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/g1StringDedup.hpp"
  37 #include "gc/g1/heapRegion.inline.hpp"
  38 #include "gc/g1/heapRegionRemSet.hpp"
  39 #include "gc/g1/heapRegionSet.inline.hpp"
  40 #include "gc/g1/suspendibleThreadSet.hpp"
  41 #include "gc/shared/gcId.hpp"
  42 #include "gc/shared/gcTimer.hpp"
  43 #include "gc/shared/gcTrace.hpp"
  44 #include "gc/shared/gcTraceTime.inline.hpp"
  45 #include "gc/shared/genOopClosures.inline.hpp"
  46 #include "gc/shared/referencePolicy.hpp"
  47 #include "gc/shared/strongRootsScope.hpp"
  48 #include "gc/shared/taskqueue.inline.hpp"
  49 #include "gc/shared/vmGCOperations.hpp"
  50 #include "logging/log.hpp"
  51 #include "memory/allocation.hpp"
  52 #include "memory/resourceArea.hpp"
  53 #include "oops/oop.inline.hpp"
  54 #include "runtime/atomic.inline.hpp"
  55 #include "runtime/handles.inline.hpp"
  56 #include "runtime/java.hpp"
  57 #include "runtime/prefetch.inline.hpp"
  58 #include "services/memTracker.hpp"
  59 
  60 // Concurrent marking bit map wrapper
  61 
  62 G1CMBitMapRO::G1CMBitMapRO(int shifter) :
  63   _bm(),
  64   _shifter(shifter) {
  65   _bmStartWord = 0;
  66   _bmWordSize = 0;
  67 }
  68 
  69 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
  70                                                  const HeapWord* limit) const {
  71   // First we must round addr *up* to a possible object boundary.
  72   addr = (HeapWord*)align_size_up((intptr_t)addr,
  73                                   HeapWordSize << _shifter);
  74   size_t addrOffset = heapWordToOffset(addr);
  75   assert(limit != NULL, "limit must not be NULL");
  76   size_t limitOffset = heapWordToOffset(limit);
  77   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  78   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  79   assert(nextAddr >= addr, "get_next_one postcondition");
  80   assert(nextAddr == limit || isMarked(nextAddr),
  81          "get_next_one postcondition");
  82   return nextAddr;
  83 }
  84 
  85 #ifndef PRODUCT
  86 bool G1CMBitMapRO::covers(MemRegion heap_rs) const {
  87   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  88   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
  89          "size inconsistency");
  90   return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
  91          _bmWordSize  == heap_rs.word_size();
  92 }
  93 #endif
  94 
  95 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
  96   _bm.print_on_error(st, prefix);
  97 }
  98 
  99 size_t G1CMBitMap::compute_size(size_t heap_size) {
 100   return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
 101 }
 102 
 103 size_t G1CMBitMap::mark_distance() {
 104   return MinObjAlignmentInBytes * BitsPerByte;
 105 }
 106 
 107 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
 108   _bmStartWord = heap.start();
 109   _bmWordSize = heap.word_size();
 110 
 111   _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
 112   _bm.set_size(_bmWordSize >> _shifter);
 113 
 114   storage->set_mapping_changed_listener(&_listener);
 115 }
 116 
 117 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
 118   if (zero_filled) {
 119     return;
 120   }
 121   // We need to clear the bitmap on commit, removing any existing information.
 122   MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
 123   _bm->clear_range(mr);
 124 }
 125 
 126 void G1CMBitMap::clear_range(MemRegion mr) {
 127   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 128   assert(!mr.is_empty(), "unexpected empty region");
 129   // convert address range into offset range
 130   _bm.at_put_range(heapWordToOffset(mr.start()),
 131                    heapWordToOffset(mr.end()), false);
 132 }
 133 
 134 G1CMMarkStack::G1CMMarkStack(G1ConcurrentMark* cm) :
 135   _base(NULL), _cm(cm)
 136 {}
 137 
 138 bool G1CMMarkStack::allocate(size_t capacity) {
 139   // allocate a stack of the requisite depth
 140   ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
 141   if (!rs.is_reserved()) {
 142     log_warning(gc)("ConcurrentMark MarkStack allocation failure");
 143     return false;
 144   }
 145   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 146   if (!_virtual_space.initialize(rs, rs.size())) {
 147     log_warning(gc)("ConcurrentMark MarkStack backing store failure");
 148     // Release the virtual memory reserved for the marking stack
 149     rs.release();
 150     return false;
 151   }
 152   assert(_virtual_space.committed_size() == rs.size(),
 153          "Didn't reserve backing store for all of G1ConcurrentMark stack?");
 154   _base = (oop*) _virtual_space.low();
 155   setEmpty();
 156   _capacity = (jint) capacity;
 157   _saved_index = -1;
 158   _should_expand = false;
 159   return true;
 160 }
 161 
 162 void G1CMMarkStack::expand() {
 163   // Called, during remark, if we've overflown the marking stack during marking.
 164   assert(isEmpty(), "stack should been emptied while handling overflow");
 165   assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
 166   // Clear expansion flag
 167   _should_expand = false;
 168   if (_capacity == (jint) MarkStackSizeMax) {
 169     log_trace(gc)("(benign) Can't expand marking stack capacity, at max size limit");
 170     return;
 171   }
 172   // Double capacity if possible
 173   jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
 174   // Do not give up existing stack until we have managed to
 175   // get the double capacity that we desired.
 176   ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
 177                                                            sizeof(oop)));
 178   if (rs.is_reserved()) {
 179     // Release the backing store associated with old stack
 180     _virtual_space.release();
 181     // Reinitialize virtual space for new stack
 182     if (!_virtual_space.initialize(rs, rs.size())) {
 183       fatal("Not enough swap for expanded marking stack capacity");
 184     }
 185     _base = (oop*)(_virtual_space.low());
 186     _index = 0;
 187     _capacity = new_capacity;
 188   } else {
 189     // Failed to double capacity, continue;
 190     log_trace(gc)("(benign) Failed to expand marking stack capacity from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
 191                   _capacity / K, new_capacity / K);
 192   }
 193 }
 194 
 195 void G1CMMarkStack::set_should_expand() {
 196   // If we're resetting the marking state because of an
 197   // marking stack overflow, record that we should, if
 198   // possible, expand the stack.
 199   _should_expand = _cm->has_overflown();
 200 }
 201 
 202 G1CMMarkStack::~G1CMMarkStack() {
 203   if (_base != NULL) {
 204     _base = NULL;
 205     _virtual_space.release();
 206   }
 207 }
 208 
 209 void G1CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 210   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 211   jint start = _index;
 212   jint next_index = start + n;
 213   if (next_index > _capacity) {
 214     _overflow = true;
 215     return;
 216   }
 217   // Otherwise.
 218   _index = next_index;
 219   for (int i = 0; i < n; i++) {
 220     int ind = start + i;
 221     assert(ind < _capacity, "By overflow test above.");
 222     _base[ind] = ptr_arr[i];
 223   }
 224 }
 225 
 226 bool G1CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 227   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 228   jint index = _index;
 229   if (index == 0) {
 230     *n = 0;
 231     return false;
 232   } else {
 233     int k = MIN2(max, index);
 234     jint  new_ind = index - k;
 235     for (int j = 0; j < k; j++) {
 236       ptr_arr[j] = _base[new_ind + j];
 237     }
 238     _index = new_ind;
 239     *n = k;
 240     return true;
 241   }
 242 }
 243 
 244 void G1CMMarkStack::note_start_of_gc() {
 245   assert(_saved_index == -1,
 246          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 247   _saved_index = _index;
 248 }
 249 
 250 void G1CMMarkStack::note_end_of_gc() {
 251   // This is intentionally a guarantee, instead of an assert. If we
 252   // accidentally add something to the mark stack during GC, it
 253   // will be a correctness issue so it's better if we crash. we'll
 254   // only check this once per GC anyway, so it won't be a performance
 255   // issue in any way.
 256   guarantee(_saved_index == _index,
 257             "saved index: %d index: %d", _saved_index, _index);
 258   _saved_index = -1;
 259 }
 260 
 261 G1CMRootRegions::G1CMRootRegions() :
 262   _young_list(NULL), _cm(NULL), _scan_in_progress(false),
 263   _should_abort(false),  _next_survivor(NULL) { }
 264 
 265 void G1CMRootRegions::init(G1CollectedHeap* g1h, G1ConcurrentMark* cm) {
 266   _young_list = g1h->young_list();
 267   _cm = cm;
 268 }
 269 
 270 void G1CMRootRegions::prepare_for_scan() {
 271   assert(!scan_in_progress(), "pre-condition");
 272 
 273   // Currently, only survivors can be root regions.
 274   assert(_next_survivor == NULL, "pre-condition");
 275   _next_survivor = _young_list->first_survivor_region();
 276   _scan_in_progress = (_next_survivor != NULL);
 277   _should_abort = false;
 278 }
 279 
 280 HeapRegion* G1CMRootRegions::claim_next() {
 281   if (_should_abort) {
 282     // If someone has set the should_abort flag, we return NULL to
 283     // force the caller to bail out of their loop.
 284     return NULL;
 285   }
 286 
 287   // Currently, only survivors can be root regions.
 288   HeapRegion* res = _next_survivor;
 289   if (res != NULL) {
 290     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 291     // Read it again in case it changed while we were waiting for the lock.
 292     res = _next_survivor;
 293     if (res != NULL) {
 294       if (res == _young_list->last_survivor_region()) {
 295         // We just claimed the last survivor so store NULL to indicate
 296         // that we're done.
 297         _next_survivor = NULL;
 298       } else {
 299         _next_survivor = res->get_next_young_region();
 300       }
 301     } else {
 302       // Someone else claimed the last survivor while we were trying
 303       // to take the lock so nothing else to do.
 304     }
 305   }
 306   assert(res == NULL || res->is_survivor(), "post-condition");
 307 
 308   return res;
 309 }
 310 
 311 void G1CMRootRegions::notify_scan_done() {
 312   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 313   _scan_in_progress = false;
 314   RootRegionScan_lock->notify_all();
 315 }
 316 
 317 void G1CMRootRegions::cancel_scan() {
 318   notify_scan_done();
 319 }
 320 
 321 void G1CMRootRegions::scan_finished() {
 322   assert(scan_in_progress(), "pre-condition");
 323 
 324   // Currently, only survivors can be root regions.
 325   if (!_should_abort) {
 326     assert(_next_survivor == NULL, "we should have claimed all survivors");
 327   }
 328   _next_survivor = NULL;
 329 
 330   notify_scan_done();
 331 }
 332 
 333 bool G1CMRootRegions::wait_until_scan_finished() {
 334   if (!scan_in_progress()) return false;
 335 
 336   {
 337     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 338     while (scan_in_progress()) {
 339       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 340     }
 341   }
 342   return true;
 343 }
 344 
 345 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 346   return MAX2((n_par_threads + 2) / 4, 1U);
 347 }
 348 
 349 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
 350   _g1h(g1h),
 351   _markBitMap1(),
 352   _markBitMap2(),
 353   _parallel_marking_threads(0),
 354   _max_parallel_marking_threads(0),
 355   _sleep_factor(0.0),
 356   _marking_task_overhead(1.0),
 357   _cleanup_list("Cleanup List"),
 358   _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
 359   _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
 360             CardTableModRefBS::card_shift,
 361             false /* in_resource_area*/),
 362 
 363   _prevMarkBitMap(&_markBitMap1),
 364   _nextMarkBitMap(&_markBitMap2),
 365 
 366   _markStack(this),
 367   // _finger set in set_non_marking_state
 368 
 369   _max_worker_id(ParallelGCThreads),
 370   // _active_tasks set in set_non_marking_state
 371   // _tasks set inside the constructor
 372   _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)),
 373   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 374 
 375   _has_overflown(false),
 376   _concurrent(false),
 377   _has_aborted(false),
 378   _restart_for_overflow(false),
 379   _concurrent_marking_in_progress(false),
 380   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 381   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 382 
 383   // _verbose_level set below
 384 
 385   _init_times(),
 386   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 387   _cleanup_times(),
 388   _total_counting_time(0.0),
 389   _total_rs_scrub_time(0.0),
 390 
 391   _parallel_workers(NULL),
 392 
 393   _count_card_bitmaps(NULL),
 394   _count_marked_bytes(NULL),
 395   _completed_initialization(false) {
 396 
 397   _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 398   _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
 399 
 400   // Create & start a ConcurrentMark thread.
 401   _cmThread = new ConcurrentMarkThread(this);
 402   assert(cmThread() != NULL, "CM Thread should have been created");
 403   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 404   if (_cmThread->osthread() == NULL) {
 405       vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 406   }
 407 
 408   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 409   assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
 410   assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
 411 
 412   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 413   satb_qs.set_buffer_size(G1SATBBufferSize);
 414 
 415   _root_regions.init(_g1h, this);
 416 
 417   if (ConcGCThreads > ParallelGCThreads) {
 418     log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).",
 419                     ConcGCThreads, ParallelGCThreads);
 420     return;
 421   }
 422   if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
 423     // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
 424     // if both are set
 425     _sleep_factor             = 0.0;
 426     _marking_task_overhead    = 1.0;
 427   } else if (G1MarkingOverheadPercent > 0) {
 428     // We will calculate the number of parallel marking threads based
 429     // on a target overhead with respect to the soft real-time goal
 430     double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 431     double overall_cm_overhead =
 432       (double) MaxGCPauseMillis * marking_overhead /
 433       (double) GCPauseIntervalMillis;
 434     double cpu_ratio = 1.0 / (double) os::processor_count();
 435     double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 436     double marking_task_overhead =
 437       overall_cm_overhead / marking_thread_num *
 438                                               (double) os::processor_count();
 439     double sleep_factor =
 440                        (1.0 - marking_task_overhead) / marking_task_overhead;
 441 
 442     FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num);
 443     _sleep_factor             = sleep_factor;
 444     _marking_task_overhead    = marking_task_overhead;
 445   } else {
 446     // Calculate the number of parallel marking threads by scaling
 447     // the number of parallel GC threads.
 448     uint marking_thread_num = scale_parallel_threads(ParallelGCThreads);
 449     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 450     _sleep_factor             = 0.0;
 451     _marking_task_overhead    = 1.0;
 452   }
 453 
 454   assert(ConcGCThreads > 0, "Should have been set");
 455   _parallel_marking_threads = ConcGCThreads;
 456   _max_parallel_marking_threads = _parallel_marking_threads;
 457 
 458   _parallel_workers = new WorkGang("G1 Marker",
 459        _max_parallel_marking_threads, false, true);
 460   if (_parallel_workers == NULL) {
 461     vm_exit_during_initialization("Failed necessary allocation.");
 462   } else {
 463     _parallel_workers->initialize_workers();
 464   }
 465 
 466   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 467     size_t mark_stack_size =
 468       MIN2(MarkStackSizeMax,
 469           MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 470     // Verify that the calculated value for MarkStackSize is in range.
 471     // It would be nice to use the private utility routine from Arguments.
 472     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 473       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 474                       "must be between 1 and " SIZE_FORMAT,
 475                       mark_stack_size, MarkStackSizeMax);
 476       return;
 477     }
 478     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 479   } else {
 480     // Verify MarkStackSize is in range.
