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/g1CollectorState.hpp"
  32 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  33 #include "gc/g1/g1HeapVerifier.hpp"
  34 #include "gc/g1/g1OopClosures.inline.hpp"
  35 #include "gc/g1/g1CardLiveData.inline.hpp"
  36 #include "gc/g1/g1Policy.hpp"
  37 #include "gc/g1/g1StringDedup.hpp"
  38 #include "gc/g1/heapRegion.inline.hpp"
  39 #include "gc/g1/heapRegionRemSet.hpp"
  40 #include "gc/g1/heapRegionSet.inline.hpp"
  41 #include "gc/g1/suspendibleThreadSet.hpp"
  42 #include "gc/shared/gcId.hpp"
  43 #include "gc/shared/gcTimer.hpp"
  44 #include "gc/shared/gcTrace.hpp"
  45 #include "gc/shared/gcTraceTime.inline.hpp"
  46 #include "gc/shared/genOopClosures.inline.hpp"
  47 #include "gc/shared/referencePolicy.hpp"
  48 #include "gc/shared/strongRootsScope.hpp"
  49 #include "gc/shared/taskqueue.inline.hpp"
  50 #include "gc/shared/vmGCOperations.hpp"
  51 #include "logging/log.hpp"
  52 #include "memory/allocation.hpp"
  53 #include "memory/resourceArea.hpp"
  54 #include "oops/oop.inline.hpp"
  55 #include "runtime/atomic.inline.hpp"
  56 #include "runtime/handles.inline.hpp"
  57 #include "runtime/java.hpp"
  58 #include "runtime/prefetch.inline.hpp"
  59 #include "services/memTracker.hpp"
  60 #include "utilities/growableArray.hpp"
  61 
  62 // Concurrent marking bit map wrapper
  63 
  64 G1CMBitMapRO::G1CMBitMapRO(int shifter) :
  65   _bm(),
  66   _shifter(shifter) {
  67   _bmStartWord = 0;
  68   _bmWordSize = 0;
  69 }
  70 
  71 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
  72                                                  const HeapWord* limit) const {
  73   // First we must round addr *up* to a possible object boundary.
  74   addr = (HeapWord*)align_size_up((intptr_t)addr,
  75                                   HeapWordSize << _shifter);
  76   size_t addrOffset = heapWordToOffset(addr);
  77   assert(limit != NULL, "limit must not be NULL");
  78   size_t limitOffset = heapWordToOffset(limit);
  79   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  80   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  81   assert(nextAddr >= addr, "get_next_one postcondition");
  82   assert(nextAddr == limit || isMarked(nextAddr),
  83          "get_next_one postcondition");
  84   return nextAddr;
  85 }
  86 
  87 #ifndef PRODUCT
  88 bool G1CMBitMapRO::covers(MemRegion heap_rs) const {
  89   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  90   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
  91          "size inconsistency");
  92   return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
  93          _bmWordSize  == heap_rs.word_size();
  94 }
  95 #endif
  96 
  97 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
  98   _bm.print_on_error(st, prefix);
  99 }
 100 
 101 size_t G1CMBitMap::compute_size(size_t heap_size) {
 102   return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
 103 }
 104 
 105 size_t G1CMBitMap::mark_distance() {
 106   return MinObjAlignmentInBytes * BitsPerByte;
 107 }
 108 
 109 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
 110   _bmStartWord = heap.start();
 111   _bmWordSize = heap.word_size();
 112 
 113   _bm = BitMapView((BitMap::bm_word_t*) storage->reserved().start(), _bmWordSize >> _shifter);
 114 
 115   storage->set_mapping_changed_listener(&_listener);
 116 }
 117 
 118 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
 119   if (zero_filled) {
 120     return;
 121   }
 122   // We need to clear the bitmap on commit, removing any existing information.
 123   MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
 124   _bm->clear_range(mr);
 125 }
 126 
 127 void G1CMBitMap::clear_range(MemRegion mr) {
 128   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 129   assert(!mr.is_empty(), "unexpected empty region");
 130   // convert address range into offset range
 131   _bm.at_put_range(heapWordToOffset(mr.start()),
 132                    heapWordToOffset(mr.end()), false);
 133 }
 134 
 135 G1CMMarkStack::G1CMMarkStack(G1ConcurrentMark* cm) :
 136   _base(NULL), _cm(cm)
 137 {}
 138 
 139 bool G1CMMarkStack::allocate(size_t capacity) {
 140   // allocate a stack of the requisite depth
 141   ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
 142   if (!rs.is_reserved()) {
 143     log_warning(gc)("ConcurrentMark MarkStack allocation failure");
 144     return false;
 145   }
 146   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 147   if (!_virtual_space.initialize(rs, rs.size())) {
 148     log_warning(gc)("ConcurrentMark MarkStack backing store failure");
 149     // Release the virtual memory reserved for the marking stack
 150     rs.release();
 151     return false;
 152   }
 153   assert(_virtual_space.committed_size() == rs.size(),
 154          "Didn't reserve backing store for all of G1ConcurrentMark stack?");
 155   _base = (oop*) _virtual_space.low();
 156   setEmpty();
 157   _capacity = (jint) capacity;
 158   _saved_index = -1;
 159   _should_expand = false;
 160   return true;
 161 }
 162 
 163 void G1CMMarkStack::expand() {
 164   // Called, during remark, if we've overflown the marking stack during marking.
 165   assert(isEmpty(), "stack should been emptied while handling overflow");
 166   assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
 167   // Clear expansion flag
 168   _should_expand = false;
 169   if (_capacity == (jint) MarkStackSizeMax) {
 170     log_trace(gc)("(benign) Can't expand marking stack capacity, at max size limit");
 171     return;
 172   }
 173   // Double capacity if possible
 174   jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
 175   // Do not give up existing stack until we have managed to
 176   // get the double capacity that we desired.
 177   ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
 178                                                            sizeof(oop)));
 179   if (rs.is_reserved()) {
 180     // Release the backing store associated with old stack
 181     _virtual_space.release();
 182     // Reinitialize virtual space for new stack
 183     if (!_virtual_space.initialize(rs, rs.size())) {
 184       fatal("Not enough swap for expanded marking stack capacity");
 185     }
 186     _base = (oop*)(_virtual_space.low());
 187     _index = 0;
 188     _capacity = new_capacity;
 189   } else {
 190     // Failed to double capacity, continue;
 191     log_trace(gc)("(benign) Failed to expand marking stack capacity from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
 192                   _capacity / K, new_capacity / K);
 193   }
 194 }
 195 
 196 void G1CMMarkStack::set_should_expand() {
 197   // If we're resetting the marking state because of an
 198   // marking stack overflow, record that we should, if
 199   // possible, expand the stack.
 200   _should_expand = _cm->has_overflown();
 201 }
 202 
 203 G1CMMarkStack::~G1CMMarkStack() {
 204   if (_base != NULL) {
 205     _base = NULL;
 206     _virtual_space.release();
 207   }
 208 }
 209 
 210 void G1CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 211   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 212   jint start = _index;
 213   jint next_index = start + n;
 214   if (next_index > _capacity) {
 215     _overflow = true;
 216     return;
 217   }
 218   // Otherwise.
 219   _index = next_index;
 220   for (int i = 0; i < n; i++) {
 221     int ind = start + i;
 222     assert(ind < _capacity, "By overflow test above.");
 223     _base[ind] = ptr_arr[i];
 224   }
 225 }
 226 
 227 bool G1CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 228   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 229   jint index = _index;
 230   if (index == 0) {
 231     *n = 0;
 232     return false;
 233   } else {
 234     int k = MIN2(max, index);
 235     jint  new_ind = index - k;
 236     for (int j = 0; j < k; j++) {
 237       ptr_arr[j] = _base[new_ind + j];
 238     }
 239     _index = new_ind;
 240     *n = k;
 241     return true;
 242   }
 243 }
 244 
 245 void G1CMMarkStack::note_start_of_gc() {
 246   assert(_saved_index == -1,
 247          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 248   _saved_index = _index;
 249 }
 250 
 251 void G1CMMarkStack::note_end_of_gc() {
 252   // This is intentionally a guarantee, instead of an assert. If we
 253   // accidentally add something to the mark stack during GC, it
 254   // will be a correctness issue so it's better if we crash. we'll
 255   // only check this once per GC anyway, so it won't be a performance
 256   // issue in any way.
 257   guarantee(_saved_index == _index,
 258             "saved index: %d index: %d", _saved_index, _index);
 259   _saved_index = -1;
 260 }
 261 
 262 G1CMRootRegions::G1CMRootRegions() :
 263   _cm(NULL), _scan_in_progress(false),
 264   _should_abort(false), _claimed_survivor_index(0) { }
 265 
 266 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
 267   _survivors = survivors;
 268   _cm = cm;
 269 }
 270 
 271 void G1CMRootRegions::prepare_for_scan() {
 272   assert(!scan_in_progress(), "pre-condition");
 273 
 274   // Currently, only survivors can be root regions.
 275   _claimed_survivor_index = 0;
 276   _scan_in_progress = _survivors->regions()->is_nonempty();
 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   const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
 289 
 290   int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
 291   if (claimed_index < survivor_regions->length()) {
 292     return survivor_regions->at(claimed_index);
 293   }
 294   return NULL;
 295 }
 296 
 297 uint G1CMRootRegions::num_root_regions() const {
 298   return (uint)_survivors->regions()->length();
 299 }
 300 
 301 void G1CMRootRegions::notify_scan_done() {
 302   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 303   _scan_in_progress = false;
 304   RootRegionScan_lock->notify_all();
 305 }
 306 
 307 void G1CMRootRegions::cancel_scan() {
 308   notify_scan_done();
 309 }
 310 
 311 void G1CMRootRegions::scan_finished() {
 312   assert(scan_in_progress(), "pre-condition");
 313 
 314   // Currently, only survivors can be root regions.
 315   if (!_should_abort) {
 316     assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
 317     assert((uint)_claimed_survivor_index >= _survivors->length(),
 318            "we should have claimed all survivors, claimed index = %u, length = %u",
 319            (uint)_claimed_survivor_index, _survivors->length());
 320   }
 321 
 322   notify_scan_done();
 323 }
 324 
 325 bool G1CMRootRegions::wait_until_scan_finished() {
 326   if (!scan_in_progress()) return false;
 327 
 328   {
 329     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 330     while (scan_in_progress()) {
 331       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 332     }
 333   }
 334   return true;
 335 }
 336 
 337 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 338   return MAX2((n_par_threads + 2) / 4, 1U);
 339 }
 340 
 341 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
 342   _g1h(g1h),
 343   _markBitMap1(),
 344   _markBitMap2(),
 345   _parallel_marking_threads(0),
 346   _max_parallel_marking_threads(0),
 347   _sleep_factor(0.0),
 348   _marking_task_overhead(1.0),
 349   _cleanup_list("Cleanup List"),
 350 
 351   _prevMarkBitMap(&_markBitMap1),
 352   _nextMarkBitMap(&_markBitMap2),
 353 
 354   _markStack(this),
 355   // _finger set in set_non_marking_state
 356 
 357   _max_worker_id(ParallelGCThreads),
 358   // _active_tasks set in set_non_marking_state
 359   // _tasks set inside the constructor
 360   _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)),
 361   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 362 
 363   _has_overflown(false),
 364   _concurrent(false),
 365   _has_aborted(false),
 366   _restart_for_overflow(false),
 367   _concurrent_marking_in_progress(false),
 368   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 369   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 370 
 371   // _verbose_level set below
 372 
 373   _init_times(),
 374   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 375   _cleanup_times(),
 376   _total_counting_time(0.0),
 377   _total_rs_scrub_time(0.0),
 378 
 379   _parallel_workers(NULL),
 380 
 381   _completed_initialization(false) {
 382 
 383   _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 384   _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
 385 
 386   // Create & start a ConcurrentMark thread.
 387   _cmThread = new ConcurrentMarkThread(this);
 388   assert(cmThread() != NULL, "CM Thread should have been created");
 389   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 390   if (_cmThread->osthread() == NULL) {
 391       vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 392   }
 393 
 394   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 395   assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
 396   assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
 397 
 398   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 399   satb_qs.set_buffer_size(G1SATBBufferSize);
 400 
 401   _root_regions.init(_g1h->survivor(), this);
 402 
 403   if (ConcGCThreads > ParallelGCThreads) {
 404     log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).",
 405                     ConcGCThreads, ParallelGCThreads);
 406     return;
 407   }
 408   if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
 409     // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
 410     // if both are set
 411     _sleep_factor             = 0.0;
 412     _marking_task_overhead    = 1.0;
 413   } else if (G1MarkingOverheadPercent > 0) {
 414     // We will calculate the number of parallel marking threads based
 415     // on a target overhead with respect to the soft real-time goal
 416     double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 417     double overall_cm_overhead =
 418       (double) MaxGCPauseMillis * marking_overhead /
 419       (double) GCPauseIntervalMillis;
 420     double cpu_ratio = 1.0 / (double) os::processor_count();
 421     double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 422     double marking_task_overhead =
 423       overall_cm_overhead / marking_thread_num *
 424                                               (double) os::processor_count();
 425     double sleep_factor =
 426                        (1.0 - marking_task_overhead) / marking_task_overhead;
 427 
 428     FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num);
 429     _sleep_factor             = sleep_factor;
 430     _marking_task_overhead    = marking_task_overhead;
 431   } else {
 432     // Calculate the number of parallel marking threads by scaling
 433     // the number of parallel GC threads.
