1 /*
   2  * Copyright (c) 2001, 2015, 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/concurrentMark.inline.hpp"
  30 #include "gc/g1/concurrentMarkThread.inline.hpp"
  31 #include "gc/g1/g1CollectedHeap.inline.hpp"
  32 #include "gc/g1/g1CollectorPolicy.hpp"
  33 #include "gc/g1/g1CollectorState.hpp"
  34 #include "gc/g1/g1ErgoVerbose.hpp"
  35 #include "gc/g1/g1Log.hpp"
  36 #include "gc/g1/g1OopClosures.inline.hpp"
  37 #include "gc/g1/g1RemSet.hpp"
  38 #include "gc/g1/g1StringDedup.hpp"
  39 #include "gc/g1/heapRegion.inline.hpp"
  40 #include "gc/g1/heapRegionManager.inline.hpp"
  41 #include "gc/g1/heapRegionRemSet.hpp"
  42 #include "gc/g1/heapRegionSet.inline.hpp"
  43 #include "gc/g1/suspendibleThreadSet.hpp"
  44 #include "gc/shared/gcId.hpp"
  45 #include "gc/shared/gcTimer.hpp"
  46 #include "gc/shared/gcTrace.hpp"
  47 #include "gc/shared/gcTraceTime.hpp"
  48 #include "gc/shared/genOopClosures.inline.hpp"
  49 #include "gc/shared/referencePolicy.hpp"
  50 #include "gc/shared/strongRootsScope.hpp"
  51 #include "gc/shared/taskqueue.inline.hpp"
  52 #include "gc/shared/vmGCOperations.hpp"
  53 #include "memory/allocation.hpp"
  54 #include "memory/resourceArea.hpp"
  55 #include "oops/oop.inline.hpp"
  56 #include "runtime/atomic.inline.hpp"
  57 #include "runtime/handles.inline.hpp"
  58 #include "runtime/java.hpp"
  59 #include "runtime/prefetch.inline.hpp"
  60 #include "services/memTracker.hpp"
  61 
  62 // Concurrent marking bit map wrapper
  63 
  64 CMBitMapRO::CMBitMapRO(int shifter) :
  65   _bm(),
  66   _shifter(shifter) {
  67   _bmStartWord = 0;
  68   _bmWordSize = 0;
  69 }
  70 
  71 HeapWord* CMBitMapRO::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   if (limit == NULL) {
  78     limit = _bmStartWord + _bmWordSize;
  79   }
  80   size_t limitOffset = heapWordToOffset(limit);
  81   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  82   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  83   assert(nextAddr >= addr, "get_next_one postcondition");
  84   assert(nextAddr == limit || isMarked(nextAddr),
  85          "get_next_one postcondition");
  86   return nextAddr;
  87 }
  88 
  89 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
  90                                                  const HeapWord* limit) const {
  91   size_t addrOffset = heapWordToOffset(addr);
  92   if (limit == NULL) {
  93     limit = _bmStartWord + _bmWordSize;
  94   }
  95   size_t limitOffset = heapWordToOffset(limit);
  96   size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
  97   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  98   assert(nextAddr >= addr, "get_next_one postcondition");
  99   assert(nextAddr == limit || !isMarked(nextAddr),
 100          "get_next_one postcondition");
 101   return nextAddr;
 102 }
 103 
 104 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
 105   assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
 106   return (int) (diff >> _shifter);
 107 }
 108 
 109 #ifndef PRODUCT
 110 bool CMBitMapRO::covers(MemRegion heap_rs) const {
 111   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
 112   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
 113          "size inconsistency");
 114   return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
 115          _bmWordSize  == heap_rs.word_size();
 116 }
 117 #endif
 118 
 119 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
 120   _bm.print_on_error(st, prefix);
 121 }
 122 
 123 size_t CMBitMap::compute_size(size_t heap_size) {
 124   return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
 125 }
 126 
 127 size_t CMBitMap::mark_distance() {
 128   return MinObjAlignmentInBytes * BitsPerByte;
 129 }
 130 
 131 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
 132   _bmStartWord = heap.start();
 133   _bmWordSize = heap.word_size();
 134 
 135   _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
 136   _bm.set_size(_bmWordSize >> _shifter);
 137 
 138   storage->set_mapping_changed_listener(&_listener);
 139 }
 140 
 141 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
 142   if (zero_filled) {
 143     return;
 144   }
 145   // We need to clear the bitmap on commit, removing any existing information.
 146   MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
 147   _bm->clearRange(mr);
 148 }
 149 
 150 // Closure used for clearing the given mark bitmap.
 151 class ClearBitmapHRClosure : public HeapRegionClosure {
 152  private:
 153   ConcurrentMark* _cm;
 154   CMBitMap* _bitmap;
 155   bool _may_yield;      // The closure may yield during iteration. If yielded, abort the iteration.
 156  public:
 157   ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
 158     assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
 159   }
 160 
 161   virtual bool doHeapRegion(HeapRegion* r) {
 162     size_t const chunk_size_in_words = M / HeapWordSize;
 163 
 164     HeapWord* cur = r->bottom();
 165     HeapWord* const end = r->end();
 166 
 167     while (cur < end) {
 168       MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 169       _bitmap->clearRange(mr);
 170 
 171       cur += chunk_size_in_words;
 172 
 173       // Abort iteration if after yielding the marking has been aborted.
 174       if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
 175         return true;
 176       }
 177       // Repeat the asserts from before the start of the closure. We will do them
 178       // as asserts here to minimize their overhead on the product. However, we
 179       // will have them as guarantees at the beginning / end of the bitmap
 180       // clearing to get some checking in the product.
 181       assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
 182       assert(!_may_yield || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
 183     }
 184 
 185     return false;
 186   }
 187 };
 188 
 189 class ParClearNextMarkBitmapTask : public AbstractGangTask {
 190   ClearBitmapHRClosure* _cl;
 191   HeapRegionClaimer     _hrclaimer;
 192   bool                  _suspendible; // If the task is suspendible, workers must join the STS.
 193 
 194 public:
 195   ParClearNextMarkBitmapTask(ClearBitmapHRClosure *cl, uint n_workers, bool suspendible) :
 196       _cl(cl), _suspendible(suspendible), AbstractGangTask("Parallel Clear Bitmap Task"), _hrclaimer(n_workers) {}
 197 
 198   void work(uint worker_id) {
 199     SuspendibleThreadSetJoiner sts_join(_suspendible);
 200     G1CollectedHeap::heap()->heap_region_par_iterate(_cl, worker_id, &_hrclaimer, true);
 201   }
 202 };
 203 
 204 void CMBitMap::clearAll() {
 205   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 206   ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
 207   uint n_workers = g1h->workers()->active_workers();
 208   ParClearNextMarkBitmapTask task(&cl, n_workers, false);
 209   g1h->workers()->run_task(&task);
 210   guarantee(cl.complete(), "Must have completed iteration.");
 211   return;
 212 }
 213 
 214 void CMBitMap::markRange(MemRegion mr) {
 215   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 216   assert(!mr.is_empty(), "unexpected empty region");
 217   assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
 218           ((HeapWord *) mr.end())),
 219          "markRange memory region end is not card aligned");
 220   // convert address range into offset range
 221   _bm.at_put_range(heapWordToOffset(mr.start()),
 222                    heapWordToOffset(mr.end()), true);
 223 }
 224 
 225 void CMBitMap::clearRange(MemRegion mr) {
 226   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 227   assert(!mr.is_empty(), "unexpected empty region");
 228   // convert address range into offset range
 229   _bm.at_put_range(heapWordToOffset(mr.start()),
 230                    heapWordToOffset(mr.end()), false);
 231 }
 232 
 233 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
 234                                             HeapWord* end_addr) {
 235   HeapWord* start = getNextMarkedWordAddress(addr);
 236   start = MIN2(start, end_addr);
 237   HeapWord* end   = getNextUnmarkedWordAddress(start);
 238   end = MIN2(end, end_addr);
 239   assert(start <= end, "Consistency check");
 240   MemRegion mr(start, end);
 241   if (!mr.is_empty()) {
 242     clearRange(mr);
 243   }
 244   return mr;
 245 }
 246 
 247 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
 248   _base(NULL), _cm(cm)
 249 #ifdef ASSERT
 250   , _drain_in_progress(false)
 251   , _drain_in_progress_yields(false)
 252 #endif
 253 {}
 254 
 255 bool CMMarkStack::allocate(size_t capacity) {
 256   // allocate a stack of the requisite depth
 257   ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
 258   if (!rs.is_reserved()) {
 259     warning("ConcurrentMark MarkStack allocation failure");
 260     return false;
 261   }
 262   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 263   if (!_virtual_space.initialize(rs, rs.size())) {
 264     warning("ConcurrentMark MarkStack backing store failure");
 265     // Release the virtual memory reserved for the marking stack
 266     rs.release();
 267     return false;
 268   }
 269   assert(_virtual_space.committed_size() == rs.size(),
 270          "Didn't reserve backing store for all of ConcurrentMark stack?");
 271   _base = (oop*) _virtual_space.low();
 272   setEmpty();
 273   _capacity = (jint) capacity;
 274   _saved_index = -1;
 275   _should_expand = false;
 276   return true;
 277 }
 278 
 279 void CMMarkStack::expand() {
 280   // Called, during remark, if we've overflown the marking stack during marking.
 281   assert(isEmpty(), "stack should been emptied while handling overflow");
 282   assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
 283   // Clear expansion flag
 284   _should_expand = false;
 285   if (_capacity == (jint) MarkStackSizeMax) {
 286     if (PrintGCDetails && Verbose) {
 287       gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
 288     }
 289     return;
 290   }
 291   // Double capacity if possible
 292   jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
 293   // Do not give up existing stack until we have managed to
 294   // get the double capacity that we desired.
 295   ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
 296                                                            sizeof(oop)));
 297   if (rs.is_reserved()) {
 298     // Release the backing store associated with old stack
 299     _virtual_space.release();
 300     // Reinitialize virtual space for new stack
 301     if (!_virtual_space.initialize(rs, rs.size())) {
 302       fatal("Not enough swap for expanded marking stack capacity");
 303     }
 304     _base = (oop*)(_virtual_space.low());
 305     _index = 0;
 306     _capacity = new_capacity;
 307   } else {
 308     if (PrintGCDetails && Verbose) {
 309       // Failed to double capacity, continue;
 310       gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
 311                           SIZE_FORMAT "K to " SIZE_FORMAT "K",
 312                           _capacity / K, new_capacity / K);
 313     }
 314   }
 315 }
 316 
 317 void CMMarkStack::set_should_expand() {
 318   // If we're resetting the marking state because of an
 319   // marking stack overflow, record that we should, if
 320   // possible, expand the stack.
 321   _should_expand = _cm->has_overflown();
 322 }
 323 
 324 CMMarkStack::~CMMarkStack() {
 325   if (_base != NULL) {
 326     _base = NULL;
 327     _virtual_space.release();
 328   }
 329 }
 330 
 331 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 332   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 333   jint start = _index;
 334   jint next_index = start + n;
 335   if (next_index > _capacity) {
 336     _overflow = true;
 337     return;
 338   }
 339   // Otherwise.
 340   _index = next_index;
 341   for (int i = 0; i < n; i++) {
 342     int ind = start + i;
 343     assert(ind < _capacity, "By overflow test above.");
 344     _base[ind] = ptr_arr[i];
 345   }
 346 }
 347 
 348 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 349   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 350   jint index = _index;
 351   if (index == 0) {
 352     *n = 0;
 353     return false;
 354   } else {
 355     int k = MIN2(max, index);
 356     jint  new_ind = index - k;
 357     for (int j = 0; j < k; j++) {
 358       ptr_arr[j] = _base[new_ind + j];
 359     }
 360     _index = new_ind;
 361     *n = k;
 362     return true;
 363   }
 364 }
 365 
 366 template<class OopClosureClass>
 367 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
 368   assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
 369          || SafepointSynchronize::is_at_safepoint(),
 370          "Drain recursion must be yield-safe.");
 371   bool res = true;
 372   debug_only(_drain_in_progress = true);
 373   debug_only(_drain_in_progress_yields = yield_after);
 374   while (!isEmpty()) {
 375     oop newOop = pop();
 376     assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
 377     assert(newOop->is_oop(), "Expected an oop");
 378     assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
 379            "only grey objects on this stack");
 380     newOop->oop_iterate(cl);
 381     if (yield_after && _cm->do_yield_check()) {
 382       res = false;
 383       break;
 384     }
 385   }
 386   debug_only(_drain_in_progress = false);
 387   return res;
 388 }
 389 
 390 void CMMarkStack::note_start_of_gc() {
 391   assert(_saved_index == -1,
 392          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 393   _saved_index = _index;
 394 }
 395 
 396 void CMMarkStack::note_end_of_gc() {
 397   // This is intentionally a guarantee, instead of an assert. If we
 398   // accidentally add something to the mark stack during GC, it
 399   // will be a correctness issue so it's better if we crash. we'll
 400   // only check this once per GC anyway, so it won't be a performance
 401   // issue in any way.
 402   guarantee(_saved_index == _index,
 403             err_msg("saved index: %d index: %d", _saved_index, _index));
 404   _saved_index = -1;
 405 }
 406 
 407 CMRootRegions::CMRootRegions() :
 408   _young_list(NULL), _cm(NULL), _scan_in_progress(false),
 409   _should_abort(false),  _next_survivor(NULL) { }
 410 
 411 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
 412   _young_list = g1h->young_list();
 413   _cm = cm;
 414 }
 415 
 416 void CMRootRegions::prepare_for_scan() {
 417   assert(!scan_in_progress(), "pre-condition");
 418 
 419   // Currently, only survivors can be root regions.
 420   assert(_next_survivor == NULL, "pre-condition");
 421   _next_survivor = _young_list->first_survivor_region();
 422   _scan_in_progress = (_next_survivor != NULL);
 423   _should_abort = false;
 424 }
 425 
 426 HeapRegion* CMRootRegions::claim_next() {
 427   if (_should_abort) {
 428     // If someone has set the should_abort flag, we return NULL to
 429     // force the caller to bail out of their loop.
 430     return NULL;
 431   }
 432 
 433   // Currently, only survivors can be root regions.
 434   HeapRegion* res = _next_survivor;
 435   if (res != NULL) {
 436     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 437     // Read it again in case it changed while we were waiting for the lock.
 438     res = _next_survivor;
 439     if (res != NULL) {
 440       if (res == _young_list->last_survivor_region()) {
 441         // We just claimed the last survivor so store NULL to indicate
 442         // that we're done.
 443         _next_survivor = NULL;
 444       } else {
 445         _next_survivor = res->get_next_young_region();
 446       }
 447     } else {
 448       // Someone else claimed the last survivor while we were trying
 449       // to take the lock so nothing else to do.
 450     }
 451   }
 452   assert(res == NULL || res->is_survivor(), "post-condition");
 453 
 454   return res;
 455 }
 456 
 457 void CMRootRegions::scan_finished() {
 458   assert(scan_in_progress(), "pre-condition");
 459 
 460   // Currently, only survivors can be root regions.
 461   if (!_should_abort) {
 462     assert(_next_survivor == NULL, "we should have claimed all survivors");
 463   }
 464   _next_survivor = NULL;
 465 
 466   {
 467     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 468     _scan_in_progress = false;
 469     RootRegionScan_lock->notify_all();
 470   }
 471 }
 472 
 473 bool CMRootRegions::wait_until_scan_finished() {
 474   if (!scan_in_progress()) return false;
 475 
 476   {
 477     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 478     while (scan_in_progress()) {
 479       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 480     }
 481   }
 482   return true;
 483 }
 484 
 485 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
 486 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
 487 #endif // _MSC_VER
 488 
 489 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 490   return MAX2((n_par_threads + 2) / 4, 1U);
 491 }
 492 
 493 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
 494   _g1h(g1h),
 495   _markBitMap1(),
 496   _markBitMap2(),
 497   _parallel_marking_threads(0),
 498   _max_parallel_marking_threads(0),
 499   _sleep_factor(0.0),
 500   _marking_task_overhead(1.0),
 501   _cleanup_sleep_factor(0.0),
 502   _cleanup_task_overhead(1.0),
 503   _cleanup_list("Cleanup List"),
 504   _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
 505   _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
 506             CardTableModRefBS::card_shift,
 507             false /* in_resource_area*/),
 508 
 509   _prevMarkBitMap(&_markBitMap1),
 510   _nextMarkBitMap(&_markBitMap2),
 511 
 512   _markStack(this),
 513   // _finger set in set_non_marking_state
 514 
 515   _max_worker_id(ParallelGCThreads),
 516   // _active_tasks set in set_non_marking_state
 517   // _tasks set inside the constructor
 518   _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
 519   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 520 
 521   _has_overflown(false),
 522   _concurrent(false),
 523   _has_aborted(false),
 524   _restart_for_overflow(false),
 525   _concurrent_marking_in_progress(false),
 526 
 527   // _verbose_level set below
 528 
 529   _init_times(),
 530   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 531   _cleanup_times(),
 532   _total_counting_time(0.0),
 533   _total_rs_scrub_time(0.0),
 534 
 535   _parallel_workers(NULL),
 536 
 537   _count_card_bitmaps(NULL),
 538   _count_marked_bytes(NULL),
 539   _completed_initialization(false) {
 540   CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
 541   if (verbose_level < no_verbose) {
 542     verbose_level = no_verbose;
 543   }
 544   if (verbose_level > high_verbose) {
 545     verbose_level = high_verbose;
 546   }
 547   _verbose_level = verbose_level;
 548 
 549   if (verbose_low()) {
 550     gclog_or_tty->print_cr("[global] init, heap start = " PTR_FORMAT ", "
 551                            "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
 552   }
 553 
 554   _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 555   _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
 556 
 557   // Create & start a ConcurrentMark thread.
 558   _cmThread = new ConcurrentMarkThread(this);
 559   assert(cmThread() != NULL, "CM Thread should have been created");
 560   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 561   if (_cmThread->osthread() == NULL) {
 562       vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 563   }
 564 
 565   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 566   assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
 567   assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
 568 
 569   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 570   satb_qs.set_buffer_size(G1SATBBufferSize);
 571 
 572   _root_regions.init(_g1h, this);
 573 
 574   if (ConcGCThreads > ParallelGCThreads) {
 575     warning("Can't have more ConcGCThreads (%u) "
 576             "than ParallelGCThreads (%u).",
 577             ConcGCThreads, ParallelGCThreads);
 578     return;
 579   }
 580   if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
 581     // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
 582     // if both are set
 583     _sleep_factor             = 0.0;
 584     _marking_task_overhead    = 1.0;
 585   } else if (G1MarkingOverheadPercent > 0) {
 586     // We will calculate the number of parallel marking threads based
 587     // on a target overhead with respect to the soft real-time goal
 588     double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 589     double overall_cm_overhead =
 590       (double) MaxGCPauseMillis * marking_overhead /
 591       (double) GCPauseIntervalMillis;
 592     double cpu_ratio = 1.0 / (double) os::processor_count();
 593     double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 594     double marking_task_overhead =
 595       overall_cm_overhead / marking_thread_num *
 596                                               (double) os::processor_count();
 597     double sleep_factor =
 598                        (1.0 - marking_task_overhead) / marking_task_overhead;
 599 
 600     FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num);
 601     _sleep_factor             = sleep_factor;
 602     _marking_task_overhead    = marking_task_overhead;
 603   } else {
 604     // Calculate the number of parallel marking threads by scaling
 605     // the number of parallel GC threads.