 481     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 482       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 483         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 484           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 485                           "must be between 1 and " SIZE_FORMAT,
 486                           MarkStackSize, MarkStackSizeMax);
 487           return;
 488         }
 489       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 490         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 491           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 492                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 493                           MarkStackSize, MarkStackSizeMax);
 494           return;
 495         }
 496       }
 497     }
 498   }
 499 
 500   if (!_markStack.allocate(MarkStackSize)) {
 501     log_warning(gc)("Failed to allocate CM marking stack");
 502     return;
 503   }
 504 
 505   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC);
 506   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 507 
 508   _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap,  _max_worker_id, mtGC);
 509   _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
 510 
 511   BitMap::idx_t card_bm_size = _card_bm.size();
 512 
 513   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 514   _active_tasks = _max_worker_id;
 515 
 516   uint max_regions = _g1h->max_regions();
 517   for (uint i = 0; i < _max_worker_id; ++i) {
 518     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 519     task_queue->initialize();
 520     _task_queues->register_queue(i, task_queue);
 521 
 522     _count_card_bitmaps[i] = BitMap(card_bm_size, false);
 523     _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
 524 
 525     _tasks[i] = new G1CMTask(i, this,
 526                              _count_marked_bytes[i],
 527                              &_count_card_bitmaps[i],
 528                              task_queue, _task_queues);
 529 
 530     _accum_task_vtime[i] = 0.0;
 531   }
 532 
 533   // Calculate the card number for the bottom of the heap. Used
 534   // in biasing indexes into the accounting card bitmaps.
 535   _heap_bottom_card_num =
 536     intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
 537                                 CardTableModRefBS::card_shift);
 538 
 539   // Clear all the liveness counting data
 540   clear_all_count_data();
 541 
 542   // so that the call below can read a sensible value
 543   _heap_start = g1h->reserved_region().start();
 544   set_non_marking_state();
 545   _completed_initialization = true;
 546 }
 547 
 548 void G1ConcurrentMark::reset() {
 549   // Starting values for these two. This should be called in a STW
 550   // phase.
 551   MemRegion reserved = _g1h->g1_reserved();
 552   _heap_start = reserved.start();
 553   _heap_end   = reserved.end();
 554 
 555   // Separated the asserts so that we know which one fires.
 556   assert(_heap_start != NULL, "heap bounds should look ok");
 557   assert(_heap_end != NULL, "heap bounds should look ok");
 558   assert(_heap_start < _heap_end, "heap bounds should look ok");
 559 
 560   // Reset all the marking data structures and any necessary flags
 561   reset_marking_state();
 562 
 563   // We do reset all of them, since different phases will use
 564   // different number of active threads. So, it's easiest to have all
 565   // of them ready.
 566   for (uint i = 0; i < _max_worker_id; ++i) {
 567     _tasks[i]->reset(_nextMarkBitMap);
 568   }
 569 
 570   // we need this to make sure that the flag is on during the evac
 571   // pause with initial mark piggy-backed
 572   set_concurrent_marking_in_progress();
 573 }
 574 
 575 
 576 void G1ConcurrentMark::reset_marking_state(bool clear_overflow) {
 577   _markStack.set_should_expand();
 578   _markStack.setEmpty();        // Also clears the _markStack overflow flag
 579   if (clear_overflow) {
 580     clear_has_overflown();
 581   } else {
 582     assert(has_overflown(), "pre-condition");
 583   }
 584   _finger = _heap_start;
 585 
 586   for (uint i = 0; i < _max_worker_id; ++i) {
 587     G1CMTaskQueue* queue = _task_queues->queue(i);
 588     queue->set_empty();
 589   }
 590 }
 591 
 592 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 593   assert(active_tasks <= _max_worker_id, "we should not have more");
 594 
 595   _active_tasks = active_tasks;
 596   // Need to update the three data structures below according to the
 597   // number of active threads for this phase.
 598   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 599   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 600   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 601 }
 602 
 603 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 604   set_concurrency(active_tasks);
 605 
 606   _concurrent = concurrent;
 607   // We propagate this to all tasks, not just the active ones.
 608   for (uint i = 0; i < _max_worker_id; ++i)
 609     _tasks[i]->set_concurrent(concurrent);
 610 
 611   if (concurrent) {
 612     set_concurrent_marking_in_progress();
 613   } else {
 614     // We currently assume that the concurrent flag has been set to
 615     // false before we start remark. At this point we should also be
 616     // in a STW phase.
 617     assert(!concurrent_marking_in_progress(), "invariant");
 618     assert(out_of_regions(),
 619            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 620            p2i(_finger), p2i(_heap_end));
 621   }
 622 }
 623 
 624 void G1ConcurrentMark::set_non_marking_state() {
 625   // We set the global marking state to some default values when we're
 626   // not doing marking.
 627   reset_marking_state();
 628   _active_tasks = 0;
 629   clear_concurrent_marking_in_progress();
 630 }
 631 
 632 G1ConcurrentMark::~G1ConcurrentMark() {
 633   // The G1ConcurrentMark instance is never freed.
 634   ShouldNotReachHere();
 635 }
 636 
 637 class G1ClearBitMapTask : public AbstractGangTask {
 638   // Heap region closure used for clearing the given mark bitmap.
 639   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 640   private:
 641     G1CMBitMap* _bitmap;
 642     G1ConcurrentMark* _cm;
 643   public:
 644     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
 645     }
 646 
 647     virtual bool doHeapRegion(HeapRegion* r) {
 648       size_t const chunk_size_in_words = M / HeapWordSize;
 649 
 650       HeapWord* cur = r->bottom();
 651       HeapWord* const end = r->end();
 652 
 653       while (cur < end) {
 654         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 655         _bitmap->clear_range(mr);
 656 
 657         cur += chunk_size_in_words;
 658 
 659         // Abort iteration if after yielding the marking has been aborted.
 660         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 661           return true;
 662         }
 663         // Repeat the asserts from before the start of the closure. We will do them
 664         // as asserts here to minimize their overhead on the product. However, we
 665         // will have them as guarantees at the beginning / end of the bitmap
 666         // clearing to get some checking in the product.
 667         assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant");
 668         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
 669       }
 670       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 671 
 672       return false;
 673     }
 674   };
 675 
 676   G1ClearBitmapHRClosure _cl;
 677   HeapRegionClaimer _hr_claimer;
 678   bool _suspendible; // If the task is suspendible, workers must join the STS.
 679 
 680 public:
 681   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 682     AbstractGangTask("Parallel Clear Bitmap Task"),
 683     _cl(bitmap, suspendible ? cm : NULL),
 684     _hr_claimer(n_workers),
 685     _suspendible(suspendible)
 686   { }
 687 
 688   void work(uint worker_id) {
 689     SuspendibleThreadSetJoiner sts_join(_suspendible);
 690     G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer, true);
 691   }
 692 
 693   bool is_complete() {
 694     return _cl.complete();
 695   }
 696 };
 697 
 698 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 699   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 700 
 701   G1ClearBitMapTask task(bitmap, this, workers->active_workers(), may_yield);
 702   workers->run_task(&task);
 703   guarantee(!may_yield || task.is_complete(), "Must have completed iteration when not yielding.");
 704 }
 705 
 706 void G1ConcurrentMark::cleanup_for_next_mark() {
 707   // Make sure that the concurrent mark thread looks to still be in
 708   // the current cycle.
 709   guarantee(cmThread()->during_cycle(), "invariant");
 710 
 711   // We are finishing up the current cycle by clearing the next
 712   // marking bitmap and getting it ready for the next cycle. During
 713   // this time no other cycle can start. So, let's make sure that this
 714   // is the case.
 715   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 716 
 717   clear_bitmap(_nextMarkBitMap, _parallel_workers, true);
 718 
 719   // Clear the liveness counting data. If the marking has been aborted, the abort()
 720   // call already did that.
 721   if (!has_aborted()) {
 722     clear_all_count_data();
 723   }
 724 
 725   // Repeat the asserts from above.
 726   guarantee(cmThread()->during_cycle(), "invariant");
 727   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 728 }
 729 
 730 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 731   assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint.");
 732   clear_bitmap((G1CMBitMap*)_prevMarkBitMap, workers, false);
 733 }
 734 
 735 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 736   G1CMBitMap* _bitmap;
 737   bool _error;
 738  public:
 739   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 740   }
 741 
 742   virtual bool doHeapRegion(HeapRegion* r) {
 743     // This closure can be called concurrently to the mutator, so we must make sure
 744     // that the result of the getNextMarkedWordAddress() call is compared to the
 745     // value passed to it as limit to detect any found bits.
 746     // end never changes in G1.
 747     HeapWord* end = r->end();
 748     return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
 749   }
 750 };
 751 
 752 bool G1ConcurrentMark::nextMarkBitmapIsClear() {
 753   CheckBitmapClearHRClosure cl(_nextMarkBitMap);
 754   _g1h->heap_region_iterate(&cl);
 755   return cl.complete();
 756 }
 757 
 758 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 759 public:
 760   bool doHeapRegion(HeapRegion* r) {
 761     r->note_start_of_marking();
 762     return false;
 763   }
 764 };
 765 
 766 void G1ConcurrentMark::checkpointRootsInitialPre() {
 767   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 768   G1CollectorPolicy* g1p = g1h->g1_policy();
 769 
 770   _has_aborted = false;
 771 
 772   // Initialize marking structures. This has to be done in a STW phase.
 773   reset();
 774 
 775   // For each region note start of marking.
 776   NoteStartOfMarkHRClosure startcl;
 777   g1h->heap_region_iterate(&startcl);
 778 }
 779 
 780 
 781 void G1ConcurrentMark::checkpointRootsInitialPost() {
 782   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 783 
 784   // Start Concurrent Marking weak-reference discovery.
 785   ReferenceProcessor* rp = g1h->ref_processor_cm();
 786   // enable ("weak") refs discovery
 787   rp->enable_discovery();
 788   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 789 
 790   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 791   // This is the start of  the marking cycle, we're expected all
 792   // threads to have SATB queues with active set to false.
 793   satb_mq_set.set_active_all_threads(true, /* new active value */
 794                                      false /* expected_active */);
 795 
 796   _root_regions.prepare_for_scan();
 797 
 798   // update_g1_committed() will be called at the end of an evac pause
 799   // when marking is on. So, it's also called at the end of the
 800   // initial-mark pause to update the heap end, if the heap expands
 801   // during it. No need to call it here.
 802 }
 803 
 804 /*
 805  * Notice that in the next two methods, we actually leave the STS
 806  * during the barrier sync and join it immediately afterwards. If we
 807  * do not do this, the following deadlock can occur: one thread could
 808  * be in the barrier sync code, waiting for the other thread to also
 809  * sync up, whereas another one could be trying to yield, while also
 810  * waiting for the other threads to sync up too.
 811  *
 812  * Note, however, that this code is also used during remark and in
 813  * this case we should not attempt to leave / enter the STS, otherwise
 814  * we'll either hit an assert (debug / fastdebug) or deadlock
 815  * (product). So we should only leave / enter the STS if we are
 816  * operating concurrently.
 817  *
 818  * Because the thread that does the sync barrier has left the STS, it
 819  * is possible to be suspended for a Full GC or an evacuation pause
 820  * could occur. This is actually safe, since the entering the sync
 821  * barrier is one of the last things do_marking_step() does, and it
 822  * doesn't manipulate any data structures afterwards.
 823  */
 824 
 825 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 826   bool barrier_aborted;
 827   {
 828     SuspendibleThreadSetLeaver sts_leave(concurrent());
 829     barrier_aborted = !_first_overflow_barrier_sync.enter();
 830   }
 831 
 832   // at this point everyone should have synced up and not be doing any
 833   // more work
 834 
 835   if (barrier_aborted) {
 836     // If the barrier aborted we ignore the overflow condition and
 837     // just abort the whole marking phase as quickly as possible.
 838     return;
 839   }
 840 
 841   // If we're executing the concurrent phase of marking, reset the marking
 842   // state; otherwise the marking state is reset after reference processing,
 843   // during the remark pause.
 844   // If we reset here as a result of an overflow during the remark we will
 845   // see assertion failures from any subsequent set_concurrency_and_phase()
 846   // calls.
 847   if (concurrent()) {
 848     // let the task associated with with worker 0 do this
 849     if (worker_id == 0) {
 850       // task 0 is responsible for clearing the global data structures
 851       // We should be here because of an overflow. During STW we should
 852       // not clear the overflow flag since we rely on it being true when
 853       // we exit this method to abort the pause and restart concurrent
 854       // marking.
 855       reset_marking_state(true /* clear_overflow */);
 856 
 857       log_info(gc, marking)("Concurrent Mark reset for overflow");
 858     }
 859   }
 860 
 861   // after this, each task should reset its own data structures then
 862   // then go into the second barrier
 863 }
 864 
 865 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 866   SuspendibleThreadSetLeaver sts_leave(concurrent());
 867   _second_overflow_barrier_sync.enter();
 868 
 869   // at this point everything should be re-initialized and ready to go
 870 }
 871 
 872 class G1CMConcurrentMarkingTask: public AbstractGangTask {
 873 private:
 874   G1ConcurrentMark*     _cm;
 875   ConcurrentMarkThread* _cmt;
 876 
 877 public:
 878   void work(uint worker_id) {
 879     assert(Thread::current()->is_ConcurrentGC_thread(),
 880            "this should only be done by a conc GC thread");
 881     ResourceMark rm;
 882 
 883     double start_vtime = os::elapsedVTime();
 884 
 885     {
 886       SuspendibleThreadSetJoiner sts_join;
 887 
 888       assert(worker_id < _cm->active_tasks(), "invariant");
 889       G1CMTask* the_task = _cm->task(worker_id);
 890       the_task->record_start_time();
 891       if (!_cm->has_aborted()) {
 892         do {
 893           double start_vtime_sec = os::elapsedVTime();
 894           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
 895 
 896           the_task->do_marking_step(mark_step_duration_ms,
 897                                     true  /* do_termination */,
 898                                     false /* is_serial*/);
 899 
 900           double end_vtime_sec = os::elapsedVTime();
 901           double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
 902           _cm->clear_has_overflown();
 903 
 904           _cm->do_yield_check(worker_id);
 905 
 906           jlong sleep_time_ms;
 907           if (!_cm->has_aborted() && the_task->has_aborted()) {
 908             sleep_time_ms =
 909               (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
 910             {
 911               SuspendibleThreadSetLeaver sts_leave;
 912               os::sleep(Thread::current(), sleep_time_ms, false);
 913             }
 914           }
 915         } while (!_cm->has_aborted() && the_task->has_aborted());
 916       }
 917       the_task->record_end_time();
 918       guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
 919     }
 920 
 921     double end_vtime = os::elapsedVTime();
 922     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 923   }
 924 
 925   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm,
 926                             ConcurrentMarkThread* cmt) :
 927       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
 928 
 929   ~G1CMConcurrentMarkingTask() { }
 930 };
 931 
 932 // Calculates the number of active workers for a concurrent
 933 // phase.
 934 uint G1ConcurrentMark::calc_parallel_marking_threads() {
 935   uint n_conc_workers = 0;
 936   if (!UseDynamicNumberOfGCThreads ||
 937       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 938        !ForceDynamicNumberOfGCThreads)) {
 939     n_conc_workers = max_parallel_marking_threads();
 940   } else {
 941     n_conc_workers =
 942       AdaptiveSizePolicy::calc_default_active_workers(
 943                                    max_parallel_marking_threads(),
 944                                    1, /* Minimum workers */
 945                                    parallel_marking_threads(),
 946                                    Threads::number_of_non_daemon_threads());
 947     // Don't scale down "n_conc_workers" by scale_parallel_threads() because
 948     // that scaling has already gone into "_max_parallel_marking_threads".
 949   }
 950   assert(n_conc_workers > 0, "Always need at least 1");
 951   return n_conc_workers;
 952 }
 953 
 954 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
 955   // Currently, only survivors can be root regions.
