1 /*
   2  * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/symbolTable.hpp"
  27 #include "gc_implementation/g1/concurrentMark.inline.hpp"
  28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
  29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
  30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
  31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
  32 #include "gc_implementation/g1/g1Log.hpp"
  33 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
  34 #include "gc_implementation/g1/g1RemSet.hpp"
  35 #include "gc_implementation/g1/heapRegion.inline.hpp"
  36 #include "gc_implementation/g1/heapRegionRemSet.hpp"
  37 #include "gc_implementation/g1/heapRegionSeq.inline.hpp"
  38 #include "gc_implementation/shared/vmGCOperations.hpp"
  39 #include "memory/genOopClosures.inline.hpp"
  40 #include "memory/referencePolicy.hpp"
  41 #include "memory/resourceArea.hpp"
  42 #include "oops/oop.inline.hpp"
  43 #include "runtime/handles.inline.hpp"
  44 #include "runtime/java.hpp"
  45 #include "services/memTracker.hpp"
  46 
  47 // Concurrent marking bit map wrapper
  48 
  49 CMBitMapRO::CMBitMapRO(int shifter) :
  50   _bm(),
  51   _shifter(shifter) {
  52   _bmStartWord = 0;
  53   _bmWordSize = 0;
  54 }
  55 
  56 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr,
  57                                                HeapWord* limit) const {
  58   // First we must round addr *up* to a possible object boundary.
  59   addr = (HeapWord*)align_size_up((intptr_t)addr,
  60                                   HeapWordSize << _shifter);
  61   size_t addrOffset = heapWordToOffset(addr);
  62   if (limit == NULL) {
  63     limit = _bmStartWord + _bmWordSize;
  64   }
  65   size_t limitOffset = heapWordToOffset(limit);
  66   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  67   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  68   assert(nextAddr >= addr, "get_next_one postcondition");
  69   assert(nextAddr == limit || isMarked(nextAddr),
  70          "get_next_one postcondition");
  71   return nextAddr;
  72 }
  73 
  74 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr,
  75                                                  HeapWord* limit) const {
  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_zero_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 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
  90   assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
  91   return (int) (diff >> _shifter);
  92 }
  93 
  94 #ifndef PRODUCT
  95 bool CMBitMapRO::covers(ReservedSpace heap_rs) const {
  96   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  97   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
  98          "size inconsistency");
  99   return _bmStartWord == (HeapWord*)(heap_rs.base()) &&
 100          _bmWordSize  == heap_rs.size()>>LogHeapWordSize;
 101 }
 102 #endif
 103 
 104 bool CMBitMap::allocate(ReservedSpace heap_rs) {
 105   _bmStartWord = (HeapWord*)(heap_rs.base());
 106   _bmWordSize  = heap_rs.size()/HeapWordSize;    // heap_rs.size() is in bytes
 107   ReservedSpace brs(ReservedSpace::allocation_align_size_up(
 108                      (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1));
 109   if (!brs.is_reserved()) {
 110     warning("ConcurrentMark marking bit map allocation failure");
 111     return false;
 112   }
 113   MemTracker::record_virtual_memory_type((address)brs.base(), mtGC);
 114   // For now we'll just commit all of the bit map up front.
 115   // Later on we'll try to be more parsimonious with swap.
 116   if (!_virtual_space.initialize(brs, brs.size())) {
 117     warning("ConcurrentMark marking bit map backing store failure");
 118     return false;
 119   }
 120   assert(_virtual_space.committed_size() == brs.size(),
 121          "didn't reserve backing store for all of concurrent marking bit map?");
 122   _bm.set_map((uintptr_t*)_virtual_space.low());
 123   assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >=
 124          _bmWordSize, "inconsistency in bit map sizing");
 125   _bm.set_size(_bmWordSize >> _shifter);
 126   return true;
 127 }
 128 
 129 void CMBitMap::clearAll() {
 130   _bm.clear();
 131   return;
 132 }
 133 
 134 void CMBitMap::markRange(MemRegion mr) {
 135   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 136   assert(!mr.is_empty(), "unexpected empty region");
 137   assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
 138           ((HeapWord *) mr.end())),
 139          "markRange memory region end is not card aligned");
 140   // convert address range into offset range
 141   _bm.at_put_range(heapWordToOffset(mr.start()),
 142                    heapWordToOffset(mr.end()), true);
 143 }
 144 
 145 void CMBitMap::clearRange(MemRegion mr) {
 146   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 147   assert(!mr.is_empty(), "unexpected empty region");
 148   // convert address range into offset range
 149   _bm.at_put_range(heapWordToOffset(mr.start()),
 150                    heapWordToOffset(mr.end()), false);
 151 }
 152 
 153 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
 154                                             HeapWord* end_addr) {
 155   HeapWord* start = getNextMarkedWordAddress(addr);
 156   start = MIN2(start, end_addr);
 157   HeapWord* end   = getNextUnmarkedWordAddress(start);
 158   end = MIN2(end, end_addr);
 159   assert(start <= end, "Consistency check");
 160   MemRegion mr(start, end);
 161   if (!mr.is_empty()) {
 162     clearRange(mr);
 163   }
 164   return mr;
 165 }
 166 
 167 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
 168   _base(NULL), _cm(cm)
 169 #ifdef ASSERT
 170   , _drain_in_progress(false)
 171   , _drain_in_progress_yields(false)
 172 #endif
 173 {}
 174 
 175 bool CMMarkStack::allocate(size_t capacity) {
 176   // allocate a stack of the requisite depth
 177   ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
 178   if (!rs.is_reserved()) {
 179     warning("ConcurrentMark MarkStack allocation failure");
 180     return false;
 181   }
 182   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 183   if (!_virtual_space.initialize(rs, rs.size())) {
 184     warning("ConcurrentMark MarkStack backing store failure");
 185     // Release the virtual memory reserved for the marking stack
 186     rs.release();
 187     return false;
 188   }
 189   assert(_virtual_space.committed_size() == rs.size(),
 190          "Didn't reserve backing store for all of ConcurrentMark stack?");
 191   _base = (oop*) _virtual_space.low();
 192   setEmpty();
 193   _capacity = (jint) capacity;
 194   _saved_index = -1;
 195   NOT_PRODUCT(_max_depth = 0);
 196   return true;
 197 }
 198 
 199 void CMMarkStack::expand() {
 200   // Called, during remark, if we've overflown the marking stack during marking.
 201   assert(isEmpty(), "stack should been emptied while handling overflow");
 202   assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
 203   // Clear expansion flag
 204   _should_expand = false;
 205   if (_capacity == (jint) MarkStackSizeMax) {
 206     if (PrintGCDetails && Verbose) {
 207       gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
 208     }
 209     return;
 210   }
 211   // Double capacity if possible
 212   jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
 213   // Do not give up existing stack until we have managed to
 214   // get the double capacity that we desired.