 434     uint marking_thread_num = scale_parallel_threads(ParallelGCThreads);
 435     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 436     _sleep_factor             = 0.0;
 437     _marking_task_overhead    = 1.0;
 438   }
 439 
 440   assert(ConcGCThreads > 0, "Should have been set");
 441   log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
 442   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 443   _parallel_marking_threads = ConcGCThreads;
 444   _max_parallel_marking_threads = _parallel_marking_threads;
 445 
 446   _parallel_workers = new WorkGang("G1 Marker",
 447        _max_parallel_marking_threads, false, true);
 448   if (_parallel_workers == NULL) {
 449     vm_exit_during_initialization("Failed necessary allocation.");
 450   } else {
 451     _parallel_workers->initialize_workers();
 452   }
 453 
 454   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 455     size_t mark_stack_size =
 456       MIN2(MarkStackSizeMax,
 457           MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 458     // Verify that the calculated value for MarkStackSize is in range.
 459     // It would be nice to use the private utility routine from Arguments.
 460     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 461       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 462                       "must be between 1 and " SIZE_FORMAT,
 463                       mark_stack_size, MarkStackSizeMax);
 464       return;
 465     }
 466     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 467   } else {
 468     // Verify MarkStackSize is in range.
 469     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 470       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 471         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 472           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 473                           "must be between 1 and " SIZE_FORMAT,
 474                           MarkStackSize, MarkStackSizeMax);
 475           return;
 476         }
 477       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 478         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 479           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 480                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 481                           MarkStackSize, MarkStackSizeMax);
 482           return;
 483         }
 484       }
 485     }
 486   }
 487 
 488   if (!_markStack.allocate(MarkStackSize)) {
 489     log_warning(gc)("Failed to allocate CM marking stack");
 490     return;
 491   }
 492 
 493   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC);
 494   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 495 
 496   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 497   _active_tasks = _max_worker_id;
 498 
 499   for (uint i = 0; i < _max_worker_id; ++i) {
 500     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 501     task_queue->initialize();
 502     _task_queues->register_queue(i, task_queue);
 503 
 504     _tasks[i] = new G1CMTask(i, this, task_queue, _task_queues);
 505 
 506     _accum_task_vtime[i] = 0.0;
 507   }
 508 
 509   // so that the call below can read a sensible value
 510   _heap_start = g1h->reserved_region().start();
 511   set_non_marking_state();
 512   _completed_initialization = true;
 513 }
 514 
 515 void G1ConcurrentMark::reset() {
 516   // Starting values for these two. This should be called in a STW
 517   // phase.
 518   MemRegion reserved = _g1h->g1_reserved();
 519   _heap_start = reserved.start();
 520   _heap_end   = reserved.end();
 521 
 522   // Separated the asserts so that we know which one fires.
 523   assert(_heap_start != NULL, "heap bounds should look ok");
 524   assert(_heap_end != NULL, "heap bounds should look ok");
 525   assert(_heap_start < _heap_end, "heap bounds should look ok");
 526 
 527   // Reset all the marking data structures and any necessary flags
 528   reset_marking_state();
 529 
 530   // We do reset all of them, since different phases will use
 531   // different number of active threads. So, it's easiest to have all
 532   // of them ready.
 533   for (uint i = 0; i < _max_worker_id; ++i) {
 534     _tasks[i]->reset(_nextMarkBitMap);
 535   }
 536 
 537   // we need this to make sure that the flag is on during the evac
 538   // pause with initial mark piggy-backed
 539   set_concurrent_marking_in_progress();
 540 }
 541 
 542 
 543 void G1ConcurrentMark::reset_marking_state(bool clear_overflow) {
 544   _markStack.set_should_expand();
 545   _markStack.setEmpty();        // Also clears the _markStack overflow flag
 546   if (clear_overflow) {
 547     clear_has_overflown();
 548   } else {
 549     assert(has_overflown(), "pre-condition");
 550   }
 551   _finger = _heap_start;
 552 
 553   for (uint i = 0; i < _max_worker_id; ++i) {
 554     G1CMTaskQueue* queue = _task_queues->queue(i);
 555     queue->set_empty();
 556   }
 557 }
 558 
 559 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 560   assert(active_tasks <= _max_worker_id, "we should not have more");
 561 
 562   _active_tasks = active_tasks;
 563   // Need to update the three data structures below according to the
 564   // number of active threads for this phase.
 565   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 566   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 567   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 568 }
 569 
 570 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 571   set_concurrency(active_tasks);
 572 
 573   _concurrent = concurrent;
 574   // We propagate this to all tasks, not just the active ones.
 575   for (uint i = 0; i < _max_worker_id; ++i)
 576     _tasks[i]->set_concurrent(concurrent);
 577 
 578   if (concurrent) {
 579     set_concurrent_marking_in_progress();
 580   } else {
 581     // We currently assume that the concurrent flag has been set to
 582     // false before we start remark. At this point we should also be
 583     // in a STW phase.
 584     assert(!concurrent_marking_in_progress(), "invariant");
 585     assert(out_of_regions(),
 586            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 587            p2i(_finger), p2i(_heap_end));
 588   }
 589 }
 590 
 591 void G1ConcurrentMark::set_non_marking_state() {
 592   // We set the global marking state to some default values when we're
 593   // not doing marking.
 594   reset_marking_state();
 595   _active_tasks = 0;
 596   clear_concurrent_marking_in_progress();
 597 }
 598 
 599 G1ConcurrentMark::~G1ConcurrentMark() {
 600   // The G1ConcurrentMark instance is never freed.
 601   ShouldNotReachHere();
 602 }
 603 
 604 class G1ClearBitMapTask : public AbstractGangTask {
 605 public:
 606   static size_t chunk_size() { return M; }
 607 
 608 private:
 609   // Heap region closure used for clearing the given mark bitmap.
 610   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 611   private:
 612     G1CMBitMap* _bitmap;
 613     G1ConcurrentMark* _cm;
 614   public:
 615     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
 616     }
 617 
 618     virtual bool doHeapRegion(HeapRegion* r) {
 619       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 620 
 621       HeapWord* cur = r->bottom();
 622       HeapWord* const end = r->end();
 623 
 624       while (cur < end) {
 625         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 626         _bitmap->clear_range(mr);
 627 
 628         cur += chunk_size_in_words;
 629 
 630         // Abort iteration if after yielding the marking has been aborted.
 631         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 632           return true;
 633         }
 634         // Repeat the asserts from before the start of the closure. We will do them
 635         // as asserts here to minimize their overhead on the product. However, we
 636         // will have them as guarantees at the beginning / end of the bitmap
 637         // clearing to get some checking in the product.
 638         assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant");
 639         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
 640       }
 641       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 642 
 643       return false;
 644     }
 645   };
 646 
 647   G1ClearBitmapHRClosure _cl;
 648   HeapRegionClaimer _hr_claimer;
 649   bool _suspendible; // If the task is suspendible, workers must join the STS.
 650 
 651 public:
 652   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 653     AbstractGangTask("G1 Clear Bitmap"),
 654     _cl(bitmap, suspendible ? cm : NULL),
 655     _hr_claimer(n_workers),
 656     _suspendible(suspendible)
 657   { }
 658 
 659   void work(uint worker_id) {
 660     SuspendibleThreadSetJoiner sts_join(_suspendible);
 661     G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer, true);
 662   }
 663 
 664   bool is_complete() {
 665     return _cl.complete();
 666   }
 667 };
 668 
 669 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 670   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 671 
 672   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 673   size_t const num_chunks = align_size_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 674 
 675   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 676 
 677   G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
 678 
 679   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 680   workers->run_task(&cl, num_workers);
 681   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 682 }
 683 
 684 void G1ConcurrentMark::cleanup_for_next_mark() {
 685   // Make sure that the concurrent mark thread looks to still be in
 686   // the current cycle.
 687   guarantee(cmThread()->during_cycle(), "invariant");
 688 
 689   // We are finishing up the current cycle by clearing the next
 690   // marking bitmap and getting it ready for the next cycle. During
 691   // this time no other cycle can start. So, let's make sure that this
 692   // is the case.
 693   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 694 
 695   clear_bitmap(_nextMarkBitMap, _parallel_workers, true);
 696 
 697   // Clear the live count data. If the marking has been aborted, the abort()
 698   // call already did that.
 699   if (!has_aborted()) {
 700     clear_live_data(_parallel_workers);
 701     DEBUG_ONLY(verify_live_data_clear());
 702   }
 703 
 704   // Repeat the asserts from above.
 705   guarantee(cmThread()->during_cycle(), "invariant");
 706   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 707 }
 708 
 709 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 710   assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint.");
 711   clear_bitmap((G1CMBitMap*)_prevMarkBitMap, workers, false);
 712 }
 713 
 714 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 715   G1CMBitMap* _bitmap;
 716   bool _error;
 717  public:
 718   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 719   }
 720 
 721   virtual bool doHeapRegion(HeapRegion* r) {
 722     // This closure can be called concurrently to the mutator, so we must make sure
 723     // that the result of the getNextMarkedWordAddress() call is compared to the
 724     // value passed to it as limit to detect any found bits.
 725     // end never changes in G1.
 726     HeapWord* end = r->end();
 727     return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
 728   }
 729 };
 730 
 731 bool G1ConcurrentMark::nextMarkBitmapIsClear() {
 732   CheckBitmapClearHRClosure cl(_nextMarkBitMap);
 733   _g1h->heap_region_iterate(&cl);
 734   return cl.complete();
 735 }
 736 
 737 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 738 public:
 739   bool doHeapRegion(HeapRegion* r) {
 740     r->note_start_of_marking();
 741     return false;
 742   }
 743 };
 744 
 745 void G1ConcurrentMark::checkpointRootsInitialPre() {
 746   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 747   G1Policy* g1p = g1h->g1_policy();
 748 
 749   _has_aborted = false;
 750 
 751   // Initialize marking structures. This has to be done in a STW phase.
 752   reset();
 753 
 754   // For each region note start of marking.
 755   NoteStartOfMarkHRClosure startcl;
 756   g1h->heap_region_iterate(&startcl);
 757 }
 758 
 759 
 760 void G1ConcurrentMark::checkpointRootsInitialPost() {
 761   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 762 
 763   // Start Concurrent Marking weak-reference discovery.
 764   ReferenceProcessor* rp = g1h->ref_processor_cm();
 765   // enable ("weak") refs discovery
 766   rp->enable_discovery();
 767   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 768 
 769   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 770   // This is the start of  the marking cycle, we're expected all
 771   // threads to have SATB queues with active set to false.
 772   satb_mq_set.set_active_all_threads(true, /* new active value */
 773                                      false /* expected_active */);
 774 
 775   _root_regions.prepare_for_scan();
 776 
 777   // update_g1_committed() will be called at the end of an evac pause
 778   // when marking is on. So, it's also called at the end of the
 779   // initial-mark pause to update the heap end, if the heap expands
 780   // during it. No need to call it here.
 781 }
 782 
 783 /*
 784  * Notice that in the next two methods, we actually leave the STS
 785  * during the barrier sync and join it immediately afterwards. If we
 786  * do not do this, the following deadlock can occur: one thread could
 787  * be in the barrier sync code, waiting for the other thread to also
 788  * sync up, whereas another one could be trying to yield, while also
 789  * waiting for the other threads to sync up too.
 790  *
 791  * Note, however, that this code is also used during remark and in
 792  * this case we should not attempt to leave / enter the STS, otherwise
 793  * we'll either hit an assert (debug / fastdebug) or deadlock
 794  * (product). So we should only leave / enter the STS if we are
 795  * operating concurrently.
 796  *
 797  * Because the thread that does the sync barrier has left the STS, it
 798  * is possible to be suspended for a Full GC or an evacuation pause
 799  * could occur. This is actually safe, since the entering the sync
 800  * barrier is one of the last things do_marking_step() does, and it
 801  * doesn't manipulate any data structures afterwards.
 802  */
 803 
 804 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 805   bool barrier_aborted;
 806   {
 807     SuspendibleThreadSetLeaver sts_leave(concurrent());
 808     barrier_aborted = !_first_overflow_barrier_sync.enter();
 809   }
 810 
 811   // at this point everyone should have synced up and not be doing any
 812   // more work
 813 
 814   if (barrier_aborted) {
 815     // If the barrier aborted we ignore the overflow condition and
 816     // just abort the whole marking phase as quickly as possible.
 817     return;
 818   }
 819 
 820   // If we're executing the concurrent phase of marking, reset the marking
 821   // state; otherwise the marking state is reset after reference processing,
 822   // during the remark pause.
 823   // If we reset here as a result of an overflow during the remark we will
 824   // see assertion failures from any subsequent set_concurrency_and_phase()
 825   // calls.
 826   if (concurrent()) {
 827     // let the task associated with with worker 0 do this
 828     if (worker_id == 0) {
 829       // task 0 is responsible for clearing the global data structures
 830       // We should be here because of an overflow. During STW we should
 831       // not clear the overflow flag since we rely on it being true when
 832       // we exit this method to abort the pause and restart concurrent
 833       // marking.