 606     uint marking_thread_num = scale_parallel_threads(ParallelGCThreads);
 607     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 608     _sleep_factor             = 0.0;
 609     _marking_task_overhead    = 1.0;
 610   }
 611 
 612   assert(ConcGCThreads > 0, "Should have been set");
 613   _parallel_marking_threads = ConcGCThreads;
 614   _max_parallel_marking_threads = _parallel_marking_threads;
 615 
 616   if (parallel_marking_threads() > 1) {
 617     _cleanup_task_overhead = 1.0;
 618   } else {
 619     _cleanup_task_overhead = marking_task_overhead();
 620   }
 621   _cleanup_sleep_factor =
 622                    (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
 623 
 624 #if 0
 625   gclog_or_tty->print_cr("Marking Threads          %d", parallel_marking_threads());
 626   gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
 627   gclog_or_tty->print_cr("CM Sleep Factor          %1.4lf", sleep_factor());
 628   gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
 629   gclog_or_tty->print_cr("CL Sleep Factor          %1.4lf", cleanup_sleep_factor());
 630 #endif
 631 
 632   _parallel_workers = new WorkGang("G1 Marker",
 633        _max_parallel_marking_threads, false, true);
 634   if (_parallel_workers == NULL) {
 635     vm_exit_during_initialization("Failed necessary allocation.");
 636   } else {
 637     _parallel_workers->initialize_workers();
 638   }
 639 
 640   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 641     size_t mark_stack_size =
 642       MIN2(MarkStackSizeMax,
 643           MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 644     // Verify that the calculated value for MarkStackSize is in range.
 645     // It would be nice to use the private utility routine from Arguments.
 646     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 647       warning("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 648               "must be between 1 and " SIZE_FORMAT,
 649               mark_stack_size, MarkStackSizeMax);
 650       return;
 651     }
 652     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 653   } else {
 654     // Verify MarkStackSize is in range.
 655     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 656       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 657         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 658           warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 659                   "must be between 1 and " SIZE_FORMAT,
 660                   MarkStackSize, MarkStackSizeMax);
 661           return;
 662         }
 663       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 664         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 665           warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 666                   " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 667                   MarkStackSize, MarkStackSizeMax);
 668           return;
 669         }
 670       }
 671     }
 672   }
 673 
 674   if (!_markStack.allocate(MarkStackSize)) {
 675     warning("Failed to allocate CM marking stack");
 676     return;
 677   }
 678 
 679   _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
 680   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 681 
 682   _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap,  _max_worker_id, mtGC);
 683   _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
 684 
 685   BitMap::idx_t card_bm_size = _card_bm.size();
 686 
 687   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 688   _active_tasks = _max_worker_id;
 689 
 690   uint max_regions = _g1h->max_regions();
 691   for (uint i = 0; i < _max_worker_id; ++i) {
 692     CMTaskQueue* task_queue = new CMTaskQueue();
 693     task_queue->initialize();
 694     _task_queues->register_queue(i, task_queue);
 695 
 696     _count_card_bitmaps[i] = BitMap(card_bm_size, false);
 697     _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
 698 
 699     _tasks[i] = new CMTask(i, this,
 700                            _count_marked_bytes[i],
 701                            &_count_card_bitmaps[i],
 702                            task_queue, _task_queues);
 703 
 704     _accum_task_vtime[i] = 0.0;
 705   }
 706 
 707   // Calculate the card number for the bottom of the heap. Used
 708   // in biasing indexes into the accounting card bitmaps.
 709   _heap_bottom_card_num =
 710     intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
 711                                 CardTableModRefBS::card_shift);
 712 
 713   // Clear all the liveness counting data
 714   clear_all_count_data();
 715 
 716   // so that the call below can read a sensible value
 717   _heap_start = g1h->reserved_region().start();
 718   set_non_marking_state();
 719   _completed_initialization = true;
 720 }
 721 
 722 void ConcurrentMark::reset() {
 723   // Starting values for these two. This should be called in a STW
 724   // phase.
 725   MemRegion reserved = _g1h->g1_reserved();
 726   _heap_start = reserved.start();
 727   _heap_end   = reserved.end();
 728 
 729   // Separated the asserts so that we know which one fires.
 730   assert(_heap_start != NULL, "heap bounds should look ok");
 731   assert(_heap_end != NULL, "heap bounds should look ok");
 732   assert(_heap_start < _heap_end, "heap bounds should look ok");
 733 
 734   // Reset all the marking data structures and any necessary flags
 735   reset_marking_state();
 736 
 737   if (verbose_low()) {
 738     gclog_or_tty->print_cr("[global] resetting");
 739   }
 740 
 741   // We do reset all of them, since different phases will use
 742   // different number of active threads. So, it's easiest to have all
 743   // of them ready.
 744   for (uint i = 0; i < _max_worker_id; ++i) {
 745     _tasks[i]->reset(_nextMarkBitMap);
 746   }
 747 
 748   // we need this to make sure that the flag is on during the evac
 749   // pause with initial mark piggy-backed
 750   set_concurrent_marking_in_progress();
 751 }
 752 
 753 
 754 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
 755   _markStack.set_should_expand();
 756   _markStack.setEmpty();        // Also clears the _markStack overflow flag
 757   if (clear_overflow) {
 758     clear_has_overflown();
 759   } else {
 760     assert(has_overflown(), "pre-condition");
 761   }
 762   _finger = _heap_start;
 763 
 764   for (uint i = 0; i < _max_worker_id; ++i) {
 765     CMTaskQueue* queue = _task_queues->queue(i);
 766     queue->set_empty();
 767   }
 768 }
 769 
 770 void ConcurrentMark::set_concurrency(uint active_tasks) {
 771   assert(active_tasks <= _max_worker_id, "we should not have more");
 772 
 773   _active_tasks = active_tasks;
 774   // Need to update the three data structures below according to the
 775   // number of active threads for this phase.
 776   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 777   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 778   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 779 }
 780 
 781 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 782   set_concurrency(active_tasks);
 783 
 784   _concurrent = concurrent;
 785   // We propagate this to all tasks, not just the active ones.
 786   for (uint i = 0; i < _max_worker_id; ++i)
 787     _tasks[i]->set_concurrent(concurrent);
 788 
 789   if (concurrent) {
 790     set_concurrent_marking_in_progress();
 791   } else {
 792     // We currently assume that the concurrent flag has been set to
 793     // false before we start remark. At this point we should also be
 794     // in a STW phase.
 795     assert(!concurrent_marking_in_progress(), "invariant");
 796     assert(out_of_regions(),
 797            err_msg("only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 798                    p2i(_finger), p2i(_heap_end)));
 799   }
 800 }
 801 
 802 void ConcurrentMark::set_non_marking_state() {
 803   // We set the global marking state to some default values when we're
 804   // not doing marking.
 805   reset_marking_state();
 806   _active_tasks = 0;
 807   clear_concurrent_marking_in_progress();
 808 }
 809 
 810 ConcurrentMark::~ConcurrentMark() {
 811   // The ConcurrentMark instance is never freed.
 812   ShouldNotReachHere();
 813 }
 814 
 815 void ConcurrentMark::clearNextBitmap() {
 816   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 817 
 818   // Make sure that the concurrent mark thread looks to still be in
 819   // the current cycle.
 820   guarantee(cmThread()->during_cycle(), "invariant");
 821 
 822   // We are finishing up the current cycle by clearing the next
 823   // marking bitmap and getting it ready for the next cycle. During
 824   // this time no other cycle can start. So, let's make sure that this
 825   // is the case.
 826   guarantee(!g1h->collector_state()->mark_in_progress(), "invariant");
 827 
 828   ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
 829   ParClearNextMarkBitmapTask task(&cl, parallel_marking_threads(), true);
 830   _parallel_workers->run_task(&task);
 831 
 832   // Clear the liveness counting data. If the marking has been aborted, the abort()
 833   // call already did that.
 834   if (cl.complete()) {
 835     clear_all_count_data();
 836   }
 837 
 838   // Repeat the asserts from above.
 839   guarantee(cmThread()->during_cycle(), "invariant");
 840   guarantee(!g1h->collector_state()->mark_in_progress(), "invariant");
 841 }
 842 
 843 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 844   CMBitMap* _bitmap;
 845   bool _error;
 846  public:
 847   CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
 848   }
 849 
 850   virtual bool doHeapRegion(HeapRegion* r) {
 851     // This closure can be called concurrently to the mutator, so we must make sure
 852     // that the result of the getNextMarkedWordAddress() call is compared to the
 853     // value passed to it as limit to detect any found bits.
 854     // We can use the region's orig_end() for the limit and the comparison value
 855     // as it always contains the "real" end of the region that never changes and
 856     // has no side effects.
 857     // Due to the latter, there can also be no problem with the compiler generating
 858     // reloads of the orig_end() call.
 859     HeapWord* end = r->orig_end();
 860     return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
 861   }
 862 };
 863 
 864 bool ConcurrentMark::nextMarkBitmapIsClear() {
 865   CheckBitmapClearHRClosure cl(_nextMarkBitMap);
 866   _g1h->heap_region_iterate(&cl);
 867   return cl.complete();
 868 }
 869 
 870 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 871 public:
 872   bool doHeapRegion(HeapRegion* r) {
 873     if (!r->is_continues_humongous()) {
 874       r->note_start_of_marking();
 875     }
 876     return false;
 877   }
 878 };
 879 
 880 void ConcurrentMark::checkpointRootsInitialPre() {
 881   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 882   G1CollectorPolicy* g1p = g1h->g1_policy();
 883 
 884   _has_aborted = false;
 885 
 886   // Initialize marking structures. This has to be done in a STW phase.
 887   reset();
 888 
 889   // For each region note start of marking.
 890   NoteStartOfMarkHRClosure startcl;
 891   g1h->heap_region_iterate(&startcl);
 892 }
 893 
 894 
 895 void ConcurrentMark::checkpointRootsInitialPost() {
 896   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 897 
 898   // If we force an overflow during remark, the remark operation will
 899   // actually abort and we'll restart concurrent marking. If we always
 900   // force an overflow during remark we'll never actually complete the
 901   // marking phase. So, we initialize this here, at the start of the
 902   // cycle, so that at the remaining overflow number will decrease at
 903   // every remark and we'll eventually not need to cause one.
 904   force_overflow_stw()->init();
 905 
 906   // Start Concurrent Marking weak-reference discovery.
 907   ReferenceProcessor* rp = g1h->ref_processor_cm();
 908   // enable ("weak") refs discovery
 909   rp->enable_discovery();
 910   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 911 
 912   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 913   // This is the start of  the marking cycle, we're expected all
 914   // threads to have SATB queues with active set to false.
 915   satb_mq_set.set_active_all_threads(true, /* new active value */
 916                                      false /* expected_active */);
 917 
 918   _root_regions.prepare_for_scan();
 919 
 920   // update_g1_committed() will be called at the end of an evac pause
 921   // when marking is on. So, it's also called at the end of the
 922   // initial-mark pause to update the heap end, if the heap expands
 923   // during it. No need to call it here.
 924 }
 925 
 926 /*
 927  * Notice that in the next two methods, we actually leave the STS
 928  * during the barrier sync and join it immediately afterwards. If we
 929  * do not do this, the following deadlock can occur: one thread could
 930  * be in the barrier sync code, waiting for the other thread to also
 931  * sync up, whereas another one could be trying to yield, while also
 932  * waiting for the other threads to sync up too.
 933  *
 934  * Note, however, that this code is also used during remark and in
 935  * this case we should not attempt to leave / enter the STS, otherwise
 936  * we'll either hit an assert (debug / fastdebug) or deadlock
 937  * (product). So we should only leave / enter the STS if we are
 938  * operating concurrently.
 939  *
 940  * Because the thread that does the sync barrier has left the STS, it
 941  * is possible to be suspended for a Full GC or an evacuation pause
 942  * could occur. This is actually safe, since the entering the sync
 943  * barrier is one of the last things do_marking_step() does, and it
 944  * doesn't manipulate any data structures afterwards.
 945  */
 946 
 947 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 948   bool barrier_aborted;
 949 
 950   if (verbose_low()) {
 951     gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
 952   }
 953 
 954   {
 955     SuspendibleThreadSetLeaver sts_leave(concurrent());
 956     barrier_aborted = !_first_overflow_barrier_sync.enter();
 957   }
 958 
 959   // at this point everyone should have synced up and not be doing any
 960   // more work
 961 
 962   if (verbose_low()) {
 963     if (barrier_aborted) {
 964       gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
 965     } else {
 966       gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
 967     }
 968   }
 969 
 970   if (barrier_aborted) {
 971     // If the barrier aborted we ignore the overflow condition and
 972     // just abort the whole marking phase as quickly as possible.
 973     return;
 974   }
 975 
 976   // If we're executing the concurrent phase of marking, reset the marking
 977   // state; otherwise the marking state is reset after reference processing,
 978   // during the remark pause.
 979   // If we reset here as a result of an overflow during the remark we will
 980   // see assertion failures from any subsequent set_concurrency_and_phase()
 981   // calls.
 982   if (concurrent()) {
 983     // let the task associated with with worker 0 do this
 984     if (worker_id == 0) {
 985       // task 0 is responsible for clearing the global data structures
 986       // We should be here because of an overflow. During STW we should
 987       // not clear the overflow flag since we rely on it being true when
 988       // we exit this method to abort the pause and restart concurrent
 989       // marking.
 990       reset_marking_state(true /* clear_overflow */);
 991       force_overflow()->update();
 992 
 993       if (G1Log::fine()) {
 994         gclog_or_tty->gclog_stamp();
 995         gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
 996       }
 997     }
 998   }
 999 
1000   // after this, each task should reset its own data structures then
1001   // then go into the second barrier
1002 }
1003 
1004 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1005   bool barrier_aborted;
1006 
1007   if (verbose_low()) {
1008     gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1009   }
1010 
1011   {
1012     SuspendibleThreadSetLeaver sts_leave(concurrent());
1013     barrier_aborted = !_second_overflow_barrier_sync.enter();
1014   }
1015 
1016   // at this point everything should be re-initialized and ready to go
1017 
1018   if (verbose_low()) {
1019     if (barrier_aborted) {
1020       gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1021     } else {
1022       gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1023     }
1024   }
1025 }
1026 
1027 #ifndef PRODUCT
1028 void ForceOverflowSettings::init() {
1029   _num_remaining = G1ConcMarkForceOverflow;
1030   _force = false;
1031   update();
1032 }
1033 
1034 void ForceOverflowSettings::update() {
1035   if (_num_remaining > 0) {
1036     _num_remaining -= 1;
1037     _force = true;
1038   } else {
1039     _force = false;
1040   }
1041 }
1042 
1043 bool ForceOverflowSettings::should_force() {
1044   if (_force) {
1045     _force = false;
1046     return true;
1047   } else {
1048     return false;
1049   }
1050 }
1051 #endif // !PRODUCT
1052 
1053 class CMConcurrentMarkingTask: public AbstractGangTask {
1054 private:
1055   ConcurrentMark*       _cm;
1056   ConcurrentMarkThread* _cmt;
1057 
1058 public:
1059   void work(uint worker_id) {
1060     assert(Thread::current()->is_ConcurrentGC_thread(),
1061            "this should only be done by a conc GC thread");
1062     ResourceMark rm;
1063 
1064     double start_vtime = os::elapsedVTime();
1065 
1066     {
1067       SuspendibleThreadSetJoiner sts_join;
1068 
1069       assert(worker_id < _cm->active_tasks(), "invariant");
1070       CMTask* the_task = _cm->task(worker_id);
1071       the_task->record_start_time();
1072       if (!_cm->has_aborted()) {
1073         do {
1074           double start_vtime_sec = os::elapsedVTime();
1075           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1076 
1077           the_task->do_marking_step(mark_step_duration_ms,
1078                                     true  /* do_termination */,
1079                                     false /* is_serial*/);
1080 
1081           double end_vtime_sec = os::elapsedVTime();
1082           double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1083           _cm->clear_has_overflown();
1084 
1085           _cm->do_yield_check(worker_id);
1086 
1087           jlong sleep_time_ms;
1088           if (!_cm->has_aborted() && the_task->has_aborted()) {
1089             sleep_time_ms =
1090               (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1091             {
1092               SuspendibleThreadSetLeaver sts_leave;
1093               os::sleep(Thread::current(), sleep_time_ms, false);
1094             }
1095           }
1096         } while (!_cm->has_aborted() && the_task->has_aborted());
1097       }
1098       the_task->record_end_time();
1099       guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1100     }
1101 
1102     double end_vtime = os::elapsedVTime();
1103     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1104   }
1105 
1106   CMConcurrentMarkingTask(ConcurrentMark* cm,
1107                           ConcurrentMarkThread* cmt) :
1108       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1109 
1110   ~CMConcurrentMarkingTask() { }
1111 };
1112 
1113 // Calculates the number of active workers for a concurrent
1114 // phase.
1115 uint ConcurrentMark::calc_parallel_marking_threads() {
1116   uint n_conc_workers = 0;
1117   if (!UseDynamicNumberOfGCThreads ||
1118       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1119        !ForceDynamicNumberOfGCThreads)) {
1120     n_conc_workers = max_parallel_marking_threads();
1121   } else {
1122     n_conc_workers =
1123       AdaptiveSizePolicy::calc_default_active_workers(
1124                                    max_parallel_marking_threads(),
1125                                    1, /* Minimum workers */
1126                                    parallel_marking_threads(),
1127                                    Threads::number_of_non_daemon_threads());
1128     // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1129     // that scaling has already gone into "_max_parallel_marking_threads".
1130   }
1131   assert(n_conc_workers > 0, "Always need at least 1");
1132   return n_conc_workers;
1133 }
1134 
1135 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1136   // Currently, only survivors can be root regions.
1137   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1138   G1RootRegionScanClosure cl(_g1h, this, worker_id);
1139 
1140   const uintx interval = PrefetchScanIntervalInBytes;
1141   HeapWord* curr = hr->bottom();
1142   const HeapWord* end = hr->top();
1143   while (curr < end) {
1144     Prefetch::read(curr, interval);
1145     oop obj = oop(curr);
1146     int size = obj->oop_iterate_size(&cl);
1147     assert(size == obj->size(), "sanity");
1148     curr += size;
1149   }
1150 }
1151 
1152 class CMRootRegionScanTask : public AbstractGangTask {
1153 private:
1154   ConcurrentMark* _cm;
1155 
1156 public:
1157   CMRootRegionScanTask(ConcurrentMark* cm) :
1158     AbstractGangTask("Root Region Scan"), _cm(cm) { }
1159 
1160   void work(uint worker_id) {
1161     assert(Thread::current()->is_ConcurrentGC_thread(),
1162            "this should only be done by a conc GC thread");
1163 
1164     CMRootRegions* root_regions = _cm->root_regions();
1165     HeapRegion* hr = root_regions->claim_next();
1166     while (hr != NULL) {
1167       _cm->scanRootRegion(hr, worker_id);
1168       hr = root_regions->claim_next();
1169     }
1170   }
1171 };
1172 
1173 void ConcurrentMark::scanRootRegions() {
1174   double scan_start = os::elapsedTime();
1175 
1176   // Start of concurrent marking.