 215   ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
 216                                                            sizeof(oop)));
 217   if (rs.is_reserved()) {
 218     // Release the backing store associated with old stack
 219     _virtual_space.release();
 220     // Reinitialize virtual space for new stack
 221     if (!_virtual_space.initialize(rs, rs.size())) {
 222       fatal("Not enough swap for expanded marking stack capacity");
 223     }
 224     _base = (oop*)(_virtual_space.low());
 225     _index = 0;
 226     _capacity = new_capacity;
 227   } else {
 228     if (PrintGCDetails && Verbose) {
 229       // Failed to double capacity, continue;
 230       gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
 231                           SIZE_FORMAT"K to " SIZE_FORMAT"K",
 232                           _capacity / K, new_capacity / K);
 233     }
 234   }
 235 }
 236 
 237 void CMMarkStack::set_should_expand() {
 238   // If we're resetting the marking state because of an
 239   // marking stack overflow, record that we should, if
 240   // possible, expand the stack.
 241   _should_expand = _cm->has_overflown();
 242 }
 243 
 244 CMMarkStack::~CMMarkStack() {
 245   if (_base != NULL) {
 246     _base = NULL;
 247     _virtual_space.release();
 248   }
 249 }
 250 
 251 void CMMarkStack::par_push(oop ptr) {
 252   while (true) {
 253     if (isFull()) {
 254       _overflow = true;
 255       return;
 256     }
 257     // Otherwise...
 258     jint index = _index;
 259     jint next_index = index+1;
 260     jint res = Atomic::cmpxchg(next_index, &_index, index);
 261     if (res == index) {
 262       _base[index] = ptr;
 263       // Note that we don't maintain this atomically.  We could, but it
 264       // doesn't seem necessary.
 265       NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 266       return;
 267     }
 268     // Otherwise, we need to try again.
 269   }
 270 }
 271 
 272 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
 273   while (true) {
 274     if (isFull()) {
 275       _overflow = true;
 276       return;
 277     }
 278     // Otherwise...
 279     jint index = _index;
 280     jint next_index = index + n;
 281     if (next_index > _capacity) {
 282       _overflow = true;
 283       return;
 284     }
 285     jint res = Atomic::cmpxchg(next_index, &_index, index);
 286     if (res == index) {
 287       for (int i = 0; i < n; i++) {
 288         int  ind = index + i;
 289         assert(ind < _capacity, "By overflow test above.");
 290         _base[ind] = ptr_arr[i];
 291       }
 292       NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 293       return;
 294     }
 295     // Otherwise, we need to try again.
 296   }
 297 }
 298 
 299 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 300   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 301   jint start = _index;
 302   jint next_index = start + n;
 303   if (next_index > _capacity) {
 304     _overflow = true;
 305     return;
 306   }
 307   // Otherwise.
 308   _index = next_index;
 309   for (int i = 0; i < n; i++) {
 310     int ind = start + i;
 311     assert(ind < _capacity, "By overflow test above.");
 312     _base[ind] = ptr_arr[i];
 313   }
 314   NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 315 }
 316 
 317 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 318   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 319   jint index = _index;
 320   if (index == 0) {
 321     *n = 0;
 322     return false;
 323   } else {
 324     int k = MIN2(max, index);
 325     jint  new_ind = index - k;
 326     for (int j = 0; j < k; j++) {
 327       ptr_arr[j] = _base[new_ind + j];
 328     }
 329     _index = new_ind;
 330     *n = k;
 331     return true;
 332   }
 333 }
 334 
 335 template<class OopClosureClass>
 336 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
 337   assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
 338          || SafepointSynchronize::is_at_safepoint(),
 339          "Drain recursion must be yield-safe.");
 340   bool res = true;
 341   debug_only(_drain_in_progress = true);
 342   debug_only(_drain_in_progress_yields = yield_after);
 343   while (!isEmpty()) {
 344     oop newOop = pop();
 345     assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
 346     assert(newOop->is_oop(), "Expected an oop");
 347     assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
 348            "only grey objects on this stack");
 349     newOop->oop_iterate(cl);
 350     if (yield_after && _cm->do_yield_check()) {
 351       res = false;
 352       break;
 353     }
 354   }
 355   debug_only(_drain_in_progress = false);
 356   return res;
 357 }
 358 
 359 void CMMarkStack::note_start_of_gc() {
 360   assert(_saved_index == -1,
 361          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 362   _saved_index = _index;
 363 }
 364 
 365 void CMMarkStack::note_end_of_gc() {
 366   // This is intentionally a guarantee, instead of an assert. If we
 367   // accidentally add something to the mark stack during GC, it
 368   // will be a correctness issue so it's better if we crash. we'll
 369   // only check this once per GC anyway, so it won't be a performance
 370   // issue in any way.
 371   guarantee(_saved_index == _index,
 372             err_msg("saved index: %d index: %d", _saved_index, _index));
 373   _saved_index = -1;
 374 }
 375 
 376 void CMMarkStack::oops_do(OopClosure* f) {
 377   assert(_saved_index == _index,
 378          err_msg("saved index: %d index: %d", _saved_index, _index));
 379   for (int i = 0; i < _index; i += 1) {
 380     f->do_oop(&_base[i]);
 381   }
 382 }
 383 
 384 bool ConcurrentMark::not_yet_marked(oop obj) const {
 385   return _g1h->is_obj_ill(obj);
 386 }
 387 
 388 CMRootRegions::CMRootRegions() :
 389   _young_list(NULL), _cm(NULL), _scan_in_progress(false),
 390   _should_abort(false),  _next_survivor(NULL) { }
 391 
 392 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
 393   _young_list = g1h->young_list();
 394   _cm = cm;
 395 }
 396 
 397 void CMRootRegions::prepare_for_scan() {
 398   assert(!scan_in_progress(), "pre-condition");
 399 
 400   // Currently, only survivors can be root regions.
 401   assert(_next_survivor == NULL, "pre-condition");
 402   _next_survivor = _young_list->first_survivor_region();
 403   _scan_in_progress = (_next_survivor != NULL);
 404   _should_abort = false;
 405 }
 406 
 407 HeapRegion* CMRootRegions::claim_next() {
 408   if (_should_abort) {
 409     // If someone has set the should_abort flag, we return NULL to
 410     // force the caller to bail out of their loop.
 411     return NULL;
 412   }
 413 
 414   // Currently, only survivors can be root regions.
 415   HeapRegion* res = _next_survivor;
 416   if (res != NULL) {
 417     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 418     // Read it again in case it changed while we were waiting for the lock.
 419     res = _next_survivor;
 420     if (res != NULL) {
 421       if (res == _young_list->last_survivor_region()) {
 422         // We just claimed the last survivor so store NULL to indicate
 423         // that we're done.
 424         _next_survivor = NULL;
 425       } else {
 426         _next_survivor = res->get_next_young_region();
 427       }
 428     } else {
 429       // Someone else claimed the last survivor while we were trying
 430       // to take the lock so nothing else to do.