 834       reset_marking_state(true /* clear_overflow */);
 835 
 836       log_info(gc, marking)("Concurrent Mark reset for overflow");
 837     }
 838   }
 839 
 840   // after this, each task should reset its own data structures then
 841   // then go into the second barrier
 842 }
 843 
 844 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 845   SuspendibleThreadSetLeaver sts_leave(concurrent());
 846   _second_overflow_barrier_sync.enter();
 847 
 848   // at this point everything should be re-initialized and ready to go
 849 }
 850 
 851 class G1CMConcurrentMarkingTask: public AbstractGangTask {
 852 private:
 853   G1ConcurrentMark*     _cm;
 854   ConcurrentMarkThread* _cmt;
 855 
 856 public:
 857   void work(uint worker_id) {
 858     assert(Thread::current()->is_ConcurrentGC_thread(),
 859            "this should only be done by a conc GC thread");
 860     ResourceMark rm;
 861 
 862     double start_vtime = os::elapsedVTime();
 863 
 864     {
 865       SuspendibleThreadSetJoiner sts_join;
 866 
 867       assert(worker_id < _cm->active_tasks(), "invariant");
 868       G1CMTask* the_task = _cm->task(worker_id);
 869       the_task->record_start_time();
 870       if (!_cm->has_aborted()) {
 871         do {
 872           double start_vtime_sec = os::elapsedVTime();
 873           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
 874 
 875           the_task->do_marking_step(mark_step_duration_ms,
 876                                     true  /* do_termination */,
 877                                     false /* is_serial*/);
 878 
 879           double end_vtime_sec = os::elapsedVTime();
 880           double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
 881           _cm->clear_has_overflown();
 882 
 883           _cm->do_yield_check();
 884 
 885           jlong sleep_time_ms;
 886           if (!_cm->has_aborted() && the_task->has_aborted()) {
 887             sleep_time_ms =
 888               (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
 889             {
 890               SuspendibleThreadSetLeaver sts_leave;
 891               os::sleep(Thread::current(), sleep_time_ms, false);
 892             }
 893           }
 894         } while (!_cm->has_aborted() && the_task->has_aborted());
 895       }
 896       the_task->record_end_time();
 897       guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
 898     }
 899 
 900     double end_vtime = os::elapsedVTime();
 901     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 902   }
 903 
 904   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm,
 905                             ConcurrentMarkThread* cmt) :
 906       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
 907 
 908   ~G1CMConcurrentMarkingTask() { }
 909 };
 910 
 911 // Calculates the number of active workers for a concurrent
 912 // phase.
 913 uint G1ConcurrentMark::calc_parallel_marking_threads() {
 914   uint n_conc_workers = 0;
 915   if (!UseDynamicNumberOfGCThreads ||
 916       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 917        !ForceDynamicNumberOfGCThreads)) {
 918     n_conc_workers = max_parallel_marking_threads();
 919   } else {
 920     n_conc_workers =
 921       AdaptiveSizePolicy::calc_default_active_workers(max_parallel_marking_threads(),
 922                                                       1, /* Minimum workers */
 923                                                       parallel_marking_threads(),
 924                                                       Threads::number_of_non_daemon_threads());
 925     // Don't scale down "n_conc_workers" by scale_parallel_threads() because
 926     // that scaling has already gone into "_max_parallel_marking_threads".
 927   }
 928   assert(n_conc_workers > 0 && n_conc_workers <= max_parallel_marking_threads(),
 929          "Calculated number of workers must be larger than zero and at most the maximum %u, but is %u",
 930          max_parallel_marking_threads(), n_conc_workers);
 931   return n_conc_workers;
 932 }
 933 
 934 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr) {
 935   // Currently, only survivors can be root regions.
 936   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 937   G1RootRegionScanClosure cl(_g1h, this);
 938 
 939   const uintx interval = PrefetchScanIntervalInBytes;
 940   HeapWord* curr = hr->bottom();
 941   const HeapWord* end = hr->top();
 942   while (curr < end) {
 943     Prefetch::read(curr, interval);
 944     oop obj = oop(curr);
 945     int size = obj->oop_iterate_size(&cl);
 946     assert(size == obj->size(), "sanity");
 947     curr += size;
 948   }
 949 }
 950 
 951 class G1CMRootRegionScanTask : public AbstractGangTask {
 952 private:
 953   G1ConcurrentMark* _cm;
 954 
 955 public:
 956   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 957     AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
 958 
 959   void work(uint worker_id) {
 960     assert(Thread::current()->is_ConcurrentGC_thread(),
 961            "this should only be done by a conc GC thread");
 962 
 963     G1CMRootRegions* root_regions = _cm->root_regions();
 964     HeapRegion* hr = root_regions->claim_next();
 965     while (hr != NULL) {
 966       _cm->scanRootRegion(hr);
 967       hr = root_regions->claim_next();
 968     }
 969   }
 970 };
 971 
 972 void G1ConcurrentMark::scan_root_regions() {
 973   // scan_in_progress() will have been set to true only if there was
 974   // at least one root region to scan. So, if it's false, we
 975   // should not attempt to do any further work.
 976   if (root_regions()->scan_in_progress()) {
 977     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 978 
 979     _parallel_marking_threads = MIN2(calc_parallel_marking_threads(),
 980                                      // We distribute work on a per-region basis, so starting
 981                                      // more threads than that is useless.
 982                                      root_regions()->num_root_regions());
 983     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
 984            "Maximum number of marking threads exceeded");
 985 
 986     G1CMRootRegionScanTask task(this);
 987     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
 988                         task.name(), _parallel_marking_threads, root_regions()->num_root_regions());
 989     _parallel_workers->run_task(&task, _parallel_marking_threads);
 990 
 991     // It's possible that has_aborted() is true here without actually
 992     // aborting the survivor scan earlier. This is OK as it's
 993     // mainly used for sanity checking.
 994     root_regions()->scan_finished();
 995   }
 996 }
 997 
 998 void G1ConcurrentMark::concurrent_cycle_start() {
 999   _gc_timer_cm->register_gc_start();
1000 
1001   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
1002 
1003   _g1h->trace_heap_before_gc(_gc_tracer_cm);
1004 }
1005 
1006 void G1ConcurrentMark::concurrent_cycle_end() {
1007   _g1h->trace_heap_after_gc(_gc_tracer_cm);
1008 
1009   if (has_aborted()) {
1010     _gc_tracer_cm->report_concurrent_mode_failure();
1011   }
1012 
1013   _gc_timer_cm->register_gc_end();
1014 
1015   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1016 }
1017 
1018 void G1ConcurrentMark::mark_from_roots() {
1019   // we might be tempted to assert that:
1020   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1021   //        "inconsistent argument?");
1022   // However that wouldn't be right, because it's possible that
1023   // a safepoint is indeed in progress as a younger generation
1024   // stop-the-world GC happens even as we mark in this generation.
1025 
1026   _restart_for_overflow = false;
1027 
1028   // _g1h has _n_par_threads
1029   _parallel_marking_threads = calc_parallel_marking_threads();
1030   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1031     "Maximum number of marking threads exceeded");
1032 
1033   uint active_workers = MAX2(1U, parallel_marking_threads());
1034   assert(active_workers > 0, "Should have been set");
1035 
1036   // Parallel task terminator is set in "set_concurrency_and_phase()"
1037   set_concurrency_and_phase(active_workers, true /* concurrent */);
1038 
1039   G1CMConcurrentMarkingTask markingTask(this, cmThread());
1040   _parallel_workers->set_active_workers(active_workers);
1041   _parallel_workers->run_task(&markingTask);
1042   print_stats();
1043 }
1044 
1045 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1046   // world is stopped at this checkpoint
1047   assert(SafepointSynchronize::is_at_safepoint(),
1048          "world should be stopped");
1049 
1050   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1051 
1052   // If a full collection has happened, we shouldn't do this.
1053   if (has_aborted()) {
1054     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1055     return;
1056   }
1057 
1058   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1059 
1060   if (VerifyDuringGC) {
1061     HandleMark hm;  // handle scope
1062     g1h->prepare_for_verify();
1063     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1064   }
1065   g1h->verifier()->check_bitmaps("Remark Start");
1066 
1067   G1Policy* g1p = g1h->g1_policy();
1068   g1p->record_concurrent_mark_remark_start();
1069 
1070   double start = os::elapsedTime();
1071 
1072   checkpointRootsFinalWork();
1073 
1074   double mark_work_end = os::elapsedTime();
1075 
1076   weakRefsWork(clear_all_soft_refs);
1077 
1078   if (has_overflown()) {
1079     // Oops.  We overflowed.  Restart concurrent marking.
1080     _restart_for_overflow = true;
1081 
1082     // Verify the heap w.r.t. the previous marking bitmap.
1083     if (VerifyDuringGC) {
1084       HandleMark hm;  // handle scope
1085       g1h->prepare_for_verify();
1086       Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)");
1087     }
1088 
1089     // Clear the marking state because we will be restarting
1090     // marking due to overflowing the global mark stack.
1091     reset_marking_state();
1092   } else {
1093     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1094     // We're done with marking.
1095     // This is the end of  the marking cycle, we're expected all
1096     // threads to have SATB queues with active set to true.
1097     satb_mq_set.set_active_all_threads(false, /* new active value */
1098                                        true /* expected_active */);
1099 
1100     if (VerifyDuringGC) {
1101       HandleMark hm;  // handle scope
1102       g1h->prepare_for_verify();
1103       Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)");
1104     }
1105     g1h->verifier()->check_bitmaps("Remark End");
1106     assert(!restart_for_overflow(), "sanity");
1107     // Completely reset the marking state since marking completed
1108     set_non_marking_state();
1109   }
1110 
1111   // Expand the marking stack, if we have to and if we can.
1112   if (_markStack.should_expand()) {
1113     _markStack.expand();
1114   }
1115 
1116   // Statistics
1117   double now = os::elapsedTime();
1118   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1119   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1120   _remark_times.add((now - start) * 1000.0);
1121 
1122   g1p->record_concurrent_mark_remark_end();
1123 
1124   G1CMIsAliveClosure is_alive(g1h);
1125   _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1126 }
1127 
1128 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1129   G1CollectedHeap* _g1;
1130   size_t _freed_bytes;
1131   FreeRegionList* _local_cleanup_list;
1132   uint _old_regions_removed;
1133   uint _humongous_regions_removed;
1134   HRRSCleanupTask* _hrrs_cleanup_task;
1135 
1136 public:
1137   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1138                              FreeRegionList* local_cleanup_list,
1139                              HRRSCleanupTask* hrrs_cleanup_task) :
1140     _g1(g1),
1141     _freed_bytes(0),
1142     _local_cleanup_list(local_cleanup_list),
1143     _old_regions_removed(0),
1144     _humongous_regions_removed(0),
1145     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1146 
1147   size_t freed_bytes() { return _freed_bytes; }
1148   const uint old_regions_removed() { return _old_regions_removed; }
1149   const uint humongous_regions_removed() { return _humongous_regions_removed; }
1150 
1151   bool doHeapRegion(HeapRegion *hr) {
1152     if (hr->is_archive()) {
1153       return false;
1154     }
1155     _g1->reset_gc_time_stamps(hr);
1156     hr->note_end_of_marking();
1157 
1158     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1159       _freed_bytes += hr->used();
1160       hr->set_containing_set(NULL);
1161       if (hr->is_humongous()) {
1162         _humongous_regions_removed++;
1163         _g1->free_humongous_region(hr, _local_cleanup_list, true /* skip_remset */);
1164       } else {
1165         _old_regions_removed++;
1166         _g1->free_region(hr, _local_cleanup_list, true /* skip_remset */);
1167       }
1168     } else {
1169       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1170     }
1171 
1172     return false;
1173   }
1174 };
1175 
1176 class G1ParNoteEndTask: public AbstractGangTask {
1177   friend class G1NoteEndOfConcMarkClosure;
1178 
1179 protected:
1180   G1CollectedHeap* _g1h;
1181   FreeRegionList* _cleanup_list;
1182   HeapRegionClaimer _hrclaimer;
1183 
1184 public:
1185   G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1186       AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1187   }
1188 
1189   void work(uint worker_id) {
1190     FreeRegionList local_cleanup_list("Local Cleanup List");
1191     HRRSCleanupTask hrrs_cleanup_task;
1192     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1193                                            &hrrs_cleanup_task);
1194     _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1195     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1196 
1197     // Now update the lists
1198     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1199     {
1200       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1201       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1202 
1203       // If we iterate over the global cleanup list at the end of
1204       // cleanup to do this printing we will not guarantee to only
1205       // generate output for the newly-reclaimed regions (the list
1206       // might not be empty at the beginning of cleanup; we might
1207       // still be working on its previous contents). So we do the
1208       // printing here, before we append the new regions to the global
1209       // cleanup list.
1210 
1211       G1HRPrinter* hr_printer = _g1h->hr_printer();
1212       if (hr_printer->is_active()) {
1213         FreeRegionListIterator iter(&local_cleanup_list);
1214         while (iter.more_available()) {
1215           HeapRegion* hr = iter.get_next();
1216           hr_printer->cleanup(hr);
1217         }
1218       }
1219 
1220       _cleanup_list->add_ordered(&local_cleanup_list);
1221       assert(local_cleanup_list.is_empty(), "post-condition");
1222 
1223       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1224     }
1225   }
1226 };
1227 
1228 void G1ConcurrentMark::cleanup() {
1229   // world is stopped at this checkpoint
1230   assert(SafepointSynchronize::is_at_safepoint(),
1231          "world should be stopped");
1232   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1233 
1234   // If a full collection has happened, we shouldn't do this.