1177   ClassLoaderDataGraph::clear_claimed_marks();
1178 
1179   // scan_in_progress() will have been set to true only if there was
1180   // at least one root region to scan. So, if it's false, we
1181   // should not attempt to do any further work.
1182   if (root_regions()->scan_in_progress()) {
1183     if (G1Log::fine()) {
1184       gclog_or_tty->gclog_stamp();
1185       gclog_or_tty->print_cr("[GC concurrent-root-region-scan-start]");
1186     }
1187 
1188     _parallel_marking_threads = calc_parallel_marking_threads();
1189     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1190            "Maximum number of marking threads exceeded");
1191     uint active_workers = MAX2(1U, parallel_marking_threads());
1192 
1193     CMRootRegionScanTask task(this);
1194     _parallel_workers->set_active_workers(active_workers);
1195     _parallel_workers->run_task(&task);
1196 
1197     if (G1Log::fine()) {
1198       gclog_or_tty->gclog_stamp();
1199       gclog_or_tty->print_cr("[GC concurrent-root-region-scan-end, %1.7lf secs]", os::elapsedTime() - scan_start);
1200     }
1201 
1202     // It's possible that has_aborted() is true here without actually
1203     // aborting the survivor scan earlier. This is OK as it's
1204     // mainly used for sanity checking.
1205     root_regions()->scan_finished();
1206   }
1207 }
1208 
1209 void ConcurrentMark::markFromRoots() {
1210   // we might be tempted to assert that:
1211   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1212   //        "inconsistent argument?");
1213   // However that wouldn't be right, because it's possible that
1214   // a safepoint is indeed in progress as a younger generation
1215   // stop-the-world GC happens even as we mark in this generation.
1216 
1217   _restart_for_overflow = false;
1218   force_overflow_conc()->init();
1219 
1220   // _g1h has _n_par_threads
1221   _parallel_marking_threads = calc_parallel_marking_threads();
1222   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1223     "Maximum number of marking threads exceeded");
1224 
1225   uint active_workers = MAX2(1U, parallel_marking_threads());
1226   assert(active_workers > 0, "Should have been set");
1227 
1228   // Parallel task terminator is set in "set_concurrency_and_phase()"
1229   set_concurrency_and_phase(active_workers, true /* concurrent */);
1230 
1231   CMConcurrentMarkingTask markingTask(this, cmThread());
1232   _parallel_workers->set_active_workers(active_workers);
1233   _parallel_workers->run_task(&markingTask);
1234   print_stats();
1235 }
1236 
1237 // Helper class to get rid of some boilerplate code.
1238 class G1CMTraceTime : public GCTraceTime {
1239   static bool doit_and_prepend(bool doit) {
1240     if (doit) {
1241       gclog_or_tty->put(' ');
1242     }
1243     return doit;
1244   }
1245 
1246  public:
1247   G1CMTraceTime(const char* title, bool doit)
1248     : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm()) {
1249   }
1250 };
1251 
1252 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1253   // world is stopped at this checkpoint
1254   assert(SafepointSynchronize::is_at_safepoint(),
1255          "world should be stopped");
1256 
1257   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1258 
1259   // If a full collection has happened, we shouldn't do this.
1260   if (has_aborted()) {
1261     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1262     return;
1263   }
1264 
1265   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1266 
1267   if (VerifyDuringGC) {
1268     HandleMark hm;  // handle scope
1269     g1h->prepare_for_verify();
1270     Universe::verify(VerifyOption_G1UsePrevMarking,
1271                      " VerifyDuringGC:(before)");
1272   }
1273   g1h->check_bitmaps("Remark Start");
1274 
1275   G1CollectorPolicy* g1p = g1h->g1_policy();
1276   g1p->record_concurrent_mark_remark_start();
1277 
1278   double start = os::elapsedTime();
1279 
1280   checkpointRootsFinalWork();
1281 
1282   double mark_work_end = os::elapsedTime();
1283 
1284   weakRefsWork(clear_all_soft_refs);
1285 
1286   if (has_overflown()) {
1287     // Oops.  We overflowed.  Restart concurrent marking.
1288     _restart_for_overflow = true;
1289     if (G1TraceMarkStackOverflow) {
1290       gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1291     }
1292 
1293     // Verify the heap w.r.t. the previous marking bitmap.
1294     if (VerifyDuringGC) {
1295       HandleMark hm;  // handle scope
1296       g1h->prepare_for_verify();
1297       Universe::verify(VerifyOption_G1UsePrevMarking,
1298                        " VerifyDuringGC:(overflow)");
1299     }
1300 
1301     // Clear the marking state because we will be restarting
1302     // marking due to overflowing the global mark stack.
1303     reset_marking_state();
1304   } else {
1305     {
1306       G1CMTraceTime trace("GC aggregate-data", G1Log::finer());
1307 
1308       // Aggregate the per-task counting data that we have accumulated
1309       // while marking.
1310       aggregate_count_data();
1311     }
1312 
1313     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1314     // We're done with marking.
1315     // This is the end of  the marking cycle, we're expected all
1316     // threads to have SATB queues with active set to true.
1317     satb_mq_set.set_active_all_threads(false, /* new active value */
1318                                        true /* expected_active */);
1319 
1320     if (VerifyDuringGC) {
1321       HandleMark hm;  // handle scope
1322       g1h->prepare_for_verify();
1323       Universe::verify(VerifyOption_G1UseNextMarking,
1324                        " VerifyDuringGC:(after)");
1325     }
1326     g1h->check_bitmaps("Remark End");
1327     assert(!restart_for_overflow(), "sanity");
1328     // Completely reset the marking state since marking completed
1329     set_non_marking_state();
1330   }
1331 
1332   // Expand the marking stack, if we have to and if we can.
1333   if (_markStack.should_expand()) {
1334     _markStack.expand();
1335   }
1336 
1337   // Statistics
1338   double now = os::elapsedTime();
1339   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1340   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1341   _remark_times.add((now - start) * 1000.0);
1342 
1343   g1p->record_concurrent_mark_remark_end();
1344 
1345   G1CMIsAliveClosure is_alive(g1h);
1346   g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1347 }
1348 
1349 // Base class of the closures that finalize and verify the
1350 // liveness counting data.
1351 class CMCountDataClosureBase: public HeapRegionClosure {
1352 protected:
1353   G1CollectedHeap* _g1h;
1354   ConcurrentMark* _cm;
1355   CardTableModRefBS* _ct_bs;
1356 
1357   BitMap* _region_bm;
1358   BitMap* _card_bm;
1359 
1360   // Takes a region that's not empty (i.e., it has at least one
1361   // live object in it and sets its corresponding bit on the region
1362   // bitmap to 1. If the region is "starts humongous" it will also set
1363   // to 1 the bits on the region bitmap that correspond to its
1364   // associated "continues humongous" regions.
1365   void set_bit_for_region(HeapRegion* hr) {
1366     assert(!hr->is_continues_humongous(), "should have filtered those out");
1367 
1368     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1369     if (!hr->is_starts_humongous()) {
1370       // Normal (non-humongous) case: just set the bit.
1371       _region_bm->par_at_put(index, true);
1372     } else {
1373       // Starts humongous case: calculate how many regions are part of
1374       // this humongous region and then set the bit range.
1375       BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1376       _region_bm->par_at_put_range(index, end_index, true);
1377     }
1378   }
1379 
1380 public:
1381   CMCountDataClosureBase(G1CollectedHeap* g1h,
1382                          BitMap* region_bm, BitMap* card_bm):
1383     _g1h(g1h), _cm(g1h->concurrent_mark()),
1384     _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
1385     _region_bm(region_bm), _card_bm(card_bm) { }
1386 };
1387 
1388 // Closure that calculates the # live objects per region. Used
1389 // for verification purposes during the cleanup pause.
1390 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1391   CMBitMapRO* _bm;
1392   size_t _region_marked_bytes;
1393 
1394 public:
1395   CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1396                          BitMap* region_bm, BitMap* card_bm) :
1397     CMCountDataClosureBase(g1h, region_bm, card_bm),
1398     _bm(bm), _region_marked_bytes(0) { }
1399 
1400   bool doHeapRegion(HeapRegion* hr) {
1401 
1402     if (hr->is_continues_humongous()) {
1403       // We will ignore these here and process them when their
1404       // associated "starts humongous" region is processed (see
1405       // set_bit_for_heap_region()). Note that we cannot rely on their
1406       // associated "starts humongous" region to have their bit set to
1407       // 1 since, due to the region chunking in the parallel region
1408       // iteration, a "continues humongous" region might be visited
1409       // before its associated "starts humongous".
1410       return false;
1411     }
1412 
1413     HeapWord* ntams = hr->next_top_at_mark_start();
1414     HeapWord* start = hr->bottom();
1415 
1416     assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1417            err_msg("Preconditions not met - "
1418                    "start: " PTR_FORMAT ", ntams: " PTR_FORMAT ", end: " PTR_FORMAT,
1419                    p2i(start), p2i(ntams), p2i(hr->end())));
1420 
1421     // Find the first marked object at or after "start".
1422     start = _bm->getNextMarkedWordAddress(start, ntams);
1423 
1424     size_t marked_bytes = 0;
1425 
1426     while (start < ntams) {
1427       oop obj = oop(start);
1428       int obj_sz = obj->size();
1429       HeapWord* obj_end = start + obj_sz;
1430 
1431       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1432       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1433 
1434       // Note: if we're looking at the last region in heap - obj_end
1435       // could be actually just beyond the end of the heap; end_idx
1436       // will then correspond to a (non-existent) card that is also
1437       // just beyond the heap.
1438       if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1439         // end of object is not card aligned - increment to cover
1440         // all the cards spanned by the object
1441         end_idx += 1;
1442       }
1443 
1444       // Set the bits in the card BM for the cards spanned by this object.
1445       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1446 
1447       // Add the size of this object to the number of marked bytes.
1448       marked_bytes += (size_t)obj_sz * HeapWordSize;
1449 
1450       // Find the next marked object after this one.
1451       start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1452     }
1453 
1454     // Mark the allocated-since-marking portion...
1455     HeapWord* top = hr->top();
1456     if (ntams < top) {
1457       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1458       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1459 
1460       // Note: if we're looking at the last region in heap - top
1461       // could be actually just beyond the end of the heap; end_idx
1462       // will then correspond to a (non-existent) card that is also
1463       // just beyond the heap.
1464       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1465         // end of object is not card aligned - increment to cover
1466         // all the cards spanned by the object
1467         end_idx += 1;
1468       }
1469       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1470 
1471       // This definitely means the region has live objects.
1472       set_bit_for_region(hr);
1473     }
1474 
1475     // Update the live region bitmap.
1476     if (marked_bytes > 0) {
1477       set_bit_for_region(hr);
1478     }
1479 
1480     // Set the marked bytes for the current region so that
1481     // it can be queried by a calling verification routine
1482     _region_marked_bytes = marked_bytes;
1483 
1484     return false;
1485   }
1486 
1487   size_t region_marked_bytes() const { return _region_marked_bytes; }
1488 };
1489 
1490 // Heap region closure used for verifying the counting data
1491 // that was accumulated concurrently and aggregated during
1492 // the remark pause. This closure is applied to the heap
1493 // regions during the STW cleanup pause.
1494 
1495 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1496   G1CollectedHeap* _g1h;
1497   ConcurrentMark* _cm;
1498   CalcLiveObjectsClosure _calc_cl;
1499   BitMap* _region_bm;   // Region BM to be verified
1500   BitMap* _card_bm;     // Card BM to be verified
1501   bool _verbose;        // verbose output?
1502 
1503   BitMap* _exp_region_bm; // Expected Region BM values
1504   BitMap* _exp_card_bm;   // Expected card BM values
1505 
1506   int _failures;
1507 
1508 public:
1509   VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1510                                 BitMap* region_bm,
1511                                 BitMap* card_bm,
1512                                 BitMap* exp_region_bm,
1513                                 BitMap* exp_card_bm,
1514                                 bool verbose) :
1515     _g1h(g1h), _cm(g1h->concurrent_mark()),
1516     _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1517     _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1518     _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1519     _failures(0) { }
1520 
1521   int failures() const { return _failures; }
1522 
1523   bool doHeapRegion(HeapRegion* hr) {
1524     if (hr->is_continues_humongous()) {
1525       // We will ignore these here and process them when their
1526       // associated "starts humongous" region is processed (see
1527       // set_bit_for_heap_region()). Note that we cannot rely on their
1528       // associated "starts humongous" region to have their bit set to
1529       // 1 since, due to the region chunking in the parallel region
1530       // iteration, a "continues humongous" region might be visited
1531       // before its associated "starts humongous".
1532       return false;
1533     }
1534 
1535     int failures = 0;
1536 
1537     // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1538     // this region and set the corresponding bits in the expected region
1539     // and card bitmaps.
1540     bool res = _calc_cl.doHeapRegion(hr);
1541     assert(res == false, "should be continuing");
1542 
1543     MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1544                     Mutex::_no_safepoint_check_flag);
1545 
1546     // Verify the marked bytes for this region.
1547     size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1548     size_t act_marked_bytes = hr->next_marked_bytes();
1549 
1550     // We're not OK if expected marked bytes > actual marked bytes. It means
1551     // we have missed accounting some objects during the actual marking.
1552     if (exp_marked_bytes > act_marked_bytes) {
1553       if (_verbose) {
1554         gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1555                                "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1556                                hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1557       }
1558       failures += 1;
1559     }
1560 
1561     // Verify the bit, for this region, in the actual and expected
1562     // (which was just calculated) region bit maps.
1563     // We're not OK if the bit in the calculated expected region
1564     // bitmap is set and the bit in the actual region bitmap is not.
1565     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1566 
1567     bool expected = _exp_region_bm->at(index);
1568     bool actual = _region_bm->at(index);
1569     if (expected && !actual) {
1570       if (_verbose) {
1571         gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1572                                "expected: %s, actual: %s",
1573                                hr->hrm_index(),
1574                                BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1575       }
1576       failures += 1;
1577     }
1578 
1579     // Verify that the card bit maps for the cards spanned by the current
1580     // region match. We have an error if we have a set bit in the expected
1581     // bit map and the corresponding bit in the actual bitmap is not set.
1582 
1583     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1584     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1585 
1586     for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1587       expected = _exp_card_bm->at(i);
1588       actual = _card_bm->at(i);
1589 
1590       if (expected && !actual) {
1591         if (_verbose) {
1592           gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1593                                  "expected: %s, actual: %s",
1594                                  hr->hrm_index(), i,
1595                                  BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1596         }
1597         failures += 1;
1598       }
1599     }
1600 
1601     if (failures > 0 && _verbose)  {
1602       gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1603                              "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1604                              HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1605                              _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1606     }
1607 
1608     _failures += failures;
1609 
1610     // We could stop iteration over the heap when we
1611     // find the first violating region by returning true.
1612     return false;
1613   }
1614 };
1615 
1616 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1617 protected:
1618   G1CollectedHeap* _g1h;
1619   ConcurrentMark* _cm;
1620   BitMap* _actual_region_bm;
1621   BitMap* _actual_card_bm;
1622 
1623   uint    _n_workers;
1624 
1625   BitMap* _expected_region_bm;
1626   BitMap* _expected_card_bm;
1627 
1628   int  _failures;
1629   bool _verbose;
1630 
1631   HeapRegionClaimer _hrclaimer;
1632 
1633 public:
1634   G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1635                             BitMap* region_bm, BitMap* card_bm,
1636                             BitMap* expected_region_bm, BitMap* expected_card_bm)
1637     : AbstractGangTask("G1 verify final counting"),
1638       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1639       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1640       _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1641       _failures(0), _verbose(false),
1642       _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1643     assert(VerifyDuringGC, "don't call this otherwise");
1644     assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1645     assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1646 
1647     _verbose = _cm->verbose_medium();
1648   }
1649 
1650   void work(uint worker_id) {
1651     assert(worker_id < _n_workers, "invariant");
1652 
1653     VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1654                                             _actual_region_bm, _actual_card_bm,
1655                                             _expected_region_bm,
1656                                             _expected_card_bm,
1657                                             _verbose);
1658 
1659     _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer);
1660 
1661     Atomic::add(verify_cl.failures(), &_failures);
1662   }
1663 
1664   int failures() const { return _failures; }
1665 };
1666 
1667 // Closure that finalizes the liveness counting data.
1668 // Used during the cleanup pause.
1669 // Sets the bits corresponding to the interval [NTAMS, top]
1670 // (which contains the implicitly live objects) in the
1671 // card liveness bitmap. Also sets the bit for each region,
1672 // containing live data, in the region liveness bitmap.
1673 
1674 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1675  public:
1676   FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1677                               BitMap* region_bm,
1678                               BitMap* card_bm) :
1679     CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1680 
1681   bool doHeapRegion(HeapRegion* hr) {
1682 
1683     if (hr->is_continues_humongous()) {
1684       // We will ignore these here and process them when their
1685       // associated "starts humongous" region is processed (see
1686       // set_bit_for_heap_region()). Note that we cannot rely on their
1687       // associated "starts humongous" region to have their bit set to
1688       // 1 since, due to the region chunking in the parallel region
1689       // iteration, a "continues humongous" region might be visited
1690       // before its associated "starts humongous".
1691       return false;
1692     }
1693 
1694     HeapWord* ntams = hr->next_top_at_mark_start();
1695     HeapWord* top   = hr->top();
1696 
1697     assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1698 
1699     // Mark the allocated-since-marking portion...
1700     if (ntams < top) {
1701       // This definitely means the region has live objects.
1702       set_bit_for_region(hr);
1703 
1704       // Now set the bits in the card bitmap for [ntams, top)
1705       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1706       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1707 
1708       // Note: if we're looking at the last region in heap - top
1709       // could be actually just beyond the end of the heap; end_idx
1710       // will then correspond to a (non-existent) card that is also
1711       // just beyond the heap.