 431     }
 432   }
 433   assert(res == NULL || res->is_survivor(), "post-condition");
 434 
 435   return res;
 436 }
 437 
 438 void CMRootRegions::scan_finished() {
 439   assert(scan_in_progress(), "pre-condition");
 440 
 441   // Currently, only survivors can be root regions.
 442   if (!_should_abort) {
 443     assert(_next_survivor == NULL, "we should have claimed all survivors");
 444   }
 445   _next_survivor = NULL;
 446 
 447   {
 448     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 449     _scan_in_progress = false;
 450     RootRegionScan_lock->notify_all();
 451   }
 452 }
 453 
 454 bool CMRootRegions::wait_until_scan_finished() {
 455   if (!scan_in_progress()) return false;
 456 
 457   {
 458     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 459     while (scan_in_progress()) {
 460       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 461     }
 462   }
 463   return true;
 464 }
 465 
 466 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
 467 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
 468 #endif // _MSC_VER
 469 
 470 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 471   return MAX2((n_par_threads + 2) / 4, 1U);
 472 }
 473 
 474 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs) :
 475   _g1h(g1h),
 476   _markBitMap1(MinObjAlignment - 1),
 477   _markBitMap2(MinObjAlignment - 1),
 478 
 479   _parallel_marking_threads(0),
 480   _max_parallel_marking_threads(0),
 481   _sleep_factor(0.0),
 482   _marking_task_overhead(1.0),
 483   _cleanup_sleep_factor(0.0),
 484   _cleanup_task_overhead(1.0),
 485   _cleanup_list("Cleanup List"),
 486   _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
 487   _card_bm((heap_rs.size() + CardTableModRefBS::card_size - 1) >>
 488             CardTableModRefBS::card_shift,
 489             false /* in_resource_area*/),
 490 
 491   _prevMarkBitMap(&_markBitMap1),
 492   _nextMarkBitMap(&_markBitMap2),
 493 
 494   _markStack(this),
 495   // _finger set in set_non_marking_state
 496 
 497   _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
 498   // _active_tasks set in set_non_marking_state
 499   // _tasks set inside the constructor
 500   _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
 501   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 502 
 503   _has_overflown(false),
 504   _concurrent(false),
 505   _has_aborted(false),
 506   _restart_for_overflow(false),
 507   _concurrent_marking_in_progress(false),
 508 
 509   // _verbose_level set below
 510 
 511   _init_times(),
 512   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 513   _cleanup_times(),
 514   _total_counting_time(0.0),
 515   _total_rs_scrub_time(0.0),
 516 
 517   _parallel_workers(NULL),
 518 
 519   _count_card_bitmaps(NULL),
 520   _count_marked_bytes(NULL),
 521   _completed_initialization(false) {
 522   CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
 523   if (verbose_level < no_verbose) {
 524     verbose_level = no_verbose;
 525   }
 526   if (verbose_level > high_verbose) {
 527     verbose_level = high_verbose;
 528   }
 529   _verbose_level = verbose_level;
 530 
 531   if (verbose_low()) {
 532     gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
 533                            "heap end = "PTR_FORMAT, _heap_start, _heap_end);
 534   }
 535 
 536   if (!_markBitMap1.allocate(heap_rs)) {
 537     warning("Failed to allocate first CM bit map");
 538     return;
 539   }
 540   if (!_markBitMap2.allocate(heap_rs)) {
 541     warning("Failed to allocate second CM bit map");
 542     return;
 543   }
 544 
 545   // Create & start a ConcurrentMark thread.
 546   _cmThread = new ConcurrentMarkThread(this);
 547   assert(cmThread() != NULL, "CM Thread should have been created");
 548   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 549 
 550   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 551   assert(_markBitMap1.covers(heap_rs), "_markBitMap1 inconsistency");
 552   assert(_markBitMap2.covers(heap_rs), "_markBitMap2 inconsistency");
 553 
 554   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 555   satb_qs.set_buffer_size(G1SATBBufferSize);
 556 
 557   _root_regions.init(_g1h, this);
 558 
 559   if (ConcGCThreads > ParallelGCThreads) {
 560     warning("Can't have more ConcGCThreads (" UINT32_FORMAT ") "
 561             "than ParallelGCThreads (" UINT32_FORMAT ").",
 562             ConcGCThreads, ParallelGCThreads);
 563     return;
 564   }
 565   if (ParallelGCThreads == 0) {
 566     // if we are not running with any parallel GC threads we will not
 567     // spawn any marking threads either
 568     _parallel_marking_threads =       0;
 569     _max_parallel_marking_threads =   0;
 570     _sleep_factor             =     0.0;
 571     _marking_task_overhead    =     1.0;
 572   } else {
 573     if (ConcGCThreads > 0) {
 574       // notice that ConcGCThreads overwrites G1MarkingOverheadPercent
 575       // if both are set
 576 
 577       _parallel_marking_threads = (uint) ConcGCThreads;
 578       _max_parallel_marking_threads = _parallel_marking_threads;
 579       _sleep_factor             = 0.0;
 580       _marking_task_overhead    = 1.0;
 581     } else if (G1MarkingOverheadPercent > 0) {
 582       // we will calculate the number of parallel marking threads
 583       // based on a target overhead with respect to the soft real-time
 584       // goal
 585 
 586       double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 587       double overall_cm_overhead =
 588         (double) MaxGCPauseMillis * marking_overhead /
 589         (double) GCPauseIntervalMillis;
 590       double cpu_ratio = 1.0 / (double) os::processor_count();
 591       double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 592       double marking_task_overhead =
 593         overall_cm_overhead / marking_thread_num *
 594                                                 (double) os::processor_count();
 595       double sleep_factor =
 596                          (1.0 - marking_task_overhead) / marking_task_overhead;
 597 
 598       _parallel_marking_threads = (uint) marking_thread_num;
 599       _max_parallel_marking_threads = _parallel_marking_threads;
 600       _sleep_factor             = sleep_factor;
 601       _marking_task_overhead    = marking_task_overhead;
 602     } else {
 603       _parallel_marking_threads = scale_parallel_threads((uint)ParallelGCThreads);
 604       _max_parallel_marking_threads = _parallel_marking_threads;
 605       _sleep_factor             = 0.0;
 606       _marking_task_overhead    = 1.0;
 607     }
 608 
 609     if (parallel_marking_threads() > 1) {
 610       _cleanup_task_overhead = 1.0;
 611     } else {
 612       _cleanup_task_overhead = marking_task_overhead();
 613     }
 614     _cleanup_sleep_factor =
 615                      (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
 616 
 617 #if 0
 618     gclog_or_tty->print_cr("Marking Threads          %d", parallel_marking_threads());
 619     gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
 620     gclog_or_tty->print_cr("CM Sleep Factor          %1.4lf", sleep_factor());
 621     gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
 622     gclog_or_tty->print_cr("CL Sleep Factor          %1.4lf", cleanup_sleep_factor());
 623 #endif
 624 
 625     guarantee(parallel_marking_threads() > 0, "peace of mind");
 626     _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
 627          _max_parallel_marking_threads, false, true);
 628     if (_parallel_workers == NULL) {
 629       vm_exit_during_initialization("Failed necessary allocation.");
 630     } else {
 631       _parallel_workers->initialize_workers();
 632     }
 633   }
 634 
 635   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 636     uintx mark_stack_size =
 637       MIN2(MarkStackSizeMax,
 638           MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 639     // Verify that the calculated value for MarkStackSize is in range.