1235   if (has_aborted()) {
1236     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1237     return;
1238   }
1239 
1240   g1h->verifier()->verify_region_sets_optional();
1241 
1242   if (VerifyDuringGC) {
1243     HandleMark hm;  // handle scope
1244     g1h->prepare_for_verify();
1245     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1246   }
1247   g1h->verifier()->check_bitmaps("Cleanup Start");
1248 
1249   G1Policy* g1p = g1h->g1_policy();
1250   g1p->record_concurrent_mark_cleanup_start();
1251 
1252   double start = os::elapsedTime();
1253 
1254   HeapRegionRemSet::reset_for_cleanup_tasks();
1255 
1256   {
1257     GCTraceTime(Debug, gc)("Finalize Live Data");
1258     finalize_live_data();
1259   }
1260 
1261   if (VerifyDuringGC) {
1262     GCTraceTime(Debug, gc)("Verify Live Data");
1263     verify_live_data();
1264   }
1265 
1266   g1h->collector_state()->set_mark_in_progress(false);
1267 
1268   double count_end = os::elapsedTime();
1269   double this_final_counting_time = (count_end - start);
1270   _total_counting_time += this_final_counting_time;
1271 
1272   if (log_is_enabled(Trace, gc, liveness)) {
1273     G1PrintRegionLivenessInfoClosure cl("Post-Marking");
1274     _g1h->heap_region_iterate(&cl);
1275   }
1276 
1277   // Install newly created mark bitMap as "prev".
1278   swapMarkBitMaps();
1279 
1280   g1h->reset_gc_time_stamp();
1281 
1282   uint n_workers = _g1h->workers()->active_workers();
1283 
1284   // Note end of marking in all heap regions.
1285   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1286   g1h->workers()->run_task(&g1_par_note_end_task);
1287   g1h->check_gc_time_stamps();
1288 
1289   if (!cleanup_list_is_empty()) {
1290     // The cleanup list is not empty, so we'll have to process it
1291     // concurrently. Notify anyone else that might be wanting free
1292     // regions that there will be more free regions coming soon.
1293     g1h->set_free_regions_coming();
1294   }
1295 
1296   // call below, since it affects the metric by which we sort the heap
1297   // regions.
1298   if (G1ScrubRemSets) {
1299     double rs_scrub_start = os::elapsedTime();
1300     g1h->scrub_rem_set();
1301     _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start);
1302   }
1303 
1304   // this will also free any regions totally full of garbage objects,
1305   // and sort the regions.
1306   g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1307 
1308   // Statistics.
1309   double end = os::elapsedTime();
1310   _cleanup_times.add((end - start) * 1000.0);
1311 
1312   // Clean up will have freed any regions completely full of garbage.
1313   // Update the soft reference policy with the new heap occupancy.
1314   Universe::update_heap_info_at_gc();
1315 
1316   if (VerifyDuringGC) {
1317     HandleMark hm;  // handle scope
1318     g1h->prepare_for_verify();
1319     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)");
1320   }
1321 
1322   g1h->verifier()->check_bitmaps("Cleanup End");
1323 
1324   g1h->verifier()->verify_region_sets_optional();
1325 
1326   // We need to make this be a "collection" so any collection pause that
1327   // races with it goes around and waits for completeCleanup to finish.
1328   g1h->increment_total_collections();
1329 
1330   // Clean out dead classes and update Metaspace sizes.
1331   if (ClassUnloadingWithConcurrentMark) {
1332     ClassLoaderDataGraph::purge();
1333   }
1334   MetaspaceGC::compute_new_size();
1335 
1336   // We reclaimed old regions so we should calculate the sizes to make
1337   // sure we update the old gen/space data.
1338   g1h->g1mm()->update_sizes();
1339   g1h->allocation_context_stats().update_after_mark();
1340 }
1341 
1342 void G1ConcurrentMark::complete_cleanup() {
1343   if (has_aborted()) return;
1344 
1345   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1346 
1347   _cleanup_list.verify_optional();
1348   FreeRegionList tmp_free_list("Tmp Free List");
1349 
1350   log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1351                                   "cleanup list has %u entries",
1352                                   _cleanup_list.length());
1353 
1354   // No one else should be accessing the _cleanup_list at this point,
1355   // so it is not necessary to take any locks
1356   while (!_cleanup_list.is_empty()) {
1357     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
1358     assert(hr != NULL, "Got NULL from a non-empty list");
1359     hr->par_clear();
1360     tmp_free_list.add_ordered(hr);
1361 
1362     // Instead of adding one region at a time to the secondary_free_list,
1363     // we accumulate them in the local list and move them a few at a
1364     // time. This also cuts down on the number of notify_all() calls
1365     // we do during this process. We'll also append the local list when
1366     // _cleanup_list is empty (which means we just removed the last
1367     // region from the _cleanup_list).
1368     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
1369         _cleanup_list.is_empty()) {
1370       log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1371                                       "appending %u entries to the secondary_free_list, "
1372                                       "cleanup list still has %u entries",
1373                                       tmp_free_list.length(),
1374                                       _cleanup_list.length());
1375 
1376       {
1377         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
1378         g1h->secondary_free_list_add(&tmp_free_list);
1379         SecondaryFreeList_lock->notify_all();
1380       }
1381 #ifndef PRODUCT
1382       if (G1StressConcRegionFreeing) {
1383         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
1384           os::sleep(Thread::current(), (jlong) 1, false);
1385         }
1386       }
1387 #endif
1388     }
1389   }
1390   assert(tmp_free_list.is_empty(), "post-condition");
1391 }
1392 
1393 // Supporting Object and Oop closures for reference discovery
1394 // and processing in during marking
1395 
1396 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1397   HeapWord* addr = (HeapWord*)obj;
1398   return addr != NULL &&
1399          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
1400 }
1401 
1402 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1403 // Uses the G1CMTask associated with a worker thread (for serial reference
1404 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1405 // trace referent objects.
1406 //
1407 // Using the G1CMTask and embedded local queues avoids having the worker
1408 // threads operating on the global mark stack. This reduces the risk
1409 // of overflowing the stack - which we would rather avoid at this late
1410 // state. Also using the tasks' local queues removes the potential
1411 // of the workers interfering with each other that could occur if
1412 // operating on the global stack.
1413 
1414 class G1CMKeepAliveAndDrainClosure: public OopClosure {
1415   G1ConcurrentMark* _cm;
1416   G1CMTask*         _task;
1417   int               _ref_counter_limit;
1418   int               _ref_counter;
1419   bool              _is_serial;
1420  public:
1421   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1422     _cm(cm), _task(task), _is_serial(is_serial),
1423     _ref_counter_limit(G1RefProcDrainInterval) {
1424     assert(_ref_counter_limit > 0, "sanity");
1425     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1426     _ref_counter = _ref_counter_limit;
1427   }
1428 
1429   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1430   virtual void do_oop(      oop* p) { do_oop_work(p); }
1431 
1432   template <class T> void do_oop_work(T* p) {
1433     if (!_cm->has_overflown()) {
1434       oop obj = oopDesc::load_decode_heap_oop(p);
1435       _task->deal_with_reference(obj);
1436       _ref_counter--;
1437 
1438       if (_ref_counter == 0) {
1439         // We have dealt with _ref_counter_limit references, pushing them
1440         // and objects reachable from them on to the local stack (and
1441         // possibly the global stack). Call G1CMTask::do_marking_step() to
1442         // process these entries.
1443         //
1444         // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1445         // there's nothing more to do (i.e. we're done with the entries that
1446         // were pushed as a result of the G1CMTask::deal_with_reference() calls
1447         // above) or we overflow.
1448         //
1449         // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1450         // flag while there may still be some work to do. (See the comment at
1451         // the beginning of G1CMTask::do_marking_step() for those conditions -
1452         // one of which is reaching the specified time target.) It is only
1453         // when G1CMTask::do_marking_step() returns without setting the
1454         // has_aborted() flag that the marking step has completed.
1455         do {
1456           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1457           _task->do_marking_step(mark_step_duration_ms,
1458                                  false      /* do_termination */,
1459                                  _is_serial);
1460         } while (_task->has_aborted() && !_cm->has_overflown());
1461         _ref_counter = _ref_counter_limit;
1462       }
1463     }
1464   }
1465 };
1466 
1467 // 'Drain' oop closure used by both serial and parallel reference processing.
1468 // Uses the G1CMTask associated with a given worker thread (for serial
1469 // reference processing the G1CMtask for worker 0 is used). Calls the
1470 // do_marking_step routine, with an unbelievably large timeout value,
1471 // to drain the marking data structures of the remaining entries
1472 // added by the 'keep alive' oop closure above.
1473 
1474 class G1CMDrainMarkingStackClosure: public VoidClosure {
1475   G1ConcurrentMark* _cm;
1476   G1CMTask*         _task;
1477   bool              _is_serial;
1478  public:
1479   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1480     _cm(cm), _task(task), _is_serial(is_serial) {
1481     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1482   }
1483 
1484   void do_void() {
1485     do {
1486       // We call G1CMTask::do_marking_step() to completely drain the local
1487       // and global marking stacks of entries pushed by the 'keep alive'
1488       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1489       //
1490       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1491       // if there's nothing more to do (i.e. we've completely drained the
1492       // entries that were pushed as a a result of applying the 'keep alive'
1493       // closure to the entries on the discovered ref lists) or we overflow
1494       // the global marking stack.
1495       //
1496       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1497       // flag while there may still be some work to do. (See the comment at
1498       // the beginning of G1CMTask::do_marking_step() for those conditions -
1499       // one of which is reaching the specified time target.) It is only
1500       // when G1CMTask::do_marking_step() returns without setting the
1501       // has_aborted() flag that the marking step has completed.
1502 
1503       _task->do_marking_step(1000000000.0 /* something very large */,
1504                              true         /* do_termination */,
1505                              _is_serial);
1506     } while (_task->has_aborted() && !_cm->has_overflown());
1507   }
1508 };
1509 
1510 // Implementation of AbstractRefProcTaskExecutor for parallel
1511 // reference processing at the end of G1 concurrent marking
1512 
1513 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
1514 private:
1515   G1CollectedHeap*  _g1h;
1516   G1ConcurrentMark* _cm;
1517   WorkGang*         _workers;
1518   uint              _active_workers;
1519 
1520 public:
1521   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1522                           G1ConcurrentMark* cm,
1523                           WorkGang* workers,
1524                           uint n_workers) :
1525     _g1h(g1h), _cm(cm),
1526     _workers(workers), _active_workers(n_workers) { }
1527 
1528   // Executes the given task using concurrent marking worker threads.
1529   virtual void execute(ProcessTask& task);
1530   virtual void execute(EnqueueTask& task);
1531 };
1532 
1533 class G1CMRefProcTaskProxy: public AbstractGangTask {
1534   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1535   ProcessTask&      _proc_task;
1536   G1CollectedHeap*  _g1h;
1537   G1ConcurrentMark* _cm;
1538 
1539 public:
1540   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1541                        G1CollectedHeap* g1h,
1542                        G1ConcurrentMark* cm) :
1543     AbstractGangTask("Process reference objects in parallel"),
1544     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1545     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1546     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1547   }
1548 
1549   virtual void work(uint worker_id) {
1550     ResourceMark rm;
1551     HandleMark hm;
1552     G1CMTask* task = _cm->task(worker_id);
1553     G1CMIsAliveClosure g1_is_alive(_g1h);
1554     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1555     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1556 
1557     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1558   }
1559 };
1560 
1561 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1562   assert(_workers != NULL, "Need parallel worker threads.");
1563   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1564 
1565   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1566 
1567   // We need to reset the concurrency level before each
1568   // proxy task execution, so that the termination protocol
1569   // and overflow handling in G1CMTask::do_marking_step() knows
1570   // how many workers to wait for.
1571   _cm->set_concurrency(_active_workers);
1572   _workers->run_task(&proc_task_proxy);
1573 }
1574 
1575 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
1576   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1577   EnqueueTask& _enq_task;
1578 
1579 public:
1580   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1581     AbstractGangTask("Enqueue reference objects in parallel"),
1582     _enq_task(enq_task) { }
1583 
1584   virtual void work(uint worker_id) {
1585     _enq_task.work(worker_id);
1586   }
1587 };
1588 
1589 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1590   assert(_workers != NULL, "Need parallel worker threads.");
1591   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1592 
1593   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1594 
1595   // Not strictly necessary but...
1596   //
1597   // We need to reset the concurrency level before each
1598   // proxy task execution, so that the termination protocol
1599   // and overflow handling in G1CMTask::do_marking_step() knows
1600   // how many workers to wait for.
1601   _cm->set_concurrency(_active_workers);
1602   _workers->run_task(&enq_task_proxy);
1603 }
1604 
1605 void G1ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
1606   G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
1607 }
1608 
1609 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
1610   if (has_overflown()) {
1611     // Skip processing the discovered references if we have
1612     // overflown the global marking stack. Reference objects
1613     // only get discovered once so it is OK to not
1614     // de-populate the discovered reference lists. We could have,
1615     // but the only benefit would be that, when marking restarts,
1616     // less reference objects are discovered.