1712       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1713         // end of object is not card aligned - increment to cover
1714         // all the cards spanned by the object
1715         end_idx += 1;
1716       }
1717 
1718       assert(end_idx <= _card_bm->size(),
1719              err_msg("oob: end_idx=  " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT,
1720                      end_idx, _card_bm->size()));
1721       assert(start_idx < _card_bm->size(),
1722              err_msg("oob: start_idx=  " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT,
1723                      start_idx, _card_bm->size()));
1724 
1725       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1726     }
1727 
1728     // Set the bit for the region if it contains live data
1729     if (hr->next_marked_bytes() > 0) {
1730       set_bit_for_region(hr);
1731     }
1732 
1733     return false;
1734   }
1735 };
1736 
1737 class G1ParFinalCountTask: public AbstractGangTask {
1738 protected:
1739   G1CollectedHeap* _g1h;
1740   ConcurrentMark* _cm;
1741   BitMap* _actual_region_bm;
1742   BitMap* _actual_card_bm;
1743 
1744   uint    _n_workers;
1745   HeapRegionClaimer _hrclaimer;
1746 
1747 public:
1748   G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1749     : AbstractGangTask("G1 final counting"),
1750       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1751       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1752       _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1753   }
1754 
1755   void work(uint worker_id) {
1756     assert(worker_id < _n_workers, "invariant");
1757 
1758     FinalCountDataUpdateClosure final_update_cl(_g1h,
1759                                                 _actual_region_bm,
1760                                                 _actual_card_bm);
1761 
1762     _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer);
1763   }
1764 };
1765 
1766 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1767   G1CollectedHeap* _g1;
1768   size_t _freed_bytes;
1769   FreeRegionList* _local_cleanup_list;
1770   HeapRegionSetCount _old_regions_removed;
1771   HeapRegionSetCount _humongous_regions_removed;
1772   HRRSCleanupTask* _hrrs_cleanup_task;
1773 
1774 public:
1775   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1776                              FreeRegionList* local_cleanup_list,
1777                              HRRSCleanupTask* hrrs_cleanup_task) :
1778     _g1(g1),
1779     _freed_bytes(0),
1780     _local_cleanup_list(local_cleanup_list),
1781     _old_regions_removed(),
1782     _humongous_regions_removed(),
1783     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1784 
1785   size_t freed_bytes() { return _freed_bytes; }
1786   const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1787   const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1788 
1789   bool doHeapRegion(HeapRegion *hr) {
1790     if (hr->is_continues_humongous() || hr->is_archive()) {
1791       return false;
1792     }
1793     // We use a claim value of zero here because all regions
1794     // were claimed with value 1 in the FinalCount task.
1795     _g1->reset_gc_time_stamps(hr);
1796     hr->note_end_of_marking();
1797 
1798     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1799       _freed_bytes += hr->used();
1800       hr->set_containing_set(NULL);
1801       if (hr->is_humongous()) {
1802         assert(hr->is_starts_humongous(), "we should only see starts humongous");
1803         _humongous_regions_removed.increment(1u, hr->capacity());
1804         _g1->free_humongous_region(hr, _local_cleanup_list, true);
1805       } else {
1806         _old_regions_removed.increment(1u, hr->capacity());
1807         _g1->free_region(hr, _local_cleanup_list, true);
1808       }
1809     } else {
1810       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1811     }
1812 
1813     return false;
1814   }
1815 };
1816 
1817 class G1ParNoteEndTask: public AbstractGangTask {
1818   friend class G1NoteEndOfConcMarkClosure;
1819 
1820 protected:
1821   G1CollectedHeap* _g1h;
1822   FreeRegionList* _cleanup_list;
1823   HeapRegionClaimer _hrclaimer;
1824 
1825 public:
1826   G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1827       AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1828   }
1829 
1830   void work(uint worker_id) {
1831     FreeRegionList local_cleanup_list("Local Cleanup List");
1832     HRRSCleanupTask hrrs_cleanup_task;
1833     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1834                                            &hrrs_cleanup_task);
1835     _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1836     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1837 
1838     // Now update the lists
1839     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1840     {
1841       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1842       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1843 
1844       // If we iterate over the global cleanup list at the end of
1845       // cleanup to do this printing we will not guarantee to only
1846       // generate output for the newly-reclaimed regions (the list
1847       // might not be empty at the beginning of cleanup; we might
1848       // still be working on its previous contents). So we do the
1849       // printing here, before we append the new regions to the global
1850       // cleanup list.
1851 
1852       G1HRPrinter* hr_printer = _g1h->hr_printer();
1853       if (hr_printer->is_active()) {
1854         FreeRegionListIterator iter(&local_cleanup_list);
1855         while (iter.more_available()) {
1856           HeapRegion* hr = iter.get_next();
1857           hr_printer->cleanup(hr);
1858         }
1859       }
1860 
1861       _cleanup_list->add_ordered(&local_cleanup_list);
1862       assert(local_cleanup_list.is_empty(), "post-condition");
1863 
1864       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1865     }
1866   }
1867 };
1868 
1869 class G1ParScrubRemSetTask: public AbstractGangTask {
1870 protected:
1871   G1RemSet* _g1rs;
1872   BitMap* _region_bm;
1873   BitMap* _card_bm;
1874   HeapRegionClaimer _hrclaimer;
1875 
1876 public:
1877   G1ParScrubRemSetTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm, uint n_workers) :
1878       AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), _region_bm(region_bm), _card_bm(card_bm), _hrclaimer(n_workers) {
1879   }
1880 
1881   void work(uint worker_id) {
1882     _g1rs->scrub(_region_bm, _card_bm, worker_id, &_hrclaimer);
1883   }
1884 
1885 };
1886 
1887 void ConcurrentMark::cleanup() {
1888   // world is stopped at this checkpoint
1889   assert(SafepointSynchronize::is_at_safepoint(),
1890          "world should be stopped");
1891   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1892 
1893   // If a full collection has happened, we shouldn't do this.
1894   if (has_aborted()) {
1895     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1896     return;
1897   }
1898 
1899   g1h->verify_region_sets_optional();
1900 
1901   if (VerifyDuringGC) {
1902     HandleMark hm;  // handle scope
1903     g1h->prepare_for_verify();
1904     Universe::verify(VerifyOption_G1UsePrevMarking,
1905                      " VerifyDuringGC:(before)");
1906   }
1907   g1h->check_bitmaps("Cleanup Start");
1908 
1909   G1CollectorPolicy* g1p = g1h->g1_policy();
1910   g1p->record_concurrent_mark_cleanup_start();
1911 
1912   double start = os::elapsedTime();
1913 
1914   HeapRegionRemSet::reset_for_cleanup_tasks();
1915 
1916   // Do counting once more with the world stopped for good measure.
1917   G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1918 
1919   g1h->workers()->run_task(&g1_par_count_task);
1920 
1921   if (VerifyDuringGC) {
1922     // Verify that the counting data accumulated during marking matches
1923     // that calculated by walking the marking bitmap.
1924 
1925     // Bitmaps to hold expected values
1926     BitMap expected_region_bm(_region_bm.size(), true);
1927     BitMap expected_card_bm(_card_bm.size(), true);
1928 
1929     G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
1930                                                  &_region_bm,
1931                                                  &_card_bm,
1932                                                  &expected_region_bm,
1933                                                  &expected_card_bm);
1934 
1935     g1h->workers()->run_task(&g1_par_verify_task);
1936 
1937     guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
1938   }
1939 
1940   size_t start_used_bytes = g1h->used();
1941   g1h->collector_state()->set_mark_in_progress(false);
1942 
1943   double count_end = os::elapsedTime();
1944   double this_final_counting_time = (count_end - start);
1945   _total_counting_time += this_final_counting_time;
1946 
1947   if (G1PrintRegionLivenessInfo) {
1948     G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
1949     _g1h->heap_region_iterate(&cl);
1950   }
1951 
1952   // Install newly created mark bitMap as "prev".
1953   swapMarkBitMaps();
1954 
1955   g1h->reset_gc_time_stamp();
1956 
1957   uint n_workers = _g1h->workers()->active_workers();
1958 
1959   // Note end of marking in all heap regions.
1960   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1961   g1h->workers()->run_task(&g1_par_note_end_task);
1962   g1h->check_gc_time_stamps();
1963 
1964   if (!cleanup_list_is_empty()) {
1965     // The cleanup list is not empty, so we'll have to process it
1966     // concurrently. Notify anyone else that might be wanting free
1967     // regions that there will be more free regions coming soon.
1968     g1h->set_free_regions_coming();
1969   }
1970 
1971   // call below, since it affects the metric by which we sort the heap
1972   // regions.
1973   if (G1ScrubRemSets) {
1974     double rs_scrub_start = os::elapsedTime();
1975     G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm, n_workers);
1976     g1h->workers()->run_task(&g1_par_scrub_rs_task);
1977 
1978     double rs_scrub_end = os::elapsedTime();
1979     double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
1980     _total_rs_scrub_time += this_rs_scrub_time;
1981   }
1982 
1983   // this will also free any regions totally full of garbage objects,
1984   // and sort the regions.
1985   g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1986 
1987   // Statistics.
1988   double end = os::elapsedTime();
1989   _cleanup_times.add((end - start) * 1000.0);
1990 
1991   if (G1Log::fine()) {
1992     g1h->g1_policy()->print_heap_transition(start_used_bytes);
1993   }
1994 
1995   // Clean up will have freed any regions completely full of garbage.
1996   // Update the soft reference policy with the new heap occupancy.
1997   Universe::update_heap_info_at_gc();
1998 
1999   if (VerifyDuringGC) {
2000     HandleMark hm;  // handle scope
2001     g1h->prepare_for_verify();
2002     Universe::verify(VerifyOption_G1UsePrevMarking,
2003                      " VerifyDuringGC:(after)");
2004   }
2005 
2006   g1h->check_bitmaps("Cleanup End");
2007 
2008   g1h->verify_region_sets_optional();
2009 
2010   // We need to make this be a "collection" so any collection pause that
2011   // races with it goes around and waits for completeCleanup to finish.
2012   g1h->increment_total_collections();
2013 
2014   // Clean out dead classes and update Metaspace sizes.
2015   if (ClassUnloadingWithConcurrentMark) {
2016     ClassLoaderDataGraph::purge();
2017   }
2018   MetaspaceGC::compute_new_size();
2019 
2020   // We reclaimed old regions so we should calculate the sizes to make
2021   // sure we update the old gen/space data.
2022   g1h->g1mm()->update_sizes();
2023   g1h->allocation_context_stats().update_after_mark();
2024 
2025   g1h->trace_heap_after_concurrent_cycle();
2026 }
2027 
2028 void ConcurrentMark::completeCleanup() {
2029   if (has_aborted()) return;
2030 
2031   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2032 
2033   _cleanup_list.verify_optional();
2034   FreeRegionList tmp_free_list("Tmp Free List");
2035 
2036   if (G1ConcRegionFreeingVerbose) {
2037     gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2038                            "cleanup list has %u entries",
2039                            _cleanup_list.length());
2040   }
2041 
2042   // No one else should be accessing the _cleanup_list at this point,
2043   // so it is not necessary to take any locks
2044   while (!_cleanup_list.is_empty()) {
2045     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2046     assert(hr != NULL, "Got NULL from a non-empty list");
2047     hr->par_clear();
2048     tmp_free_list.add_ordered(hr);
2049 
2050     // Instead of adding one region at a time to the secondary_free_list,
2051     // we accumulate them in the local list and move them a few at a
2052     // time. This also cuts down on the number of notify_all() calls
2053     // we do during this process. We'll also append the local list when
2054     // _cleanup_list is empty (which means we just removed the last
2055     // region from the _cleanup_list).
2056     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2057         _cleanup_list.is_empty()) {
2058       if (G1ConcRegionFreeingVerbose) {
2059         gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2060                                "appending %u entries to the secondary_free_list, "
2061                                "cleanup list still has %u entries",
2062                                tmp_free_list.length(),
2063                                _cleanup_list.length());
2064       }
2065 
2066       {
2067         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2068         g1h->secondary_free_list_add(&tmp_free_list);
2069         SecondaryFreeList_lock->notify_all();
2070       }
2071 #ifndef PRODUCT
2072       if (G1StressConcRegionFreeing) {
2073         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2074           os::sleep(Thread::current(), (jlong) 1, false);
2075         }
2076       }
2077 #endif
2078     }
2079   }
2080   assert(tmp_free_list.is_empty(), "post-condition");
2081 }
2082 
2083 // Supporting Object and Oop closures for reference discovery
2084 // and processing in during marking
2085 
2086 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2087   HeapWord* addr = (HeapWord*)obj;
2088   return addr != NULL &&
2089          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2090 }
2091 
2092 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2093 // Uses the CMTask associated with a worker thread (for serial reference
2094 // processing the CMTask for worker 0 is used) to preserve (mark) and
2095 // trace referent objects.
2096 //
2097 // Using the CMTask and embedded local queues avoids having the worker
2098 // threads operating on the global mark stack. This reduces the risk
2099 // of overflowing the stack - which we would rather avoid at this late
2100 // state. Also using the tasks' local queues removes the potential
2101 // of the workers interfering with each other that could occur if
2102 // operating on the global stack.
2103 
2104 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2105   ConcurrentMark* _cm;
2106   CMTask*         _task;
2107   int             _ref_counter_limit;
2108   int             _ref_counter;
2109   bool            _is_serial;
2110  public:
2111   G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2112     _cm(cm), _task(task), _is_serial(is_serial),
2113     _ref_counter_limit(G1RefProcDrainInterval) {
2114     assert(_ref_counter_limit > 0, "sanity");
2115     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2116     _ref_counter = _ref_counter_limit;
2117   }
2118 
2119   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2120   virtual void do_oop(      oop* p) { do_oop_work(p); }
2121 
2122   template <class T> void do_oop_work(T* p) {
2123     if (!_cm->has_overflown()) {
2124       oop obj = oopDesc::load_decode_heap_oop(p);
2125       if (_cm->verbose_high()) {
2126         gclog_or_tty->print_cr("\t[%u] we're looking at location "
2127                                "*" PTR_FORMAT " = " PTR_FORMAT,
2128                                _task->worker_id(), p2i(p), p2i((void*) obj));
2129       }
2130 
2131       _task->deal_with_reference(obj);
2132       _ref_counter--;
2133 
2134       if (_ref_counter == 0) {
2135         // We have dealt with _ref_counter_limit references, pushing them
2136         // and objects reachable from them on to the local stack (and
2137         // possibly the global stack). Call CMTask::do_marking_step() to
2138         // process these entries.
2139         //
2140         // We call CMTask::do_marking_step() in a loop, which we'll exit if
2141         // there's nothing more to do (i.e. we're done with the entries that
2142         // were pushed as a result of the CMTask::deal_with_reference() calls
2143         // above) or we overflow.
2144         //
2145         // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2146         // flag while there may still be some work to do. (See the comment at
2147         // the beginning of CMTask::do_marking_step() for those conditions -
2148         // one of which is reaching the specified time target.) It is only
2149         // when CMTask::do_marking_step() returns without setting the
2150         // has_aborted() flag that the marking step has completed.
2151         do {
2152           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2153           _task->do_marking_step(mark_step_duration_ms,
2154                                  false      /* do_termination */,
2155                                  _is_serial);
2156         } while (_task->has_aborted() && !_cm->has_overflown());
2157         _ref_counter = _ref_counter_limit;
2158       }
2159     } else {
2160       if (_cm->verbose_high()) {
2161          gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2162       }
2163     }
2164   }
2165 };
2166 
2167 // 'Drain' oop closure used by both serial and parallel reference processing.
2168 // Uses the CMTask associated with a given worker thread (for serial
2169 // reference processing the CMtask for worker 0 is used). Calls the
2170 // do_marking_step routine, with an unbelievably large timeout value,
2171 // to drain the marking data structures of the remaining entries
2172 // added by the 'keep alive' oop closure above.
2173 
2174 class G1CMDrainMarkingStackClosure: public VoidClosure {
2175   ConcurrentMark* _cm;
2176   CMTask*         _task;
2177   bool            _is_serial;
2178  public:
2179   G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2180     _cm(cm), _task(task), _is_serial(is_serial) {
2181     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2182   }
2183 
2184   void do_void() {
2185     do {
2186       if (_cm->verbose_high()) {
2187         gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2188                                _task->worker_id(), BOOL_TO_STR(_is_serial));
2189       }
2190 
2191       // We call CMTask::do_marking_step() to completely drain the local
2192       // and global marking stacks of entries pushed by the 'keep alive'
2193       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2194       //
2195       // CMTask::do_marking_step() is called in a loop, which we'll exit
2196       // if there's nothing more to do (i.e. we've completely drained the
2197       // entries that were pushed as a a result of applying the 'keep alive'
2198       // closure to the entries on the discovered ref lists) or we overflow
2199       // the global marking stack.
2200       //
2201       // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2202       // flag while there may still be some work to do. (See the comment at
2203       // the beginning of CMTask::do_marking_step() for those conditions -
2204       // one of which is reaching the specified time target.) It is only
2205       // when CMTask::do_marking_step() returns without setting the
2206       // has_aborted() flag that the marking step has completed.
2207 
2208       _task->do_marking_step(1000000000.0 /* something very large */,
2209                              true         /* do_termination */,
2210                              _is_serial);
2211     } while (_task->has_aborted() && !_cm->has_overflown());
2212   }
2213 };
2214 
2215 // Implementation of AbstractRefProcTaskExecutor for parallel
2216 // reference processing at the end of G1 concurrent marking
2217 
2218 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2219 private:
2220   G1CollectedHeap* _g1h;
2221   ConcurrentMark*  _cm;
2222   WorkGang*        _workers;
2223   uint             _active_workers;
2224 
2225 public:
2226   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2227                           ConcurrentMark* cm,
2228                           WorkGang* workers,
2229                           uint n_workers) :
2230     _g1h(g1h), _cm(cm),
2231     _workers(workers), _active_workers(n_workers) { }
2232 
2233   // Executes the given task using concurrent marking worker threads.
2234   virtual void execute(ProcessTask& task);
2235   virtual void execute(EnqueueTask& task);
2236 };
2237 
2238 class G1CMRefProcTaskProxy: public AbstractGangTask {
2239   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2240   ProcessTask&     _proc_task;
2241   G1CollectedHeap* _g1h;
2242   ConcurrentMark*  _cm;
2243 
2244 public:
2245   G1CMRefProcTaskProxy(ProcessTask& proc_task,
2246                      G1CollectedHeap* g1h,
2247                      ConcurrentMark* cm) :
2248     AbstractGangTask("Process reference objects in parallel"),
2249     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2250     ReferenceProcessor* rp = _g1h->ref_processor_cm();
2251     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2252   }
2253 
2254   virtual void work(uint worker_id) {
2255     ResourceMark rm;
2256     HandleMark hm;
2257     CMTask* task = _cm->task(worker_id);
2258     G1CMIsAliveClosure g1_is_alive(_g1h);
2259     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2260     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2261 
2262     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2263   }
2264 };
2265 
2266 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2267   assert(_workers != NULL, "Need parallel worker threads.");
2268   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2269 
2270   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2271 
2272   // We need to reset the concurrency level before each
2273   // proxy task execution, so that the termination protocol
2274   // and overflow handling in CMTask::do_marking_step() knows
2275   // how many workers to wait for.
2276   _cm->set_concurrency(_active_workers);
2277   _workers->run_task(&proc_task_proxy);
2278 }
2279 
2280 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2281   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2282   EnqueueTask& _enq_task;
2283 
2284 public:
2285   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2286     AbstractGangTask("Enqueue reference objects in parallel"),
2287     _enq_task(enq_task) { }
2288 
2289   virtual void work(uint worker_id) {
2290     _enq_task.work(worker_id);
2291   }
2292 };
2293 
2294 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2295   assert(_workers != NULL, "Need parallel worker threads.");
2296   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2297 
2298   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2299 
2300   // Not strictly necessary but...
2301   //
2302   // We need to reset the concurrency level before each
2303   // proxy task execution, so that the termination protocol
2304   // and overflow handling in CMTask::do_marking_step() knows
2305   // how many workers to wait for.