 956   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 957   G1RootRegionScanClosure cl(_g1h, this, worker_id);
 958 
 959   const uintx interval = PrefetchScanIntervalInBytes;
 960   HeapWord* curr = hr->bottom();
 961   const HeapWord* end = hr->top();
 962   while (curr < end) {
 963     Prefetch::read(curr, interval);
 964     oop obj = oop(curr);
 965     int size = obj->oop_iterate_size(&cl);
 966     assert(size == obj->size(), "sanity");
 967     curr += size;
 968   }
 969 }
 970 
 971 class G1CMRootRegionScanTask : public AbstractGangTask {
 972 private:
 973   G1ConcurrentMark* _cm;
 974 
 975 public:
 976   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 977     AbstractGangTask("Root Region Scan"), _cm(cm) { }
 978 
 979   void work(uint worker_id) {
 980     assert(Thread::current()->is_ConcurrentGC_thread(),
 981            "this should only be done by a conc GC thread");
 982 
 983     G1CMRootRegions* root_regions = _cm->root_regions();
 984     HeapRegion* hr = root_regions->claim_next();
 985     while (hr != NULL) {
 986       _cm->scanRootRegion(hr, worker_id);
 987       hr = root_regions->claim_next();
 988     }
 989   }
 990 };
 991 
 992 void G1ConcurrentMark::scan_root_regions() {
 993   // scan_in_progress() will have been set to true only if there was
 994   // at least one root region to scan. So, if it's false, we
 995   // should not attempt to do any further work.
 996   if (root_regions()->scan_in_progress()) {
 997     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 998 
 999     _parallel_marking_threads = calc_parallel_marking_threads();
1000     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1001            "Maximum number of marking threads exceeded");
1002     uint active_workers = MAX2(1U, parallel_marking_threads());
1003 
1004     G1CMRootRegionScanTask task(this);
1005     _parallel_workers->set_active_workers(active_workers);
1006     _parallel_workers->run_task(&task);
1007 
1008     // It's possible that has_aborted() is true here without actually
1009     // aborting the survivor scan earlier. This is OK as it's
1010     // mainly used for sanity checking.
1011     root_regions()->scan_finished();
1012   }
1013 }
1014 
1015 void G1ConcurrentMark::concurrent_cycle_start() {
1016   _gc_timer_cm->register_gc_start();
1017 
1018   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
1019 
1020   _g1h->trace_heap_before_gc(_gc_tracer_cm);
1021 }
1022 
1023 void G1ConcurrentMark::concurrent_cycle_end() {
1024   _g1h->trace_heap_after_gc(_gc_tracer_cm);
1025 
1026   if (has_aborted()) {
1027     _gc_tracer_cm->report_concurrent_mode_failure();
1028   }
1029 
1030   _gc_timer_cm->register_gc_end();
1031 
1032   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1033 }
1034 
1035 void G1ConcurrentMark::mark_from_roots() {
1036   // we might be tempted to assert that:
1037   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1038   //        "inconsistent argument?");
1039   // However that wouldn't be right, because it's possible that
1040   // a safepoint is indeed in progress as a younger generation
1041   // stop-the-world GC happens even as we mark in this generation.
1042 
1043   _restart_for_overflow = false;
1044 
1045   // _g1h has _n_par_threads
1046   _parallel_marking_threads = calc_parallel_marking_threads();
1047   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1048     "Maximum number of marking threads exceeded");
1049 
1050   uint active_workers = MAX2(1U, parallel_marking_threads());
1051   assert(active_workers > 0, "Should have been set");
1052 
1053   // Parallel task terminator is set in "set_concurrency_and_phase()"
1054   set_concurrency_and_phase(active_workers, true /* concurrent */);
1055 
1056   G1CMConcurrentMarkingTask markingTask(this, cmThread());
1057   _parallel_workers->set_active_workers(active_workers);
1058   _parallel_workers->run_task(&markingTask);
1059   print_stats();
1060 }
1061 
1062 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1063   // world is stopped at this checkpoint
1064   assert(SafepointSynchronize::is_at_safepoint(),
1065          "world should be stopped");
1066 
1067   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1068 
1069   // If a full collection has happened, we shouldn't do this.
1070   if (has_aborted()) {
1071     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1072     return;
1073   }
1074 
1075   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1076 
1077   if (VerifyDuringGC) {
1078     HandleMark hm;  // handle scope
1079     g1h->prepare_for_verify();
1080     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1081   }
1082   g1h->verifier()->check_bitmaps("Remark Start");
1083 
1084   G1CollectorPolicy* g1p = g1h->g1_policy();
1085   g1p->record_concurrent_mark_remark_start();
1086 
1087   double start = os::elapsedTime();
1088 
1089   checkpointRootsFinalWork();
1090 
1091   double mark_work_end = os::elapsedTime();
1092 
1093   weakRefsWork(clear_all_soft_refs);
1094 
1095   if (has_overflown()) {
1096     // Oops.  We overflowed.  Restart concurrent marking.
1097     _restart_for_overflow = true;
1098 
1099     // Verify the heap w.r.t. the previous marking bitmap.
1100     if (VerifyDuringGC) {
1101       HandleMark hm;  // handle scope
1102       g1h->prepare_for_verify();
1103       Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)");
1104     }
1105 
1106     // Clear the marking state because we will be restarting
1107     // marking due to overflowing the global mark stack.
1108     reset_marking_state();
1109   } else {
1110     {
1111       GCTraceTime(Debug, gc, phases) trace("Aggregate Data", _gc_timer_cm);
1112 
1113       // Aggregate the per-task counting data that we have accumulated
1114       // while marking.
1115       aggregate_count_data();
1116     }
1117 
1118     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1119     // We're done with marking.
1120     // This is the end of  the marking cycle, we're expected all
1121     // threads to have SATB queues with active set to true.
1122     satb_mq_set.set_active_all_threads(false, /* new active value */
1123                                        true /* expected_active */);
1124 
1125     if (VerifyDuringGC) {
1126       HandleMark hm;  // handle scope
1127       g1h->prepare_for_verify();
1128       Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)");
1129     }
1130     g1h->verifier()->check_bitmaps("Remark End");
1131     assert(!restart_for_overflow(), "sanity");
1132     // Completely reset the marking state since marking completed
1133     set_non_marking_state();
1134   }
1135 
1136   // Expand the marking stack, if we have to and if we can.
1137   if (_markStack.should_expand()) {
1138     _markStack.expand();
1139   }
1140 
1141   // Statistics
1142   double now = os::elapsedTime();
1143   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1144   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1145   _remark_times.add((now - start) * 1000.0);
1146 
1147   g1p->record_concurrent_mark_remark_end();
1148 
1149   G1CMIsAliveClosure is_alive(g1h);
1150   _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1151 }
1152 
1153 // Base class of the closures that finalize and verify the
1154 // liveness counting data.
1155 class G1CMCountDataClosureBase: public HeapRegionClosure {
1156 protected:
1157   G1CollectedHeap* _g1h;
1158   G1ConcurrentMark* _cm;
1159   CardTableModRefBS* _ct_bs;
1160 
1161   BitMap* _region_bm;
1162   BitMap* _card_bm;
1163 
1164   // Takes a region that's not empty (i.e., it has at least one
1165   // live object in it and sets its corresponding bit on the region
1166   // bitmap to 1.
1167   void set_bit_for_region(HeapRegion* hr) {
1168     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1169     _region_bm->par_at_put(index, true);
1170   }
1171 
1172 public:
1173   G1CMCountDataClosureBase(G1CollectedHeap* g1h,
1174                            BitMap* region_bm, BitMap* card_bm):
1175     _g1h(g1h), _cm(g1h->concurrent_mark()),
1176     _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
1177     _region_bm(region_bm), _card_bm(card_bm) { }
1178 };
1179 
1180 // Closure that calculates the # live objects per region. Used
1181 // for verification purposes during the cleanup pause.
1182 class CalcLiveObjectsClosure: public G1CMCountDataClosureBase {
1183   G1CMBitMapRO* _bm;
1184   size_t _region_marked_bytes;
1185 
1186 public:
1187   CalcLiveObjectsClosure(G1CMBitMapRO *bm, G1CollectedHeap* g1h,
1188                          BitMap* region_bm, BitMap* card_bm) :
1189     G1CMCountDataClosureBase(g1h, region_bm, card_bm),
1190     _bm(bm), _region_marked_bytes(0) { }
1191 
1192   bool doHeapRegion(HeapRegion* hr) {
1193     HeapWord* ntams = hr->next_top_at_mark_start();
1194     HeapWord* start = hr->bottom();
1195 
1196     assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1197            "Preconditions not met - "
1198            "start: " PTR_FORMAT ", ntams: " PTR_FORMAT ", end: " PTR_FORMAT,
1199            p2i(start), p2i(ntams), p2i(hr->end()));
1200 
1201     // Find the first marked object at or after "start".
1202     start = _bm->getNextMarkedWordAddress(start, ntams);
1203 
1204     size_t marked_bytes = 0;
1205 
1206     while (start < ntams) {
1207       oop obj = oop(start);
1208       int obj_sz = obj->size();
1209       HeapWord* obj_end = start + obj_sz;
1210 
1211       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1212       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1213 
1214       // Note: if we're looking at the last region in heap - obj_end
1215       // could be actually just beyond the end of the heap; end_idx
1216       // will then correspond to a (non-existent) card that is also
1217       // just beyond the heap.
1218       if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1219         // end of object is not card aligned - increment to cover
1220         // all the cards spanned by the object
1221         end_idx += 1;
1222       }
1223 
1224       // Set the bits in the card BM for the cards spanned by this object.
1225       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1226 
1227       // Add the size of this object to the number of marked bytes.
1228       marked_bytes += (size_t)obj_sz * HeapWordSize;
1229 
1230       // This will happen if we are handling a humongous object that spans
1231       // several heap regions.
1232       if (obj_end > hr->end()) {
1233         break;
1234       }
1235       // Find the next marked object after this one.
1236       start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1237     }
1238 
1239     // Mark the allocated-since-marking portion...
1240     HeapWord* top = hr->top();
1241     if (ntams < top) {
1242       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1243       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1244 
1245       // Note: if we're looking at the last region in heap - top
1246       // could be actually just beyond the end of the heap; end_idx
1247       // will then correspond to a (non-existent) card that is also
1248       // just beyond the heap.
1249       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1250         // end of object is not card aligned - increment to cover
1251         // all the cards spanned by the object
1252         end_idx += 1;
1253       }
1254       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1255 
1256       // This definitely means the region has live objects.
1257       set_bit_for_region(hr);
1258     }
1259 
1260     // Update the live region bitmap.
1261     if (marked_bytes > 0) {
1262       set_bit_for_region(hr);
1263     }
1264 
1265     // Set the marked bytes for the current region so that
1266     // it can be queried by a calling verification routine
1267     _region_marked_bytes = marked_bytes;
1268 
1269     return false;
1270   }
1271 
1272   size_t region_marked_bytes() const { return _region_marked_bytes; }
1273 };
1274 
1275 // Heap region closure used for verifying the counting data
1276 // that was accumulated concurrently and aggregated during
1277 // the remark pause. This closure is applied to the heap
1278 // regions during the STW cleanup pause.
1279 
1280 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1281   G1CollectedHeap* _g1h;
1282   G1ConcurrentMark* _cm;
1283   CalcLiveObjectsClosure _calc_cl;
1284   BitMap* _region_bm;   // Region BM to be verified
1285   BitMap* _card_bm;     // Card BM to be verified
1286 
1287   BitMap* _exp_region_bm; // Expected Region BM values
1288   BitMap* _exp_card_bm;   // Expected card BM values
1289 
1290   int _failures;
1291 
1292 public:
1293   VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1294                                 BitMap* region_bm,
1295                                 BitMap* card_bm,
1296                                 BitMap* exp_region_bm,
1297                                 BitMap* exp_card_bm) :
1298     _g1h(g1h), _cm(g1h->concurrent_mark()),
1299     _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1300     _region_bm(region_bm), _card_bm(card_bm),
1301     _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1302     _failures(0) { }
1303 
1304   int failures() const { return _failures; }
1305 
1306   bool doHeapRegion(HeapRegion* hr) {
1307     int failures = 0;
1308 
1309     // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1310     // this region and set the corresponding bits in the expected region
1311     // and card bitmaps.
1312     bool res = _calc_cl.doHeapRegion(hr);
1313     assert(res == false, "should be continuing");
1314 
1315     // Verify the marked bytes for this region.
1316     size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1317     size_t act_marked_bytes = hr->next_marked_bytes();
1318 
1319     if (exp_marked_bytes > act_marked_bytes) {
1320       if (hr->is_starts_humongous()) {
1321         // For start_humongous regions, the size of the whole object will be
1322         // in exp_marked_bytes.
1323         HeapRegion* region = hr;
1324         int num_regions;
1325         for (num_regions = 0; region != NULL; num_regions++) {
1326           region = _g1h->next_region_in_humongous(region);
1327         }
1328         if ((num_regions-1) * HeapRegion::GrainBytes >= exp_marked_bytes) {
1329           failures += 1;
1330         } else if (num_regions * HeapRegion::GrainBytes < exp_marked_bytes) {
1331           failures += 1;
1332         }
1333       } else {
1334         // We're not OK if expected marked bytes > actual marked bytes. It means
1335         // we have missed accounting some objects during the actual marking.
1336         failures += 1;
1337       }
1338     }
1339 
1340     // Verify the bit, for this region, in the actual and expected
1341     // (which was just calculated) region bit maps.
1342     // We're not OK if the bit in the calculated expected region
1343     // bitmap is set and the bit in the actual region bitmap is not.
1344     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1345 
1346     bool expected = _exp_region_bm->at(index);
1347     bool actual = _region_bm->at(index);
1348     if (expected && !actual) {
1349       failures += 1;
1350     }
1351 
1352     // Verify that the card bit maps for the cards spanned by the current
1353     // region match. We have an error if we have a set bit in the expected
1354     // bit map and the corresponding bit in the actual bitmap is not set.
1355 
1356     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1357     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1358 
1359     for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1360       expected = _exp_card_bm->at(i);
1361       actual = _card_bm->at(i);
1362 
1363       if (expected && !actual) {
1364         failures += 1;
1365       }
1366     }
1367 
1368     _failures += failures;
1369 
1370     // We could stop iteration over the heap when we
1371     // find the first violating region by returning true.
1372     return false;
1373   }
1374 };
1375 
1376 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1377 protected:
1378   G1CollectedHeap* _g1h;
1379   G1ConcurrentMark* _cm;
1380   BitMap* _actual_region_bm;
1381   BitMap* _actual_card_bm;
1382 
1383   uint    _n_workers;
1384 
1385   BitMap* _expected_region_bm;
1386   BitMap* _expected_card_bm;
1387 
1388   int  _failures;
1389 
1390   HeapRegionClaimer _hrclaimer;
1391 
1392 public:
1393   G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1394                             BitMap* region_bm, BitMap* card_bm,
1395                             BitMap* expected_region_bm, BitMap* expected_card_bm)
1396     : AbstractGangTask("G1 verify final counting"),
1397       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1398       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1399       _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1400       _failures(0),
1401       _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1402     assert(VerifyDuringGC, "don't call this otherwise");
1403     assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1404     assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1405   }
1406 
1407   void work(uint worker_id) {
1408     assert(worker_id < _n_workers, "invariant");
1409 
1410     VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1411                                             _actual_region_bm, _actual_card_bm,
1412                                             _expected_region_bm,
1413                                             _expected_card_bm);
1414 
1415     _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer);
1416 
1417     Atomic::add(verify_cl.failures(), &_failures);
1418   }
1419 
1420   int failures() const { return _failures; }
1421 };
1422 
1423 // Closure that finalizes the liveness counting data.
1424 // Used during the cleanup pause.