 640     // It would be nice to use the private utility routine from Arguments.
 641     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 642       warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
 643               "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
 644               mark_stack_size, 1, MarkStackSizeMax);
 645       return;
 646     }
 647     FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
 648   } else {
 649     // Verify MarkStackSize is in range.
 650     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 651       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 652         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 653           warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
 654                   "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
 655                   MarkStackSize, 1, MarkStackSizeMax);
 656           return;
 657         }
 658       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 659         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 660           warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
 661                   " or for MarkStackSizeMax (" UINTX_FORMAT ")",
 662                   MarkStackSize, MarkStackSizeMax);
 663           return;
 664         }
 665       }
 666     }
 667   }
 668 
 669   if (!_markStack.allocate(MarkStackSize)) {
 670     warning("Failed to allocate CM marking stack");
 671     return;
 672   }
 673 
 674   _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
 675   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 676 
 677   _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap,  _max_worker_id, mtGC);
 678   _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
 679 
 680   BitMap::idx_t card_bm_size = _card_bm.size();
 681 
 682   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 683   _active_tasks = _max_worker_id;
 684 
 685   size_t max_regions = (size_t) _g1h->max_regions();
 686   for (uint i = 0; i < _max_worker_id; ++i) {
 687     CMTaskQueue* task_queue = new CMTaskQueue();
 688     task_queue->initialize();
 689     _task_queues->register_queue(i, task_queue);
 690 
 691     _count_card_bitmaps[i] = BitMap(card_bm_size, false);
 692     _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
 693 
 694     _tasks[i] = new CMTask(i, this,
 695                            _count_marked_bytes[i],
 696                            &_count_card_bitmaps[i],
 697                            task_queue, _task_queues);
 698 
 699     _accum_task_vtime[i] = 0.0;
 700   }
 701 
 702   // Calculate the card number for the bottom of the heap. Used
 703   // in biasing indexes into the accounting card bitmaps.
 704   _heap_bottom_card_num =
 705     intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
 706                                 CardTableModRefBS::card_shift);
 707 
 708   // Clear all the liveness counting data
 709   clear_all_count_data();
 710 
 711   // so that the call below can read a sensible value
 712   _heap_start = (HeapWord*) heap_rs.base();
 713   set_non_marking_state();
 714   _completed_initialization = true;
 715 }
 716 
 717 void ConcurrentMark::update_g1_committed(bool force) {
 718   // If concurrent marking is not in progress, then we do not need to
 719   // update _heap_end.
 720   if (!concurrent_marking_in_progress() && !force) return;
 721 
 722   MemRegion committed = _g1h->g1_committed();
 723   assert(committed.start() == _heap_start, "start shouldn't change");
 724   HeapWord* new_end = committed.end();
 725   if (new_end > _heap_end) {
 726     // The heap has been expanded.
 727 
 728     _heap_end = new_end;
 729   }
 730   // Notice that the heap can also shrink. However, this only happens
 731   // during a Full GC (at least currently) and the entire marking
 732   // phase will bail out and the task will not be restarted. So, let's
 733   // do nothing.
 734 }
 735 
 736 void ConcurrentMark::reset() {
 737   // Starting values for these two. This should be called in a STW
 738   // phase. CM will be notified of any future g1_committed expansions
 739   // will be at the end of evacuation pauses, when tasks are
 740   // inactive.
 741   MemRegion committed = _g1h->g1_committed();
 742   _heap_start = committed.start();
 743   _heap_end   = committed.end();
 744 
 745   // Separated the asserts so that we know which one fires.
 746   assert(_heap_start != NULL, "heap bounds should look ok");
 747   assert(_heap_end != NULL, "heap bounds should look ok");
 748   assert(_heap_start < _heap_end, "heap bounds should look ok");
 749 
 750   // reset all the marking data structures and any necessary flags
 751   clear_marking_state();
 752 
 753   if (verbose_low()) {
 754     gclog_or_tty->print_cr("[global] resetting");
 755   }
 756 
 757   // We do reset all of them, since different phases will use
 758   // different number of active threads. So, it's easiest to have all
 759   // of them ready.
 760   for (uint i = 0; i < _max_worker_id; ++i) {
 761     _tasks[i]->reset(_nextMarkBitMap);
 762   }
 763 
 764   // we need this to make sure that the flag is on during the evac
 765   // pause with initial mark piggy-backed
 766   set_concurrent_marking_in_progress();
 767 }
 768 
 769 void ConcurrentMark::set_phase(uint active_tasks, bool concurrent) {
 770   assert(active_tasks <= _max_worker_id, "we should not have more");
 771 
 772   _active_tasks = active_tasks;
 773   // Need to update the three data structures below according to the
 774   // number of active threads for this phase.
 775   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 776   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 777   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 778 
 779   _concurrent = concurrent;
 780   // We propagate this to all tasks, not just the active ones.
 781   for (uint i = 0; i < _max_worker_id; ++i)
 782     _tasks[i]->set_concurrent(concurrent);
 783 
 784   if (concurrent) {
 785     set_concurrent_marking_in_progress();
 786   } else {
 787     // We currently assume that the concurrent flag has been set to
 788     // false before we start remark. At this point we should also be
 789     // in a STW phase.
 790     assert(!concurrent_marking_in_progress(), "invariant");
 791     assert(_finger == _heap_end, "only way to get here");
 792     update_g1_committed(true);
 793   }
 794 }
 795 
 796 void ConcurrentMark::set_non_marking_state() {
 797   // We set the global marking state to some default values when we're
 798   // not doing marking.
 799   clear_marking_state();
 800   _active_tasks = 0;
 801   clear_concurrent_marking_in_progress();
 802 }
 803 
 804 ConcurrentMark::~ConcurrentMark() {
 805   // The ConcurrentMark instance is never freed.
 806   ShouldNotReachHere();
 807 }
 808 
 809 void ConcurrentMark::clearNextBitmap() {
 810   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 811   G1CollectorPolicy* g1p = g1h->g1_policy();
 812 
 813   // Make sure that the concurrent mark thread looks to still be in
 814   // the current cycle.
 815   guarantee(cmThread()->during_cycle(), "invariant");
 816 
 817   // We are finishing up the current cycle by clearing the next
 818   // marking bitmap and getting it ready for the next cycle. During
 819   // this time no other cycle can start. So, let's make sure that this
 820   // is the case.
 821   guarantee(!g1h->mark_in_progress(), "invariant");
 822 
 823   // clear the mark bitmap (no grey objects to start with).