2306   _cm->set_concurrency(_active_workers);
2307   _workers->run_task(&enq_task_proxy);
2308 }
2309 
2310 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2311   G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2312 }
2313 
2314 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2315   if (has_overflown()) {
2316     // Skip processing the discovered references if we have
2317     // overflown the global marking stack. Reference objects
2318     // only get discovered once so it is OK to not
2319     // de-populate the discovered reference lists. We could have,
2320     // but the only benefit would be that, when marking restarts,
2321     // less reference objects are discovered.
2322     return;
2323   }
2324 
2325   ResourceMark rm;
2326   HandleMark   hm;
2327 
2328   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2329 
2330   // Is alive closure.
2331   G1CMIsAliveClosure g1_is_alive(g1h);
2332 
2333   // Inner scope to exclude the cleaning of the string and symbol
2334   // tables from the displayed time.
2335   {
2336     G1CMTraceTime t("GC ref-proc", G1Log::finer());
2337 
2338     ReferenceProcessor* rp = g1h->ref_processor_cm();
2339 
2340     // See the comment in G1CollectedHeap::ref_processing_init()
2341     // about how reference processing currently works in G1.
2342 
2343     // Set the soft reference policy
2344     rp->setup_policy(clear_all_soft_refs);
2345     assert(_markStack.isEmpty(), "mark stack should be empty");
2346 
2347     // Instances of the 'Keep Alive' and 'Complete GC' closures used
2348     // in serial reference processing. Note these closures are also
2349     // used for serially processing (by the the current thread) the
2350     // JNI references during parallel reference processing.
2351     //
2352     // These closures do not need to synchronize with the worker
2353     // threads involved in parallel reference processing as these
2354     // instances are executed serially by the current thread (e.g.
2355     // reference processing is not multi-threaded and is thus
2356     // performed by the current thread instead of a gang worker).
2357     //
2358     // The gang tasks involved in parallel reference processing create
2359     // their own instances of these closures, which do their own
2360     // synchronization among themselves.
2361     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2362     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2363 
2364     // We need at least one active thread. If reference processing
2365     // is not multi-threaded we use the current (VMThread) thread,
2366     // otherwise we use the work gang from the G1CollectedHeap and
2367     // we utilize all the worker threads we can.
2368     bool processing_is_mt = rp->processing_is_mt();
2369     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2370     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2371 
2372     // Parallel processing task executor.
2373     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2374                                               g1h->workers(), active_workers);
2375     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2376 
2377     // Set the concurrency level. The phase was already set prior to
2378     // executing the remark task.
2379     set_concurrency(active_workers);
2380 
2381     // Set the degree of MT processing here.  If the discovery was done MT,
2382     // the number of threads involved during discovery could differ from
2383     // the number of active workers.  This is OK as long as the discovered
2384     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2385     rp->set_active_mt_degree(active_workers);
2386 
2387     // Process the weak references.
2388     const ReferenceProcessorStats& stats =
2389         rp->process_discovered_references(&g1_is_alive,
2390                                           &g1_keep_alive,
2391                                           &g1_drain_mark_stack,
2392                                           executor,
2393                                           g1h->gc_timer_cm());
2394     g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2395 
2396     // The do_oop work routines of the keep_alive and drain_marking_stack
2397     // oop closures will set the has_overflown flag if we overflow the
2398     // global marking stack.
2399 
2400     assert(_markStack.overflow() || _markStack.isEmpty(),
2401             "mark stack should be empty (unless it overflowed)");
2402 
2403     if (_markStack.overflow()) {
2404       // This should have been done already when we tried to push an
2405       // entry on to the global mark stack. But let's do it again.
2406       set_has_overflown();
2407     }
2408 
2409     assert(rp->num_q() == active_workers, "why not");
2410 
2411     rp->enqueue_discovered_references(executor);
2412 
2413     rp->verify_no_references_recorded();
2414     assert(!rp->discovery_enabled(), "Post condition");
2415   }
2416 
2417   if (has_overflown()) {
2418     // We can not trust g1_is_alive if the marking stack overflowed
2419     return;
2420   }
2421 
2422   assert(_markStack.isEmpty(), "Marking should have completed");
2423 
2424   // Unload Klasses, String, Symbols, Code Cache, etc.
2425   {
2426     G1CMTraceTime trace("Unloading", G1Log::finer());
2427 
2428     if (ClassUnloadingWithConcurrentMark) {
2429       bool purged_classes;
2430 
2431       {
2432         G1CMTraceTime trace("System Dictionary Unloading", G1Log::finest());
2433         purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2434       }
2435 
2436       {
2437         G1CMTraceTime trace("Parallel Unloading", G1Log::finest());
2438         weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2439       }
2440     }
2441 
2442     if (G1StringDedup::is_enabled()) {
2443       G1CMTraceTime trace("String Deduplication Unlink", G1Log::finest());
2444       G1StringDedup::unlink(&g1_is_alive);
2445     }
2446   }
2447 }
2448 
2449 void ConcurrentMark::swapMarkBitMaps() {
2450   CMBitMapRO* temp = _prevMarkBitMap;
2451   _prevMarkBitMap  = (CMBitMapRO*)_nextMarkBitMap;
2452   _nextMarkBitMap  = (CMBitMap*)  temp;
2453 }
2454 
2455 // Closure for marking entries in SATB buffers.
2456 class CMSATBBufferClosure : public SATBBufferClosure {
2457 private:
2458   CMTask* _task;
2459   G1CollectedHeap* _g1h;
2460 
2461   // This is very similar to CMTask::deal_with_reference, but with
2462   // more relaxed requirements for the argument, so this must be more
2463   // circumspect about treating the argument as an object.
2464   void do_entry(void* entry) const {
2465     _task->increment_refs_reached();
2466     HeapRegion* hr = _g1h->heap_region_containing_raw(entry);
2467     if (entry < hr->next_top_at_mark_start()) {
2468       // Until we get here, we don't know whether entry refers to a valid
2469       // object; it could instead have been a stale reference.
2470       oop obj = static_cast<oop>(entry);
2471       assert(obj->is_oop(true /* ignore mark word */),
2472              err_msg("Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj)));
2473       _task->make_reference_grey(obj, hr);
2474     }
2475   }
2476 
2477 public:
2478   CMSATBBufferClosure(CMTask* task, G1CollectedHeap* g1h)
2479     : _task(task), _g1h(g1h) { }
2480 
2481   virtual void do_buffer(void** buffer, size_t size) {
2482     for (size_t i = 0; i < size; ++i) {
2483       do_entry(buffer[i]);
2484     }
2485   }
2486 };
2487 
2488 class G1RemarkThreadsClosure : public ThreadClosure {
2489   CMSATBBufferClosure _cm_satb_cl;
2490   G1CMOopClosure _cm_cl;
2491   MarkingCodeBlobClosure _code_cl;
2492   int _thread_parity;
2493 
2494  public:
2495   G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task) :
2496     _cm_satb_cl(task, g1h),
2497     _cm_cl(g1h, g1h->concurrent_mark(), task),
2498     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2499     _thread_parity(Threads::thread_claim_parity()) {}
2500 
2501   void do_thread(Thread* thread) {
2502     if (thread->is_Java_thread()) {
2503       if (thread->claim_oops_do(true, _thread_parity)) {
2504         JavaThread* jt = (JavaThread*)thread;
2505 
2506         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2507         // however the liveness of oops reachable from nmethods have very complex lifecycles:
2508         // * Alive if on the stack of an executing method
2509         // * Weakly reachable otherwise
2510         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2511         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2512         jt->nmethods_do(&_code_cl);
2513 
2514         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
2515       }
2516     } else if (thread->is_VM_thread()) {
2517       if (thread->claim_oops_do(true, _thread_parity)) {
2518         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
2519       }
2520     }
2521   }
2522 };
2523 
2524 class CMRemarkTask: public AbstractGangTask {
2525 private:
2526   ConcurrentMark* _cm;
2527 public:
2528   void work(uint worker_id) {
2529     // Since all available tasks are actually started, we should
2530     // only proceed if we're supposed to be active.
2531     if (worker_id < _cm->active_tasks()) {
2532       CMTask* task = _cm->task(worker_id);
2533       task->record_start_time();
2534       {
2535         ResourceMark rm;
2536         HandleMark hm;
2537 
2538         G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
2539         Threads::threads_do(&threads_f);
2540       }
2541 
2542       do {
2543         task->do_marking_step(1000000000.0 /* something very large */,
2544                               true         /* do_termination       */,
2545                               false        /* is_serial            */);
2546       } while (task->has_aborted() && !_cm->has_overflown());
2547       // If we overflow, then we do not want to restart. We instead
2548       // want to abort remark and do concurrent marking again.
2549       task->record_end_time();
2550     }
2551   }
2552 
2553   CMRemarkTask(ConcurrentMark* cm, uint active_workers) :
2554     AbstractGangTask("Par Remark"), _cm(cm) {
2555     _cm->terminator()->reset_for_reuse(active_workers);
2556   }
2557 };
2558 
2559 void ConcurrentMark::checkpointRootsFinalWork() {
2560   ResourceMark rm;
2561   HandleMark   hm;
2562   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2563 
2564   G1CMTraceTime trace("Finalize Marking", G1Log::finer());
2565 
2566   g1h->ensure_parsability(false);
2567 
2568   // this is remark, so we'll use up all active threads
2569   uint active_workers = g1h->workers()->active_workers();
2570   set_concurrency_and_phase(active_workers, false /* concurrent */);
2571   // Leave _parallel_marking_threads at it's
2572   // value originally calculated in the ConcurrentMark
2573   // constructor and pass values of the active workers
2574   // through the gang in the task.
2575 
2576   {
2577     StrongRootsScope srs(active_workers);
2578 
2579     CMRemarkTask remarkTask(this, active_workers);
2580     // We will start all available threads, even if we decide that the
2581     // active_workers will be fewer. The extra ones will just bail out
2582     // immediately.
2583     g1h->workers()->run_task(&remarkTask);
2584   }
2585 
2586   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2587   guarantee(has_overflown() ||
2588             satb_mq_set.completed_buffers_num() == 0,
2589             err_msg("Invariant: has_overflown = %s, num buffers = %d",
2590                     BOOL_TO_STR(has_overflown()),
2591                     satb_mq_set.completed_buffers_num()));
2592 
2593   print_stats();
2594 }
2595 
2596 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2597   // Note we are overriding the read-only view of the prev map here, via
2598   // the cast.
2599   ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2600 }
2601 
2602 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2603   _nextMarkBitMap->clearRange(mr);
2604 }
2605 
2606 HeapRegion*
2607 ConcurrentMark::claim_region(uint worker_id) {
2608   // "checkpoint" the finger
2609   HeapWord* finger = _finger;
2610 
2611   // _heap_end will not change underneath our feet; it only changes at
2612   // yield points.
2613   while (finger < _heap_end) {
2614     assert(_g1h->is_in_g1_reserved(finger), "invariant");
2615 
2616     // Note on how this code handles humongous regions. In the
2617     // normal case the finger will reach the start of a "starts
2618     // humongous" (SH) region. Its end will either be the end of the
2619     // last "continues humongous" (CH) region in the sequence, or the
2620     // standard end of the SH region (if the SH is the only region in
2621     // the sequence). That way claim_region() will skip over the CH
2622     // regions. However, there is a subtle race between a CM thread
2623     // executing this method and a mutator thread doing a humongous
2624     // object allocation. The two are not mutually exclusive as the CM
2625     // thread does not need to hold the Heap_lock when it gets
2626     // here. So there is a chance that claim_region() will come across
2627     // a free region that's in the progress of becoming a SH or a CH
2628     // region. In the former case, it will either
2629     //   a) Miss the update to the region's end, in which case it will
2630     //      visit every subsequent CH region, will find their bitmaps
2631     //      empty, and do nothing, or
2632     //   b) Will observe the update of the region's end (in which case
2633     //      it will skip the subsequent CH regions).
2634     // If it comes across a region that suddenly becomes CH, the
2635     // scenario will be similar to b). So, the race between
2636     // claim_region() and a humongous object allocation might force us
2637     // to do a bit of unnecessary work (due to some unnecessary bitmap
2638     // iterations) but it should not introduce and correctness issues.
2639     HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2640 
2641     // Above heap_region_containing_raw may return NULL as we always scan claim
2642     // until the end of the heap. In this case, just jump to the next region.
2643     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2644 
2645     // Is the gap between reading the finger and doing the CAS too long?
2646     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2647     if (res == finger && curr_region != NULL) {
2648       // we succeeded
2649       HeapWord*   bottom        = curr_region->bottom();
2650       HeapWord*   limit         = curr_region->next_top_at_mark_start();
2651 
2652       if (verbose_low()) {
2653         gclog_or_tty->print_cr("[%u] curr_region = " PTR_FORMAT " "
2654                                "[" PTR_FORMAT ", " PTR_FORMAT "), "
2655                                "limit = " PTR_FORMAT,
2656                                worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2657       }
2658 
2659       // notice that _finger == end cannot be guaranteed here since,
2660       // someone else might have moved the finger even further
2661       assert(_finger >= end, "the finger should have moved forward");
2662 
2663       if (verbose_low()) {
2664         gclog_or_tty->print_cr("[%u] we were successful with region = "
2665                                PTR_FORMAT, worker_id, p2i(curr_region));
2666       }
2667 
2668       if (limit > bottom) {
2669         if (verbose_low()) {
2670           gclog_or_tty->print_cr("[%u] region " PTR_FORMAT " is not empty, "
2671                                  "returning it ", worker_id, p2i(curr_region));
2672         }
2673         return curr_region;
2674       } else {
2675         assert(limit == bottom,
2676                "the region limit should be at bottom");
2677         if (verbose_low()) {
2678           gclog_or_tty->print_cr("[%u] region " PTR_FORMAT " is empty, "
2679                                  "returning NULL", worker_id, p2i(curr_region));
2680         }
2681         // we return NULL and the caller should try calling
2682         // claim_region() again.
2683         return NULL;
2684       }
2685     } else {
2686       assert(_finger > finger, "the finger should have moved forward");
2687       if (verbose_low()) {
2688         if (curr_region == NULL) {
2689           gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
2690                                  "global finger = " PTR_FORMAT ", "
2691                                  "our finger = " PTR_FORMAT,
2692                                  worker_id, p2i(_finger), p2i(finger));
2693         } else {
2694           gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2695                                  "global finger = " PTR_FORMAT ", "
2696                                  "our finger = " PTR_FORMAT,
2697                                  worker_id, p2i(_finger), p2i(finger));
2698         }
2699       }
2700 
2701       // read it again
2702       finger = _finger;
2703     }
2704   }
2705 
2706   return NULL;
2707 }
2708 
2709 #ifndef PRODUCT
2710 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
2711 private:
2712   G1CollectedHeap* _g1h;
2713   const char* _phase;
2714   int _info;
2715 
2716 public:
2717   VerifyNoCSetOops(const char* phase, int info = -1) :
2718     _g1h(G1CollectedHeap::heap()),
2719     _phase(phase),
2720     _info(info)
2721   { }
2722 
2723   void operator()(oop obj) const {
2724     guarantee(obj->is_oop(),
2725               err_msg("Non-oop " PTR_FORMAT ", phase: %s, info: %d",
2726                       p2i(obj), _phase, _info));
2727     guarantee(!_g1h->obj_in_cs(obj),
2728               err_msg("obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
2729                       p2i(obj), _phase, _info));
2730   }
2731 };
2732 
2733 void ConcurrentMark::verify_no_cset_oops() {
2734   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2735   if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
2736     return;
2737   }
2738 
2739   // Verify entries on the global mark stack
2740   _markStack.iterate(VerifyNoCSetOops("Stack"));
2741 
2742   // Verify entries on the task queues
2743   for (uint i = 0; i < _max_worker_id; ++i) {
2744     CMTaskQueue* queue = _task_queues->queue(i);
2745     queue->iterate(VerifyNoCSetOops("Queue", i));
2746   }
2747 
2748   // Verify the global finger
2749   HeapWord* global_finger = finger();
2750   if (global_finger != NULL && global_finger < _heap_end) {
2751     // The global finger always points to a heap region boundary. We
2752     // use heap_region_containing_raw() to get the containing region
2753     // given that the global finger could be pointing to a free region
2754     // which subsequently becomes continues humongous. If that
2755     // happens, heap_region_containing() will return the bottom of the
2756     // corresponding starts humongous region and the check below will
2757     // not hold any more.
2758     // Since we always iterate over all regions, we might get a NULL HeapRegion
2759     // here.
2760     HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
2761     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
2762               err_msg("global finger: " PTR_FORMAT " region: " HR_FORMAT,
2763                       p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
2764   }
2765 
2766   // Verify the task fingers
2767   assert(parallel_marking_threads() <= _max_worker_id, "sanity");
2768   for (uint i = 0; i < parallel_marking_threads(); ++i) {
2769     CMTask* task = _tasks[i];
2770     HeapWord* task_finger = task->finger();
2771     if (task_finger != NULL && task_finger < _heap_end) {
2772       // See above note on the global finger verification.
2773       HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
2774       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2775                 !task_hr->in_collection_set(),
2776                 err_msg("task finger: " PTR_FORMAT " region: " HR_FORMAT,
2777                         p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
2778     }
2779   }
2780 }
2781 #endif // PRODUCT
2782 
2783 // Aggregate the counting data that was constructed concurrently
2784 // with marking.
2785 class AggregateCountDataHRClosure: public HeapRegionClosure {
2786   G1CollectedHeap* _g1h;
2787   ConcurrentMark* _cm;
2788   CardTableModRefBS* _ct_bs;
2789   BitMap* _cm_card_bm;
2790   uint _max_worker_id;
2791 
2792  public:
2793   AggregateCountDataHRClosure(G1CollectedHeap* g1h,
2794                               BitMap* cm_card_bm,
2795                               uint max_worker_id) :
2796     _g1h(g1h), _cm(g1h->concurrent_mark()),
2797     _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
2798     _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
2799 
2800   bool doHeapRegion(HeapRegion* hr) {
2801     if (hr->is_continues_humongous()) {
2802       // We will ignore these here and process them when their
2803       // associated "starts humongous" region is processed.
2804       // Note that we cannot rely on their associated
2805       // "starts humongous" region to have their bit set to 1
2806       // since, due to the region chunking in the parallel region
2807       // iteration, a "continues humongous" region might be visited
2808       // before its associated "starts humongous".
2809       return false;
2810     }
2811 
2812     HeapWord* start = hr->bottom();
2813     HeapWord* limit = hr->next_top_at_mark_start();
2814     HeapWord* end = hr->end();
2815 
2816     assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
2817            err_msg("Preconditions not met - "
2818                    "start: " PTR_FORMAT ", limit: " PTR_FORMAT ", "
2819                    "top: " PTR_FORMAT ", end: " PTR_FORMAT,
2820                    p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
2821 
2822     assert(hr->next_marked_bytes() == 0, "Precondition");
2823 
2824     if (start == limit) {
2825       // NTAMS of this region has not been set so nothing to do.
2826       return false;
2827     }
2828 
2829     // 'start' should be in the heap.
2830     assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
2831     // 'end' *may* be just beyond the end of the heap (if hr is the last region)
2832     assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
2833 
2834     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
2835     BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
2836     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
2837 
2838     // If ntams is not card aligned then we bump card bitmap index
2839     // for limit so that we get the all the cards spanned by
2840     // the object ending at ntams.