1425 // Sets the bits corresponding to the interval [NTAMS, top]
1426 // (which contains the implicitly live objects) in the
1427 // card liveness bitmap. Also sets the bit for each region,
1428 // containing live data, in the region liveness bitmap.
1429 
1430 class FinalCountDataUpdateClosure: public G1CMCountDataClosureBase {
1431  public:
1432   FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1433                               BitMap* region_bm,
1434                               BitMap* card_bm) :
1435     G1CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1436 
1437   bool doHeapRegion(HeapRegion* hr) {
1438     HeapWord* ntams = hr->next_top_at_mark_start();
1439     HeapWord* top   = hr->top();
1440 
1441     assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1442 
1443     // Mark the allocated-since-marking portion...
1444     if (ntams < top) {
1445       // This definitely means the region has live objects.
1446       set_bit_for_region(hr);
1447 
1448       // Now set the bits in the card bitmap for [ntams, top)
1449       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1450       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1451 
1452       // Note: if we're looking at the last region in heap - top
1453       // could be actually just beyond the end of the heap; end_idx
1454       // will then correspond to a (non-existent) card that is also
1455       // just beyond the heap.
1456       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1457         // end of object is not card aligned - increment to cover
1458         // all the cards spanned by the object
1459         end_idx += 1;
1460       }
1461 
1462       assert(end_idx <= _card_bm->size(),
1463              "oob: end_idx=  " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT,
1464              end_idx, _card_bm->size());
1465       assert(start_idx < _card_bm->size(),
1466              "oob: start_idx=  " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT,
1467              start_idx, _card_bm->size());
1468 
1469       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1470     }
1471 
1472     // Set the bit for the region if it contains live data
1473     if (hr->next_marked_bytes() > 0) {
1474       set_bit_for_region(hr);
1475     }
1476 
1477     return false;
1478   }
1479 };
1480 
1481 class G1ParFinalCountTask: public AbstractGangTask {
1482 protected:
1483   G1CollectedHeap* _g1h;
1484   G1ConcurrentMark* _cm;
1485   BitMap* _actual_region_bm;
1486   BitMap* _actual_card_bm;
1487 
1488   uint    _n_workers;
1489   HeapRegionClaimer _hrclaimer;
1490 
1491 public:
1492   G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1493     : AbstractGangTask("G1 final counting"),
1494       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1495       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1496       _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1497   }
1498 
1499   void work(uint worker_id) {
1500     assert(worker_id < _n_workers, "invariant");
1501 
1502     FinalCountDataUpdateClosure final_update_cl(_g1h,
1503                                                 _actual_region_bm,
1504                                                 _actual_card_bm);
1505 
1506     _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer);
1507   }
1508 };
1509 
1510 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1511   G1CollectedHeap* _g1;
1512   size_t _freed_bytes;
1513   FreeRegionList* _local_cleanup_list;
1514   uint _old_regions_removed;
1515   uint _humongous_regions_removed;
1516   HRRSCleanupTask* _hrrs_cleanup_task;
1517 
1518 public:
1519   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1520                              FreeRegionList* local_cleanup_list,
1521                              HRRSCleanupTask* hrrs_cleanup_task) :
1522     _g1(g1),
1523     _freed_bytes(0),
1524     _local_cleanup_list(local_cleanup_list),
1525     _old_regions_removed(0),
1526     _humongous_regions_removed(0),
1527     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1528 
1529   size_t freed_bytes() { return _freed_bytes; }
1530   const uint old_regions_removed() { return _old_regions_removed; }
1531   const uint humongous_regions_removed() { return _humongous_regions_removed; }
1532 
1533   bool doHeapRegion(HeapRegion *hr) {
1534     if (hr->is_archive()) {
1535       return false;
1536     }
1537     // We use a claim value of zero here because all regions
1538     // were claimed with value 1 in the FinalCount task.
1539     _g1->reset_gc_time_stamps(hr);
1540     hr->note_end_of_marking();
1541 
1542     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1543       _freed_bytes += hr->used();
1544       hr->set_containing_set(NULL);
1545       if (hr->is_humongous()) {
1546         _humongous_regions_removed++;
1547         _g1->free_humongous_region(hr, _local_cleanup_list, true);
1548       } else {
1549         _old_regions_removed++;
1550         _g1->free_region(hr, _local_cleanup_list, true);
1551       }
1552     } else {
1553       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1554     }
1555 
1556     return false;
1557   }
1558 };
1559 
1560 class G1ParNoteEndTask: public AbstractGangTask {
1561   friend class G1NoteEndOfConcMarkClosure;
1562 
1563 protected:
1564   G1CollectedHeap* _g1h;
1565   FreeRegionList* _cleanup_list;
1566   HeapRegionClaimer _hrclaimer;
1567 
1568 public:
1569   G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1570       AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1571   }
1572 
1573   void work(uint worker_id) {
1574     FreeRegionList local_cleanup_list("Local Cleanup List");
1575     HRRSCleanupTask hrrs_cleanup_task;
1576     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1577                                            &hrrs_cleanup_task);
1578     _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1579     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1580 
1581     // Now update the lists
1582     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1583     {
1584       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1585       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1586 
1587       // If we iterate over the global cleanup list at the end of
1588       // cleanup to do this printing we will not guarantee to only
1589       // generate output for the newly-reclaimed regions (the list
1590       // might not be empty at the beginning of cleanup; we might
1591       // still be working on its previous contents). So we do the
1592       // printing here, before we append the new regions to the global
1593       // cleanup list.
1594 
1595       G1HRPrinter* hr_printer = _g1h->hr_printer();
1596       if (hr_printer->is_active()) {
1597         FreeRegionListIterator iter(&local_cleanup_list);
1598         while (iter.more_available()) {
1599           HeapRegion* hr = iter.get_next();
1600           hr_printer->cleanup(hr);
1601         }
1602       }
1603 
1604       _cleanup_list->add_ordered(&local_cleanup_list);
1605       assert(local_cleanup_list.is_empty(), "post-condition");
1606 
1607       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1608     }
1609   }
1610 };
1611 
1612 void G1ConcurrentMark::cleanup() {
1613   // world is stopped at this checkpoint
1614   assert(SafepointSynchronize::is_at_safepoint(),
1615          "world should be stopped");
1616   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1617 
1618   // If a full collection has happened, we shouldn't do this.
1619   if (has_aborted()) {
1620     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1621     return;
1622   }
1623 
1624   g1h->verifier()->verify_region_sets_optional();
1625 
1626   if (VerifyDuringGC) {
1627     HandleMark hm;  // handle scope
1628     g1h->prepare_for_verify();
1629     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1630   }
1631   g1h->verifier()->check_bitmaps("Cleanup Start");
1632 
1633   G1CollectorPolicy* g1p = g1h->g1_policy();
1634   g1p->record_concurrent_mark_cleanup_start();
1635 
1636   double start = os::elapsedTime();
1637 
1638   HeapRegionRemSet::reset_for_cleanup_tasks();
1639 
1640   // Do counting once more with the world stopped for good measure.
1641   G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1642 
1643   g1h->workers()->run_task(&g1_par_count_task);
1644 
1645   if (VerifyDuringGC) {
1646     // Verify that the counting data accumulated during marking matches
1647     // that calculated by walking the marking bitmap.
1648 
1649     // Bitmaps to hold expected values
1650     BitMap expected_region_bm(_region_bm.size(), true);
1651     BitMap expected_card_bm(_card_bm.size(), true);
1652 
1653     G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
1654                                                  &_region_bm,
1655                                                  &_card_bm,
1656                                                  &expected_region_bm,
1657                                                  &expected_card_bm);
1658 
1659     g1h->workers()->run_task(&g1_par_verify_task);
1660 
1661     guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
1662   }
1663 
1664   size_t start_used_bytes = g1h->used();
1665   g1h->collector_state()->set_mark_in_progress(false);
1666 
1667   double count_end = os::elapsedTime();
1668   double this_final_counting_time = (count_end - start);
1669   _total_counting_time += this_final_counting_time;
1670 
1671   if (log_is_enabled(Trace, gc, liveness)) {
1672     G1PrintRegionLivenessInfoClosure cl("Post-Marking");
1673     _g1h->heap_region_iterate(&cl);
1674   }
1675 
1676   // Install newly created mark bitMap as "prev".
1677   swapMarkBitMaps();
1678 
1679   g1h->reset_gc_time_stamp();
1680 
1681   uint n_workers = _g1h->workers()->active_workers();
1682 
1683   // Note end of marking in all heap regions.
1684   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1685   g1h->workers()->run_task(&g1_par_note_end_task);
1686   g1h->check_gc_time_stamps();
1687 
1688   if (!cleanup_list_is_empty()) {
1689     // The cleanup list is not empty, so we'll have to process it
1690     // concurrently. Notify anyone else that might be wanting free
1691     // regions that there will be more free regions coming soon.
1692     g1h->set_free_regions_coming();
1693   }
1694 
1695   // call below, since it affects the metric by which we sort the heap
1696   // regions.
1697   if (G1ScrubRemSets) {
1698     double rs_scrub_start = os::elapsedTime();
1699     g1h->scrub_rem_set(&_region_bm, &_card_bm);
1700     _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start);
1701   }
1702 
1703   // this will also free any regions totally full of garbage objects,
1704   // and sort the regions.
1705   g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1706 
1707   // Statistics.
1708   double end = os::elapsedTime();
1709   _cleanup_times.add((end - start) * 1000.0);
1710 
1711   // Clean up will have freed any regions completely full of garbage.
1712   // Update the soft reference policy with the new heap occupancy.
1713   Universe::update_heap_info_at_gc();
1714 
1715   if (VerifyDuringGC) {
1716     HandleMark hm;  // handle scope
1717     g1h->prepare_for_verify();
1718     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)");
1719   }
1720 
1721   g1h->verifier()->check_bitmaps("Cleanup End");
1722 
1723   g1h->verifier()->verify_region_sets_optional();
1724 
1725   // We need to make this be a "collection" so any collection pause that
1726   // races with it goes around and waits for completeCleanup to finish.
1727   g1h->increment_total_collections();
1728 
1729   // Clean out dead classes and update Metaspace sizes.
1730   if (ClassUnloadingWithConcurrentMark) {
1731     ClassLoaderDataGraph::purge();
1732   }
1733   MetaspaceGC::compute_new_size();
1734 
1735   // We reclaimed old regions so we should calculate the sizes to make
1736   // sure we update the old gen/space data.
1737   g1h->g1mm()->update_sizes();
1738   g1h->allocation_context_stats().update_after_mark();
1739 }
1740 
1741 void G1ConcurrentMark::complete_cleanup() {
1742   if (has_aborted()) return;
1743 
1744   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1745 
1746   _cleanup_list.verify_optional();
1747   FreeRegionList tmp_free_list("Tmp Free List");
1748 
1749   log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1750                                   "cleanup list has %u entries",
1751                                   _cleanup_list.length());
1752 
1753   // No one else should be accessing the _cleanup_list at this point,
1754   // so it is not necessary to take any locks
1755   while (!_cleanup_list.is_empty()) {
1756     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
1757     assert(hr != NULL, "Got NULL from a non-empty list");
1758     hr->par_clear();
1759     tmp_free_list.add_ordered(hr);
1760 
1761     // Instead of adding one region at a time to the secondary_free_list,
1762     // we accumulate them in the local list and move them a few at a
1763     // time. This also cuts down on the number of notify_all() calls
1764     // we do during this process. We'll also append the local list when
1765     // _cleanup_list is empty (which means we just removed the last
1766     // region from the _cleanup_list).
1767     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
1768         _cleanup_list.is_empty()) {
1769       log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1770                                       "appending %u entries to the secondary_free_list, "
1771                                       "cleanup list still has %u entries",
1772                                       tmp_free_list.length(),
1773                                       _cleanup_list.length());
1774 
1775       {
1776         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
1777         g1h->secondary_free_list_add(&tmp_free_list);
1778         SecondaryFreeList_lock->notify_all();
1779       }
1780 #ifndef PRODUCT
1781       if (G1StressConcRegionFreeing) {
1782         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
1783           os::sleep(Thread::current(), (jlong) 1, false);
1784         }
1785       }
1786 #endif
1787     }
1788   }
1789   assert(tmp_free_list.is_empty(), "post-condition");
1790 }
1791 
1792 // Supporting Object and Oop closures for reference discovery
1793 // and processing in during marking
1794 
1795 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1796   HeapWord* addr = (HeapWord*)obj;
1797   return addr != NULL &&
1798          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
1799 }
1800 
1801 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1802 // Uses the G1CMTask associated with a worker thread (for serial reference
1803 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1804 // trace referent objects.
1805 //
1806 // Using the G1CMTask and embedded local queues avoids having the worker
1807 // threads operating on the global mark stack. This reduces the risk
1808 // of overflowing the stack - which we would rather avoid at this late
1809 // state. Also using the tasks' local queues removes the potential
1810 // of the workers interfering with each other that could occur if
1811 // operating on the global stack.
1812 
1813 class G1CMKeepAliveAndDrainClosure: public OopClosure {
1814   G1ConcurrentMark* _cm;
1815   G1CMTask*         _task;
1816   int               _ref_counter_limit;
1817   int               _ref_counter;
1818   bool              _is_serial;
1819  public:
1820   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1821     _cm(cm), _task(task), _is_serial(is_serial),
1822     _ref_counter_limit(G1RefProcDrainInterval) {
1823     assert(_ref_counter_limit > 0, "sanity");
1824     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1825     _ref_counter = _ref_counter_limit;
1826   }
1827 
1828   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1829   virtual void do_oop(      oop* p) { do_oop_work(p); }
1830 
1831   template <class T> void do_oop_work(T* p) {
1832     if (!_cm->has_overflown()) {
1833       oop obj = oopDesc::load_decode_heap_oop(p);
1834       _task->deal_with_reference(obj);
1835       _ref_counter--;
1836 
1837       if (_ref_counter == 0) {
1838         // We have dealt with _ref_counter_limit references, pushing them
1839         // and objects reachable from them on to the local stack (and
1840         // possibly the global stack). Call G1CMTask::do_marking_step() to
1841         // process these entries.
1842         //
1843         // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1844         // there's nothing more to do (i.e. we're done with the entries that
1845         // were pushed as a result of the G1CMTask::deal_with_reference() calls
1846         // above) or we overflow.
1847         //
1848         // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1849         // flag while there may still be some work to do. (See the comment at
1850         // the beginning of G1CMTask::do_marking_step() for those conditions -
1851         // one of which is reaching the specified time target.) It is only
1852         // when G1CMTask::do_marking_step() returns without setting the
1853         // has_aborted() flag that the marking step has completed.
1854         do {
1855           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1856           _task->do_marking_step(mark_step_duration_ms,
1857                                  false      /* do_termination */,
1858                                  _is_serial);
1859         } while (_task->has_aborted() && !_cm->has_overflown());
1860         _ref_counter = _ref_counter_limit;
1861       }
1862     }
1863   }
1864 };
1865 
1866 // 'Drain' oop closure used by both serial and parallel reference processing.
1867 // Uses the G1CMTask associated with a given worker thread (for serial
1868 // reference processing the G1CMtask for worker 0 is used). Calls the
1869 // do_marking_step routine, with an unbelievably large timeout value,
1870 // to drain the marking data structures of the remaining entries
1871 // added by the 'keep alive' oop closure above.