 824   // We need to do this in chunks and offer to yield in between
 825   // each chunk.
 826   HeapWord* start  = _nextMarkBitMap->startWord();
 827   HeapWord* end    = _nextMarkBitMap->endWord();
 828   HeapWord* cur    = start;
 829   size_t chunkSize = M;
 830   while (cur < end) {
 831     HeapWord* next = cur + chunkSize;
 832     if (next > end) {
 833       next = end;
 834     }
 835     MemRegion mr(cur,next);
 836     _nextMarkBitMap->clearRange(mr);
 837     cur = next;
 838     do_yield_check();
 839 
 840     // Repeat the asserts from above. We'll do them as asserts here to
 841     // minimize their overhead on the product. However, we'll have
 842     // them as guarantees at the beginning / end of the bitmap
 843     // clearing to get some checking in the product.
 844     assert(cmThread()->during_cycle(), "invariant");
 845     assert(!g1h->mark_in_progress(), "invariant");
 846   }
 847 
 848   // Clear the liveness counting data
 849   clear_all_count_data();
 850 
 851   // Repeat the asserts from above.
 852   guarantee(cmThread()->during_cycle(), "invariant");
 853   guarantee(!g1h->mark_in_progress(), "invariant");
 854 }
 855 
 856 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 857 public:
 858   bool doHeapRegion(HeapRegion* r) {
 859     if (!r->continuesHumongous()) {
 860       r->note_start_of_marking();
 861     }
 862     return false;
 863   }
 864 };
 865 
 866 void ConcurrentMark::checkpointRootsInitialPre() {
 867   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 868   G1CollectorPolicy* g1p = g1h->g1_policy();
 869 
 870   _has_aborted = false;
 871 
 872 #ifndef PRODUCT
 873   if (G1PrintReachableAtInitialMark) {
 874     print_reachable("at-cycle-start",
 875                     VerifyOption_G1UsePrevMarking, true /* all */);
 876   }
 877 #endif
 878 
 879   // Initialise marking structures. This has to be done in a STW phase.
 880   reset();
 881 
 882   // For each region note start of marking.
 883   NoteStartOfMarkHRClosure startcl;
 884   g1h->heap_region_iterate(&startcl);
 885 }
 886 
 887 
 888 void ConcurrentMark::checkpointRootsInitialPost() {
 889   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 890 
 891   // If we force an overflow during remark, the remark operation will
 892   // actually abort and we'll restart concurrent marking. If we always
 893   // force an oveflow during remark we'll never actually complete the
 894   // marking phase. So, we initilize this here, at the start of the
 895   // cycle, so that at the remaining overflow number will decrease at
 896   // every remark and we'll eventually not need to cause one.
 897   force_overflow_stw()->init();
 898 
 899   // Start Concurrent Marking weak-reference discovery.
 900   ReferenceProcessor* rp = g1h->ref_processor_cm();
 901   // enable ("weak") refs discovery
 902   rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
 903   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 904 
 905   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 906   // This is the start of  the marking cycle, we're expected all
 907   // threads to have SATB queues with active set to false.
 908   satb_mq_set.set_active_all_threads(true, /* new active value */
 909                                      false /* expected_active */);
 910 
 911   _root_regions.prepare_for_scan();
 912 
 913   // update_g1_committed() will be called at the end of an evac pause
 914   // when marking is on. So, it's also called at the end of the
 915   // initial-mark pause to update the heap end, if the heap expands
 916   // during it. No need to call it here.
 917 }
 918 
 919 /*
 920  * Notice that in the next two methods, we actually leave the STS
 921  * during the barrier sync and join it immediately afterwards. If we
 922  * do not do this, the following deadlock can occur: one thread could
 923  * be in the barrier sync code, waiting for the other thread to also
 924  * sync up, whereas another one could be trying to yield, while also
 925  * waiting for the other threads to sync up too.
 926  *
 927  * Note, however, that this code is also used during remark and in
 928  * this case we should not attempt to leave / enter the STS, otherwise
 929  * we'll either hit an asseert (debug / fastdebug) or deadlock
 930  * (product). So we should only leave / enter the STS if we are
 931  * operating concurrently.
 932  *
 933  * Because the thread that does the sync barrier has left the STS, it
 934  * is possible to be suspended for a Full GC or an evacuation pause
 935  * could occur. This is actually safe, since the entering the sync
 936  * barrier is one of the last things do_marking_step() does, and it
 937  * doesn't manipulate any data structures afterwards.
 938  */
 939 
 940 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 941   if (verbose_low()) {
 942     gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
 943   }
 944 
 945   if (concurrent()) {
 946     ConcurrentGCThread::stsLeave();
 947   }
 948   _first_overflow_barrier_sync.enter();
 949   if (concurrent()) {
 950     ConcurrentGCThread::stsJoin();
 951   }
 952   // at this point everyone should have synced up and not be doing any
 953   // more work
 954 
 955   if (verbose_low()) {
 956     gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
 957   }
 958 
 959   // let the task associated with with worker 0 do this
 960   if (worker_id == 0) {
 961     // task 0 is responsible for clearing the global data structures
 962     // We should be here because of an overflow. During STW we should
 963     // not clear the overflow flag since we rely on it being true when
 964     // we exit this method to abort the pause and restart concurent
 965     // marking.
