks might be pushing objects to it concurrently.
2783     assert(_cm->out_of_regions(),
2784            "at this point we should be out of regions");
2785     // Try to reduce the number of available SATB buffers so that
2786     // remark has less work to do.
2787     drain_satb_buffers();
2788   }
2789 
2790   // Since we've done everything else, we can now totally drain the
2791   // local queue and global stack.
2792   drain_local_queue(false);
2793   drain_global_stack(false);
2794 
2795   // Attempt at work stealing from other task's queues.
2796   if (do_stealing && !has_aborted()) {
2797     // We have not aborted. This means that we have finished all that
2798     // we could. Let's try to do some stealing...
2799 
2800     // We cannot check whether the global stack is empty, since other
2801     // tasks might be pushing objects to it concurrently.
2802     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2803            "only way to reach here");
2804     while (!has_aborted()) {
2805       oop obj;
2806       if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
2807         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
2808                "any stolen object should be marked");
2809         scan_object(obj);
2810 
2811         // And since we're towards the end, let's totally drain the
2812         // local queue and global stack.
2813         drain_local_queue(false);
2814         drain_global_stack(false);
2815       } else {
2816         break;
2817       }
2818     }
2819   }
2820 
2821   // We still haven't aborted. Now, let's try to get into the
2822   // termination protocol.
2823   if (do_termination && !has_aborted()) {
2824     // We cannot check whether the global stack is empty, since other
2825     // tasks might be concurrently pushing objects on it.
2826     // Separated the asserts so that we know which one fires.
2827     assert(_cm->out_of_regions(), "only way to reach here");
2828     assert(_task_queue->size() == 0, "only way to reach here");
2829     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2830 
2831     // The G1CMTask class also extends the TerminatorTerminator class,
2832     // hence its should_exit_termination() method will also decide
2833     // whether to exit the termination protocol or not.
2834     bool finished = (is_serial ||
2835                      _cm->terminator()->offer_termination(this));
2836     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2837     _termination_time_ms +=
2838       termination_end_time_ms - _termination_start_time_ms;
2839 
2840     if (finished) {
2841       // We're all done.
2842 
2843       if (_worker_id == 0) {
2844         // let's allow task 0 to do this
2845         if (concurrent()) {
2846           assert(_cm->concurrent_marking_in_progress(), "invariant");
2847           // we need to set this to false before the next
2848           // safepoint. This way we ensure that the marking phase
2849           // doesn't observe any more heap expansions.
2850           _cm->clear_concurrent_marking_in_progress();
2851         }
2852       }
2853 
2854       // We can now guarantee that the global stack is empty, since
2855       // all other tasks have finished. We separated the guarantees so
2856       // that, if a condition is false, we can immediately find out
2857       // which one.
2858       guarantee(_cm->out_of_regions(), "only way to reach here");
2859       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2860       guarantee(_task_queue->size() == 0, "only way to reach here");
2861       guarantee(!_cm->has_overflown(), "only way to reach here");
2862       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
2863     } else {
2864       // Apparently there's more work to do. Let's abort this task. It
2865       // will restart it and we can hopefully find more things to do.
2866       set_has_aborted();
2867     }
2868   }
2869 
2870   // Mainly for debugging purposes to make sure that a pointer to the
2871   // closure which was statically allocated in this frame doesn't
2872   // escape it by accident.
2873   set_cm_oop_closure(NULL);
2874   double end_time_ms = os::elapsedVTime() * 1000.0;
2875   double elapsed_time_ms = end_time_ms - _start_time_ms;
2876   // Update the step history.
2877   _step_times_ms.add(elapsed_time_ms);
2878 
2879   if (has_aborted()) {
2880     // The task was aborted for some reason.
2881     if (_has_timed_out) {
2882       double diff_ms = elapsed_time_ms - _time_target_ms;
2883       // Keep statistics of how well we did with respect to hitting
2884       // our target only if we actually timed out (if we aborted for
2885       // other reasons, then the results might get skewed).
2886       _marking_step_diffs_ms.add(diff_ms);
2887     }
2888 
2889     if (_cm->has_overflown()) {
2890       // This is the interesting one. We aborted because a global
2891       // overflow was raised. This means we have to restart the
2892       // marking phase and start iterating over regions. However, in
2893       // order to do this we have to make sure that all tasks stop
2894       // what they are doing and re-initialize in a safe manner. We
2895       // will achieve this with the use of two barrier sync points.
2896 
2897       if (!is_serial) {
2898         // We only need to enter the sync barrier if being called
2899         // from a parallel context
2900         _cm->enter_first_sync_barrier(_worker_id);
2901 
2902         // When we exit this sync barrier we know that all tasks have
2903         // stopped doing marking work. So, it's now safe to
2904         // re-initialize our data structures. At the end of this method,
2905         // task 0 will clear the global data structures.
2906       }
2907 
2908       // We clear the local state of this task...
2909       clear_region_fields();
2910 
2911       if (!is_serial) {
2912         // ...and enter the second barrier.
2913         _cm->enter_second_sync_barrier(_worker_id);
2914       }
2915       // At this point, if we're during the concurrent phase of
2916       // marking, everything has been re-initialized and we're
2917       // ready to restart.