1872 
1873 class G1CMDrainMarkingStackClosure: public VoidClosure {
1874   G1ConcurrentMark* _cm;
1875   G1CMTask*         _task;
1876   bool              _is_serial;
1877  public:
1878   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1879     _cm(cm), _task(task), _is_serial(is_serial) {
1880     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1881   }
1882 
1883   void do_void() {
1884     do {
1885       // We call G1CMTask::do_marking_step() to completely drain the local
1886       // and global marking stacks of entries pushed by the 'keep alive'
1887       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1888       //
1889       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1890       // if there's nothing more to do (i.e. we've completely drained the
1891       // entries that were pushed as a a result of applying the 'keep alive'
1892       // closure to the entries on the discovered ref lists) or we overflow
1893       // the global marking stack.
1894       //
1895       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1896       // flag while there may still be some work to do. (See the comment at
1897       // the beginning of G1CMTask::do_marking_step() for those conditions -
1898       // one of which is reaching the specified time target.) It is only
1899       // when G1CMTask::do_marking_step() returns without setting the
1900       // has_aborted() flag that the marking step has completed.
1901 
1902       _task->do_marking_step(1000000000.0 /* something very large */,
1903                              true         /* do_termination */,
1904                              _is_serial);
1905     } while (_task->has_aborted() && !_cm->has_overflown());
1906   }
1907 };
1908 
1909 // Implementation of AbstractRefProcTaskExecutor for parallel
1910 // reference processing at the end of G1 concurrent marking
1911 
1912 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
1913 private:
1914   G1CollectedHeap*  _g1h;
1915   G1ConcurrentMark* _cm;
1916   WorkGang*         _workers;
1917   uint              _active_workers;
1918 
1919 public:
1920   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1921                           G1ConcurrentMark* cm,
1922                           WorkGang* workers,
1923                           uint n_workers) :
1924     _g1h(g1h), _cm(cm),
1925     _workers(workers), _active_workers(n_workers) { }
1926 
1927   // Executes the given task using concurrent marking worker threads.
1928   virtual void execute(ProcessTask& task);
1929   virtual void execute(EnqueueTask& task);
1930 };
1931 
1932 class G1CMRefProcTaskProxy: public AbstractGangTask {
1933   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1934   ProcessTask&      _proc_task;
1935   G1CollectedHeap*  _g1h;
1936   G1ConcurrentMark* _cm;
1937 
1938 public:
1939   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1940                        G1CollectedHeap* g1h,
1941                        G1ConcurrentMark* cm) :
1942     AbstractGangTask("Process reference objects in parallel"),
1943     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1944     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1945     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1946   }
1947 
1948   virtual void work(uint worker_id) {
1949     ResourceMark rm;
1950     HandleMark hm;
1951     G1CMTask* task = _cm->task(worker_id);
1952     G1CMIsAliveClosure g1_is_alive(_g1h);
1953     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1954     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1955 
1956     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1957   }
1958 };
1959 
1960 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1961   assert(_workers != NULL, "Need parallel worker threads.");
1962   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1963 
1964   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1965 
1966   // We need to reset the concurrency level before each
1967   // proxy task execution, so that the termination protocol
1968   // and overflow handling in G1CMTask::do_marking_step() knows
1969   // how many workers to wait for.
1970   _cm->set_concurrency(_active_workers);
1971   _workers->run_task(&proc_task_proxy);
1972 }
1973 
1974 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
1975   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1976   EnqueueTask& _enq_task;
1977 
1978 public:
1979   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1980     AbstractGangTask("Enqueue reference objects in parallel"),
1981     _enq_task(enq_task) { }
1982 
1983   virtual void work(uint worker_id) {
1984     _enq_task.work(worker_id);
1985   }
1986 };
1987 
1988 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1989   assert(_workers != NULL, "Need parallel worker threads.");
1990   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1991 
1992   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1993 
1994   // Not strictly necessary but...
1995   //
1996   // We need to reset the concurrency level before each
1997   // proxy task execution, so that the termination protocol
1998   // and overflow handling in G1CMTask::do_marking_step() knows
1999   // how many workers to wait for.
2000   _cm->set_concurrency(_active_workers);
2001   _workers->run_task(&enq_task_proxy);
2002 }
2003 
2004 void G1ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2005   G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2006 }
2007 
2008 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2009   if (has_overflown()) {
2010     // Skip processing the discovered references if we have
2011     // overflown the global marking stack. Reference objects
2012     // only get discovered once so it is OK to not
2013     // de-populate the discovered reference lists. We could have,
2014     // but the only benefit would be that, when marking restarts,
2015     // less reference objects are discovered.
2016     return;
2017   }
2018 
2019   ResourceMark rm;
2020   HandleMark   hm;
2021 
2022   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2023 
2024   // Is alive closure.
2025   G1CMIsAliveClosure g1_is_alive(g1h);
2026 
2027   // Inner scope to exclude the cleaning of the string and symbol
2028   // tables from the displayed time.
2029   {
2030     GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm);
2031 
2032     ReferenceProcessor* rp = g1h->ref_processor_cm();
2033 
2034     // See the comment in G1CollectedHeap::ref_processing_init()
2035     // about how reference processing currently works in G1.
2036 
2037     // Set the soft reference policy
2038     rp->setup_policy(clear_all_soft_refs);
2039     assert(_markStack.isEmpty(), "mark stack should be empty");
2040 
2041     // Instances of the 'Keep Alive' and 'Complete GC' closures used
2042     // in serial reference processing. Note these closures are also
2043     // used for serially processing (by the the current thread) the
2044     // JNI references during parallel reference processing.
2045     //
2046     // These closures do not need to synchronize with the worker
2047     // threads involved in parallel reference processing as these
2048     // instances are executed serially by the current thread (e.g.
2049     // reference processing is not multi-threaded and is thus
2050     // performed by the current thread instead of a gang worker).
2051     //
2052     // The gang tasks involved in parallel reference processing create
2053     // their own instances of these closures, which do their own
2054     // synchronization among themselves.
2055     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2056     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2057 
2058     // We need at least one active thread. If reference processing
2059     // is not multi-threaded we use the current (VMThread) thread,
2060     // otherwise we use the work gang from the G1CollectedHeap and
2061     // we utilize all the worker threads we can.
2062     bool processing_is_mt = rp->processing_is_mt();
2063     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2064     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2065 
2066     // Parallel processing task executor.
2067     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2068                                               g1h->workers(), active_workers);
2069     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2070 
2071     // Set the concurrency level. The phase was already set prior to
2072     // executing the remark task.
2073     set_concurrency(active_workers);
2074 
2075     // Set the degree of MT processing here.  If the discovery was done MT,
2076     // the number of threads involved during discovery could differ from
2077     // the number of active workers.  This is OK as long as the discovered
2078     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2079     rp->set_active_mt_degree(active_workers);
2080 
2081     // Process the weak references.
2082     const ReferenceProcessorStats& stats =
2083         rp->process_discovered_references(&g1_is_alive,
2084                                           &g1_keep_alive,
2085                                           &g1_drain_mark_stack,
2086                                           executor,
2087                                           _gc_timer_cm);
2088     _gc_tracer_cm->report_gc_reference_stats(stats);
2089 
2090     // The do_oop work routines of the keep_alive and drain_marking_stack
2091     // oop closures will set the has_overflown flag if we overflow the
2092     // global marking stack.
2093 
2094     assert(_markStack.overflow() || _markStack.isEmpty(),
2095             "mark stack should be empty (unless it overflowed)");
2096 
2097     if (_markStack.overflow()) {
2098       // This should have been done already when we tried to push an
2099       // entry on to the global mark stack. But let's do it again.
2100       set_has_overflown();
2101     }
2102 
2103     assert(rp->num_q() == active_workers, "why not");
2104 
2105     rp->enqueue_discovered_references(executor);
2106 
2107     rp->verify_no_references_recorded();
2108     assert(!rp->discovery_enabled(), "Post condition");
2109   }
2110 
2111   if (has_overflown()) {
2112     // We can not trust g1_is_alive if the marking stack overflowed
2113     return;
2114   }
2115 
2116   assert(_markStack.isEmpty(), "Marking should have completed");
2117 
2118   // Unload Klasses, String, Symbols, Code Cache, etc.
2119   if (ClassUnloadingWithConcurrentMark) {
2120     bool purged_classes;
2121 
2122     {
2123       GCTraceTime(Debug, gc, phases) trace("System Dictionary Unloading", _gc_timer_cm);
2124       purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2125     }
2126 
2127     {
2128       GCTraceTime(Debug, gc, phases) trace("Parallel Unloading", _gc_timer_cm);
2129       weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2130     }
2131   }
2132 
2133   if (G1StringDedup::is_enabled()) {
2134     GCTraceTime(Debug, gc, phases) trace("String Deduplication Unlink", _gc_timer_cm);
2135     G1StringDedup::unlink(&g1_is_alive);
2136   }
2137 }
2138 
2139 void G1ConcurrentMark::swapMarkBitMaps() {
2140   G1CMBitMapRO* temp = _prevMarkBitMap;
2141   _prevMarkBitMap    = (G1CMBitMapRO*)_nextMarkBitMap;
2142   _nextMarkBitMap    = (G1CMBitMap*)  temp;
2143 }
2144 
2145 // Closure for marking entries in SATB buffers.
2146 class G1CMSATBBufferClosure : public SATBBufferClosure {
2147 private:
2148   G1CMTask* _task;
2149   G1CollectedHeap* _g1h;
2150 
2151   // This is very similar to G1CMTask::deal_with_reference, but with
2152   // more relaxed requirements for the argument, so this must be more
2153   // circumspect about treating the argument as an object.
2154   void do_entry(void* entry) const {
2155     _task->increment_refs_reached();
2156     HeapRegion* hr = _g1h->heap_region_containing(entry);
2157     if (entry < hr->next_top_at_mark_start()) {
2158       // Until we get here, we don't know whether entry refers to a valid
2159       // object; it could instead have been a stale reference.
2160       oop obj = static_cast<oop>(entry);
2161       assert(obj->is_oop(true /* ignore mark word */),
2162              "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj));
2163       _task->make_reference_grey(obj, hr);
2164     }
2165   }
2166 
2167 public:
2168   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
2169     : _task(task), _g1h(g1h) { }
2170 
2171   virtual void do_buffer(void** buffer, size_t size) {
2172     for (size_t i = 0; i < size; ++i) {
2173       do_entry(buffer[i]);
2174     }
2175   }
2176 };
2177 
2178 class G1RemarkThreadsClosure : public ThreadClosure {
2179   G1CMSATBBufferClosure _cm_satb_cl;
2180   G1CMOopClosure _cm_cl;
2181   MarkingCodeBlobClosure _code_cl;
2182   int _thread_parity;
2183 
2184  public:
2185   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
2186     _cm_satb_cl(task, g1h),
2187     _cm_cl(g1h, g1h->concurrent_mark(), task),
2188     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2189     _thread_parity(Threads::thread_claim_parity()) {}
2190 
2191   void do_thread(Thread* thread) {
2192     if (thread->is_Java_thread()) {
2193       if (thread->claim_oops_do(true, _thread_parity)) {
2194         JavaThread* jt = (JavaThread*)thread;
2195 
2196         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2197         // however the liveness of oops reachable from nmethods have very complex lifecycles:
2198         // * Alive if on the stack of an executing method
2199         // * Weakly reachable otherwise
2200         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2201         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2202         jt->nmethods_do(&_code_cl);
2203 
2204         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
2205       }
2206     } else if (thread->is_VM_thread()) {
2207       if (thread->claim_oops_do(true, _thread_parity)) {
2208         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
2209       }
2210     }
2211   }
2212 };
2213 
2214 class G1CMRemarkTask: public AbstractGangTask {
2215 private:
2216   G1ConcurrentMark* _cm;
2217 public:
2218   void work(uint worker_id) {
2219     // Since all available tasks are actually started, we should
2220     // only proceed if we're supposed to be active.
2221     if (worker_id < _cm->active_tasks()) {
2222       G1CMTask* task = _cm->task(worker_id);
2223       task->record_start_time();
2224       {
2225         ResourceMark rm;
2226         HandleMark hm;
2227 
2228         G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
2229         Threads::threads_do(&threads_f);
2230       }
2231 
2232       do {
2233         task->do_marking_step(1000000000.0 /* something very large */,
2234                               true         /* do_termination       */,
2235                               false        /* is_serial            */);
2236       } while (task->has_aborted() && !_cm->has_overflown());
2237       // If we overflow, then we do not want to restart. We instead
2238       // want to abort remark and do concurrent marking again.
2239       task->record_end_time();
2240     }
2241   }
2242 
2243   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
2244     AbstractGangTask("Par Remark"), _cm(cm) {
2245     _cm->terminator()->reset_for_reuse(active_workers);
2246   }
2247 };
2248 
2249 void G1ConcurrentMark::checkpointRootsFinalWork() {
2250   ResourceMark rm;
2251   HandleMark   hm;
2252   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2253 
2254   GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm);
2255 
2256   g1h->ensure_parsability(false);
2257 
2258   // this is remark, so we'll use up all active threads
2259   uint active_workers = g1h->workers()->active_workers();
2260   set_concurrency_and_phase(active_workers, false /* concurrent */);
2261   // Leave _parallel_marking_threads at it's
2262   // value originally calculated in the G1ConcurrentMark
2263   // constructor and pass values of the active workers
2264   // through the gang in the task.
2265 
2266   {
2267     StrongRootsScope srs(active_workers);
2268 
2269     G1CMRemarkTask remarkTask(this, active_workers);
2270     // We will start all available threads, even if we decide that the
2271     // active_workers will be fewer. The extra ones will just bail out
2272     // immediately.
2273     g1h->workers()->run_task(&remarkTask);
2274   }
2275 
2276   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2277   guarantee(has_overflown() ||
2278             satb_mq_set.completed_buffers_num() == 0,
2279             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
2280             BOOL_TO_STR(has_overflown()),
2281             satb_mq_set.completed_buffers_num());
2282 
2283   print_stats();
2284 }
2285 
2286 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2287   // Note we are overriding the read-only view of the prev map here, via
2288   // the cast.
2289   ((G1CMBitMap*)_prevMarkBitMap)->clear_range(mr);
2290 }
2291 
2292 HeapRegion*
2293 G1ConcurrentMark::claim_region(uint worker_id) {
2294   // "checkpoint" the finger
2295   HeapWord* finger = _finger;
2296 
2297   // _heap_end will not change underneath our feet; it only changes at
2298   // yield points.
2299   while (finger < _heap_end) {
2300     assert(_g1h->is_in_g1_reserved(finger), "invariant");
2301 
2302     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
2303 
2304     // Above heap_region_containing may return NULL as we always scan claim
2305     // until the end of the heap. In this case, just jump to the next region.
2306     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2307 
2308     // Is the gap between reading the finger and doing the CAS too long?
2309     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2310     if (res == finger && curr_region != NULL) {
2311       // we succeeded
2312       HeapWord*   bottom        = curr_region->bottom();
2313       HeapWord*   limit         = curr_region->next_top_at_mark_start();
2314 
2315       // notice that _finger == end cannot be guaranteed here since,
2316       // someone else might have moved the finger even further
2317       assert(_finger >= end, "the finger should have moved forward");
2318 
2319       if (limit > bottom) {
2320         return curr_region;
2321       } else {
2322         assert(limit == bottom,
2323                "the region limit should be at bottom");
2324         // we return NULL and the caller should try calling
2325         // claim_region() again.