 966     clear_marking_state(concurrent() /* clear_overflow */);
 967     force_overflow()->update();
 968 
 969     if (G1Log::fine()) {
 970       gclog_or_tty->date_stamp(PrintGCDateStamps);
 971       gclog_or_tty->stamp(PrintGCTimeStamps);
 972       gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
 973     }
 974   }
 975 
 976   // after this, each task should reset its own data structures then
 977   // then go into the second barrier
 978 }
 979 
 980 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 981   if (verbose_low()) {
 982     gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
 983   }
 984 
 985   if (concurrent()) {
 986     ConcurrentGCThread::stsLeave();
 987   }
 988   _second_overflow_barrier_sync.enter();
 989   if (concurrent()) {
 990     ConcurrentGCThread::stsJoin();
 991   }
 992   // at this point everything should be re-initialised and ready to go
 993 
 994   if (verbose_low()) {
 995     gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
 996   }
 997 }
 998 
 999 #ifndef PRODUCT
1000 void ForceOverflowSettings::init() {
1001   _num_remaining = G1ConcMarkForceOverflow;
1002   _force = false;
1003   update();
1004 }
1005 
1006 void ForceOverflowSettings::update() {
1007   if (_num_remaining > 0) {
1008     _num_remaining -= 1;
1009     _force = true;
1010   } else {
1011     _force = false;
1012   }
1013 }
1014 
1015 bool ForceOverflowSettings::should_force() {
1016   if (_force) {
1017     _force = false;
1018     return true;
1019   } else {
1020     return false;
1021   }
1022 }
1023 #endif // !PRODUCT
1024 
1025 class CMConcurrentMarkingTask: public AbstractGangTask {
1026 private:
1027   ConcurrentMark*       _cm;
1028   ConcurrentMarkThread* _cmt;
1029 
1030 public:
1031   void work(uint worker_id) {
1032     assert(Thread::current()->is_ConcurrentGC_thread(),
1033            "this should only be done by a conc GC thread");
1034     ResourceMark rm;
1035 
1036     double start_vtime = os::elapsedVTime();
1037 
1038     ConcurrentGCThread::stsJoin();
1039 
1040     assert(worker_id < _cm->active_tasks(), "invariant");
1041     CMTask* the_task = _cm->task(worker_id);
1042     the_task->record_start_time();
1043     if (!_cm->has_aborted()) {
1044       do {
1045         double start_vtime_sec = os::elapsedVTime();
1046         double start_time_sec = os::elapsedTime();
1047         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1048 
1049         the_task->do_marking_step(mark_step_duration_ms,
1050                                   true /* do_stealing    */,
1051                                   true /* do_termination */);
1052 
1053         double end_time_sec = os::elapsedTime();
1054         double end_vtime_sec = os::elapsedVTime();
1055         double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1056         double elapsed_time_sec = end_time_sec - start_time_sec;
1057         _cm->clear_has_overflown();
1058 
1059         bool ret = _cm->do_yield_check(worker_id);
1060 
1061         jlong sleep_time_ms;
1062         if (!_cm->has_aborted() && the_task->has_aborted()) {
1063           sleep_time_ms =
1064             (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1065           ConcurrentGCThread::stsLeave();
1066           os::sleep(Thread::current(), sleep_time_ms, false);
1067           ConcurrentGCThread::stsJoin();
1068         }
1069         double end_time2_sec = os::elapsedTime();
1070         double elapsed_time2_sec = end_time2_sec - start_time_sec;
1071 
1072 #if 0
1073           gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, "
1074                                  "overhead %1.4lf",
1075                                  elapsed_vtime_sec * 1000.0, (double) sleep_time_ms,
1076                                  the_task->conc_overhead(os::elapsedTime()) * 8.0);
1077           gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms",
1078                                  elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0);
1079 #endif
1080       } while (!_cm->has_aborted() && the_task->has_aborted());
1081     }
1082     the_task->record_end_time();
1083     guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1084 
1085     ConcurrentGCThread::stsLeave();
1086 
1087     double end_vtime = os::elapsedVTime();
1088     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1089   }
1090 
1091   CMConcurrentMarkingTask(ConcurrentMark* cm,
1092                           ConcurrentMarkThread* cmt) :
1093       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1094 
1095   ~CMConcurrentMarkingTask() { }
1096 };
1097 
1098 // Calculates the number of active workers for a concurrent
1099 // phase.
1100 uint ConcurrentMark::calc_parallel_marking_threads() {
1101   if (G1CollectedHeap::use_parallel_gc_threads()) {
1102     uint n_conc_workers = 0;
1103     if (!UseDynamicNumberOfGCThreads ||
1104         (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1105          !ForceDynamicNumberOfGCThreads)) {
1106       n_conc_workers = max_parallel_marking_threads();
1107     } else {
1108       n_conc_workers =
1109         AdaptiveSizePolicy::calc_default_active_workers(
1110                                      max_parallel_marking_threads(),
1111                                      1, /* Minimum workers */
1112                                      parallel_marking_threads(),
1113                                      Threads::number_of_non_daemon_threads());
1114       // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1115       // that scaling has already gone into "_max_parallel_marking_threads".
1116     }
1117     assert(n_conc_workers > 0, "Always need at least 1");
1118     return n_conc_workers;
1119   }
1120   // If we are not running with any parallel GC threads we will not
1121   // have spawned any marking threads either. Hence the number of
1122   // concurrent workers should be 0.
1123   return 0;
1124 }
1125 
1126 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1127   // Currently, only survivors can be root regions.
1128   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1129   G1RootRegionScanClosure cl(_g1h, this, worker_id);
1130 
1131   const uintx interval = PrefetchScanIntervalInBytes;
1132   HeapWord* curr = hr->bottom();
1133   const HeapWord* end = hr->top();
1134   while (curr < end) {
1135     Prefetch::read(curr, interval);
1136     oop obj = oop(curr);
1137     int size = obj->oop_iterate(&cl);
1138     assert(size == obj->size(), "sanity");
1139     curr += size;
1140   }
1141 }
1142 
1143 class CMRootRegionScanTask : public AbstractGangTask {
1144 private:
1145   ConcurrentMark* _cm;
1146 
1147 public:
1148   CMRootRegionScanTask(ConcurrentMark* cm) :
1149     AbstractGangTask("Root Region Scan"), _cm(cm) { }
1150 
1151   void work(uint worker_id) {
1152     assert(Thread::current()->is_ConcurrentGC_thread(),
1153            "this should only be done by a conc GC thread");
1154 
1155     CMRootRegions* root_regions = _cm->root_regions();
1156     HeapRegion* hr = root_regions->claim_next();
1157     while (hr != NULL) {
1158       _cm->scanRootRegion(hr, worker_id);
1159       hr = root_regions->claim_next();
1160     }
1161   }
1162 };
1163 
1164 void ConcurrentMark::scanRootRegions() {
1165   // scan_in_progress() will have been set to true only if there was
1166   // at least one root region to scan. So, if it's false, we
1167   // should not attempt to do any further work.
1168   if (root_regions()->scan_in_progress()) {
1169     _parallel_marking_threads = calc_parallel_marking_threads();
1170     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1171            "Maximum number of marking threads exceeded");
1172     uint active_workers = MAX2(1U, parallel_marking_threads());
1173 
1174     CMRootRegionScanTask task(this);
1175     if (parallel_marking_threads() > 0) {
1176       _parallel_workers->set_active_workers((int) active_workers);
1177       _parallel_workers->run_task(&task);
1178     } else {
1179       task.work(0);
1180     }
1181 
1182     // It's possible that has_aborted() is true here without actually
1183     // aborting the survivor scan earlier. This is OK as it's
1184     // mainly used for sanity checking.
1185     root_regions()->scan_finished();
1186   }
1187 }
1188 
1189 void ConcurrentMark::markFromRoots() {
1190   // we might be tempted to assert that:
1191   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1192   //        "inconsistent argument?");
1193   // However that wouldn't be right, because it's possible that
1194   // a safepoint is indeed in progress as a younger generation
1195   // stop-the-world GC happens even as we mark in this generation.
1196 
1197   _restart_for_overflow = false;
1198   force_overflow_conc()->init();
1199 
1200   // _g1h has _n_par_threads
1201   _parallel_marking_threads = calc_parallel_marking_threads();
1202   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1203     "Maximum number of marking threads exceeded");
1204 
1205   uint active_workers = MAX2(1U, parallel_marking_threads());
1206 
1207   // Parallel task terminator is set in "set_phase()"
1208   set_phase(active_workers, true /* concurrent */);
1209 
1210   CMConcurrentMarkingTask markingTask(this, cmThread());
1211   if (parallel_marking_threads() > 0) {
1212     _parallel_workers->set_active_workers((int)active_workers);
1213     // Don't set _n_par_threads because it affects MT in proceess_strong_roots()
1214     // and the decisions on that MT processing is made elsewhere.