1617     return;
1618   }
1619 
1620   ResourceMark rm;
1621   HandleMark   hm;
1622 
1623   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1624 
1625   // Is alive closure.
1626   G1CMIsAliveClosure g1_is_alive(g1h);
1627 
1628   // Inner scope to exclude the cleaning of the string and symbol
1629   // tables from the displayed time.
1630   {
1631     GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm);
1632 
1633     ReferenceProcessor* rp = g1h->ref_processor_cm();
1634 
1635     // See the comment in G1CollectedHeap::ref_processing_init()
1636     // about how reference processing currently works in G1.
1637 
1638     // Set the soft reference policy
1639     rp->setup_policy(clear_all_soft_refs);
1640     assert(_markStack.isEmpty(), "mark stack should be empty");
1641 
1642     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1643     // in serial reference processing. Note these closures are also
1644     // used for serially processing (by the the current thread) the
1645     // JNI references during parallel reference processing.
1646     //
1647     // These closures do not need to synchronize with the worker
1648     // threads involved in parallel reference processing as these
1649     // instances are executed serially by the current thread (e.g.
1650     // reference processing is not multi-threaded and is thus
1651     // performed by the current thread instead of a gang worker).
1652     //
1653     // The gang tasks involved in parallel reference processing create
1654     // their own instances of these closures, which do their own
1655     // synchronization among themselves.
1656     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1657     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1658 
1659     // We need at least one active thread. If reference processing
1660     // is not multi-threaded we use the current (VMThread) thread,
1661     // otherwise we use the work gang from the G1CollectedHeap and
1662     // we utilize all the worker threads we can.
1663     bool processing_is_mt = rp->processing_is_mt();
1664     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
1665     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
1666 
1667     // Parallel processing task executor.
1668     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
1669                                               g1h->workers(), active_workers);
1670     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1671 
1672     // Set the concurrency level. The phase was already set prior to
1673     // executing the remark task.
1674     set_concurrency(active_workers);
1675 
1676     // Set the degree of MT processing here.  If the discovery was done MT,
1677     // the number of threads involved during discovery could differ from
1678     // the number of active workers.  This is OK as long as the discovered
1679     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1680     rp->set_active_mt_degree(active_workers);
1681 
1682     // Process the weak references.
1683     const ReferenceProcessorStats& stats =
1684         rp->process_discovered_references(&g1_is_alive,
1685                                           &g1_keep_alive,
1686                                           &g1_drain_mark_stack,
1687                                           executor,
1688                                           _gc_timer_cm);
1689     _gc_tracer_cm->report_gc_reference_stats(stats);
1690 
1691     // The do_oop work routines of the keep_alive and drain_marking_stack
1692     // oop closures will set the has_overflown flag if we overflow the
1693     // global marking stack.
1694 
1695     assert(_markStack.overflow() || _markStack.isEmpty(),
1696             "mark stack should be empty (unless it overflowed)");
1697 
1698     if (_markStack.overflow()) {
1699       // This should have been done already when we tried to push an
1700       // entry on to the global mark stack. But let's do it again.
1701       set_has_overflown();
1702     }
1703 
1704     assert(rp->num_q() == active_workers, "why not");
1705 
1706     rp->enqueue_discovered_references(executor);
1707 
1708     rp->verify_no_references_recorded();
1709     assert(!rp->discovery_enabled(), "Post condition");
1710   }
1711 
1712   if (has_overflown()) {
1713     // We can not trust g1_is_alive if the marking stack overflowed
1714     return;
1715   }
1716 
1717   assert(_markStack.isEmpty(), "Marking should have completed");
1718 
1719   // Unload Klasses, String, Symbols, Code Cache, etc.
1720   if (ClassUnloadingWithConcurrentMark) {
1721     bool purged_classes;
1722 
1723     {
1724       GCTraceTime(Debug, gc, phases) trace("System Dictionary Unloading", _gc_timer_cm);
1725       purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
1726     }
1727 
1728     {
1729       GCTraceTime(Debug, gc, phases) trace("Parallel Unloading", _gc_timer_cm);
1730       weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
1731     }
1732   }
1733 
1734   if (G1StringDedup::is_enabled()) {
1735     GCTraceTime(Debug, gc, phases) trace("String Deduplication Unlink", _gc_timer_cm);
1736     G1StringDedup::unlink(&g1_is_alive);
1737   }
1738 }
1739 
1740 void G1ConcurrentMark::swapMarkBitMaps() {
1741   G1CMBitMapRO* temp = _prevMarkBitMap;
1742   _prevMarkBitMap    = (G1CMBitMapRO*)_nextMarkBitMap;
1743   _nextMarkBitMap    = (G1CMBitMap*)  temp;
1744 }
1745 
1746 // Closure for marking entries in SATB buffers.
1747 class G1CMSATBBufferClosure : public SATBBufferClosure {
1748 private:
1749   G1CMTask* _task;
1750   G1CollectedHeap* _g1h;
1751 
1752   // This is very similar to G1CMTask::deal_with_reference, but with
1753   // more relaxed requirements for the argument, so this must be more
1754   // circumspect about treating the argument as an object.
1755   void do_entry(void* entry) const {
1756     _task->increment_refs_reached();
1757     HeapRegion* hr = _g1h->heap_region_containing(entry);
1758     if (entry < hr->next_top_at_mark_start()) {
1759       // Until we get here, we don't know whether entry refers to a valid
1760       // object; it could instead have been a stale reference.
1761       oop obj = static_cast<oop>(entry);
1762       assert(obj->is_oop(true /* ignore mark word */),
1763              "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj));
1764       _task->make_reference_grey(obj);
1765     }
1766   }
1767 
1768 public:
1769   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1770     : _task(task), _g1h(g1h) { }
1771 
1772   virtual void do_buffer(void** buffer, size_t size) {
1773     for (size_t i = 0; i < size; ++i) {
1774       do_entry(buffer[i]);
1775     }
1776   }
1777 };
1778 
1779 class G1RemarkThreadsClosure : public ThreadClosure {
1780   G1CMSATBBufferClosure _cm_satb_cl;
1781   G1CMOopClosure _cm_cl;
1782   MarkingCodeBlobClosure _code_cl;
1783   int _thread_parity;
1784 
1785  public:
1786   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1787     _cm_satb_cl(task, g1h),
1788     _cm_cl(g1h, g1h->concurrent_mark(), task),
1789     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1790     _thread_parity(Threads::thread_claim_parity()) {}
1791 
1792   void do_thread(Thread* thread) {
1793     if (thread->is_Java_thread()) {
1794       if (thread->claim_oops_do(true, _thread_parity)) {
1795         JavaThread* jt = (JavaThread*)thread;
1796 
1797         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1798         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1799         // * Alive if on the stack of an executing method
1800         // * Weakly reachable otherwise
1801         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1802         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1803         jt->nmethods_do(&_code_cl);
1804 
1805         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
1806       }
1807     } else if (thread->is_VM_thread()) {
1808       if (thread->claim_oops_do(true, _thread_parity)) {
1809         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1810       }
1811     }
1812   }
1813 };
1814 
1815 class G1CMRemarkTask: public AbstractGangTask {
1816 private:
1817   G1ConcurrentMark* _cm;
1818 public:
1819   void work(uint worker_id) {
1820     // Since all available tasks are actually started, we should
1821     // only proceed if we're supposed to be active.
1822     if (worker_id < _cm->active_tasks()) {
1823       G1CMTask* task = _cm->task(worker_id);
1824       task->record_start_time();
1825       {
1826         ResourceMark rm;
1827         HandleMark hm;
1828 
1829         G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1830         Threads::threads_do(&threads_f);
1831       }
1832 
1833       do {
1834         task->do_marking_step(1000000000.0 /* something very large */,
1835                               true         /* do_termination       */,
1836                               false        /* is_serial            */);
1837       } while (task->has_aborted() && !_cm->has_overflown());
1838       // If we overflow, then we do not want to restart. We instead
1839       // want to abort remark and do concurrent marking again.
1840       task->record_end_time();
1841     }
1842   }
1843 
1844   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1845     AbstractGangTask("Par Remark"), _cm(cm) {
1846     _cm->terminator()->reset_for_reuse(active_workers);
1847   }
1848 };
1849 
1850 void G1ConcurrentMark::checkpointRootsFinalWork() {
1851   ResourceMark rm;
1852   HandleMark   hm;
1853   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1854 
1855   GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm);
1856 
1857   g1h->ensure_parsability(false);
1858 
1859   // this is remark, so we'll use up all active threads
1860   uint active_workers = g1h->workers()->active_workers();
1861   set_concurrency_and_phase(active_workers, false /* concurrent */);
1862   // Leave _parallel_marking_threads at it's
1863   // value originally calculated in the G1ConcurrentMark
1864   // constructor and pass values of the active workers
1865   // through the gang in the task.
1866 
1867   {
1868     StrongRootsScope srs(active_workers);
1869 
1870     G1CMRemarkTask remarkTask(this, active_workers);
1871     // We will start all available threads, even if we decide that the
1872     // active_workers will be fewer. The extra ones will just bail out
1873     // immediately.
1874     g1h->workers()->run_task(&remarkTask);
1875   }
1876 
1877   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1878   guarantee(has_overflown() ||
1879             satb_mq_set.completed_buffers_num() == 0,
1880             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1881             BOOL_TO_STR(has_overflown()),
1882             satb_mq_set.completed_buffers_num());
1883 
1884   print_stats();
1885 }
1886 
1887 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
1888   // Note we are overriding the read-only view of the prev map here, via
1889   // the cast.
1890   ((G1CMBitMap*)_prevMarkBitMap)->clear_range(mr);
1891 }
1892 
1893 HeapRegion*
1894 G1ConcurrentMark::claim_region(uint worker_id) {
1895   // "checkpoint" the finger
1896   HeapWord* finger = _finger;
1897 
1898   // _heap_end will not change underneath our feet; it only changes at
1899   // yield points.
1900   while (finger < _heap_end) {
1901     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1902 
1903     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1904 
1905     // Above heap_region_containing may return NULL as we always scan claim
1906     // until the end of the heap. In this case, just jump to the next region.
1907     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1908 
1909     // Is the gap between reading the finger and doing the CAS too long?
1910     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
1911     if (res == finger && curr_region != NULL) {
1912       // we succeeded
1913       HeapWord*   bottom        = curr_region->bottom();
1914       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1915 
1916       // notice that _finger == end cannot be guaranteed here since,
1917       // someone else might have moved the finger even further
1918       assert(_finger >= end, "the finger should have moved forward");
1919 
1920       if (limit > bottom) {
1921         return curr_region;
1922       } else {
1923         assert(limit == bottom,
1924                "the region limit should be at bottom");
1925         // we return NULL and the caller should try calling
1926         // claim_region() again.
1927         return NULL;
1928       }
1929     } else {
1930       assert(_finger > finger, "the finger should have moved forward");
1931       // read it again
1932       finger = _finger;
1933     }
1934   }
1935 
1936   return NULL;
1937 }
1938 
1939 #ifndef PRODUCT
1940 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
1941 private:
1942   G1CollectedHeap* _g1h;
1943   const char* _phase;
1944   int _info;
1945 
1946 public:
1947   VerifyNoCSetOops(const char* phase, int info = -1) :
1948     _g1h(G1CollectedHeap::heap()),
1949     _phase(phase),
1950     _info(info)
1951   { }
1952 
1953   void operator()(oop obj) const {
1954     guarantee(obj->is_oop(),
1955               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1956               p2i(obj), _phase, _info);
1957     guarantee(!_g1h->obj_in_cs(obj),
1958               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1959               p2i(obj), _phase, _info);
1960   }
1961 };
1962 
1963 void G1ConcurrentMark::verify_no_cset_oops() {
1964   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1965   if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
1966     return;
1967   }
1968 
1969   // Verify entries on the global mark stack
1970   _markStack.iterate(VerifyNoCSetOops("Stack"));
1971 
1972   // Verify entries on the task queues
1973   for (uint i = 0; i < _max_worker_id; ++i) {
1974     G1CMTaskQueue* queue = _task_queues->queue(i);
1975     queue->iterate(VerifyNoCSetOops("Queue", i));
1976   }
1977 
1978   // Verify the global finger
1979   HeapWord* global_finger = finger();
1980   if (global_finger != NULL && global_finger < _heap_end) {
1981     // Since we always iterate over all regions, we might get a NULL HeapRegion
1982     // here.
1983     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1984     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1985               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1986               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1987   }
1988 
1989   // Verify the task fingers
1990   assert(parallel_marking_threads() <= _max_worker_id, "sanity");
1991   for (uint i = 0; i < parallel_marking_threads(); ++i) {
1992     G1CMTask* task = _tasks[i];
1993     HeapWord* task_finger = task->finger();
1994     if (task_finger != NULL && task_finger < _heap_end) {
1995       // See above note on the global finger verification.