2326         return NULL;
2327       }
2328     } else {
2329       assert(_finger > finger, "the finger should have moved forward");
2330       // read it again
2331       finger = _finger;
2332     }
2333   }
2334 
2335   return NULL;
2336 }
2337 
2338 #ifndef PRODUCT
2339 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
2340 private:
2341   G1CollectedHeap* _g1h;
2342   const char* _phase;
2343   int _info;
2344 
2345 public:
2346   VerifyNoCSetOops(const char* phase, int info = -1) :
2347     _g1h(G1CollectedHeap::heap()),
2348     _phase(phase),
2349     _info(info)
2350   { }
2351 
2352   void operator()(oop obj) const {
2353     guarantee(obj->is_oop(),
2354               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
2355               p2i(obj), _phase, _info);
2356     guarantee(!_g1h->obj_in_cs(obj),
2357               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
2358               p2i(obj), _phase, _info);
2359   }
2360 };
2361 
2362 void G1ConcurrentMark::verify_no_cset_oops() {
2363   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2364   if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
2365     return;
2366   }
2367 
2368   // Verify entries on the global mark stack
2369   _markStack.iterate(VerifyNoCSetOops("Stack"));
2370 
2371   // Verify entries on the task queues
2372   for (uint i = 0; i < _max_worker_id; ++i) {
2373     G1CMTaskQueue* queue = _task_queues->queue(i);
2374     queue->iterate(VerifyNoCSetOops("Queue", i));
2375   }
2376 
2377   // Verify the global finger
2378   HeapWord* global_finger = finger();
2379   if (global_finger != NULL && global_finger < _heap_end) {
2380     // Since we always iterate over all regions, we might get a NULL HeapRegion
2381     // here.
2382     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
2383     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
2384               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
2385               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
2386   }
2387 
2388   // Verify the task fingers
2389   assert(parallel_marking_threads() <= _max_worker_id, "sanity");
2390   for (uint i = 0; i < parallel_marking_threads(); ++i) {
2391     G1CMTask* task = _tasks[i];
2392     HeapWord* task_finger = task->finger();
2393     if (task_finger != NULL && task_finger < _heap_end) {
2394       // See above note on the global finger verification.
2395       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
2396       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2397                 !task_hr->in_collection_set(),
2398                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2399                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
2400     }
2401   }
2402 }
2403 #endif // PRODUCT
2404 
2405 // Aggregate the counting data that was constructed concurrently
2406 // with marking.
2407 class AggregateCountDataHRClosure: public HeapRegionClosure {
2408   G1CollectedHeap* _g1h;
2409   G1ConcurrentMark* _cm;
2410   CardTableModRefBS* _ct_bs;
2411   BitMap* _cm_card_bm;
2412   uint _max_worker_id;
2413 
2414  public:
2415   AggregateCountDataHRClosure(G1CollectedHeap* g1h,
2416                               BitMap* cm_card_bm,
2417                               uint max_worker_id) :
2418     _g1h(g1h), _cm(g1h->concurrent_mark()),
2419     _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
2420     _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
2421 
2422   bool doHeapRegion(HeapRegion* hr) {
2423     HeapWord* start = hr->bottom();
2424     HeapWord* limit = hr->next_top_at_mark_start();
2425     HeapWord* end = hr->end();
2426 
2427     assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
2428            "Preconditions not met - "
2429            "start: " PTR_FORMAT ", limit: " PTR_FORMAT ", "
2430            "top: " PTR_FORMAT ", end: " PTR_FORMAT,
2431            p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end()));
2432 
2433     assert(hr->next_marked_bytes() == 0, "Precondition");
2434 
2435     if (start == limit) {
2436       // NTAMS of this region has not been set so nothing to do.
2437       return false;
2438     }
2439 
2440     // 'start' should be in the heap.
2441     assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
2442     // 'end' *may* be just beyond the end of the heap (if hr is the last region)
2443     assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
2444 
2445     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
2446     BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
2447     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
2448 
2449     // If ntams is not card aligned then we bump card bitmap index
2450     // for limit so that we get the all the cards spanned by
2451     // the object ending at ntams.
2452     // Note: if this is the last region in the heap then ntams
2453     // could be actually just beyond the end of the the heap;
2454     // limit_idx will then  correspond to a (non-existent) card
2455     // that is also outside the heap.
2456     if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
2457       limit_idx += 1;
2458     }
2459 
2460     assert(limit_idx <= end_idx, "or else use atomics");
2461 
2462     // Aggregate the "stripe" in the count data associated with hr.
2463     uint hrm_index = hr->hrm_index();
2464     size_t marked_bytes = 0;
2465 
2466     for (uint i = 0; i < _max_worker_id; i += 1) {
2467       size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
2468       BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
2469 
2470       // Fetch the marked_bytes in this region for task i and
2471       // add it to the running total for this region.
2472       marked_bytes += marked_bytes_array[hrm_index];
2473 
2474       // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
2475       // into the global card bitmap.
2476       BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
2477 
2478       while (scan_idx < limit_idx) {
2479         assert(task_card_bm->at(scan_idx) == true, "should be");
2480         _cm_card_bm->set_bit(scan_idx);
2481         assert(_cm_card_bm->at(scan_idx) == true, "should be");
2482 
2483         // BitMap::get_next_one_offset() can handle the case when
2484         // its left_offset parameter is greater than its right_offset
2485         // parameter. It does, however, have an early exit if
2486         // left_offset == right_offset. So let's limit the value
2487         // passed in for left offset here.
2488         BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
2489         scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
2490       }
2491     }
2492 
2493     // Update the marked bytes for this region.
2494     hr->add_to_marked_bytes(marked_bytes);
2495 
2496     // Next heap region
2497     return false;
2498   }
2499 };
2500 
2501 class G1AggregateCountDataTask: public AbstractGangTask {
2502 protected:
2503   G1CollectedHeap* _g1h;
2504   G1ConcurrentMark* _cm;
2505   BitMap* _cm_card_bm;
2506   uint _max_worker_id;
2507   uint _active_workers;
2508   HeapRegionClaimer _hrclaimer;
2509 
2510 public:
2511   G1AggregateCountDataTask(G1CollectedHeap* g1h,
2512                            G1ConcurrentMark* cm,
2513                            BitMap* cm_card_bm,
2514                            uint max_worker_id,
2515                            uint n_workers) :
2516       AbstractGangTask("Count Aggregation"),
2517       _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
2518       _max_worker_id(max_worker_id),
2519       _active_workers(n_workers),
2520       _hrclaimer(_active_workers) {
2521   }
2522 
2523   void work(uint worker_id) {
2524     AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
2525 
2526     _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer);
2527   }
2528 };
2529 
2530 
2531 void G1ConcurrentMark::aggregate_count_data() {
2532   uint n_workers = _g1h->workers()->active_workers();
2533 
2534   G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
2535                                            _max_worker_id, n_workers);
2536 
2537   _g1h->workers()->run_task(&g1_par_agg_task);
2538 }
2539 
2540 // Clear the per-worker arrays used to store the per-region counting data
2541 void G1ConcurrentMark::clear_all_count_data() {
2542   // Clear the global card bitmap - it will be filled during
2543   // liveness count aggregation (during remark) and the
2544   // final counting task.
2545   _card_bm.clear();
2546 
2547   // Clear the global region bitmap - it will be filled as part
2548   // of the final counting task.
2549   _region_bm.clear();
2550 
2551   uint max_regions = _g1h->max_regions();
2552   assert(_max_worker_id > 0, "uninitialized");
2553 
2554   for (uint i = 0; i < _max_worker_id; i += 1) {
2555     BitMap* task_card_bm = count_card_bitmap_for(i);
2556     size_t* marked_bytes_array = count_marked_bytes_array_for(i);
2557 
2558     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
2559     assert(marked_bytes_array != NULL, "uninitialized");
2560 
2561     memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
2562     task_card_bm->clear();
2563   }
2564 }
2565 
2566 void G1ConcurrentMark::print_stats() {
2567   if (!log_is_enabled(Debug, gc, stats)) {
2568     return;
2569   }
2570   log_debug(gc, stats)("---------------------------------------------------------------------");
2571   for (size_t i = 0; i < _active_tasks; ++i) {
2572     _tasks[i]->print_stats();
2573     log_debug(gc, stats)("---------------------------------------------------------------------");
2574   }
2575 }
2576 
2577 // abandon current marking iteration due to a Full GC
2578 void G1ConcurrentMark::abort() {
2579   if (!cmThread()->during_cycle() || _has_aborted) {
2580     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2581     return;
2582   }
2583 
2584   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2585   // concurrent bitmap clearing.
2586   clear_bitmap(_nextMarkBitMap, _g1h->workers(), false);
2587 
2588   // Note we cannot clear the previous marking bitmap here
2589   // since VerifyDuringGC verifies the objects marked during
2590   // a full GC against the previous bitmap.
2591 
2592   // Clear the liveness counting data
2593   clear_all_count_data();
2594   // Empty mark stack
2595   reset_marking_state();
2596   for (uint i = 0; i < _max_worker_id; ++i) {
2597     _tasks[i]->clear_region_fields();
2598   }
2599   _first_overflow_barrier_sync.abort();
2600   _second_overflow_barrier_sync.abort();
2601   _has_aborted = true;
2602 
2603   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2604   satb_mq_set.abandon_partial_marking();
2605   // This can be called either during or outside marking, we'll read
2606   // the expected_active value from the SATB queue set.
2607   satb_mq_set.set_active_all_threads(
2608                                  false, /* new active value */
2609                                  satb_mq_set.is_active() /* expected_active */);
2610 }
2611 
2612 static void print_ms_time_info(const char* prefix, const char* name,
2613                                NumberSeq& ns) {
2614   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2615                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2616   if (ns.num() > 0) {
2617     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2618                            prefix, ns.sd(), ns.maximum());
2619   }
2620 }
2621 
2622 void G1ConcurrentMark::print_summary_info() {
2623   LogHandle(gc, marking) log;
2624   if (!log.is_trace()) {
2625     return;
2626   }
2627 
2628   log.trace(" Concurrent marking:");
2629   print_ms_time_info("  ", "init marks", _init_times);
2630   print_ms_time_info("  ", "remarks", _remark_times);
2631   {
2632     print_ms_time_info("     ", "final marks", _remark_mark_times);
2633     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2634 
2635   }
2636   print_ms_time_info("  ", "cleanups", _cleanup_times);
2637   log.trace("    Final counting total time = %8.2f s (avg = %8.2f ms).",
2638             _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2639   if (G1ScrubRemSets) {
2640     log.trace("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
2641               _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2642   }
2643   log.trace("  Total stop_world time = %8.2f s.",
2644             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2645   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2646             cmThread()->vtime_accum(), cmThread()->vtime_mark_accum());
2647 }
2648 
2649 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2650   _parallel_workers->print_worker_threads_on(st);
2651 }
2652 
2653 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2654   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2655       p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
2656   _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
2657   _nextMarkBitMap->print_on_error(st, " Next Bits: ");
2658 }
2659 
2660 // We take a break if someone is trying to stop the world.
2661 bool G1ConcurrentMark::do_yield_check(uint worker_id) {
2662   if (SuspendibleThreadSet::should_yield()) {
2663     SuspendibleThreadSet::yield();
2664     return true;
2665   } else {
2666     return false;
2667   }
2668 }
2669 
2670 // Closure for iteration over bitmaps
2671 class G1CMBitMapClosure : public BitMapClosure {
2672 private:
2673   // the bitmap that is being iterated over
2674   G1CMBitMap*                 _nextMarkBitMap;
2675   G1ConcurrentMark*           _cm;
2676   G1CMTask*                   _task;
2677 
2678 public:
2679   G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm, G1CMBitMap* nextMarkBitMap) :
2680     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
2681 
2682   bool do_bit(size_t offset) {
2683     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
2684     assert(_nextMarkBitMap->isMarked(addr), "invariant");
2685     assert( addr < _cm->finger(), "invariant");
2686     assert(addr >= _task->finger(), "invariant");
2687 
2688     // We move that task's local finger along.
2689     _task->move_finger_to(addr);
2690 
2691     _task->scan_object(oop(addr));
2692     // we only partially drain the local queue and global stack
2693     _task->drain_local_queue(true);
2694     _task->drain_global_stack(true);
2695 
2696     // if the has_aborted flag has been raised, we need to bail out of
2697     // the iteration
2698     return !_task->has_aborted();
2699   }
2700 };
2701 
2702 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2703   ReferenceProcessor* result = g1h->ref_processor_cm();
2704   assert(result != NULL, "CM reference processor should not be NULL");
2705   return result;
2706 }
2707 
2708 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2709                                G1ConcurrentMark* cm,
2710                                G1CMTask* task)
2711   : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2712     _g1h(g1h), _cm(cm), _task(task)
2713 { }
2714 
2715 void G1CMTask::setup_for_region(HeapRegion* hr) {
2716   assert(hr != NULL,
2717         "claim_region() should have filtered out NULL regions");
2718   _curr_region  = hr;
2719   _finger       = hr->bottom();
2720   update_region_limit();
2721 }
2722 
2723 void G1CMTask::update_region_limit() {
2724   HeapRegion* hr            = _curr_region;
2725   HeapWord* bottom          = hr->bottom();
2726   HeapWord* limit           = hr->next_top_at_mark_start();
2727 
2728   if (limit == bottom) {
2729     // The region was collected underneath our feet.
2730     // We set the finger to bottom to ensure that the bitmap
2731     // iteration that will follow this will not do anything.
2732     // (this is not a condition that holds when we set the region up,
2733     // as the region is not supposed to be empty in the first place)
2734     _finger = bottom;
2735   } else if (limit >= _region_limit) {
2736     assert(limit >= _finger, "peace of mind");
2737   } else {
2738     assert(limit < _region_limit, "only way to get here");
2739     // This can happen under some pretty unusual circumstances.  An
2740     // evacuation pause empties the region underneath our feet (NTAMS
2741     // at bottom). We then do some allocation in the region (NTAMS
2742     // stays at bottom), followed by the region being used as a GC
2743     // alloc region (NTAMS will move to top() and the objects
2744     // originally below it will be grayed). All objects now marked in
2745     // the region are explicitly grayed, if below the global finger,
2746     // and we do not need in fact to scan anything else. So, we simply
2747     // set _finger to be limit to ensure that the bitmap iteration
2748     // doesn't do anything.
2749     _finger = limit;
2750   }
2751 
2752   _region_limit = limit;
2753 }
2754 
2755 void G1CMTask::giveup_current_region() {
2756   assert(_curr_region != NULL, "invariant");
2757   clear_region_fields();
2758 }
2759 
2760 void G1CMTask::clear_region_fields() {
2761   // Values for these three fields that indicate that we're not
2762   // holding on to a region.
2763   _curr_region   = NULL;
2764   _finger        = NULL;
2765   _region_limit  = NULL;
2766 }
2767 
2768 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2769   if (cm_oop_closure == NULL) {
2770     assert(_cm_oop_closure != NULL, "invariant");
2771   } else {
2772     assert(_cm_oop_closure == NULL, "invariant");
2773   }
2774   _cm_oop_closure = cm_oop_closure;
2775 }
2776 
2777 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) {
2778   guarantee(nextMarkBitMap != NULL, "invariant");
2779   _nextMarkBitMap                = nextMarkBitMap;
2780   clear_region_fields();
2781 
2782   _calls                         = 0;
2783   _elapsed_time_ms               = 0.0;
2784   _termination_time_ms           = 0.0;
2785   _termination_start_time_ms     = 0.0;
2786 }
2787 
2788 bool G1CMTask::should_exit_termination() {
2789   regular_clock_call();
2790   // This is called when we are in the termination protocol. We should
2791   // quit if, for some reason, this task wants to abort or the global
2792   // stack is not empty (this means that we can get work from it).
2793   return !_cm->mark_stack_empty() || has_aborted();
2794 }
2795 
2796 void G1CMTask::reached_limit() {
2797   assert(_words_scanned >= _words_scanned_limit ||
2798          _refs_reached >= _refs_reached_limit ,
2799          "shouldn't have been called otherwise");
2800   regular_clock_call();
2801 }
2802 
2803 void G1CMTask::regular_clock_call() {
2804   if (has_aborted()) return;
2805 
2806   // First, we need to recalculate the words scanned and refs reached
2807   // limits for the next clock call.