1215     assert(_parallel_workers->active_workers() > 0, "Should have been set");
1216     _parallel_workers->run_task(&markingTask);
1217   } else {
1218     markingTask.work(0);
1219   }
1220   print_stats();
1221 }
1222 
1223 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1224   // world is stopped at this checkpoint
1225   assert(SafepointSynchronize::is_at_safepoint(),
1226          "world should be stopped");
1227 
1228   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1229 
1230   // If a full collection has happened, we shouldn't do this.
1231   if (has_aborted()) {
1232     g1h->set_marking_complete(); // So bitmap clearing isn't confused
1233     return;
1234   }
1235 
1236   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1237 
1238   if (VerifyDuringGC) {
1239     HandleMark hm;  // handle scope
1240     gclog_or_tty->print(" VerifyDuringGC:(before)");
1241     Universe::heap()->prepare_for_verify();
1242     Universe::verify(/* silent      */ false,
1243                      /* option      */ VerifyOption_G1UsePrevMarking);
1244   }
1245 
1246   G1CollectorPolicy* g1p = g1h->g1_policy();
1247   g1p->record_concurrent_mark_remark_start();
1248 
1249   double start = os::elapsedTime();
1250 
1251   checkpointRootsFinalWork();
1252 
1253   double mark_work_end = os::elapsedTime();
1254 
1255   weakRefsWork(clear_all_soft_refs);
1256 
1257   if (has_overflown()) {
1258     // Oops.  We overflowed.  Restart concurrent marking.
1259     _restart_for_overflow = true;
1260     // Clear the flag. We do not need it any more.
1261     clear_has_overflown();
1262     if (G1TraceMarkStackOverflow) {
1263       gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1264     }
1265   } else {
1266     // Aggregate the per-task counting data that we have accumulated
1267     // while marking.
1268     aggregate_count_data();
1269 
1270     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1271     // We're done with marking.
1272     // This is the end of  the marking cycle, we're expected all
1273     // threads to have SATB queues with active set to true.
1274     satb_mq_set.set_active_all_threads(false, /* new active value */
1275                                        true /* expected_active */);
1276 
1277     if (VerifyDuringGC) {
1278       HandleMark hm;  // handle scope
1279       gclog_or_tty->print(" VerifyDuringGC:(after)");
1280       Universe::heap()->prepare_for_verify();
1281       Universe::verify(/* silent      */ false,
1282                        /* option      */ VerifyOption_G1UseNextMarking);
1283     }
1284     assert(!restart_for_overflow(), "sanity");
1285   }
1286 
1287   // Expand the marking stack, if we have to and if we can.
1288   if (_markStack.should_expand()) {
1289     _markStack.expand();
1290   }
1291 
1292   // Reset the marking state if marking completed
1293   if (!restart_for_overflow()) {
1294     set_non_marking_state();
1295   }
1296 
1297 #if VERIFY_OBJS_PROCESSED
1298   _scan_obj_cl.objs_processed = 0;
1299   ThreadLocalObjQueue::objs_enqueued = 0;
1300 #endif
1301 
1302   // Statistics
1303   double now = os::elapsedTime();
1304   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1305   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1306   _remark_times.add((now - start) * 1000.0);
1307 
1308   g1p->record_concurrent_mark_remark_end();
1309 }
1310 
1311 // Base class of the closures that finalize and verify the
1312 // liveness counting data.
1313 class CMCountDataClosureBase: public HeapRegionClosure {
1314 protected:
1315   G1CollectedHeap* _g1h;
1316   ConcurrentMark* _cm;
1317   CardTableModRefBS* _ct_bs;
1318 
1319   BitMap* _region_bm;
1320   BitMap* _card_bm;
1321 
1322   // Takes a region that's not empty (i.e., it has at least one
1323   // live object in it and sets its corresponding bit on the region
1324   // bitmap to 1. If the region is "starts humongous" it will also set
1325   // to 1 the bits on the region bitmap that correspond to its
1326   // associated "continues humongous" regions.
1327   void set_bit_for_region(HeapRegion* hr) {
1328     assert(!hr->continuesHumongous(), "should have filtered those out");
1329 
1330     BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1331     if (!hr->startsHumongous()) {
1332       // Normal (non-humongous) case: just set the bit.
1333       _region_bm->par_at_put(index, true);
1334     } else {
1335       // Starts humongous case: calculate how many regions are part of
1336       // this humongous region and then set the bit range.
1337       BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1338       _region_bm->par_at_put_range(index, end_index, true);
1339     }
1340   }
1341 
1342 public:
1343   CMCountDataClosureBase(G1CollectedHeap* g1h,
1344                          BitMap* region_bm, BitMap* card_bm):
1345     _g1h(g1h), _cm(g1h->concurrent_mark()),
1346     _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1347     _region_bm(region_bm), _card_bm(card_bm) { }
1348 };
1349 
1350 // Closure that calculates the # live objects per region. Used
1351 // for verification purposes during the cleanup pause.
1352 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1353   CMBitMapRO* _bm;
1354   size_t _region_marked_bytes;
1355 
1356 public:
1357   CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1358                          BitMap* region_bm, BitMap* card_bm) :
1359     CMCountDataClosureBase(g1h, region_bm, card_bm),
1360     _bm(bm), _region_marked_bytes(0) { }
1361 
1362   bool doHeapRegion(HeapRegion* hr) {
1363 
1364     if (hr->continuesHumongous()) {
1365       // We will ignore these here and process them when their
1366       // associated "starts humongous" region is processed (see
1367       // set_bit_for_heap_region()). Note that we cannot rely on their
1368       // associated "starts humongous" region to have their bit set to
1369       // 1 since, due to the region chunking in the parallel region
1370       // iteration, a "continues humongous" region might be visited
1371       // before its associated "starts humongous".
1372       return false;
1373     }
1374 
1375     HeapWord* ntams = hr->next_top_at_mark_start();
1376     HeapWord* start = hr->bottom();
1377 
1378     assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1379            err_msg("Preconditions not met - "
1380                    "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1381                    start, ntams, hr->end()));
1382 
1383     // Find the first marked object at or after "start".
1384     start = _bm->getNextMarkedWordAddress(start, ntams);
1385 
1386     size_t marked_bytes = 0;
1387 
1388     while (start < ntams) {
1389       oop obj = oop(start);
1390       int obj_sz = obj->size();
1391       HeapWord* obj_end = start + obj_sz;
1392 
1393       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1394       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1395 
1396       // Note: if we're looking at the last region in heap - obj_end
1397       // could be actually just beyond the end of the heap; end_idx
1398       // will then correspond to a (non-existent) card that is also
1399       // just beyond the heap.