2841     // Note: if this is the last region in the heap then ntams
2842     // could be actually just beyond the end of the the heap;
2843     // limit_idx will then  correspond to a (non-existent) card
2844     // that is also outside the heap.
2845     if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
2846       limit_idx += 1;
2847     }
2848 
2849     assert(limit_idx <= end_idx, "or else use atomics");
2850 
2851     // Aggregate the "stripe" in the count data associated with hr.
2852     uint hrm_index = hr->hrm_index();
2853     size_t marked_bytes = 0;
2854 
2855     for (uint i = 0; i < _max_worker_id; i += 1) {
2856       size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
2857       BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
2858 
2859       // Fetch the marked_bytes in this region for task i and
2860       // add it to the running total for this region.
2861       marked_bytes += marked_bytes_array[hrm_index];
2862 
2863       // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
2864       // into the global card bitmap.
2865       BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
2866 
2867       while (scan_idx < limit_idx) {
2868         assert(task_card_bm->at(scan_idx) == true, "should be");
2869         _cm_card_bm->set_bit(scan_idx);
2870         assert(_cm_card_bm->at(scan_idx) == true, "should be");
2871 
2872         // BitMap::get_next_one_offset() can handle the case when
2873         // its left_offset parameter is greater than its right_offset
2874         // parameter. It does, however, have an early exit if
2875         // left_offset == right_offset. So let's limit the value
2876         // passed in for left offset here.
2877         BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
2878         scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
2879       }
2880     }
2881 
2882     // Update the marked bytes for this region.
2883     hr->add_to_marked_bytes(marked_bytes);
2884 
2885     // Next heap region
2886     return false;
2887   }
2888 };
2889 
2890 class G1AggregateCountDataTask: public AbstractGangTask {
2891 protected:
2892   G1CollectedHeap* _g1h;
2893   ConcurrentMark* _cm;
2894   BitMap* _cm_card_bm;
2895   uint _max_worker_id;
2896   uint _active_workers;
2897   HeapRegionClaimer _hrclaimer;
2898 
2899 public:
2900   G1AggregateCountDataTask(G1CollectedHeap* g1h,
2901                            ConcurrentMark* cm,
2902                            BitMap* cm_card_bm,
2903                            uint max_worker_id,
2904                            uint n_workers) :
2905       AbstractGangTask("Count Aggregation"),
2906       _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
2907       _max_worker_id(max_worker_id),
2908       _active_workers(n_workers),
2909       _hrclaimer(_active_workers) {
2910   }
2911 
2912   void work(uint worker_id) {
2913     AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
2914 
2915     _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer);
2916   }
2917 };
2918 
2919 
2920 void ConcurrentMark::aggregate_count_data() {
2921   uint n_workers = _g1h->workers()->active_workers();
2922 
2923   G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
2924                                            _max_worker_id, n_workers);
2925 
2926   _g1h->workers()->run_task(&g1_par_agg_task);
2927 }
2928 
2929 // Clear the per-worker arrays used to store the per-region counting data
2930 void ConcurrentMark::clear_all_count_data() {
2931   // Clear the global card bitmap - it will be filled during
2932   // liveness count aggregation (during remark) and the
2933   // final counting task.
2934   _card_bm.clear();
2935 
2936   // Clear the global region bitmap - it will be filled as part
2937   // of the final counting task.
2938   _region_bm.clear();
2939 
2940   uint max_regions = _g1h->max_regions();
2941   assert(_max_worker_id > 0, "uninitialized");
2942 
2943   for (uint i = 0; i < _max_worker_id; i += 1) {
2944     BitMap* task_card_bm = count_card_bitmap_for(i);
2945     size_t* marked_bytes_array = count_marked_bytes_array_for(i);
2946 
2947     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
2948     assert(marked_bytes_array != NULL, "uninitialized");
2949 
2950     memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
2951     task_card_bm->clear();
2952   }
2953 }
2954 
2955 void ConcurrentMark::print_stats() {
2956   if (verbose_stats()) {
2957     gclog_or_tty->print_cr("---------------------------------------------------------------------");
2958     for (size_t i = 0; i < _active_tasks; ++i) {
2959       _tasks[i]->print_stats();
2960       gclog_or_tty->print_cr("---------------------------------------------------------------------");
2961     }
2962   }
2963 }
2964 
2965 // abandon current marking iteration due to a Full GC
2966 void ConcurrentMark::abort() {
2967   if (!cmThread()->during_cycle() || _has_aborted) {
2968     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2969     return;
2970   }
2971   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2972   // concurrent bitmap clearing.
2973   _nextMarkBitMap->clearAll();
2974 
2975   // Note we cannot clear the previous marking bitmap here
2976   // since VerifyDuringGC verifies the objects marked during
2977   // a full GC against the previous bitmap.
2978 
2979   // Clear the liveness counting data
2980   clear_all_count_data();
2981   // Empty mark stack
2982   reset_marking_state();
2983   for (uint i = 0; i < _max_worker_id; ++i) {
2984     _tasks[i]->clear_region_fields();
2985   }
2986   _first_overflow_barrier_sync.abort();
2987   _second_overflow_barrier_sync.abort();
2988   _has_aborted = true;
2989 
2990   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2991   satb_mq_set.abandon_partial_marking();
2992   // This can be called either during or outside marking, we'll read
2993   // the expected_active value from the SATB queue set.
2994   satb_mq_set.set_active_all_threads(
2995                                  false, /* new active value */
2996                                  satb_mq_set.is_active() /* expected_active */);
2997 
2998   _g1h->trace_heap_after_concurrent_cycle();
2999   _g1h->register_concurrent_cycle_end();
3000 }
3001 
3002 static void print_ms_time_info(const char* prefix, const char* name,
3003                                NumberSeq& ns) {
3004   gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3005                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3006   if (ns.num() > 0) {
3007     gclog_or_tty->print_cr("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
3008                            prefix, ns.sd(), ns.maximum());
3009   }
3010 }
3011 
3012 void ConcurrentMark::print_summary_info() {
3013   gclog_or_tty->print_cr(" Concurrent marking:");
3014   print_ms_time_info("  ", "init marks", _init_times);
3015   print_ms_time_info("  ", "remarks", _remark_times);
3016   {
3017     print_ms_time_info("     ", "final marks", _remark_mark_times);
3018     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
3019 
3020   }
3021   print_ms_time_info("  ", "cleanups", _cleanup_times);
3022   gclog_or_tty->print_cr("    Final counting total time = %8.2f s (avg = %8.2f ms).",
3023                          _total_counting_time,
3024                          (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3025                           (double)_cleanup_times.num()
3026                          : 0.0));
3027   if (G1ScrubRemSets) {
3028     gclog_or_tty->print_cr("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
3029                            _total_rs_scrub_time,
3030                            (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3031                             (double)_cleanup_times.num()
3032                            : 0.0));
3033   }
3034   gclog_or_tty->print_cr("  Total stop_world time = %8.2f s.",
3035                          (_init_times.sum() + _remark_times.sum() +
3036                           _cleanup_times.sum())/1000.0);
3037   gclog_or_tty->print_cr("  Total concurrent time = %8.2f s "
3038                 "(%8.2f s marking).",
3039                 cmThread()->vtime_accum(),
3040                 cmThread()->vtime_mark_accum());
3041 }
3042 
3043 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3044   _parallel_workers->print_worker_threads_on(st);
3045 }
3046 
3047 void ConcurrentMark::print_on_error(outputStream* st) const {
3048   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3049       p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3050   _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3051   _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3052 }
3053 
3054 // We take a break if someone is trying to stop the world.
3055 bool ConcurrentMark::do_yield_check(uint worker_id) {
3056   if (SuspendibleThreadSet::should_yield()) {
3057     if (worker_id == 0) {
3058       _g1h->g1_policy()->record_concurrent_pause();
3059     }
3060     SuspendibleThreadSet::yield();
3061     return true;
3062   } else {
3063     return false;
3064   }
3065 }
3066 
3067 #ifndef PRODUCT
3068 // for debugging purposes
3069 void ConcurrentMark::print_finger() {
3070   gclog_or_tty->print_cr("heap [" PTR_FORMAT ", " PTR_FORMAT "), global finger = " PTR_FORMAT,
3071                          p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3072   for (uint i = 0; i < _max_worker_id; ++i) {
3073     gclog_or_tty->print("   %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3074   }
3075   gclog_or_tty->cr();
3076 }
3077 #endif
3078 
3079 // Closure for iteration over bitmaps
3080 class CMBitMapClosure : public BitMapClosure {
3081 private:
3082   // the bitmap that is being iterated over
3083   CMBitMap*                   _nextMarkBitMap;
3084   ConcurrentMark*             _cm;
3085   CMTask*                     _task;
3086 
3087 public:
3088   CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3089     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3090 
3091   bool do_bit(size_t offset) {
3092     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3093     assert(_nextMarkBitMap->isMarked(addr), "invariant");
3094     assert( addr < _cm->finger(), "invariant");
3095 
3096     statsOnly( _task->increase_objs_found_on_bitmap() );
3097     assert(addr >= _task->finger(), "invariant");
3098 
3099     // We move that task's local finger along.
3100     _task->move_finger_to(addr);
3101 
3102     _task->scan_object(oop(addr));
3103     // we only partially drain the local queue and global stack
3104     _task->drain_local_queue(true);
3105     _task->drain_global_stack(true);
3106 
3107     // if the has_aborted flag has been raised, we need to bail out of
3108     // the iteration
3109     return !_task->has_aborted();
3110   }
3111 };
3112 
3113 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3114                                ConcurrentMark* cm,
3115                                CMTask* task)
3116   : _g1h(g1h), _cm(cm), _task(task) {
3117   assert(_ref_processor == NULL, "should be initialized to NULL");
3118 
3119   if (G1UseConcMarkReferenceProcessing) {
3120     _ref_processor = g1h->ref_processor_cm();
3121     assert(_ref_processor != NULL, "should not be NULL");
3122   }
3123 }
3124 
3125 void CMTask::setup_for_region(HeapRegion* hr) {
3126   assert(hr != NULL,
3127         "claim_region() should have filtered out NULL regions");
3128   assert(!hr->is_continues_humongous(),
3129         "claim_region() should have filtered out continues humongous regions");
3130 
3131   if (_cm->verbose_low()) {
3132     gclog_or_tty->print_cr("[%u] setting up for region " PTR_FORMAT,
3133                            _worker_id, p2i(hr));
3134   }
3135 
3136   _curr_region  = hr;
3137   _finger       = hr->bottom();
3138   update_region_limit();
3139 }
3140 
3141 void CMTask::update_region_limit() {
3142   HeapRegion* hr            = _curr_region;
3143   HeapWord* bottom          = hr->bottom();
3144   HeapWord* limit           = hr->next_top_at_mark_start();
3145 
3146   if (limit == bottom) {
3147     if (_cm->verbose_low()) {
3148       gclog_or_tty->print_cr("[%u] found an empty region "
3149                              "[" PTR_FORMAT ", " PTR_FORMAT ")",
3150                              _worker_id, p2i(bottom), p2i(limit));
3151     }
3152     // The region was collected underneath our feet.
3153     // We set the finger to bottom to ensure that the bitmap
3154     // iteration that will follow this will not do anything.
3155     // (this is not a condition that holds when we set the region up,
3156     // as the region is not supposed to be empty in the first place)
3157     _finger = bottom;
3158   } else if (limit >= _region_limit) {
3159     assert(limit >= _finger, "peace of mind");
3160   } else {
3161     assert(limit < _region_limit, "only way to get here");
3162     // This can happen under some pretty unusual circumstances.  An
3163     // evacuation pause empties the region underneath our feet (NTAMS
3164     // at bottom). We then do some allocation in the region (NTAMS
3165     // stays at bottom), followed by the region being used as a GC
3166     // alloc region (NTAMS will move to top() and the objects
3167     // originally below it will be grayed). All objects now marked in
3168     // the region are explicitly grayed, if below the global finger,
3169     // and we do not need in fact to scan anything else. So, we simply
3170     // set _finger to be limit to ensure that the bitmap iteration
3171     // doesn't do anything.
3172     _finger = limit;
3173   }
3174 
3175   _region_limit = limit;
3176 }
3177 
3178 void CMTask::giveup_current_region() {
3179   assert(_curr_region != NULL, "invariant");
3180   if (_cm->verbose_low()) {
3181     gclog_or_tty->print_cr("[%u] giving up region " PTR_FORMAT,
3182                            _worker_id, p2i(_curr_region));
3183   }
3184   clear_region_fields();
3185 }
3186 
3187 void CMTask::clear_region_fields() {
3188   // Values for these three fields that indicate that we're not
3189   // holding on to a region.
3190   _curr_region   = NULL;
3191   _finger        = NULL;
3192   _region_limit  = NULL;
3193 }
3194 
3195 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3196   if (cm_oop_closure == NULL) {
3197     assert(_cm_oop_closure != NULL, "invariant");
3198   } else {
3199     assert(_cm_oop_closure == NULL, "invariant");
3200   }
3201   _cm_oop_closure = cm_oop_closure;
3202 }
3203 
3204 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3205   guarantee(nextMarkBitMap != NULL, "invariant");
3206 
3207   if (_cm->verbose_low()) {
3208     gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3209   }
3210 
3211   _nextMarkBitMap                = nextMarkBitMap;
3212   clear_region_fields();
3213 
3214   _calls                         = 0;
3215   _elapsed_time_ms               = 0.0;
3216   _termination_time_ms           = 0.0;
3217   _termination_start_time_ms     = 0.0;
3218 
3219 #if _MARKING_STATS_
3220   _aborted                       = 0;
3221   _aborted_overflow              = 0;
3222   _aborted_cm_aborted            = 0;
3223   _aborted_yield                 = 0;
3224   _aborted_timed_out             = 0;
3225   _aborted_satb                  = 0;
3226   _aborted_termination           = 0;
3227   _steal_attempts                = 0;
3228   _steals                        = 0;
3229   _local_pushes                  = 0;
3230   _local_pops                    = 0;
3231   _local_max_size                = 0;
3232   _objs_scanned                  = 0;
3233   _global_pushes                 = 0;
3234   _global_pops                   = 0;
3235   _global_max_size               = 0;
3236   _global_transfers_to           = 0;
3237   _global_transfers_from         = 0;
3238   _regions_claimed               = 0;
3239   _objs_found_on_bitmap          = 0;
3240   _satb_buffers_processed        = 0;
3241 #endif // _MARKING_STATS_
3242 }
3243 
3244 bool CMTask::should_exit_termination() {
3245   regular_clock_call();
3246   // This is called when we are in the termination protocol. We should
3247   // quit if, for some reason, this task wants to abort or the global
3248   // stack is not empty (this means that we can get work from it).
3249   return !_cm->mark_stack_empty() || has_aborted();
3250 }
3251 
3252 void CMTask::reached_limit() {
3253   assert(_words_scanned >= _words_scanned_limit ||
3254          _refs_reached >= _refs_reached_limit ,
3255          "shouldn't have been called otherwise");
3256   regular_clock_call();
3257 }
3258 
3259 void CMTask::regular_clock_call() {
3260   if (has_aborted()) return;
3261 
3262   // First, we need to recalculate the words scanned and refs reached
3263   // limits for the next clock call.
3264   recalculate_limits();
3265 
3266   // During the regular clock call we do the following
3267 
3268   // (1) If an overflow has been flagged, then we abort.
3269   if (_cm->has_overflown()) {
3270     set_has_aborted();
3271     return;
3272   }
3273 
3274   // If we are not concurrent (i.e. we're doing remark) we don't need
3275   // to check anything else. The other steps are only needed during
3276   // the concurrent marking phase.
3277   if (!concurrent()) return;
3278 
3279   // (2) If marking has been aborted for Full GC, then we also abort.
3280   if (_cm->has_aborted()) {
3281     set_has_aborted();
3282     statsOnly( ++_aborted_cm_aborted );
3283     return;
3284   }
3285 
3286   double curr_time_ms = os::elapsedVTime() * 1000.0;
3287 
3288   // (3) If marking stats are enabled, then we update the step history.
3289 #if _MARKING_STATS_
3290   if (_words_scanned >= _words_scanned_limit) {
3291     ++_clock_due_to_scanning;
3292   }
3293   if (_refs_reached >= _refs_reached_limit) {
3294     ++_clock_due_to_marking;
3295   }
3296 
3297   double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3298   _interval_start_time_ms = curr_time_ms;
3299   _all_clock_intervals_ms.add(last_interval_ms);
3300 
3301   if (_cm->verbose_medium()) {
3302       gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3303                         "scanned = " SIZE_FORMAT "%s, refs reached = " SIZE_FORMAT "%s",
3304                         _worker_id, last_interval_ms,
3305                         _words_scanned,
3306                         (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3307                         _refs_reached,
3308                         (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3309   }
3310 #endif // _MARKING_STATS_
3311 
3312   // (4) We check whether we should yield. If we have to, then we abort.
3313   if (SuspendibleThreadSet::should_yield()) {
3314     // We should yield. To do this we abort the task. The caller is
3315     // responsible for yielding.
3316     set_has_aborted();
3317     statsOnly( ++_aborted_yield );
3318     return;
3319   }
3320 
3321   // (5) We check whether we've reached our time quota. If we have,
3322   // then we abort.
3323   double elapsed_time_ms = curr_time_ms - _start_time_ms;
3324   if (elapsed_time_ms > _time_target_ms) {
3325     set_has_aborted();
3326     _has_timed_out = true;
3327     statsOnly( ++_aborted_timed_out );
3328     return;
3329   }
3330 
3331   // (6) Finally, we check whether there are enough completed STAB
3332   // buffers available for processing. If there are, we abort.
3333   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3334   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3335     if (_cm->verbose_low()) {
3336       gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3337                              _worker_id);
3338     }
3339     // we do need to process SATB buffers, we'll abort and restart
3340     // the marking task to do so
3341     set_has_aborted();
3342     statsOnly( ++_aborted_satb );
3343     return;
3344   }
3345 }
3346 
3347 void CMTask::recalculate_limits() {
3348   _real_words_scanned_limit = _words_scanned + words_scanned_period;
3349   _words_scanned_limit      = _real_words_scanned_limit;
3350 
3351   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
3352   _refs_reached_limit       = _real_refs_reached_limit;
3353 }
3354 
3355 void CMTask::decrease_limits() {
3356   // This is called when we believe that we're going to do an infrequent
3357   // operation which will increase the per byte scanned cost (i.e. move
3358   // entries to/from the global stack). It basically tries to decrease the
3359   // scanning limit so that the clock is called earlier.