2808   recalculate_limits();
2809 
2810   // During the regular clock call we do the following
2811 
2812   // (1) If an overflow has been flagged, then we abort.
2813   if (_cm->has_overflown()) {
2814     set_has_aborted();
2815     return;
2816   }
2817 
2818   // If we are not concurrent (i.e. we're doing remark) we don't need
2819   // to check anything else. The other steps are only needed during
2820   // the concurrent marking phase.
2821   if (!concurrent()) return;
2822 
2823   // (2) If marking has been aborted for Full GC, then we also abort.
2824   if (_cm->has_aborted()) {
2825     set_has_aborted();
2826     return;
2827   }
2828 
2829   double curr_time_ms = os::elapsedVTime() * 1000.0;
2830 
2831   // (4) We check whether we should yield. If we have to, then we abort.
2832   if (SuspendibleThreadSet::should_yield()) {
2833     // We should yield. To do this we abort the task. The caller is
2834     // responsible for yielding.
2835     set_has_aborted();
2836     return;
2837   }
2838 
2839   // (5) We check whether we've reached our time quota. If we have,
2840   // then we abort.
2841   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2842   if (elapsed_time_ms > _time_target_ms) {
2843     set_has_aborted();
2844     _has_timed_out = true;
2845     return;
2846   }
2847 
2848   // (6) Finally, we check whether there are enough completed STAB
2849   // buffers available for processing. If there are, we abort.
2850   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2851   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2852     // we do need to process SATB buffers, we'll abort and restart
2853     // the marking task to do so
2854     set_has_aborted();
2855     return;
2856   }
2857 }
2858 
2859 void G1CMTask::recalculate_limits() {
2860   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2861   _words_scanned_limit      = _real_words_scanned_limit;
2862 
2863   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2864   _refs_reached_limit       = _real_refs_reached_limit;
2865 }
2866 
2867 void G1CMTask::decrease_limits() {
2868   // This is called when we believe that we're going to do an infrequent
2869   // operation which will increase the per byte scanned cost (i.e. move
2870   // entries to/from the global stack). It basically tries to decrease the
2871   // scanning limit so that the clock is called earlier.
2872 
2873   _words_scanned_limit = _real_words_scanned_limit -
2874     3 * words_scanned_period / 4;
2875   _refs_reached_limit  = _real_refs_reached_limit -
2876     3 * refs_reached_period / 4;
2877 }
2878 
2879 void G1CMTask::move_entries_to_global_stack() {
2880   // local array where we'll store the entries that will be popped
2881   // from the local queue
2882   oop buffer[global_stack_transfer_size];
2883 
2884   int n = 0;
2885   oop obj;
2886   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
2887     buffer[n] = obj;
2888     ++n;
2889   }
2890 
2891   if (n > 0) {
2892     // we popped at least one entry from the local queue
2893 
2894     if (!_cm->mark_stack_push(buffer, n)) {
2895       set_has_aborted();
2896     }
2897   }
2898 
2899   // this operation was quite expensive, so decrease the limits
2900   decrease_limits();
2901 }
2902 
2903 void G1CMTask::get_entries_from_global_stack() {
2904   // local array where we'll store the entries that will be popped
2905   // from the global stack.
2906   oop buffer[global_stack_transfer_size];
2907   int n;
2908   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
2909   assert(n <= global_stack_transfer_size,
2910          "we should not pop more than the given limit");
2911   if (n > 0) {
2912     // yes, we did actually pop at least one entry
2913     for (int i = 0; i < n; ++i) {
2914       bool success = _task_queue->push(buffer[i]);
2915       // We only call this when the local queue is empty or under a
2916       // given target limit. So, we do not expect this push to fail.
2917       assert(success, "invariant");
2918     }
2919   }
2920 
2921   // this operation was quite expensive, so decrease the limits
2922   decrease_limits();
2923 }
2924 
2925 void G1CMTask::drain_local_queue(bool partially) {
2926   if (has_aborted()) return;
2927 
2928   // Decide what the target size is, depending whether we're going to
2929   // drain it partially (so that other tasks can steal if they run out
2930   // of things to do) or totally (at the very end).
2931   size_t target_size;
2932   if (partially) {
2933     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2934   } else {
2935     target_size = 0;
2936   }
2937 
2938   if (_task_queue->size() > target_size) {
2939     oop obj;
2940     bool ret = _task_queue->pop_local(obj);
2941     while (ret) {
2942       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
2943       assert(!_g1h->is_on_master_free_list(
2944                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
2945 
2946       scan_object(obj);
2947 
2948       if (_task_queue->size() <= target_size || has_aborted()) {
2949         ret = false;
2950       } else {
2951         ret = _task_queue->pop_local(obj);
2952       }
2953     }
2954   }
2955 }
2956 
2957 void G1CMTask::drain_global_stack(bool partially) {
2958   if (has_aborted()) return;
2959 
2960   // We have a policy to drain the local queue before we attempt to
2961   // drain the global stack.
2962   assert(partially || _task_queue->size() == 0, "invariant");
2963 
2964   // Decide what the target size is, depending whether we're going to
2965   // drain it partially (so that other tasks can steal if they run out
2966   // of things to do) or totally (at the very end).  Notice that,
2967   // because we move entries from the global stack in chunks or
2968   // because another task might be doing the same, we might in fact
2969   // drop below the target. But, this is not a problem.
2970   size_t target_size;
2971   if (partially) {
2972     target_size = _cm->partial_mark_stack_size_target();
2973   } else {
2974     target_size = 0;
2975   }
2976 
2977   if (_cm->mark_stack_size() > target_size) {
2978     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2979       get_entries_from_global_stack();
2980       drain_local_queue(partially);
2981     }
2982   }
2983 }
2984 
2985 // SATB Queue has several assumptions on whether to call the par or
2986 // non-par versions of the methods. this is why some of the code is
2987 // replicated. We should really get rid of the single-threaded version
2988 // of the code to simplify things.
2989 void G1CMTask::drain_satb_buffers() {
2990   if (has_aborted()) return;
2991 
2992   // We set this so that the regular clock knows that we're in the
2993   // middle of draining buffers and doesn't set the abort flag when it
2994   // notices that SATB buffers are available for draining. It'd be
2995   // very counter productive if it did that. :-)
2996   _draining_satb_buffers = true;
2997 
2998   G1CMSATBBufferClosure satb_cl(this, _g1h);
2999   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3000 
3001   // This keeps claiming and applying the closure to completed buffers
3002   // until we run out of buffers or we need to abort.
3003   while (!has_aborted() &&
3004          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
3005     regular_clock_call();
3006   }
3007 
3008   _draining_satb_buffers = false;
3009 
3010   assert(has_aborted() ||
3011          concurrent() ||
3012          satb_mq_set.completed_buffers_num() == 0, "invariant");
3013 
3014   // again, this was a potentially expensive operation, decrease the
3015   // limits to get the regular clock call early
3016   decrease_limits();
3017 }
3018 
3019 void G1CMTask::print_stats() {
3020   log_debug(gc, stats)("Marking Stats, task = %u, calls = %d",
3021                        _worker_id, _calls);
3022   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3023                        _elapsed_time_ms, _termination_time_ms);
3024   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3025                        _step_times_ms.num(), _step_times_ms.avg(),
3026                        _step_times_ms.sd());
3027   log_debug(gc, stats)("                    max = %1.2lfms, total = %1.2lfms",
3028                        _step_times_ms.maximum(), _step_times_ms.sum());
3029 }
3030 
3031 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) {
3032   return _task_queues->steal(worker_id, hash_seed, obj);
3033 }
3034 
3035 /*****************************************************************************
3036 
3037     The do_marking_step(time_target_ms, ...) method is the building
3038     block of the parallel marking framework. It can be called in parallel
3039     with other invocations of do_marking_step() on different tasks
3040     (but only one per task, obviously) and concurrently with the
3041     mutator threads, or during remark, hence it eliminates the need
3042     for two versions of the code. When called during remark, it will
3043     pick up from where the task left off during the concurrent marking
3044     phase. Interestingly, tasks are also claimable during evacuation
3045     pauses too, since do_marking_step() ensures that it aborts before
3046     it needs to yield.
3047 
3048     The data structures that it uses to do marking work are the
3049     following:
3050 
3051       (1) Marking Bitmap. If there are gray objects that appear only
3052       on the bitmap (this happens either when dealing with an overflow
3053       or when the initial marking phase has simply marked the roots
3054       and didn't push them on the stack), then tasks claim heap
3055       regions whose bitmap they then scan to find gray objects. A
3056       global finger indicates where the end of the last claimed region
3057       is. A local finger indicates how far into the region a task has
3058       scanned. The two fingers are used to determine how to gray an
3059       object (i.e. whether simply marking it is OK, as it will be
3060       visited by a task in the future, or whether it needs to be also
3061       pushed on a stack).
3062 
3063       (2) Local Queue. The local queue of the task which is accessed
3064       reasonably efficiently by the task. Other tasks can steal from
3065       it when they run out of work. Throughout the marking phase, a
3066       task attempts to keep its local queue short but not totally
3067       empty, so that entries are available for stealing by other
3068       tasks. Only when there is no more work, a task will totally
3069       drain its local queue.
3070 
3071       (3) Global Mark Stack. This handles local queue overflow. During
3072       marking only sets of entries are moved between it and the local
3073       queues, as access to it requires a mutex and more fine-grain
3074       interaction with it which might cause contention. If it
3075       overflows, then the marking phase should restart and iterate
3076       over the bitmap to identify gray objects. Throughout the marking
3077       phase, tasks attempt to keep the global mark stack at a small
3078       length but not totally empty, so that entries are available for
3079       popping by other tasks. Only when there is no more work, tasks
3080       will totally drain the global mark stack.
3081 
3082       (4) SATB Buffer Queue. This is where completed SATB buffers are
3083       made available. Buffers are regularly removed from this queue
3084       and scanned for roots, so that the queue doesn't get too
3085       long. During remark, all completed buffers are processed, as
3086       well as the filled in parts of any uncompleted buffers.
3087 
3088     The do_marking_step() method tries to abort when the time target
3089     has been reached. There are a few other cases when the
3090     do_marking_step() method also aborts:
3091 
3092       (1) When the marking phase has been aborted (after a Full GC).
3093 
3094       (2) When a global overflow (on the global stack) has been
3095       triggered. Before the task aborts, it will actually sync up with
3096       the other tasks to ensure that all the marking data structures
3097       (local queues, stacks, fingers etc.)  are re-initialized so that
3098       when do_marking_step() completes, the marking phase can
3099       immediately restart.
3100 
3101       (3) When enough completed SATB buffers are available. The
3102       do_marking_step() method only tries to drain SATB buffers right
3103       at the beginning. So, if enough buffers are available, the
3104       marking step aborts and the SATB buffers are processed at
3105       the beginning of the next invocation.
3106 
3107       (4) To yield. when we have to yield then we abort and yield
3108       right at the end of do_marking_step(). This saves us from a lot
3109       of hassle as, by yielding we might allow a Full GC. If this
3110       happens then objects will be compacted underneath our feet, the
3111       heap might shrink, etc. We save checking for this by just
3112       aborting and doing the yield right at the end.
3113 
3114     From the above it follows that the do_marking_step() method should
3115     be called in a loop (or, otherwise, regularly) until it completes.
3116 
3117     If a marking step completes without its has_aborted() flag being
3118     true, it means it has completed the current marking phase (and
3119     also all other marking tasks have done so and have all synced up).
3120 
3121     A method called regular_clock_call() is invoked "regularly" (in
3122     sub ms intervals) throughout marking. It is this clock method that
3123     checks all the abort conditions which were mentioned above and
3124     decides when the task should abort. A work-based scheme is used to
3125     trigger this clock method: when the number of object words the
3126     marking phase has scanned or the number of references the marking
3127     phase has visited reach a given limit. Additional invocations to
3128     the method clock have been planted in a few other strategic places
3129     too. The initial reason for the clock method was to avoid calling
3130     vtime too regularly, as it is quite expensive. So, once it was in
3131     place, it was natural to piggy-back all the other conditions on it
3132     too and not constantly check them throughout the code.
3133 
3134     If do_termination is true then do_marking_step will enter its
3135     termination protocol.
3136 
3137     The value of is_serial must be true when do_marking_step is being
3138     called serially (i.e. by the VMThread) and do_marking_step should
3139     skip any synchronization in the termination and overflow code.
3140     Examples include the serial remark code and the serial reference
3141     processing closures.
3142 
3143     The value of is_serial must be false when do_marking_step is
3144     being called by any of the worker threads in a work gang.
3145     Examples include the concurrent marking code (CMMarkingTask),
3146     the MT remark code, and the MT reference processing closures.
3147 
3148  *****************************************************************************/
3149 
3150 void G1CMTask::do_marking_step(double time_target_ms,
3151                                bool do_termination,
3152                                bool is_serial) {
3153   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
3154   assert(concurrent() == _cm->concurrent(), "they should be the same");
3155 
3156   G1CollectorPolicy* g1_policy = _g1h->g1_policy();
3157   assert(_task_queues != NULL, "invariant");
3158   assert(_task_queue != NULL, "invariant");
3159   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
3160 
3161   assert(!_claimed,
3162          "only one thread should claim this task at any one time");
3163 
3164   // OK, this doesn't safeguard again all possible scenarios, as it is
3165   // possible for two threads to set the _claimed flag at the same
3166   // time. But it is only for debugging purposes anyway and it will
3167   // catch most problems.
3168   _claimed = true;
3169 
3170   _start_time_ms = os::elapsedVTime() * 1000.0;
3171 
3172   // If do_stealing is true then do_marking_step will attempt to
3173   // steal work from the other G1CMTasks. It only makes sense to
3174   // enable stealing when the termination protocol is enabled
3175   // and do_marking_step() is not being called serially.
3176   bool do_stealing = do_termination && !is_serial;
3177 
3178   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
3179   _time_target_ms = time_target_ms - diff_prediction_ms;
3180 
3181   // set up the variables that are used in the work-based scheme to
3182   // call the regular clock method
3183   _words_scanned = 0;
3184   _refs_reached  = 0;
3185   recalculate_limits();
3186 
3187   // clear all flags
3188   clear_has_aborted();
3189   _has_timed_out = false;
3190   _draining_satb_buffers = false;
3191 
3192   ++_calls;
3193 
3194   // Set up the bitmap and oop closures. Anything that uses them is
3195   // eventually called from this method, so it is OK to allocate these
3196   // statically.
3197   G1CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
3198   G1CMOopClosure    cm_oop_closure(_g1h, _cm, this);
3199   set_cm_oop_closure(&cm_oop_closure);
3200 
3201   if (_cm->has_overflown()) {
3202     // This can happen if the mark stack overflows during a GC pause
3203     // and this task, after a yield point, restarts. We have to abort
3204     // as we need to get into the overflow protocol which happens
3205     // right at the end of this task.
3206     set_has_aborted();
3207   }
3208 
3209   // First drain any available SATB buffers. After this, we will not
3210   // look at SATB buffers before the next invocation of this method.
3211   // If enough completed SATB buffers are queued up, the regular clock
3212   // will abort this task so that it restarts.
3213   drain_satb_buffers();
3214   // ...then partially drain the local queue and the global stack
3215   drain_local_queue(true);
3216   drain_global_stack(true);
3217 
3218   do {
3219     if (!has_aborted() && _curr_region != NULL) {
3220       // This means that we're already holding on to a region.