1996       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
1997       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
1998                 !task_hr->in_collection_set(),
1999                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2000                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
2001     }
2002   }
2003 }
2004 #endif // PRODUCT
2005 void G1ConcurrentMark::create_live_data() {
2006   _g1h->g1_rem_set()->create_card_live_data(_parallel_workers, _nextMarkBitMap);
2007 }
2008 
2009 void G1ConcurrentMark::finalize_live_data() {
2010   _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _nextMarkBitMap);
2011 }
2012 
2013 void G1ConcurrentMark::verify_live_data() {
2014   _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _nextMarkBitMap);
2015 }
2016 
2017 void G1ConcurrentMark::clear_live_data(WorkGang* workers) {
2018   _g1h->g1_rem_set()->clear_card_live_data(workers);
2019 }
2020 
2021 #ifdef ASSERT
2022 void G1ConcurrentMark::verify_live_data_clear() {
2023   _g1h->g1_rem_set()->verify_card_live_data_is_clear();
2024 }
2025 #endif
2026 
2027 void G1ConcurrentMark::print_stats() {
2028   if (!log_is_enabled(Debug, gc, stats)) {
2029     return;
2030   }
2031   log_debug(gc, stats)("---------------------------------------------------------------------");
2032   for (size_t i = 0; i < _active_tasks; ++i) {
2033     _tasks[i]->print_stats();
2034     log_debug(gc, stats)("---------------------------------------------------------------------");
2035   }
2036 }
2037 
2038 void G1ConcurrentMark::abort() {
2039   if (!cmThread()->during_cycle() || _has_aborted) {
2040     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2041     return;
2042   }
2043 
2044   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2045   // concurrent bitmap clearing.
2046   {
2047     GCTraceTime(Debug, gc)("Clear Next Bitmap");
2048     clear_bitmap(_nextMarkBitMap, _g1h->workers(), false);
2049   }
2050   // Note we cannot clear the previous marking bitmap here
2051   // since VerifyDuringGC verifies the objects marked during
2052   // a full GC against the previous bitmap.
2053 
2054   {
2055     GCTraceTime(Debug, gc)("Clear Live Data");
2056     clear_live_data(_g1h->workers());
2057   }
2058   DEBUG_ONLY({
2059     GCTraceTime(Debug, gc)("Verify Live Data Clear");
2060     verify_live_data_clear();
2061   })
2062   // Empty mark stack
2063   reset_marking_state();
2064   for (uint i = 0; i < _max_worker_id; ++i) {
2065     _tasks[i]->clear_region_fields();
2066   }
2067   _first_overflow_barrier_sync.abort();
2068   _second_overflow_barrier_sync.abort();
2069   _has_aborted = true;
2070 
2071   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2072   satb_mq_set.abandon_partial_marking();
2073   // This can be called either during or outside marking, we'll read
2074   // the expected_active value from the SATB queue set.
2075   satb_mq_set.set_active_all_threads(
2076                                  false, /* new active value */
2077                                  satb_mq_set.is_active() /* expected_active */);
2078 }
2079 
2080 static void print_ms_time_info(const char* prefix, const char* name,
2081                                NumberSeq& ns) {
2082   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2083                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2084   if (ns.num() > 0) {
2085     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2086                            prefix, ns.sd(), ns.maximum());
2087   }
2088 }
2089 
2090 void G1ConcurrentMark::print_summary_info() {
2091   Log(gc, marking) log;
2092   if (!log.is_trace()) {
2093     return;
2094   }
2095 
2096   log.trace(" Concurrent marking:");
2097   print_ms_time_info("  ", "init marks", _init_times);
2098   print_ms_time_info("  ", "remarks", _remark_times);
2099   {
2100     print_ms_time_info("     ", "final marks", _remark_mark_times);
2101     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2102 
2103   }
2104   print_ms_time_info("  ", "cleanups", _cleanup_times);
2105   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2106             _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2107   if (G1ScrubRemSets) {
2108     log.trace("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
2109               _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2110   }
2111   log.trace("  Total stop_world time = %8.2f s.",
2112             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2113   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2114             cmThread()->vtime_accum(), cmThread()->vtime_mark_accum());
2115 }
2116 
2117 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2118   _parallel_workers->print_worker_threads_on(st);
2119 }
2120 
2121 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2122   _parallel_workers->threads_do(tc);
2123 }
2124 
2125 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2126   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2127       p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
2128   _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
2129   _nextMarkBitMap->print_on_error(st, " Next Bits: ");
2130 }
2131 
2132 // Closure for iteration over bitmaps
2133 class G1CMBitMapClosure : public BitMapClosure {
2134 private:
2135   // the bitmap that is being iterated over
2136   G1CMBitMap*                 _nextMarkBitMap;
2137   G1ConcurrentMark*           _cm;
2138   G1CMTask*                   _task;
2139 
2140 public:
2141   G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm, G1CMBitMap* nextMarkBitMap) :
2142     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
2143 
2144   bool do_bit(size_t offset) {
2145     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
2146     assert(_nextMarkBitMap->isMarked(addr), "invariant");
2147     assert( addr < _cm->finger(), "invariant");
2148     assert(addr >= _task->finger(), "invariant");
2149 
2150     // We move that task's local finger along.
2151     _task->move_finger_to(addr);
2152 
2153     _task->scan_object(oop(addr));
2154     // we only partially drain the local queue and global stack
2155     _task->drain_local_queue(true);
2156     _task->drain_global_stack(true);
2157 
2158     // if the has_aborted flag has been raised, we need to bail out of
2159     // the iteration
2160     return !_task->has_aborted();
2161   }
2162 };
2163 
2164 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2165   ReferenceProcessor* result = g1h->ref_processor_cm();
2166   assert(result != NULL, "CM reference processor should not be NULL");
2167   return result;
2168 }
2169 
2170 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2171                                G1ConcurrentMark* cm,
2172                                G1CMTask* task)
2173   : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2174     _g1h(g1h), _cm(cm), _task(task)
2175 { }
2176 
2177 void G1CMTask::setup_for_region(HeapRegion* hr) {
2178   assert(hr != NULL,
2179         "claim_region() should have filtered out NULL regions");
2180   _curr_region  = hr;
2181   _finger       = hr->bottom();
2182   update_region_limit();
2183 }
2184 
2185 void G1CMTask::update_region_limit() {
2186   HeapRegion* hr            = _curr_region;
2187   HeapWord* bottom          = hr->bottom();
2188   HeapWord* limit           = hr->next_top_at_mark_start();
2189 
2190   if (limit == bottom) {
2191     // The region was collected underneath our feet.
2192     // We set the finger to bottom to ensure that the bitmap
2193     // iteration that will follow this will not do anything.
2194     // (this is not a condition that holds when we set the region up,
2195     // as the region is not supposed to be empty in the first place)
2196     _finger = bottom;
2197   } else if (limit >= _region_limit) {
2198     assert(limit >= _finger, "peace of mind");
2199   } else {
2200     assert(limit < _region_limit, "only way to get here");
2201     // This can happen under some pretty unusual circumstances.  An
2202     // evacuation pause empties the region underneath our feet (NTAMS
2203     // at bottom). We then do some allocation in the region (NTAMS
2204     // stays at bottom), followed by the region being used as a GC
2205     // alloc region (NTAMS will move to top() and the objects
2206     // originally below it will be grayed). All objects now marked in
2207     // the region are explicitly grayed, if below the global finger,
2208     // and we do not need in fact to scan anything else. So, we simply
2209     // set _finger to be limit to ensure that the bitmap iteration
2210     // doesn't do anything.
2211     _finger = limit;
2212   }
2213 
2214   _region_limit = limit;
2215 }
2216 
2217 void G1CMTask::giveup_current_region() {
2218   assert(_curr_region != NULL, "invariant");
2219   clear_region_fields();
2220 }
2221 
2222 void G1CMTask::clear_region_fields() {
2223   // Values for these three fields that indicate that we're not
2224   // holding on to a region.
2225   _curr_region   = NULL;
2226   _finger        = NULL;
2227   _region_limit  = NULL;
2228 }
2229 
2230 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2231   if (cm_oop_closure == NULL) {
2232     assert(_cm_oop_closure != NULL, "invariant");
2233   } else {
2234     assert(_cm_oop_closure == NULL, "invariant");
2235   }
2236   _cm_oop_closure = cm_oop_closure;
2237 }
2238 
2239 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) {
2240   guarantee(nextMarkBitMap != NULL, "invariant");
2241   _nextMarkBitMap                = nextMarkBitMap;
2242   clear_region_fields();
2243 
2244   _calls                         = 0;
2245   _elapsed_time_ms               = 0.0;
2246   _termination_time_ms           = 0.0;
2247   _termination_start_time_ms     = 0.0;
2248 }
2249 
2250 bool G1CMTask::should_exit_termination() {
2251   regular_clock_call();
2252   // This is called when we are in the termination protocol. We should
2253   // quit if, for some reason, this task wants to abort or the global
2254   // stack is not empty (this means that we can get work from it).
2255   return !_cm->mark_stack_empty() || has_aborted();
2256 }
2257 
2258 void G1CMTask::reached_limit() {
2259   assert(_words_scanned >= _words_scanned_limit ||
2260          _refs_reached >= _refs_reached_limit ,
2261          "shouldn't have been called otherwise");
2262   regular_clock_call();
2263 }
2264 
2265 void G1CMTask::regular_clock_call() {
2266   if (has_aborted()) return;
2267 
2268   // First, we need to recalculate the words scanned and refs reached
2269   // limits for the next clock call.
2270   recalculate_limits();
2271 
2272   // During the regular clock call we do the following
2273 
2274   // (1) If an overflow has been flagged, then we abort.
2275   if (_cm->has_overflown()) {
2276     set_has_aborted();
2277     return;
2278   }
2279 
2280   // If we are not concurrent (i.e. we're doing remark) we don't need
2281   // to check anything else. The other steps are only needed during
2282   // the concurrent marking phase.
2283   if (!concurrent()) return;
2284 
2285   // (2) If marking has been aborted for Full GC, then we also abort.
2286   if (_cm->has_aborted()) {
2287     set_has_aborted();
2288     return;
2289   }
2290 
2291   double curr_time_ms = os::elapsedVTime() * 1000.0;
2292 
2293   // (4) We check whether we should yield. If we have to, then we abort.
2294   if (SuspendibleThreadSet::should_yield()) {
2295     // We should yield. To do this we abort the task. The caller is
2296     // responsible for yielding.
2297     set_has_aborted();
2298     return;
2299   }
2300 
2301   // (5) We check whether we've reached our time quota. If we have,
2302   // then we abort.
2303   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2304   if (elapsed_time_ms > _time_target_ms) {
2305     set_has_aborted();
2306     _has_timed_out = true;
2307     return;
2308   }
2309 
2310   // (6) Finally, we check whether there are enough completed STAB
2311   // buffers available for processing. If there are, we abort.
2312   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2313   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2314     // we do need to process SATB buffers, we'll abort and restart
2315     // the marking task to do so
2316     set_has_aborted();
2317     return;
2318   }
2319 }
2320 
2321 void G1CMTask::recalculate_limits() {
2322   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2323   _words_scanned_limit      = _real_words_scanned_limit;
2324 
2325   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2326   _refs_reached_limit       = _real_refs_reached_limit;
2327 }
2328 
2329 void G1CMTask::decrease_limits() {
2330   // This is called when we believe that we're going to do an infrequent
2331   // operation which will increase the per byte scanned cost (i.e. move
2332   // entries to/from the global stack). It basically tries to decrease the
2333   // scanning limit so that the clock is called earlier.
2334 
2335   _words_scanned_limit = _real_words_scanned_limit -
2336     3 * words_scanned_period / 4;
2337   _refs_reached_limit  = _real_refs_reached_limit -
2338     3 * refs_reached_period / 4;
2339 }
2340 
2341 void G1CMTask::move_entries_to_global_stack() {
2342   // local array where we'll store the entries that will be popped
2343   // from the local queue
2344   oop buffer[global_stack_transfer_size];
2345 
2346   int n = 0;
2347   oop obj;
2348   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
2349     buffer[n] = obj;
2350     ++n;
2351   }
2352 
2353   if (n > 0) {
2354     // we popped at least one entry from the local queue
2355 
2356     if (!_cm->mark_stack_push(buffer, n)) {
2357       set_has_aborted();
2358     }
2359   }
2360 
2361   // this operation was quite expensive, so decrease the limits
2362   decrease_limits();
2363 }
2364 
2365 void G1CMTask::get_entries_from_global_stack() {
2366   // local array where we'll store the entries that will be popped
2367   // from the global stack.
2368   oop buffer[global_stack_transfer_size];
2369   int n;
2370   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
2371   assert(n <= global_stack_transfer_size,
2372          "we should not pop more than the given limit");
2373   if (n > 0) {
2374     // yes, we did actually pop at least one entry
2375     for (int i = 0; i < n; ++i) {
2376       bool success = _task_queue->push(buffer[i]);
2377       // We only call this when the local queue is empty or under a
2378       // given target limit. So, we do not expect this push to fail.
2379       assert(success, "invariant");
2380     }
2381   }
2382 
2383   // this operation was quite expensive, so decrease the limits
2384   decrease_limits();
2385 }
2386 
2387 void G1CMTask::drain_local_queue(bool partially) {
2388   if (has_aborted()) return;
2389 
2390   // Decide what the target size is, depending whether we're going to
2391   // drain it partially (so that other tasks can steal if they run out
2392   // of things to do) or totally (at the very end).