3360 
3361   if (_cm->verbose_medium()) {
3362     gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3363   }
3364 
3365   _words_scanned_limit = _real_words_scanned_limit -
3366     3 * words_scanned_period / 4;
3367   _refs_reached_limit  = _real_refs_reached_limit -
3368     3 * refs_reached_period / 4;
3369 }
3370 
3371 void CMTask::move_entries_to_global_stack() {
3372   // local array where we'll store the entries that will be popped
3373   // from the local queue
3374   oop buffer[global_stack_transfer_size];
3375 
3376   int n = 0;
3377   oop obj;
3378   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3379     buffer[n] = obj;
3380     ++n;
3381   }
3382 
3383   if (n > 0) {
3384     // we popped at least one entry from the local queue
3385 
3386     statsOnly( ++_global_transfers_to; _local_pops += n );
3387 
3388     if (!_cm->mark_stack_push(buffer, n)) {
3389       if (_cm->verbose_low()) {
3390         gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3391                                _worker_id);
3392       }
3393       set_has_aborted();
3394     } else {
3395       // the transfer was successful
3396 
3397       if (_cm->verbose_medium()) {
3398         gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3399                                _worker_id, n);
3400       }
3401       statsOnly( size_t tmp_size = _cm->mark_stack_size();
3402                  if (tmp_size > _global_max_size) {
3403                    _global_max_size = tmp_size;
3404                  }
3405                  _global_pushes += n );
3406     }
3407   }
3408 
3409   // this operation was quite expensive, so decrease the limits
3410   decrease_limits();
3411 }
3412 
3413 void CMTask::get_entries_from_global_stack() {
3414   // local array where we'll store the entries that will be popped
3415   // from the global stack.
3416   oop buffer[global_stack_transfer_size];
3417   int n;
3418   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3419   assert(n <= global_stack_transfer_size,
3420          "we should not pop more than the given limit");
3421   if (n > 0) {
3422     // yes, we did actually pop at least one entry
3423 
3424     statsOnly( ++_global_transfers_from; _global_pops += n );
3425     if (_cm->verbose_medium()) {
3426       gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3427                              _worker_id, n);
3428     }
3429     for (int i = 0; i < n; ++i) {
3430       bool success = _task_queue->push(buffer[i]);
3431       // We only call this when the local queue is empty or under a
3432       // given target limit. So, we do not expect this push to fail.
3433       assert(success, "invariant");
3434     }
3435 
3436     statsOnly( size_t tmp_size = (size_t)_task_queue->size();
3437                if (tmp_size > _local_max_size) {
3438                  _local_max_size = tmp_size;
3439                }
3440                _local_pushes += n );
3441   }
3442 
3443   // this operation was quite expensive, so decrease the limits
3444   decrease_limits();
3445 }
3446 
3447 void CMTask::drain_local_queue(bool partially) {
3448   if (has_aborted()) return;
3449 
3450   // Decide what the target size is, depending whether we're going to
3451   // drain it partially (so that other tasks can steal if they run out
3452   // of things to do) or totally (at the very end).
3453   size_t target_size;
3454   if (partially) {
3455     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3456   } else {
3457     target_size = 0;
3458   }
3459 
3460   if (_task_queue->size() > target_size) {
3461     if (_cm->verbose_high()) {
3462       gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3463                              _worker_id, target_size);
3464     }
3465 
3466     oop obj;
3467     bool ret = _task_queue->pop_local(obj);
3468     while (ret) {
3469       statsOnly( ++_local_pops );
3470 
3471       if (_cm->verbose_high()) {
3472         gclog_or_tty->print_cr("[%u] popped " PTR_FORMAT, _worker_id,
3473                                p2i((void*) obj));
3474       }
3475 
3476       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3477       assert(!_g1h->is_on_master_free_list(
3478                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3479 
3480       scan_object(obj);
3481 
3482       if (_task_queue->size() <= target_size || has_aborted()) {
3483         ret = false;
3484       } else {
3485         ret = _task_queue->pop_local(obj);
3486       }
3487     }
3488 
3489     if (_cm->verbose_high()) {
3490       gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3491                              _worker_id, _task_queue->size());
3492     }
3493   }
3494 }
3495 
3496 void CMTask::drain_global_stack(bool partially) {
3497   if (has_aborted()) return;
3498 
3499   // We have a policy to drain the local queue before we attempt to
3500   // drain the global stack.
3501   assert(partially || _task_queue->size() == 0, "invariant");
3502 
3503   // Decide what the target size is, depending whether we're going to
3504   // drain it partially (so that other tasks can steal if they run out
3505   // of things to do) or totally (at the very end).  Notice that,
3506   // because we move entries from the global stack in chunks or
3507   // because another task might be doing the same, we might in fact
3508   // drop below the target. But, this is not a problem.
3509   size_t target_size;
3510   if (partially) {
3511     target_size = _cm->partial_mark_stack_size_target();
3512   } else {
3513     target_size = 0;
3514   }
3515 
3516   if (_cm->mark_stack_size() > target_size) {
3517     if (_cm->verbose_low()) {
3518       gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3519                              _worker_id, target_size);
3520     }
3521 
3522     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3523       get_entries_from_global_stack();
3524       drain_local_queue(partially);
3525     }
3526 
3527     if (_cm->verbose_low()) {
3528       gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3529                              _worker_id, _cm->mark_stack_size());
3530     }
3531   }
3532 }
3533 
3534 // SATB Queue has several assumptions on whether to call the par or
3535 // non-par versions of the methods. this is why some of the code is
3536 // replicated. We should really get rid of the single-threaded version
3537 // of the code to simplify things.
3538 void CMTask::drain_satb_buffers() {
3539   if (has_aborted()) return;
3540 
3541   // We set this so that the regular clock knows that we're in the
3542   // middle of draining buffers and doesn't set the abort flag when it
3543   // notices that SATB buffers are available for draining. It'd be
3544   // very counter productive if it did that. :-)
3545   _draining_satb_buffers = true;
3546 
3547   CMSATBBufferClosure satb_cl(this, _g1h);
3548   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3549 
3550   // This keeps claiming and applying the closure to completed buffers
3551   // until we run out of buffers or we need to abort.
3552   while (!has_aborted() &&
3553          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
3554     if (_cm->verbose_medium()) {
3555       gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3556     }
3557     statsOnly( ++_satb_buffers_processed );
3558     regular_clock_call();
3559   }
3560 
3561   _draining_satb_buffers = false;
3562 
3563   assert(has_aborted() ||
3564          concurrent() ||
3565          satb_mq_set.completed_buffers_num() == 0, "invariant");
3566 
3567   // again, this was a potentially expensive operation, decrease the
3568   // limits to get the regular clock call early
3569   decrease_limits();
3570 }
3571 
3572 void CMTask::print_stats() {
3573   gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3574                          _worker_id, _calls);
3575   gclog_or_tty->print_cr("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3576                          _elapsed_time_ms, _termination_time_ms);
3577   gclog_or_tty->print_cr("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3578                          _step_times_ms.num(), _step_times_ms.avg(),
3579                          _step_times_ms.sd());
3580   gclog_or_tty->print_cr("                    max = %1.2lfms, total = %1.2lfms",
3581                          _step_times_ms.maximum(), _step_times_ms.sum());
3582 
3583 #if _MARKING_STATS_
3584   gclog_or_tty->print_cr("  Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3585                          _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3586                          _all_clock_intervals_ms.sd());
3587   gclog_or_tty->print_cr("                         max = %1.2lfms, total = %1.2lfms",
3588                          _all_clock_intervals_ms.maximum(),
3589                          _all_clock_intervals_ms.sum());
3590   gclog_or_tty->print_cr("  Clock Causes (cum): scanning = " SIZE_FORMAT ", marking = " SIZE_FORMAT,
3591                          _clock_due_to_scanning, _clock_due_to_marking);
3592   gclog_or_tty->print_cr("  Objects: scanned = " SIZE_FORMAT ", found on the bitmap = " SIZE_FORMAT,
3593                          _objs_scanned, _objs_found_on_bitmap);
3594   gclog_or_tty->print_cr("  Local Queue:  pushes = " SIZE_FORMAT ", pops = " SIZE_FORMAT ", max size = " SIZE_FORMAT,
3595                          _local_pushes, _local_pops, _local_max_size);
3596   gclog_or_tty->print_cr("  Global Stack: pushes = " SIZE_FORMAT ", pops = " SIZE_FORMAT ", max size = " SIZE_FORMAT,
3597                          _global_pushes, _global_pops, _global_max_size);
3598   gclog_or_tty->print_cr("                transfers to = " SIZE_FORMAT ", transfers from = " SIZE_FORMAT,
3599                          _global_transfers_to,_global_transfers_from);
3600   gclog_or_tty->print_cr("  Regions: claimed = " SIZE_FORMAT, _regions_claimed);
3601   gclog_or_tty->print_cr("  SATB buffers: processed = " SIZE_FORMAT, _satb_buffers_processed);
3602   gclog_or_tty->print_cr("  Steals: attempts = " SIZE_FORMAT ", successes = " SIZE_FORMAT,
3603                          _steal_attempts, _steals);
3604   gclog_or_tty->print_cr("  Aborted: " SIZE_FORMAT ", due to", _aborted);
3605   gclog_or_tty->print_cr("    overflow: " SIZE_FORMAT ", global abort: " SIZE_FORMAT ", yield: " SIZE_FORMAT,
3606                          _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3607   gclog_or_tty->print_cr("    time out: " SIZE_FORMAT ", SATB: " SIZE_FORMAT ", termination: " SIZE_FORMAT,
3608                          _aborted_timed_out, _aborted_satb, _aborted_termination);
3609 #endif // _MARKING_STATS_
3610 }
3611 
3612 bool ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) {
3613   return _task_queues->steal(worker_id, hash_seed, obj);
3614 }
3615 
3616 /*****************************************************************************
3617 
3618     The do_marking_step(time_target_ms, ...) method is the building
3619     block of the parallel marking framework. It can be called in parallel
3620     with other invocations of do_marking_step() on different tasks
3621     (but only one per task, obviously) and concurrently with the
3622     mutator threads, or during remark, hence it eliminates the need
3623     for two versions of the code. When called during remark, it will
3624     pick up from where the task left off during the concurrent marking
3625     phase. Interestingly, tasks are also claimable during evacuation
3626     pauses too, since do_marking_step() ensures that it aborts before
3627     it needs to yield.
3628 
3629     The data structures that it uses to do marking work are the
3630     following:
3631 
3632       (1) Marking Bitmap. If there are gray objects that appear only
3633       on the bitmap (this happens either when dealing with an overflow
3634       or when the initial marking phase has simply marked the roots
3635       and didn't push them on the stack), then tasks claim heap
3636       regions whose bitmap they then scan to find gray objects. A
3637       global finger indicates where the end of the last claimed region
3638       is. A local finger indicates how far into the region a task has
3639       scanned. The two fingers are used to determine how to gray an
3640       object (i.e. whether simply marking it is OK, as it will be
3641       visited by a task in the future, or whether it needs to be also
3642       pushed on a stack).
3643 
3644       (2) Local Queue. The local queue of the task which is accessed
3645       reasonably efficiently by the task. Other tasks can steal from
3646       it when they run out of work. Throughout the marking phase, a
3647       task attempts to keep its local queue short but not totally
3648       empty, so that entries are available for stealing by other
3649       tasks. Only when there is no more work, a task will totally
3650       drain its local queue.
3651 
3652       (3) Global Mark Stack. This handles local queue overflow. During
3653       marking only sets of entries are moved between it and the local
3654       queues, as access to it requires a mutex and more fine-grain
3655       interaction with it which might cause contention. If it
3656       overflows, then the marking phase should restart and iterate
3657       over the bitmap to identify gray objects. Throughout the marking
3658       phase, tasks attempt to keep the global mark stack at a small
3659       length but not totally empty, so that entries are available for
3660       popping by other tasks. Only when there is no more work, tasks
3661       will totally drain the global mark stack.
3662 
3663       (4) SATB Buffer Queue. This is where completed SATB buffers are
3664       made available. Buffers are regularly removed from this queue
3665       and scanned for roots, so that the queue doesn't get too
3666       long. During remark, all completed buffers are processed, as
3667       well as the filled in parts of any uncompleted buffers.
3668 
3669     The do_marking_step() method tries to abort when the time target
3670     has been reached. There are a few other cases when the
3671     do_marking_step() method also aborts:
3672 
3673       (1) When the marking phase has been aborted (after a Full GC).
3674 
3675       (2) When a global overflow (on the global stack) has been
3676       triggered. Before the task aborts, it will actually sync up with
3677       the other tasks to ensure that all the marking data structures
3678       (local queues, stacks, fingers etc.)  are re-initialized so that
3679       when do_marking_step() completes, the marking phase can
3680       immediately restart.
3681 
3682       (3) When enough completed SATB buffers are available. The
3683       do_marking_step() method only tries to drain SATB buffers right
3684       at the beginning. So, if enough buffers are available, the
3685       marking step aborts and the SATB buffers are processed at
3686       the beginning of the next invocation.
3687 
3688       (4) To yield. when we have to yield then we abort and yield
3689       right at the end of do_marking_step(). This saves us from a lot
3690       of hassle as, by yielding we might allow a Full GC. If this
3691       happens then objects will be compacted underneath our feet, the
3692       heap might shrink, etc. We save checking for this by just
3693       aborting and doing the yield right at the end.
3694 
3695     From the above it follows that the do_marking_step() method should
3696     be called in a loop (or, otherwise, regularly) until it completes.
3697 
3698     If a marking step completes without its has_aborted() flag being
3699     true, it means it has completed the current marking phase (and
3700     also all other marking tasks have done so and have all synced up).
3701 
3702     A method called regular_clock_call() is invoked "regularly" (in
3703     sub ms intervals) throughout marking. It is this clock method that
3704     checks all the abort conditions which were mentioned above and
3705     decides when the task should abort. A work-based scheme is used to
3706     trigger this clock method: when the number of object words the
3707     marking phase has scanned or the number of references the marking
3708     phase has visited reach a given limit. Additional invocations to
3709     the method clock have been planted in a few other strategic places
3710     too. The initial reason for the clock method was to avoid calling
3711     vtime too regularly, as it is quite expensive. So, once it was in
3712     place, it was natural to piggy-back all the other conditions on it
3713     too and not constantly check them throughout the code.
3714 
3715     If do_termination is true then do_marking_step will enter its
3716     termination protocol.
3717 
3718     The value of is_serial must be true when do_marking_step is being
3719     called serially (i.e. by the VMThread) and do_marking_step should
3720     skip any synchronization in the termination and overflow code.
3721     Examples include the serial remark code and the serial reference
3722     processing closures.
3723 
3724     The value of is_serial must be false when do_marking_step is
3725     being called by any of the worker threads in a work gang.
3726     Examples include the concurrent marking code (CMMarkingTask),
3727     the MT remark code, and the MT reference processing closures.
3728 
3729  *****************************************************************************/
3730 
3731 void CMTask::do_marking_step(double time_target_ms,
3732                              bool do_termination,
3733                              bool is_serial) {
3734   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
3735   assert(concurrent() == _cm->concurrent(), "they should be the same");
3736 
3737   G1CollectorPolicy* g1_policy = _g1h->g1_policy();
3738   assert(_task_queues != NULL, "invariant");
3739   assert(_task_queue != NULL, "invariant");
3740   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
3741 
3742   assert(!_claimed,
3743          "only one thread should claim this task at any one time");
3744 
3745   // OK, this doesn't safeguard again all possible scenarios, as it is
3746   // possible for two threads to set the _claimed flag at the same
3747   // time. But it is only for debugging purposes anyway and it will
3748   // catch most problems.
3749   _claimed = true;
3750 
3751   _start_time_ms = os::elapsedVTime() * 1000.0;
3752   statsOnly( _interval_start_time_ms = _start_time_ms );
3753 
3754   // If do_stealing is true then do_marking_step will attempt to
3755   // steal work from the other CMTasks. It only makes sense to
3756   // enable stealing when the termination protocol is enabled
3757   // and do_marking_step() is not being called serially.
3758   bool do_stealing = do_termination && !is_serial;
3759 
3760   double diff_prediction_ms =
3761     g1_policy->get_new_prediction(&_marking_step_diffs_ms);
3762   _time_target_ms = time_target_ms - diff_prediction_ms;
3763 
3764   // set up the variables that are used in the work-based scheme to
3765   // call the regular clock method
3766   _words_scanned = 0;
3767   _refs_reached  = 0;
3768   recalculate_limits();
3769 
3770   // clear all flags
3771   clear_has_aborted();
3772   _has_timed_out = false;
3773   _draining_satb_buffers = false;
3774 
3775   ++_calls;
3776 
3777   if (_cm->verbose_low()) {
3778     gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
3779                            "target = %1.2lfms >>>>>>>>>>",
3780                            _worker_id, _calls, _time_target_ms);
3781   }
3782 
3783   // Set up the bitmap and oop closures. Anything that uses them is
3784   // eventually called from this method, so it is OK to allocate these
3785   // statically.
3786   CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
3787   G1CMOopClosure  cm_oop_closure(_g1h, _cm, this);
3788   set_cm_oop_closure(&cm_oop_closure);
3789 
3790   if (_cm->has_overflown()) {
3791     // This can happen if the mark stack overflows during a GC pause
3792     // and this task, after a yield point, restarts. We have to abort
3793     // as we need to get into the overflow protocol which happens
3794     // right at the end of this task.
3795     set_has_aborted();
3796   }
3797 
3798   // First drain any available SATB buffers. After this, we will not
3799   // look at SATB buffers before the next invocation of this method.
3800   // If enough completed SATB buffers are queued up, the regular clock
3801   // will abort this task so that it restarts.
3802   drain_satb_buffers();
3803   // ...then partially drain the local queue and the global stack
3804   drain_local_queue(true);
3805   drain_global_stack(true);
3806 
3807   do {
3808     if (!has_aborted() && _curr_region != NULL) {
3809       // This means that we're already holding on to a region.
3810       assert(_finger != NULL, "if region is not NULL, then the finger "
3811              "should not be NULL either");
3812 
3813       // We might have restarted this task after an evacuation pause
3814       // which might have evacuated the region we're holding on to
3815       // underneath our feet. Let's read its limit again to make sure
3816       // that we do not iterate over a region of the heap that
3817       // contains garbage (update_region_limit() will also move
3818       // _finger to the start of the region if it is found empty).
3819       update_region_limit();
3820       // We will start from _finger not from the start of the region,
3821       // as we might be restarting this task after aborting half-way
3822       // through scanning this region. In this case, _finger points to
3823       // the address where we last found a marked object. If this is a
3824       // fresh region, _finger points to start().
3825       MemRegion mr = MemRegion(_finger, _region_limit);
3826 
3827       if (_cm->verbose_low()) {
3828         gclog_or_tty->print_cr("[%u] we're scanning part "
3829                                "[" PTR_FORMAT ", " PTR_FORMAT ") "
3830                                "of region " HR_FORMAT,
3831                                _worker_id, p2i(_finger), p2i(_region_limit),
3832                                HR_FORMAT_PARAMS(_curr_region));
3833       }
3834 
3835       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
3836              "humongous regions should go around loop once only");
3837 
3838       // Some special cases:
3839       // If the memory region is empty, we can just give up the region.
3840       // If the current region is humongous then we only need to check
3841       // the bitmap for the bit associated with the start of the object,
3842       // scan the object if it's live, and give up the region.
3843       // Otherwise, let's iterate over the bitmap of the part of the region
3844       // that is left.
3845       // If the iteration is successful, give up the region.
3846       if (mr.is_empty()) {
3847         giveup_current_region();
3848         regular_clock_call();
3849       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
3850         if (_nextMarkBitMap->isMarked(mr.start())) {
3851           // The object is marked - apply the closure
3852           BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
3853           bitmap_closure.do_bit(offset);
3854         }
3855         // Even if this task aborted while scanning the humongous object
3856         // we can (and should) give up the current region.