3221       assert(_finger != NULL, "if region is not NULL, then the finger "
3222              "should not be NULL either");
3223 
3224       // We might have restarted this task after an evacuation pause
3225       // which might have evacuated the region we're holding on to
3226       // underneath our feet. Let's read its limit again to make sure
3227       // that we do not iterate over a region of the heap that
3228       // contains garbage (update_region_limit() will also move
3229       // _finger to the start of the region if it is found empty).
3230       update_region_limit();
3231       // We will start from _finger not from the start of the region,
3232       // as we might be restarting this task after aborting half-way
3233       // through scanning this region. In this case, _finger points to
3234       // the address where we last found a marked object. If this is a
3235       // fresh region, _finger points to start().
3236       MemRegion mr = MemRegion(_finger, _region_limit);
3237 
3238       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
3239              "humongous regions should go around loop once only");
3240 
3241       // Some special cases:
3242       // If the memory region is empty, we can just give up the region.
3243       // If the current region is humongous then we only need to check
3244       // the bitmap for the bit associated with the start of the object,
3245       // scan the object if it's live, and give up the region.
3246       // Otherwise, let's iterate over the bitmap of the part of the region
3247       // that is left.
3248       // If the iteration is successful, give up the region.
3249       if (mr.is_empty()) {
3250         giveup_current_region();
3251         regular_clock_call();
3252       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
3253         if (_nextMarkBitMap->isMarked(mr.start())) {
3254           // The object is marked - apply the closure
3255           BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
3256           bitmap_closure.do_bit(offset);
3257         }
3258         // Even if this task aborted while scanning the humongous object
3259         // we can (and should) give up the current region.
3260         giveup_current_region();
3261         regular_clock_call();
3262       } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
3263         giveup_current_region();
3264         regular_clock_call();
3265       } else {
3266         assert(has_aborted(), "currently the only way to do so");
3267         // The only way to abort the bitmap iteration is to return
3268         // false from the do_bit() method. However, inside the
3269         // do_bit() method we move the _finger to point to the
3270         // object currently being looked at. So, if we bail out, we
3271         // have definitely set _finger to something non-null.
3272         assert(_finger != NULL, "invariant");
3273 
3274         // Region iteration was actually aborted. So now _finger
3275         // points to the address of the object we last scanned. If we
3276         // leave it there, when we restart this task, we will rescan
3277         // the object. It is easy to avoid this. We move the finger by
3278         // enough to point to the next possible object header (the
3279         // bitmap knows by how much we need to move it as it knows its
3280         // granularity).
3281         assert(_finger < _region_limit, "invariant");
3282         HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
3283         // Check if bitmap iteration was aborted while scanning the last object
3284         if (new_finger >= _region_limit) {
3285           giveup_current_region();
3286         } else {
3287           move_finger_to(new_finger);
3288         }
3289       }
3290     }
3291     // At this point we have either completed iterating over the
3292     // region we were holding on to, or we have aborted.
3293 
3294     // We then partially drain the local queue and the global stack.
3295     // (Do we really need this?)
3296     drain_local_queue(true);
3297     drain_global_stack(true);
3298 
3299     // Read the note on the claim_region() method on why it might
3300     // return NULL with potentially more regions available for
3301     // claiming and why we have to check out_of_regions() to determine
3302     // whether we're done or not.
3303     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
3304       // We are going to try to claim a new region. We should have
3305       // given up on the previous one.
3306       // Separated the asserts so that we know which one fires.
3307       assert(_curr_region  == NULL, "invariant");
3308       assert(_finger       == NULL, "invariant");
3309       assert(_region_limit == NULL, "invariant");
3310       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
3311       if (claimed_region != NULL) {
3312         // Yes, we managed to claim one
3313         setup_for_region(claimed_region);
3314         assert(_curr_region == claimed_region, "invariant");
3315       }
3316       // It is important to call the regular clock here. It might take
3317       // a while to claim a region if, for example, we hit a large
3318       // block of empty regions. So we need to call the regular clock
3319       // method once round the loop to make sure it's called
3320       // frequently enough.
3321       regular_clock_call();
3322     }
3323 
3324     if (!has_aborted() && _curr_region == NULL) {
3325       assert(_cm->out_of_regions(),
3326              "at this point we should be out of regions");
3327     }
3328   } while ( _curr_region != NULL && !has_aborted());
3329 
3330   if (!has_aborted()) {
3331     // We cannot check whether the global stack is empty, since other
3332     // tasks might be pushing objects to it concurrently.
3333     assert(_cm->out_of_regions(),
3334            "at this point we should be out of regions");
3335     // Try to reduce the number of available SATB buffers so that
3336     // remark has less work to do.
3337     drain_satb_buffers();
3338   }
3339 
3340   // Since we've done everything else, we can now totally drain the
3341   // local queue and global stack.
3342   drain_local_queue(false);
3343   drain_global_stack(false);
3344 
3345   // Attempt at work stealing from other task's queues.
3346   if (do_stealing && !has_aborted()) {
3347     // We have not aborted. This means that we have finished all that
3348     // we could. Let's try to do some stealing...
3349 
3350     // We cannot check whether the global stack is empty, since other
3351     // tasks might be pushing objects to it concurrently.
3352     assert(_cm->out_of_regions() && _task_queue->size() == 0,
3353            "only way to reach here");
3354     while (!has_aborted()) {
3355       oop obj;
3356       if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
3357         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
3358                "any stolen object should be marked");
3359         scan_object(obj);
3360 
3361         // And since we're towards the end, let's totally drain the
3362         // local queue and global stack.
3363         drain_local_queue(false);
3364         drain_global_stack(false);
3365       } else {
3366         break;
3367       }
3368     }
3369   }
3370 
3371   // We still haven't aborted. Now, let's try to get into the
3372   // termination protocol.
3373   if (do_termination && !has_aborted()) {
3374     // We cannot check whether the global stack is empty, since other
3375     // tasks might be concurrently pushing objects on it.
3376     // Separated the asserts so that we know which one fires.
3377     assert(_cm->out_of_regions(), "only way to reach here");
3378     assert(_task_queue->size() == 0, "only way to reach here");
3379     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
3380 
3381     // The G1CMTask class also extends the TerminatorTerminator class,
3382     // hence its should_exit_termination() method will also decide
3383     // whether to exit the termination protocol or not.
3384     bool finished = (is_serial ||
3385                      _cm->terminator()->offer_termination(this));
3386     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
3387     _termination_time_ms +=
3388       termination_end_time_ms - _termination_start_time_ms;
3389 
3390     if (finished) {
3391       // We're all done.
3392 
3393       if (_worker_id == 0) {
3394         // let's allow task 0 to do this
3395         if (concurrent()) {
3396           assert(_cm->concurrent_marking_in_progress(), "invariant");
3397           // we need to set this to false before the next
3398           // safepoint. This way we ensure that the marking phase
3399           // doesn't observe any more heap expansions.
3400           _cm->clear_concurrent_marking_in_progress();
3401         }
3402       }
3403 
3404       // We can now guarantee that the global stack is empty, since
3405       // all other tasks have finished. We separated the guarantees so
3406       // that, if a condition is false, we can immediately find out
3407       // which one.
3408       guarantee(_cm->out_of_regions(), "only way to reach here");
3409       guarantee(_cm->mark_stack_empty(), "only way to reach here");
3410       guarantee(_task_queue->size() == 0, "only way to reach here");
3411       guarantee(!_cm->has_overflown(), "only way to reach here");
3412       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
3413     } else {
3414       // Apparently there's more work to do. Let's abort this task. It
3415       // will restart it and we can hopefully find more things to do.
3416       set_has_aborted();
3417     }
3418   }
3419 
3420   // Mainly for debugging purposes to make sure that a pointer to the
3421   // closure which was statically allocated in this frame doesn't
3422   // escape it by accident.
3423   set_cm_oop_closure(NULL);
3424   double end_time_ms = os::elapsedVTime() * 1000.0;
3425   double elapsed_time_ms = end_time_ms - _start_time_ms;
3426   // Update the step history.
3427   _step_times_ms.add(elapsed_time_ms);
3428 
3429   if (has_aborted()) {
3430     // The task was aborted for some reason.
3431     if (_has_timed_out) {
3432       double diff_ms = elapsed_time_ms - _time_target_ms;
3433       // Keep statistics of how well we did with respect to hitting
3434       // our target only if we actually timed out (if we aborted for
3435       // other reasons, then the results might get skewed).
3436       _marking_step_diffs_ms.add(diff_ms);
3437     }
3438 
3439     if (_cm->has_overflown()) {
3440       // This is the interesting one. We aborted because a global
3441       // overflow was raised. This means we have to restart the
3442       // marking phase and start iterating over regions. However, in
3443       // order to do this we have to make sure that all tasks stop
3444       // what they are doing and re-initialize in a safe manner. We
3445       // will achieve this with the use of two barrier sync points.
3446 
3447       if (!is_serial) {
3448         // We only need to enter the sync barrier if being called
3449         // from a parallel context
3450         _cm->enter_first_sync_barrier(_worker_id);
3451 
3452         // When we exit this sync barrier we know that all tasks have
3453         // stopped doing marking work. So, it's now safe to
3454         // re-initialize our data structures. At the end of this method,
3455         // task 0 will clear the global data structures.
3456       }
3457 
3458       // We clear the local state of this task...
3459       clear_region_fields();
3460 
3461       if (!is_serial) {
3462         // ...and enter the second barrier.
3463         _cm->enter_second_sync_barrier(_worker_id);
3464       }
3465       // At this point, if we're during the concurrent phase of
3466       // marking, everything has been re-initialized and we're
3467       // ready to restart.
3468     }
3469   }
3470 
3471   _claimed = false;
3472 }
3473 
3474 G1CMTask::G1CMTask(uint worker_id,
3475                    G1ConcurrentMark* cm,
3476                    size_t* marked_bytes,
3477                    BitMap* card_bm,
3478                    G1CMTaskQueue* task_queue,
3479                    G1CMTaskQueueSet* task_queues)
3480   : _g1h(G1CollectedHeap::heap()),
3481     _worker_id(worker_id), _cm(cm),
3482     _claimed(false),
3483     _nextMarkBitMap(NULL), _hash_seed(17),
3484     _task_queue(task_queue),
3485     _task_queues(task_queues),
3486     _cm_oop_closure(NULL),
3487     _marked_bytes_array(marked_bytes),
3488     _card_bm(card_bm) {
3489   guarantee(task_queue != NULL, "invariant");
3490   guarantee(task_queues != NULL, "invariant");
3491 
3492   _marking_step_diffs_ms.add(0.5);
3493 }
3494 
3495 // These are formatting macros that are used below to ensure
3496 // consistent formatting. The *_H_* versions are used to format the
3497 // header for a particular value and they should be kept consistent
3498 // with the corresponding macro. Also note that most of the macros add
3499 // the necessary white space (as a prefix) which makes them a bit
3500 // easier to compose.
3501 
3502 // All the output lines are prefixed with this string to be able to
3503 // identify them easily in a large log file.
3504 #define G1PPRL_LINE_PREFIX            "###"
3505 
3506 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
3507 #ifdef _LP64
3508 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
3509 #else // _LP64
3510 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
3511 #endif // _LP64
3512 
3513 // For per-region info
3514 #define G1PPRL_TYPE_FORMAT            "   %-4s"
3515 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
3516 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
3517 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
3518 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
3519 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
3520 
3521 // For summary info
3522 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
3523 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
3524 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
3525 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
3526 
3527 G1PrintRegionLivenessInfoClosure::
3528 G1PrintRegionLivenessInfoClosure(const char* phase_name)
3529   : _total_used_bytes(0), _total_capacity_bytes(0),
3530     _total_prev_live_bytes(0), _total_next_live_bytes(0),
3531     _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
3532   G1CollectedHeap* g1h = G1CollectedHeap::heap();
3533   MemRegion g1_reserved = g1h->g1_reserved();
3534   double now = os::elapsedTime();
3535 
3536   // Print the header of the output.
3537   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
3538   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
3539                           G1PPRL_SUM_ADDR_FORMAT("reserved")
3540                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
3541                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
3542                           HeapRegion::GrainBytes);
3543   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3544   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3545                           G1PPRL_TYPE_H_FORMAT
3546                           G1PPRL_ADDR_BASE_H_FORMAT
3547                           G1PPRL_BYTE_H_FORMAT
3548                           G1PPRL_BYTE_H_FORMAT
3549                           G1PPRL_BYTE_H_FORMAT
3550                           G1PPRL_DOUBLE_H_FORMAT
3551                           G1PPRL_BYTE_H_FORMAT
3552                           G1PPRL_BYTE_H_FORMAT,
3553                           "type", "address-range",
3554                           "used", "prev-live", "next-live", "gc-eff",
3555                           "remset", "code-roots");
3556   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3557                           G1PPRL_TYPE_H_FORMAT
3558                           G1PPRL_ADDR_BASE_H_FORMAT
3559                           G1PPRL_BYTE_H_FORMAT
3560                           G1PPRL_BYTE_H_FORMAT
3561                           G1PPRL_BYTE_H_FORMAT
3562                           G1PPRL_DOUBLE_H_FORMAT
3563                           G1PPRL_BYTE_H_FORMAT
3564                           G1PPRL_BYTE_H_FORMAT,
3565                           "", "",
3566                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3567                           "(bytes)", "(bytes)");
3568 }
3569 
3570 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
3571   const char* type       = r->get_type_str();
3572   HeapWord* bottom       = r->bottom();
3573   HeapWord* end          = r->end();
3574   size_t capacity_bytes  = r->capacity();
3575   size_t used_bytes      = r->used();
3576   size_t prev_live_bytes = r->live_bytes();
3577   size_t next_live_bytes = r->next_live_bytes();
3578   double gc_eff          = r->gc_efficiency();
3579   size_t remset_bytes    = r->rem_set()->mem_size();
3580   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3581 
3582   _total_used_bytes      += used_bytes;
3583   _total_capacity_bytes  += capacity_bytes;
3584   _total_prev_live_bytes += prev_live_bytes;
3585   _total_next_live_bytes += next_live_bytes;
3586   _total_remset_bytes    += remset_bytes;
3587   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3588 
3589   // Print a line for this particular region.
3590   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3591                           G1PPRL_TYPE_FORMAT
3592                           G1PPRL_ADDR_BASE_FORMAT
3593                           G1PPRL_BYTE_FORMAT
3594                           G1PPRL_BYTE_FORMAT
3595                           G1PPRL_BYTE_FORMAT
3596                           G1PPRL_DOUBLE_FORMAT
3597                           G1PPRL_BYTE_FORMAT
3598                           G1PPRL_BYTE_FORMAT,
3599                           type, p2i(bottom), p2i(end),
3600                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3601                           remset_bytes, strong_code_roots_bytes);
3602 
3603   return false;
3604 }
3605 
3606 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3607   // add static memory usages to remembered set sizes
3608   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3609   // Print the footer of the output.
3610   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3611   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3612                          " SUMMARY"
3613                          G1PPRL_SUM_MB_FORMAT("capacity")
3614                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3615                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3616                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3617                          G1PPRL_SUM_MB_FORMAT("remset")
3618                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3619                          bytes_to_mb(_total_capacity_bytes),
3620                          bytes_to_mb(_total_used_bytes),
3621                          perc(_total_used_bytes, _total_capacity_bytes),
3622                          bytes_to_mb(_total_prev_live_bytes),
3623                          perc(_total_prev_live_bytes, _total_capacity_bytes),
3624                          bytes_to_mb(_total_next_live_bytes),
3625                          perc(_total_next_live_bytes, _total_capacity_bytes),
3626                          bytes_to_mb(_total_remset_bytes),
3627                          bytes_to_mb(_total_strong_code_roots_bytes));
3628 }