2393   size_t target_size;
2394   if (partially) {
2395     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2396   } else {
2397     target_size = 0;
2398   }
2399 
2400   if (_task_queue->size() > target_size) {
2401     oop obj;
2402     bool ret = _task_queue->pop_local(obj);
2403     while (ret) {
2404       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
2405       assert(!_g1h->is_on_master_free_list(
2406                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
2407 
2408       scan_object(obj);
2409 
2410       if (_task_queue->size() <= target_size || has_aborted()) {
2411         ret = false;
2412       } else {
2413         ret = _task_queue->pop_local(obj);
2414       }
2415     }
2416   }
2417 }
2418 
2419 void G1CMTask::drain_global_stack(bool partially) {
2420   if (has_aborted()) return;
2421 
2422   // We have a policy to drain the local queue before we attempt to
2423   // drain the global stack.
2424   assert(partially || _task_queue->size() == 0, "invariant");
2425 
2426   // Decide what the target size is, depending whether we're going to
2427   // drain it partially (so that other tasks can steal if they run out
2428   // of things to do) or totally (at the very end).  Notice that,
2429   // because we move entries from the global stack in chunks or
2430   // because another task might be doing the same, we might in fact
2431   // drop below the target. But, this is not a problem.
2432   size_t target_size;
2433   if (partially) {
2434     target_size = _cm->partial_mark_stack_size_target();
2435   } else {
2436     target_size = 0;
2437   }
2438 
2439   if (_cm->mark_stack_size() > target_size) {
2440     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2441       get_entries_from_global_stack();
2442       drain_local_queue(partially);
2443     }
2444   }
2445 }
2446 
2447 // SATB Queue has several assumptions on whether to call the par or
2448 // non-par versions of the methods. this is why some of the code is
2449 // replicated. We should really get rid of the single-threaded version
2450 // of the code to simplify things.
2451 void G1CMTask::drain_satb_buffers() {
2452   if (has_aborted()) return;
2453 
2454   // We set this so that the regular clock knows that we're in the
2455   // middle of draining buffers and doesn't set the abort flag when it
2456   // notices that SATB buffers are available for draining. It'd be
2457   // very counter productive if it did that. :-)
2458   _draining_satb_buffers = true;
2459 
2460   G1CMSATBBufferClosure satb_cl(this, _g1h);
2461   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2462 
2463   // This keeps claiming and applying the closure to completed buffers
2464   // until we run out of buffers or we need to abort.
2465   while (!has_aborted() &&
2466          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2467     regular_clock_call();
2468   }
2469 
2470   _draining_satb_buffers = false;
2471 
2472   assert(has_aborted() ||
2473          concurrent() ||
2474          satb_mq_set.completed_buffers_num() == 0, "invariant");
2475 
2476   // again, this was a potentially expensive operation, decrease the
2477   // limits to get the regular clock call early
2478   decrease_limits();
2479 }
2480 
2481 void G1CMTask::print_stats() {
2482   log_debug(gc, stats)("Marking Stats, task = %u, calls = %d",
2483                        _worker_id, _calls);
2484   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2485                        _elapsed_time_ms, _termination_time_ms);
2486   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
2487                        _step_times_ms.num(), _step_times_ms.avg(),
2488                        _step_times_ms.sd());
2489   log_debug(gc, stats)("                    max = %1.2lfms, total = %1.2lfms",
2490                        _step_times_ms.maximum(), _step_times_ms.sum());
2491 }
2492 
2493 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) {
2494   return _task_queues->steal(worker_id, hash_seed, obj);
2495 }
2496 
2497 /*****************************************************************************
2498 
2499     The do_marking_step(time_target_ms, ...) method is the building
2500     block of the parallel marking framework. It can be called in parallel
2501     with other invocations of do_marking_step() on different tasks
2502     (but only one per task, obviously) and concurrently with the
2503     mutator threads, or during remark, hence it eliminates the need
2504     for two versions of the code. When called during remark, it will
2505     pick up from where the task left off during the concurrent marking
2506     phase. Interestingly, tasks are also claimable during evacuation
2507     pauses too, since do_marking_step() ensures that it aborts before
2508     it needs to yield.
2509 
2510     The data structures that it uses to do marking work are the
2511     following:
2512 
2513       (1) Marking Bitmap. If there are gray objects that appear only
2514       on the bitmap (this happens either when dealing with an overflow
2515       or when the initial marking phase has simply marked the roots
2516       and didn't push them on the stack), then tasks claim heap
2517       regions whose bitmap they then scan to find gray objects. A
2518       global finger indicates where the end of the last claimed region
2519       is. A local finger indicates how far into the region a task has
2520       scanned. The two fingers are used to determine how to gray an
2521       object (i.e. whether simply marking it is OK, as it will be
2522       visited by a task in the future, or whether it needs to be also
2523       pushed on a stack).
2524 
2525       (2) Local Queue. The local queue of the task which is accessed
2526       reasonably efficiently by the task. Other tasks can steal from
2527       it when they run out of work. Throughout the marking phase, a
2528       task attempts to keep its local queue short but not totally
2529       empty, so that entries are available for stealing by other
2530       tasks. Only when there is no more work, a task will totally
2531       drain its local queue.
2532 
2533       (3) Global Mark Stack. This handles local queue overflow. During
2534       marking only sets of entries are moved between it and the local
2535       queues, as access to it requires a mutex and more fine-grain
2536       interaction with it which might cause contention. If it
2537       overflows, then the marking phase should restart and iterate
2538       over the bitmap to identify gray objects. Throughout the marking
2539       phase, tasks attempt to keep the global mark stack at a small
2540       length but not totally empty, so that entries are available for
2541       popping by other tasks. Only when there is no more work, tasks
2542       will totally drain the global mark stack.
2543 
2544       (4) SATB Buffer Queue. This is where completed SATB buffers are
2545       made available. Buffers are regularly removed from this queue
2546       and scanned for roots, so that the queue doesn't get too
2547       long. During remark, all completed buffers are processed, as
2548       well as the filled in parts of any uncompleted buffers.
2549 
2550     The do_marking_step() method tries to abort when the time target
2551     has been reached. There are a few other cases when the
2552     do_marking_step() method also aborts:
2553 
2554       (1) When the marking phase has been aborted (after a Full GC).
2555 
2556       (2) When a global overflow (on the global stack) has been
2557       triggered. Before the task aborts, it will actually sync up with
2558       the other tasks to ensure that all the marking data structures
2559       (local queues, stacks, fingers etc.)  are re-initialized so that
2560       when do_marking_step() completes, the marking phase can
2561       immediately restart.
2562 
2563       (3) When enough completed SATB buffers are available. The
2564       do_marking_step() method only tries to drain SATB buffers right
2565       at the beginning. So, if enough buffers are available, the
2566       marking step aborts and the SATB buffers are processed at
2567       the beginning of the next invocation.
2568 
2569       (4) To yield. when we have to yield then we abort and yield
2570       right at the end of do_marking_step(). This saves us from a lot
2571       of hassle as, by yielding we might allow a Full GC. If this
2572       happens then objects will be compacted underneath our feet, the
2573       heap might shrink, etc. We save checking for this by just
2574       aborting and doing the yield right at the end.
2575 
2576     From the above it follows that the do_marking_step() method should
2577     be called in a loop (or, otherwise, regularly) until it completes.
2578 
2579     If a marking step completes without its has_aborted() flag being
2580     true, it means it has completed the current marking phase (and
2581     also all other marking tasks have done so and have all synced up).
2582 
2583     A method called regular_clock_call() is invoked "regularly" (in
2584     sub ms intervals) throughout marking. It is this clock method that
2585     checks all the abort conditions which were mentioned above and
2586     decides when the task should abort. A work-based scheme is used to
2587     trigger this clock method: when the number of object words the
2588     marking phase has scanned or the number of references the marking
2589     phase has visited reach a given limit. Additional invocations to
2590     the method clock have been planted in a few other strategic places
2591     too. The initial reason for the clock method was to avoid calling
2592     vtime too regularly, as it is quite expensive. So, once it was in
2593     place, it was natural to piggy-back all the other conditions on it
2594     too and not constantly check them throughout the code.
2595 
2596     If do_termination is true then do_marking_step will enter its
2597     termination protocol.
2598 
2599     The value of is_serial must be true when do_marking_step is being
2600     called serially (i.e. by the VMThread) and do_marking_step should
2601     skip any synchronization in the termination and overflow code.
2602     Examples include the serial remark code and the serial reference
2603     processing closures.
2604 
2605     The value of is_serial must be false when do_marking_step is
2606     being called by any of the worker threads in a work gang.
2607     Examples include the concurrent marking code (CMMarkingTask),
2608     the MT remark code, and the MT reference processing closures.
2609 
2610  *****************************************************************************/
2611 
2612 void G1CMTask::do_marking_step(double time_target_ms,
2613                                bool do_termination,
2614                                bool is_serial) {
2615   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2616   assert(concurrent() == _cm->concurrent(), "they should be the same");
2617 
2618   G1Policy* g1_policy = _g1h->g1_policy();
2619   assert(_task_queues != NULL, "invariant");
2620   assert(_task_queue != NULL, "invariant");
2621   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
2622 
2623   assert(!_claimed,
2624          "only one thread should claim this task at any one time");
2625 
2626   // OK, this doesn't safeguard again all possible scenarios, as it is
2627   // possible for two threads to set the _claimed flag at the same
2628   // time. But it is only for debugging purposes anyway and it will
2629   // catch most problems.
2630   _claimed = true;
2631 
2632   _start_time_ms = os::elapsedVTime() * 1000.0;
2633 
2634   // If do_stealing is true then do_marking_step will attempt to
2635   // steal work from the other G1CMTasks. It only makes sense to
2636   // enable stealing when the termination protocol is enabled
2637   // and do_marking_step() is not being called serially.
2638   bool do_stealing = do_termination && !is_serial;
2639 
2640   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2641   _time_target_ms = time_target_ms - diff_prediction_ms;
2642 
2643   // set up the variables that are used in the work-based scheme to
2644   // call the regular clock method
2645   _words_scanned = 0;
2646   _refs_reached  = 0;
2647   recalculate_limits();
2648 
2649   // clear all flags
2650   clear_has_aborted();
2651   _has_timed_out = false;
2652   _draining_satb_buffers = false;
2653 
2654   ++_calls;
2655 
2656   // Set up the bitmap and oop closures. Anything that uses them is
2657   // eventually called from this method, so it is OK to allocate these
2658   // statically.
2659   G1CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
2660   G1CMOopClosure    cm_oop_closure(_g1h, _cm, this);
2661   set_cm_oop_closure(&cm_oop_closure);
2662 
2663   if (_cm->has_overflown()) {
2664     // This can happen if the mark stack overflows during a GC pause
2665     // and this task, after a yield point, restarts. We have to abort
2666     // as we need to get into the overflow protocol which happens
2667     // right at the end of this task.
2668     set_has_aborted();
2669   }
2670 
2671   // First drain any available SATB buffers. After this, we will not
2672   // look at SATB buffers before the next invocation of this method.
2673   // If enough completed SATB buffers are queued up, the regular clock
2674   // will abort this task so that it restarts.
2675   drain_satb_buffers();
2676   // ...then partially drain the local queue and the global stack
2677   drain_local_queue(true);
2678   drain_global_stack(true);
2679 
2680   do {
2681     if (!has_aborted() && _curr_region != NULL) {
2682       // This means that we're already holding on to a region.
2683       assert(_finger != NULL, "if region is not NULL, then the finger "
2684              "should not be NULL either");
2685 
2686       // We might have restarted this task after an evacuation pause
2687       // which might have evacuated the region we're holding on to
2688       // underneath our feet. Let's read its limit again to make sure
2689       // that we do not iterate over a region of the heap that
2690       // contains garbage (update_region_limit() will also move
2691       // _finger to the start of the region if it is found empty).
2692       update_region_limit();
2693       // We will start from _finger not from the start of the region,
2694       // as we might be restarting this task after aborting half-way
2695       // through scanning this region. In this case, _finger points to
2696       // the address where we last found a marked object. If this is a
2697       // fresh region, _finger points to start().
2698       MemRegion mr = MemRegion(_finger, _region_limit);
2699 
2700       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2701              "humongous regions should go around loop once only");
2702 
2703       // Some special cases:
2704       // If the memory region is empty, we can just give up the region.
2705       // If the current region is humongous then we only need to check
2706       // the bitmap for the bit associated with the start of the object,
2707       // scan the object if it's live, and give up the region.
2708       // Otherwise, let's iterate over the bitmap of the part of the region
2709       // that is left.
2710       // If the iteration is successful, give up the region.
2711       if (mr.is_empty()) {
2712         giveup_current_region();
2713         regular_clock_call();
2714       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2715         if (_nextMarkBitMap->isMarked(mr.start())) {
2716           // The object is marked - apply the closure
2717           BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
2718           bitmap_closure.do_bit(offset);
2719         }
2720         // Even if this task aborted while scanning the humongous object
2721         // we can (and should) give up the current region.
2722         giveup_current_region();
2723         regular_clock_call();
2724       } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
2725         giveup_current_region();
2726         regular_clock_call();
2727       } else {
2728         assert(has_aborted(), "currently the only way to do so");
2729         // The only way to abort the bitmap iteration is to return
2730         // false from the do_bit() method. However, inside the
2731         // do_bit() method we move the _finger to point to the
2732         // object currently being looked at. So, if we bail out, we
2733         // have definitely set _finger to something non-null.