2918     }
2919   }
2920 
2921   _claimed = false;
2922 }
2923 
2924 G1CMTask::G1CMTask(uint worker_id,
2925                    G1ConcurrentMark* cm,
2926                    G1CMTaskQueue* task_queue,
2927                    G1CMTaskQueueSet* task_queues)
2928   : _g1h(G1CollectedHeap::heap()),
2929     _worker_id(worker_id), _cm(cm),
2930     _claimed(false),
2931     _nextMarkBitMap(NULL), _hash_seed(17),
2932     _task_queue(task_queue),
2933     _task_queues(task_queues),
2934     _cm_oop_closure(NULL) {
2935   guarantee(task_queue != NULL, "invariant");
2936   guarantee(task_queues != NULL, "invariant");
2937 
2938   _marking_step_diffs_ms.add(0.5);
2939 }
2940 
2941 // These are formatting macros that are used below to ensure
2942 // consistent formatting. The *_H_* versions are used to format the
2943 // header for a particular value and they should be kept consistent
2944 // with the corresponding macro. Also note that most of the macros add
2945 // the necessary white space (as a prefix) which makes them a bit
2946 // easier to compose.
2947 
2948 // All the output lines are prefixed with this string to be able to
2949 // identify them easily in a large log file.
2950 #define G1PPRL_LINE_PREFIX            "###"
2951 
2952 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2953 #ifdef _LP64
2954 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2955 #else // _LP64
2956 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2957 #endif // _LP64
2958 
2959 // For per-region info
2960 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2961 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2962 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2963 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2964 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2965 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2966 
2967 // For summary info
2968 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2969 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2970 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2971 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2972 
2973 G1PrintRegionLivenessInfoClosure::
2974 G1PrintRegionLivenessInfoClosure(const char* phase_name)
2975   : _total_used_bytes(0), _total_capacity_bytes(0),
2976     _total_prev_live_bytes(0), _total_next_live_bytes(0),
2977     _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
2978   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2979   MemRegion g1_reserved = g1h->g1_reserved();
2980   double now = os::elapsedTime();
2981 
2982   // Print the header of the output.
2983   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2984   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2985                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2986                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2987                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2988                           HeapRegion::GrainBytes);
2989   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2990   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2991                           G1PPRL_TYPE_H_FORMAT
2992                           G1PPRL_ADDR_BASE_H_FORMAT
2993                           G1PPRL_BYTE_H_FORMAT
2994                           G1PPRL_BYTE_H_FORMAT
2995                           G1PPRL_BYTE_H_FORMAT
2996                           G1PPRL_DOUBLE_H_FORMAT
2997                           G1PPRL_BYTE_H_FORMAT
2998                           G1PPRL_BYTE_H_FORMAT,
2999                           "type", "address-range",
3000                           "used", "prev-live", "next-live", "gc-eff",
3001                           "remset", "code-roots");
3002   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3003                           G1PPRL_TYPE_H_FORMAT
3004                           G1PPRL_ADDR_BASE_H_FORMAT
3005                           G1PPRL_BYTE_H_FORMAT
3006                           G1PPRL_BYTE_H_FORMAT
3007                           G1PPRL_BYTE_H_FORMAT
3008                           G1PPRL_DOUBLE_H_FORMAT
3009                           G1PPRL_BYTE_H_FORMAT
3010                           G1PPRL_BYTE_H_FORMAT,
3011                           "", "",
3012                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3013                           "(bytes)", "(bytes)");
3014 }
3015 
3016 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
3017   const char* type       = r->get_type_str();
3018   HeapWord* bottom       = r->bottom();
3019   HeapWord* end          = r->end();
3020   size_t capacity_bytes  = r->capacity();
3021   size_t used_bytes      = r->used();
3022   size_t prev_live_bytes = r->live_bytes();
3023   size_t next_live_bytes = r->next_live_bytes();
3024   double gc_eff          = r->gc_efficiency();
3025   size_t remset_bytes    = r->rem_set()->mem_size();
3026   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3027 
3028   _total_used_bytes      += used_bytes;
3029   _total_capacity_bytes  += capacity_bytes;
3030   _total_prev_live_bytes += prev_live_bytes;
3031   _total_next_live_bytes += next_live_bytes;
3032   _total_remset_bytes    += remset_bytes;
3033   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3034 
3035   // Print a line for this particular region.
3036   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3037                           G1PPRL_TYPE_FORMAT
3038                           G1PPRL_ADDR_BASE_FORMAT
3039                           G1PPRL_BYTE_FORMAT
3040                           G1PPRL_BYTE_FORMAT
3041                           G1PPRL_BYTE_FORMAT
3042                           G1PPRL_DOUBLE_FORMAT
3043                           G1PPRL_BYTE_FORMAT
3044                           G1PPRL_BYTE_FORMAT,
3045                           type, p2i(bottom), p2i(end),
3046                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3047                           remset_bytes, strong_code_roots_bytes);
3048 
3049   return false;
3050 }
3051 
3052 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3053   // add static memory usages to remembered set sizes
3054   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3055   // Print the footer of the output.
3056   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3057   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3058                          " SUMMARY"
3059                          G1PPRL_SUM_MB_FORMAT("capacity")
3060                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3061                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3062                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3063                          G1PPRL_SUM_MB_FORMAT("remset")
3064                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3065                          bytes_to_mb(_total_capacity_bytes),
3066                          bytes_to_mb(_total_used_bytes),
3067                          perc(_total_used_bytes, _total_capacity_bytes),
3068                          bytes_to_mb(_total_prev_live_bytes),
3069                          perc(_total_prev_live_bytes, _total_capacity_bytes),
3070                          bytes_to_mb(_total_next_live_bytes),
3071                          perc(_total_next_live_bytes, _total_capacity_bytes),
3072                          bytes_to_mb(_total_remset_bytes),
3073                          bytes_to_mb(_total_strong_code_roots_bytes));
3074 }