3857         giveup_current_region();
3858         regular_clock_call();
3859       } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
3860         giveup_current_region();
3861         regular_clock_call();
3862       } else {
3863         assert(has_aborted(), "currently the only way to do so");
3864         // The only way to abort the bitmap iteration is to return
3865         // false from the do_bit() method. However, inside the
3866         // do_bit() method we move the _finger to point to the
3867         // object currently being looked at. So, if we bail out, we
3868         // have definitely set _finger to something non-null.
3869         assert(_finger != NULL, "invariant");
3870 
3871         // Region iteration was actually aborted. So now _finger
3872         // points to the address of the object we last scanned. If we
3873         // leave it there, when we restart this task, we will rescan
3874         // the object. It is easy to avoid this. We move the finger by
3875         // enough to point to the next possible object header (the
3876         // bitmap knows by how much we need to move it as it knows its
3877         // granularity).
3878         assert(_finger < _region_limit, "invariant");
3879         HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
3880         // Check if bitmap iteration was aborted while scanning the last object
3881         if (new_finger >= _region_limit) {
3882           giveup_current_region();
3883         } else {
3884           move_finger_to(new_finger);
3885         }
3886       }
3887     }
3888     // At this point we have either completed iterating over the
3889     // region we were holding on to, or we have aborted.
3890 
3891     // We then partially drain the local queue and the global stack.
3892     // (Do we really need this?)
3893     drain_local_queue(true);
3894     drain_global_stack(true);
3895 
3896     // Read the note on the claim_region() method on why it might
3897     // return NULL with potentially more regions available for
3898     // claiming and why we have to check out_of_regions() to determine
3899     // whether we're done or not.
3900     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
3901       // We are going to try to claim a new region. We should have
3902       // given up on the previous one.
3903       // Separated the asserts so that we know which one fires.
3904       assert(_curr_region  == NULL, "invariant");
3905       assert(_finger       == NULL, "invariant");
3906       assert(_region_limit == NULL, "invariant");
3907       if (_cm->verbose_low()) {
3908         gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
3909       }
3910       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
3911       if (claimed_region != NULL) {
3912         // Yes, we managed to claim one
3913         statsOnly( ++_regions_claimed );
3914 
3915         if (_cm->verbose_low()) {
3916           gclog_or_tty->print_cr("[%u] we successfully claimed "
3917                                  "region " PTR_FORMAT,
3918                                  _worker_id, p2i(claimed_region));
3919         }
3920 
3921         setup_for_region(claimed_region);
3922         assert(_curr_region == claimed_region, "invariant");
3923       }
3924       // It is important to call the regular clock here. It might take
3925       // a while to claim a region if, for example, we hit a large
3926       // block of empty regions. So we need to call the regular clock
3927       // method once round the loop to make sure it's called
3928       // frequently enough.
3929       regular_clock_call();
3930     }
3931 
3932     if (!has_aborted() && _curr_region == NULL) {
3933       assert(_cm->out_of_regions(),
3934              "at this point we should be out of regions");
3935     }
3936   } while ( _curr_region != NULL && !has_aborted());
3937 
3938   if (!has_aborted()) {
3939     // We cannot check whether the global stack is empty, since other
3940     // tasks might be pushing objects to it concurrently.
3941     assert(_cm->out_of_regions(),
3942            "at this point we should be out of regions");
3943 
3944     if (_cm->verbose_low()) {
3945       gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
3946     }
3947 
3948     // Try to reduce the number of available SATB buffers so that
3949     // remark has less work to do.
3950     drain_satb_buffers();
3951   }
3952 
3953   // Since we've done everything else, we can now totally drain the
3954   // local queue and global stack.
3955   drain_local_queue(false);
3956   drain_global_stack(false);
3957 
3958   // Attempt at work stealing from other task's queues.
3959   if (do_stealing && !has_aborted()) {
3960     // We have not aborted. This means that we have finished all that
3961     // we could. Let's try to do some stealing...
3962 
3963     // We cannot check whether the global stack is empty, since other
3964     // tasks might be pushing objects to it concurrently.
3965     assert(_cm->out_of_regions() && _task_queue->size() == 0,
3966            "only way to reach here");
3967 
3968     if (_cm->verbose_low()) {
3969       gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
3970     }
3971 
3972     while (!has_aborted()) {
3973       oop obj;
3974       statsOnly( ++_steal_attempts );
3975 
3976       if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
3977         if (_cm->verbose_medium()) {
3978           gclog_or_tty->print_cr("[%u] stolen " PTR_FORMAT " successfully",
3979                                  _worker_id, p2i((void*) obj));
3980         }
3981 
3982         statsOnly( ++_steals );
3983 
3984         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
3985                "any stolen object should be marked");
3986         scan_object(obj);
3987 
3988         // And since we're towards the end, let's totally drain the
3989         // local queue and global stack.
3990         drain_local_queue(false);
3991         drain_global_stack(false);
3992       } else {
3993         break;
3994       }
3995     }
3996   }
3997 
3998   // If we are about to wrap up and go into termination, check if we
3999   // should raise the overflow flag.
4000   if (do_termination && !has_aborted()) {
4001     if (_cm->force_overflow()->should_force()) {
4002       _cm->set_has_overflown();
4003       regular_clock_call();
4004     }
4005   }
4006 
4007   // We still haven't aborted. Now, let's try to get into the
4008   // termination protocol.
4009   if (do_termination && !has_aborted()) {
4010     // We cannot check whether the global stack is empty, since other
4011     // tasks might be concurrently pushing objects on it.
4012     // Separated the asserts so that we know which one fires.
4013     assert(_cm->out_of_regions(), "only way to reach here");
4014     assert(_task_queue->size() == 0, "only way to reach here");
4015 
4016     if (_cm->verbose_low()) {
4017       gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4018     }
4019 
4020     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4021 
4022     // The CMTask class also extends the TerminatorTerminator class,
4023     // hence its should_exit_termination() method will also decide
4024     // whether to exit the termination protocol or not.
4025     bool finished = (is_serial ||
4026                      _cm->terminator()->offer_termination(this));
4027     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4028     _termination_time_ms +=
4029       termination_end_time_ms - _termination_start_time_ms;
4030 
4031     if (finished) {
4032       // We're all done.
4033 
4034       if (_worker_id == 0) {
4035         // let's allow task 0 to do this
4036         if (concurrent()) {
4037           assert(_cm->concurrent_marking_in_progress(), "invariant");
4038           // we need to set this to false before the next
4039           // safepoint. This way we ensure that the marking phase
4040           // doesn't observe any more heap expansions.
4041           _cm->clear_concurrent_marking_in_progress();
4042         }
4043       }
4044 
4045       // We can now guarantee that the global stack is empty, since
4046       // all other tasks have finished. We separated the guarantees so
4047       // that, if a condition is false, we can immediately find out
4048       // which one.
4049       guarantee(_cm->out_of_regions(), "only way to reach here");
4050       guarantee(_cm->mark_stack_empty(), "only way to reach here");
4051       guarantee(_task_queue->size() == 0, "only way to reach here");
4052       guarantee(!_cm->has_overflown(), "only way to reach here");
4053       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4054 
4055       if (_cm->verbose_low()) {
4056         gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4057       }
4058     } else {
4059       // Apparently there's more work to do. Let's abort this task. It
4060       // will restart it and we can hopefully find more things to do.
4061 
4062       if (_cm->verbose_low()) {
4063         gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4064                                _worker_id);
4065       }
4066 
4067       set_has_aborted();
4068       statsOnly( ++_aborted_termination );
4069     }
4070   }
4071 
4072   // Mainly for debugging purposes to make sure that a pointer to the
4073   // closure which was statically allocated in this frame doesn't
4074   // escape it by accident.
4075   set_cm_oop_closure(NULL);
4076   double end_time_ms = os::elapsedVTime() * 1000.0;
4077   double elapsed_time_ms = end_time_ms - _start_time_ms;
4078   // Update the step history.
4079   _step_times_ms.add(elapsed_time_ms);
4080 
4081   if (has_aborted()) {
4082     // The task was aborted for some reason.
4083 
4084     statsOnly( ++_aborted );
4085 
4086     if (_has_timed_out) {
4087       double diff_ms = elapsed_time_ms - _time_target_ms;
4088       // Keep statistics of how well we did with respect to hitting
4089       // our target only if we actually timed out (if we aborted for
4090       // other reasons, then the results might get skewed).
4091       _marking_step_diffs_ms.add(diff_ms);
4092     }
4093 
4094     if (_cm->has_overflown()) {
4095       // This is the interesting one. We aborted because a global
4096       // overflow was raised. This means we have to restart the
4097       // marking phase and start iterating over regions. However, in
4098       // order to do this we have to make sure that all tasks stop
4099       // what they are doing and re-initialize in a safe manner. We
4100       // will achieve this with the use of two barrier sync points.
4101 
4102       if (_cm->verbose_low()) {
4103         gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4104       }
4105 
4106       if (!is_serial) {
4107         // We only need to enter the sync barrier if being called
4108         // from a parallel context
4109         _cm->enter_first_sync_barrier(_worker_id);
4110 
4111         // When we exit this sync barrier we know that all tasks have
4112         // stopped doing marking work. So, it's now safe to
4113         // re-initialize our data structures. At the end of this method,
4114         // task 0 will clear the global data structures.
4115       }
4116 
4117       statsOnly( ++_aborted_overflow );
4118 
4119       // We clear the local state of this task...
4120       clear_region_fields();
4121 
4122       if (!is_serial) {
4123         // ...and enter the second barrier.
4124         _cm->enter_second_sync_barrier(_worker_id);
4125       }
4126       // At this point, if we're during the concurrent phase of
4127       // marking, everything has been re-initialized and we're
4128       // ready to restart.
4129     }
4130 
4131     if (_cm->verbose_low()) {
4132       gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4133                              "elapsed = %1.2lfms <<<<<<<<<<",
4134                              _worker_id, _time_target_ms, elapsed_time_ms);
4135       if (_cm->has_aborted()) {
4136         gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4137                                _worker_id);
4138       }
4139     }
4140   } else {
4141     if (_cm->verbose_low()) {
4142       gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4143                              "elapsed = %1.2lfms <<<<<<<<<<",
4144                              _worker_id, _time_target_ms, elapsed_time_ms);
4145     }
4146   }
4147 
4148   _claimed = false;
4149 }
4150 
4151 CMTask::CMTask(uint worker_id,
4152                ConcurrentMark* cm,
4153                size_t* marked_bytes,
4154                BitMap* card_bm,
4155                CMTaskQueue* task_queue,
4156                CMTaskQueueSet* task_queues)
4157   : _g1h(G1CollectedHeap::heap()),
4158     _worker_id(worker_id), _cm(cm),
4159     _claimed(false),
4160     _nextMarkBitMap(NULL), _hash_seed(17),
4161     _task_queue(task_queue),
4162     _task_queues(task_queues),
4163     _cm_oop_closure(NULL),
4164     _marked_bytes_array(marked_bytes),
4165     _card_bm(card_bm) {
4166   guarantee(task_queue != NULL, "invariant");
4167   guarantee(task_queues != NULL, "invariant");
4168 
4169   statsOnly( _clock_due_to_scanning = 0;
4170              _clock_due_to_marking  = 0 );
4171 
4172   _marking_step_diffs_ms.add(0.5);
4173 }
4174 
4175 // These are formatting macros that are used below to ensure
4176 // consistent formatting. The *_H_* versions are used to format the
4177 // header for a particular value and they should be kept consistent
4178 // with the corresponding macro. Also note that most of the macros add
4179 // the necessary white space (as a prefix) which makes them a bit
4180 // easier to compose.
4181 
4182 // All the output lines are prefixed with this string to be able to
4183 // identify them easily in a large log file.
4184 #define G1PPRL_LINE_PREFIX            "###"
4185 
4186 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
4187 #ifdef _LP64
4188 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
4189 #else // _LP64
4190 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
4191 #endif // _LP64
4192 
4193 // For per-region info
4194 #define G1PPRL_TYPE_FORMAT            "   %-4s"
4195 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
4196 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
4197 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
4198 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
4199 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
4200 
4201 // For summary info
4202 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
4203 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
4204 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
4205 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
4206 
4207 G1PrintRegionLivenessInfoClosure::
4208 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4209   : _out(out),
4210     _total_used_bytes(0), _total_capacity_bytes(0),
4211     _total_prev_live_bytes(0), _total_next_live_bytes(0),
4212     _hum_used_bytes(0), _hum_capacity_bytes(0),
4213     _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4214     _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4215   G1CollectedHeap* g1h = G1CollectedHeap::heap();
4216   MemRegion g1_reserved = g1h->g1_reserved();
4217   double now = os::elapsedTime();
4218 
4219   // Print the header of the output.
4220   _out->cr();
4221   _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4222   _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4223                  G1PPRL_SUM_ADDR_FORMAT("reserved")
4224                  G1PPRL_SUM_BYTE_FORMAT("region-size"),
4225                  p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4226                  HeapRegion::GrainBytes);
4227   _out->print_cr(G1PPRL_LINE_PREFIX);
4228   _out->print_cr(G1PPRL_LINE_PREFIX
4229                 G1PPRL_TYPE_H_FORMAT
4230                 G1PPRL_ADDR_BASE_H_FORMAT
4231                 G1PPRL_BYTE_H_FORMAT
4232                 G1PPRL_BYTE_H_FORMAT
4233                 G1PPRL_BYTE_H_FORMAT
4234                 G1PPRL_DOUBLE_H_FORMAT
4235                 G1PPRL_BYTE_H_FORMAT
4236                 G1PPRL_BYTE_H_FORMAT,
4237                 "type", "address-range",
4238                 "used", "prev-live", "next-live", "gc-eff",
4239                 "remset", "code-roots");
4240   _out->print_cr(G1PPRL_LINE_PREFIX
4241                 G1PPRL_TYPE_H_FORMAT
4242                 G1PPRL_ADDR_BASE_H_FORMAT
4243                 G1PPRL_BYTE_H_FORMAT
4244                 G1PPRL_BYTE_H_FORMAT
4245                 G1PPRL_BYTE_H_FORMAT
4246                 G1PPRL_DOUBLE_H_FORMAT
4247                 G1PPRL_BYTE_H_FORMAT
4248                 G1PPRL_BYTE_H_FORMAT,
4249                 "", "",
4250                 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4251                 "(bytes)", "(bytes)");
4252 }
4253 
4254 // It takes as a parameter a reference to one of the _hum_* fields, it
4255 // deduces the corresponding value for a region in a humongous region
4256 // series (either the region size, or what's left if the _hum_* field
4257 // is < the region size), and updates the _hum_* field accordingly.
4258 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4259   size_t bytes = 0;
4260   // The > 0 check is to deal with the prev and next live bytes which
4261   // could be 0.
4262   if (*hum_bytes > 0) {
4263     bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4264     *hum_bytes -= bytes;
4265   }
4266   return bytes;
4267 }
4268 
4269 // It deduces the values for a region in a humongous region series
4270 // from the _hum_* fields and updates those accordingly. It assumes
4271 // that that _hum_* fields have already been set up from the "starts
4272 // humongous" region and we visit the regions in address order.
4273 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4274                                                      size_t* capacity_bytes,
4275                                                      size_t* prev_live_bytes,
4276                                                      size_t* next_live_bytes) {
4277   assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4278   *used_bytes      = get_hum_bytes(&_hum_used_bytes);
4279   *capacity_bytes  = get_hum_bytes(&_hum_capacity_bytes);
4280   *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4281   *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4282 }
4283 
4284 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4285   const char* type       = r->get_type_str();
4286   HeapWord* bottom       = r->bottom();
4287   HeapWord* end          = r->end();
4288   size_t capacity_bytes  = r->capacity();
4289   size_t used_bytes      = r->used();
4290   size_t prev_live_bytes = r->live_bytes();
4291   size_t next_live_bytes = r->next_live_bytes();
4292   double gc_eff          = r->gc_efficiency();
4293   size_t remset_bytes    = r->rem_set()->mem_size();
4294   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4295 
4296   if (r->is_starts_humongous()) {
4297     assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4298            _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4299            "they should have been zeroed after the last time we used them");
4300     // Set up the _hum_* fields.
4301     _hum_capacity_bytes  = capacity_bytes;
4302     _hum_used_bytes      = used_bytes;
4303     _hum_prev_live_bytes = prev_live_bytes;
4304     _hum_next_live_bytes = next_live_bytes;
4305     get_hum_bytes(&used_bytes, &capacity_bytes,
4306                   &prev_live_bytes, &next_live_bytes);
4307     end = bottom + HeapRegion::GrainWords;
4308   } else if (r->is_continues_humongous()) {
4309     get_hum_bytes(&used_bytes, &capacity_bytes,
4310                   &prev_live_bytes, &next_live_bytes);
4311     assert(end == bottom + HeapRegion::GrainWords, "invariant");
4312   }
4313 
4314   _total_used_bytes      += used_bytes;
4315   _total_capacity_bytes  += capacity_bytes;
4316   _total_prev_live_bytes += prev_live_bytes;
4317   _total_next_live_bytes += next_live_bytes;
4318   _total_remset_bytes    += remset_bytes;
4319   _total_strong_code_roots_bytes += strong_code_roots_bytes;
4320 
4321   // Print a line for this particular region.
4322   _out->print_cr(G1PPRL_LINE_PREFIX
4323                  G1PPRL_TYPE_FORMAT
4324                  G1PPRL_ADDR_BASE_FORMAT
4325                  G1PPRL_BYTE_FORMAT
4326                  G1PPRL_BYTE_FORMAT
4327                  G1PPRL_BYTE_FORMAT
4328                  G1PPRL_DOUBLE_FORMAT
4329                  G1PPRL_BYTE_FORMAT
4330                  G1PPRL_BYTE_FORMAT,
4331                  type, p2i(bottom), p2i(end),
4332                  used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4333                  remset_bytes, strong_code_roots_bytes);
4334 
4335   return false;
4336 }
4337 
4338 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4339   // add static memory usages to remembered set sizes
4340   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4341   // Print the footer of the output.
4342   _out->print_cr(G1PPRL_LINE_PREFIX);
4343   _out->print_cr(G1PPRL_LINE_PREFIX
4344                  " SUMMARY"
4345                  G1PPRL_SUM_MB_FORMAT("capacity")
4346                  G1PPRL_SUM_MB_PERC_FORMAT("used")
4347                  G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4348                  G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4349                  G1PPRL_SUM_MB_FORMAT("remset")
4350                  G1PPRL_SUM_MB_FORMAT("code-roots"),
4351                  bytes_to_mb(_total_capacity_bytes),
4352                  bytes_to_mb(_total_used_bytes),
4353                  perc(_total_used_bytes, _total_capacity_bytes),
4354                  bytes_to_mb(_total_prev_live_bytes),
4355                  perc(_total_prev_live_bytes, _total_capacity_bytes),
4356                  bytes_to_mb(_total_next_live_bytes),
4357                  perc(_total_next_live_bytes, _total_capacity_bytes),
4358                  bytes_to_mb(_total_remset_bytes),
4359                  bytes_to_mb(_total_strong_code_roots_bytes));
4360   _out->cr();
4361 }
--- EOF ---