2734         assert(_finger != NULL, "invariant");
2735 
2736         // Region iteration was actually aborted. So now _finger
2737         // points to the address of the object we last scanned. If we
2738         // leave it there, when we restart this task, we will rescan
2739         // the object. It is easy to avoid this. We move the finger by
2740         // enough to point to the next possible object header (the
2741         // bitmap knows by how much we need to move it as it knows its
2742         // granularity).
2743         assert(_finger < _region_limit, "invariant");
2744         HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
2745         // Check if bitmap iteration was aborted while scanning the last object
2746         if (new_finger >= _region_limit) {
2747           giveup_current_region();
2748         } else {
2749           move_finger_to(new_finger);
2750         }
2751       }
2752     }
2753     // At this point we have either completed iterating over the
2754     // region we were holding on to, or we have aborted.
2755 
2756     // We then partially drain the local queue and the global stack.
2757     // (Do we really need this?)
2758     drain_local_queue(true);
2759     drain_global_stack(true);
2760 
2761     // Read the note on the claim_region() method on why it might
2762     // return NULL with potentially more regions available for
2763     // claiming and why we have to check out_of_regions() to determine
2764     // whether we're done or not.
2765     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2766       // We are going to try to claim a new region. We should have
2767       // given up on the previous one.
2768       // Separated the asserts so that we know which one fires.
2769       assert(_curr_region  == NULL, "invariant");
2770       assert(_finger       == NULL, "invariant");
2771       assert(_region_limit == NULL, "invariant");
2772       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2773       if (claimed_region != NULL) {
2774         // Yes, we managed to claim one
2775         setup_for_region(claimed_region);
2776         assert(_curr_region == claimed_region, "invariant");
2777       }
2778       // It is important to call the regular clock here. It might take
2779       // a while to claim a region if, for example, we hit a large
2780       // block of empty regions. So we need to call the regular clock
2781       // method once round the loop to make sure it's called
2782       // frequently enough.
2783       regular_clock_call();
2784     }
2785 
2786     if (!has_aborted() && _curr_region == NULL) {
2787       assert(_cm->out_of_regions(),
2788              "at this point we should be out of regions");
2789     }
2790   } while ( _curr_region != NULL && !has_aborted());
2791 
2792   if (!has_aborted()) {
2793     // We cannot check whether the global stack is empty, since other
2794     // tasks might be pushing objects to it concurrently.
2795     assert(_cm->out_of_regions(),
2796            "at this point we should be out of regions");
2797     // Try to reduce the number of available SATB buffers so that
2798     // remark has less work to do.
2799     drain_satb_buffers();
2800   }
2801 
2802   // Since we've done everything else, we can now totally drain the
2803   // local queue and global stack.
2804   drain_local_queue(false);
2805   drain_global_stack(false);
2806 
2807   // Attempt at work stealing from other task's queues.
2808   if (do_stealing && !has_aborted()) {
2809     // We have not aborted. This means that we have finished all that
2810     // we could. Let's try to do some stealing...
2811 
2812     // We cannot check whether the global stack is empty, since other
2813     // tasks might be pushing objects to it concurrently.
2814     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2815            "only way to reach here");
2816     while (!has_aborted()) {
2817       oop obj;
2818       if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
2819         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
2820                "any stolen object should be marked");
2821         scan_object(obj);
2822 
2823         // And since we're towards the end, let's totally drain the
2824         // local queue and global stack.
2825         drain_local_queue(false);
2826         drain_global_stack(false);
2827       } else {
2828         break;
2829       }
2830     }
2831   }
2832 
2833   // We still haven't aborted. Now, let's try to get into the
2834   // termination protocol.
2835   if (do_termination && !has_aborted()) {
2836     // We cannot check whether the global stack is empty, since other
2837     // tasks might be concurrently pushing objects on it.
2838     // Separated the asserts so that we know which one fires.
2839     assert(_cm->out_of_regions(), "only way to reach here");
2840     assert(_task_queue->size() == 0, "only way to reach here");
2841     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2842 
2843     // The G1CMTask class also extends the TerminatorTerminator class,
2844     // hence its should_exit_termination() method will also decide
2845     // whether to exit the termination protocol or not.
2846     bool finished = (is_serial ||
2847                      _cm->terminator()->offer_termination(this));
2848     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2849     _termination_time_ms +=
2850       termination_end_time_ms - _termination_start_time_ms;
2851 
2852     if (finished) {
2853       // We're all done.
2854 
2855       if (_worker_id == 0) {
2856         // let's allow task 0 to do this
2857         if (concurrent()) {
2858           assert(_cm->concurrent_marking_in_progress(), "invariant");
2859           // we need to set this to false before the next
2860           // safepoint. This way we ensure that the marking phase
2861           // doesn't observe any more heap expansions.
2862           _cm->clear_concurrent_marking_in_progress();
2863         }
2864       }
2865 
2866       // We can now guarantee that the global stack is empty, since
2867       // all other tasks have finished. We separated the guarantees so
2868       // that, if a condition is false, we can immediately find out
2869       // which one.
2870       guarantee(_cm->out_of_regions(), "only way to reach here");
2871       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2872       guarantee(_task_queue->size() == 0, "only way to reach here");
2873       guarantee(!_cm->has_overflown(), "only way to reach here");
2874       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
2875     } else {
2876       // Apparently there's more work to do. Let's abort this task. It
2877       // will restart it and we can hopefully find more things to do.
2878       set_has_aborted();
2879     }
2880   }
2881 
2882   // Mainly for debugging purposes to make sure that a pointer to the
2883   // closure which was statically allocated in this frame doesn't
2884   // escape it by accident.
2885   set_cm_oop_closure(NULL);
2886   double end_time_ms = os::elapsedVTime() * 1000.0;
2887   double elapsed_time_ms = end_time_ms - _start_time_ms;
2888   // Update the step history.
2889   _step_times_ms.add(elapsed_time_ms);
2890 
2891   if (has_aborted()) {
2892     // The task was aborted for some reason.
2893     if (_has_timed_out) {
2894       double diff_ms = elapsed_time_ms - _time_target_ms;
2895       // Keep statistics of how well we did with respect to hitting
2896       // our target only if we actually timed out (if we aborted for
2897       // other reasons, then the results might get skewed).
2898       _marking_step_diffs_ms.add(diff_ms);
2899     }
2900 
2901     if (_cm->has_overflown()) {
2902       // This is the interesting one. We aborted because a global
2903       // overflow was raised. This means we have to restart the
2904       // marking phase and start iterating over regions. However, in
2905       // order to do this we have to make sure that all tasks stop
2906       // what they are doing and re-initialize in a safe manner. We
2907       // will achieve this with the use of two barrier sync points.
2908 
2909       if (!is_serial) {
2910         // We only need to enter the sync barrier if being called
2911         // from a parallel context
2912         _cm->enter_first_sync_barrier(_worker_id);
2913 
2914         // When we exit this sync barrier we know that all tasks have
2915         // stopped doing marking work. So, it's now safe to
2916         // re-initialize our data structures. At the end of this method,
2917         // task 0 will clear the global data structures.
2918       }
2919 
2920       // We clear the local state of this task...
2921       clear_region_fields();
2922 
2923       if (!is_serial) {
2924         // ...and enter the second barrier.
2925         _cm->enter_second_sync_barrier(_worker_id);
2926       }
2927       // At this point, if we're during the concurrent phase of
2928       // marking, everything has been re-initialized and we're
2929       // ready to restart.
2930     }
2931   }
2932 
2933   _claimed = false;
2934 }
2935 
2936 G1CMTask::G1CMTask(uint worker_id,
2937                    G1ConcurrentMark* cm,
2938                    G1CMTaskQueue* task_queue,
2939                    G1CMTaskQueueSet* task_queues)
2940   : _g1h(G1CollectedHeap::heap()),
2941     _worker_id(worker_id), _cm(cm),
2942     _claimed(false),
2943     _nextMarkBitMap(NULL), _hash_seed(17),
2944     _task_queue(task_queue),
2945     _task_queues(task_queues),
2946     _cm_oop_closure(NULL) {
2947   guarantee(task_queue != NULL, "invariant");
2948   guarantee(task_queues != NULL, "invariant");
2949 
2950   _marking_step_diffs_ms.add(0.5);
2951 }
2952 
2953 // These are formatting macros that are used below to ensure
2954 // consistent formatting. The *_H_* versions are used to format the
2955 // header for a particular value and they should be kept consistent
2956 // with the corresponding macro. Also note that most of the macros add
2957 // the necessary white space (as a prefix) which makes them a bit
2958 // easier to compose.
2959 
2960 // All the output lines are prefixed with this string to be able to
2961 // identify them easily in a large log file.
2962 #define G1PPRL_LINE_PREFIX            "###"
2963 
2964 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2965 #ifdef _LP64
2966 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2967 #else // _LP64
2968 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2969 #endif // _LP64
2970 
2971 // For per-region info
2972 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2973 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2974 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2975 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2976 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2977 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2978 
2979 // For summary info
2980 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2981 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2982 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2983 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2984 
2985 G1PrintRegionLivenessInfoClosure::
2986 G1PrintRegionLivenessInfoClosure(const char* phase_name)
2987   : _total_used_bytes(0), _total_capacity_bytes(0),
2988     _total_prev_live_bytes(0), _total_next_live_bytes(0),
2989     _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
2990   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2991   MemRegion g1_reserved = g1h->g1_reserved();
2992   double now = os::elapsedTime();
2993 
2994   // Print the header of the output.
2995   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2996   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2997                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2998                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2999                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
3000                           HeapRegion::GrainBytes);
3001   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3002   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3003                           G1PPRL_TYPE_H_FORMAT
3004                           G1PPRL_ADDR_BASE_H_FORMAT
3005                           G1PPRL_BYTE_H_FORMAT
3006                           G1PPRL_BYTE_H_FORMAT
3007                           G1PPRL_BYTE_H_FORMAT
3008                           G1PPRL_DOUBLE_H_FORMAT
3009                           G1PPRL_BYTE_H_FORMAT
3010                           G1PPRL_BYTE_H_FORMAT,
3011                           "type", "address-range",
3012                           "used", "prev-live", "next-live", "gc-eff",
3013                           "remset", "code-roots");
3014   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3015                           G1PPRL_TYPE_H_FORMAT
3016                           G1PPRL_ADDR_BASE_H_FORMAT
3017                           G1PPRL_BYTE_H_FORMAT
3018                           G1PPRL_BYTE_H_FORMAT
3019                           G1PPRL_BYTE_H_FORMAT
3020                           G1PPRL_DOUBLE_H_FORMAT
3021                           G1PPRL_BYTE_H_FORMAT
3022                           G1PPRL_BYTE_H_FORMAT,
3023                           "", "",
3024                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3025                           "(bytes)", "(bytes)");
3026 }
3027 
3028 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
3029   const char* type       = r->get_type_str();
3030   HeapWord* bottom       = r->bottom();
3031   HeapWord* end          = r->end();
3032   size_t capacity_bytes  = r->capacity();
3033   size_t used_bytes      = r->used();
3034   size_t prev_live_bytes = r->live_bytes();
3035   size_t next_live_bytes = r->next_live_bytes();
3036   double gc_eff          = r->gc_efficiency();
3037   size_t remset_bytes    = r->rem_set()->mem_size();
3038   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3039 
3040   _total_used_bytes      += used_bytes;
3041   _total_capacity_bytes  += capacity_bytes;
3042   _total_prev_live_bytes += prev_live_bytes;
3043   _total_next_live_bytes += next_live_bytes;
3044   _total_remset_bytes    += remset_bytes;
3045   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3046 
3047   // Print a line for this particular region.
3048   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3049                           G1PPRL_TYPE_FORMAT
3050                           G1PPRL_ADDR_BASE_FORMAT
3051                           G1PPRL_BYTE_FORMAT
3052                           G1PPRL_BYTE_FORMAT
3053                           G1PPRL_BYTE_FORMAT
3054                           G1PPRL_DOUBLE_FORMAT
3055                           G1PPRL_BYTE_FORMAT
3056                           G1PPRL_BYTE_FORMAT,
3057                           type, p2i(bottom), p2i(end),
3058                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3059                           remset_bytes, strong_code_roots_bytes);
3060 
3061   return false;
3062 }
3063 
3064 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3065   // add static memory usages to remembered set sizes
3066   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3067   // Print the footer of the output.
3068   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3069   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3070                          " SUMMARY"
3071                          G1PPRL_SUM_MB_FORMAT("capacity")
3072                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3073                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3074                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3075                          G1PPRL_SUM_MB_FORMAT("remset")
3076                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3077                          bytes_to_mb(_total_capacity_bytes),
3078                          bytes_to_mb(_total_used_bytes),
3079                          perc(_total_used_bytes, _total_capacity_bytes),
3080                          bytes_to_mb(_total_prev_live_bytes),
3081                          perc(_total_prev_live_bytes, _total_capacity_bytes),
3082                          bytes_to_mb(_total_next_live_bytes),
3083                          perc(_total_next_live_bytes, _total_capacity_bytes),
3084                          bytes_to_mb(_total_remset_bytes),
3085                          bytes_to_mb(_total_strong_code_roots_bytes));
3086 }