1400       if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1401         // end of object is not card aligned - increment to cover
1402         // all the cards spanned by the object
1403         end_idx += 1;
1404       }
1405 
1406       // Set the bits in the card BM for the cards spanned by this object.
1407       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1408 
1409       // Add the size of this object to the number of marked bytes.
1410       marked_bytes += (size_t)obj_sz * HeapWordSize;
1411 
1412       // Find the next marked object after this one.
1413       start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1414     }
1415 
1416     // Mark the allocated-since-marking portion...
1417     HeapWord* top = hr->top();
1418     if (ntams < top) {
1419       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1420       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1421 
1422       // Note: if we're looking at the last region in heap - top
1423       // could be actually just beyond the end of the heap; end_idx
1424       // will then correspond to a (non-existent) card that is also
1425       // just beyond the heap.
1426       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1427         // end of object is not card aligned - increment to cover
1428         // all the cards spanned by the object
1429         end_idx += 1;
1430       }
1431       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1432 
1433       // This definitely means the region has live objects.
1434       set_bit_for_region(hr);
1435     }
1436 
1437     // Update the live region bitmap.
1438     if (marked_bytes > 0) {
1439       set_bit_for_region(hr);
1440     }
1441 
1442     // Set the marked bytes for the current region so that
1443     // it can be queried by a calling verificiation routine
1444     _region_marked_bytes = marked_bytes;
1445 
1446     return false;
1447   }
1448 
1449   size_t region_marked_bytes() const { return _region_marked_bytes; }
1450 };
1451 
1452 // Heap region closure used for verifying the counting data
1453 // that was accumulated concurrently and aggregated during
1454 // the remark pause. This closure is applied to the heap
1455 // regions during the STW cleanup pause.
1456 
1457 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1458   G1CollectedHeap* _g1h;
1459   ConcurrentMark* _cm;
1460   CalcLiveObjectsClosure _calc_cl;
1461   BitMap* _region_bm;   // Region BM to be verified
1462   BitMap* _card_bm;     // Card BM to be verified
1463   bool _verbose;        // verbose output?
1464 
1465   BitMap* _exp_region_bm; // Expected Region BM values
1466   BitMap* _exp_card_bm;   // Expected card BM values
1467 
1468   int _failures;
1469 
1470 public:
1471   VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1472                                 BitMap* region_bm,
1473                                 BitMap* card_bm,
1474                                 BitMap* exp_region_bm,
1475                                 BitMap* exp_card_bm,
1476                                 bool verbose) :
1477     _g1h(g1h), _cm(g1h->concurrent_mark()),
1478     _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1479     _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1480     _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1481     _failures(0) { }
1482 
1483   int failures() const { return _failures; }
1484 
1485   bool doHeapRegion(HeapRegion* hr) {
1486     if (hr->continuesHumongous()) {
1487       // We will ignore these here and process them when their
1488       // associated "starts humongous" region is processed (see
1489       // set_bit_for_heap_region()). Note that we cannot rely on their
1490       // associated "starts humongous" region to have their bit set to
1491       // 1 since, due to the region chunking in the parallel region
1492       // iteration, a "continues humongous" region might be visited
1493       // before its associated "starts humongous".
1494       return false;
1495     }
1496 
1497     int failures = 0;
1498 
1499     // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1500     // this region and set the corresponding bits in the expected region
1501     // and card bitmaps.
1502     bool res = _calc_cl.doHeapRegion(hr);
1503     assert(res == false, "should be continuing");
1504 
1505     MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1506                     Mutex::_no_safepoint_check_flag);
1507 
1508     // Verify the marked bytes for this region.
1509     size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1510     size_t act_marked_bytes = hr->next_marked_bytes();
1511 
1512     // We're not OK if expected marked bytes > actual marked bytes. It means
1513     // we have missed accounting some objects during the actual marking.
1514     if (exp_marked_bytes > act_marked_bytes) {
1515       if (_verbose) {
1516         gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1517                                "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1518                                hr->hrs_index(), exp_marked_bytes, act_marked_bytes);
1519       }
1520       failures += 1;
1521     }
1522 
1523     // Verify the bit, for this region, in the actual and expected
1524     // (which was just calculated) region bit maps.
1525     // We're not OK if the bit in the calculated expected region
1526     // bitmap is set and the bit in the actual region bitmap is not.
1527     BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index();
1528 
1529     bool expected = _exp_region_bm->at(index);
1530     bool actual = _region_bm->at(index);
1531     if (expected && !actual) {
1532       if (_verbose) {
1533         gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1534                                "expected: %s, actual: %s",
1535                                hr->hrs_index(),
1536                                BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1537       }
1538       failures += 1;
1539     }
1540 
1541     // Verify that the card bit maps for the cards spanned by the current
1542     // region match. We have an error if we have a set bit in the expected
1543     // bit map and the corresponding bit in the actual bitmap is not set.
1544 
1545     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1546     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1547 
1548     for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1549       expected = _exp_card_bm->at(i);
1550       actual = _card_bm->at(i);
1551 
1552       if (expected && !actual) {
1553         if (_verbose) {
1554           gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1555                                  "expected: %s, actual: %s",
1556                                  hr->hrs_index(), i,
1557                                  BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1558         }
1559         failures += 1;
1560       }
1561     }
1562 
1563     if (failures > 0 && _verbose)  {
1564       gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1565                              "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1566                              HR_FORMAT_PARAMS(hr), hr->next_top_at_mark_start(),
1567                              _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1568     }
1569 
1570     _failures += failures;
1571 
1572     // We could stop iteration over the heap when we
1573     // find the first violating region by returning true.
1574     return false;
1575   }
1576 };
1577 
1578 
1579 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1580 protected:
1581   G1CollectedHeap* _g1h;
1582   ConcurrentMark* _cm;
1583   BitMap* _actual_region_bm;
1584   BitMap* _actual_card_bm;
1585 
1586   uint    _n_workers;
1587 
1588   BitMap* _expected_region_bm;
1589   BitMap* _expected_card_bm;
1590 
1591   int  _failures;
1592   bool _verbose;
1593 
1594 public:
1595   G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1596                             BitMap* region_bm, BitMap* card_bm,
1597                             BitMap* expected_region_bm, BitMap* expected_card_bm)
1598     : AbstractGangTask("G1 verify final counting"),
1599       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1600       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1601       _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1602       _failures(0), _verbose(false),
1603       _n_workers(0) {
1604     assert(VerifyDuringGC, "don't call this otherwise");
1605 
1606     // Use the value already set as the number of active threads
1607     // in the call to run_task().
1608     if (G1CollectedHeap::use_parallel_gc_threads()) {
1609       assert( _g1h->workers()->active_workers() > 0,
1610         "Should have been previously set");
1611       _n_workers = _g1h->workers()->active_workers();
1612     } else {
1613       _n_workers = 1;
1614     }
1615 
1616     assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1617     assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1618 
1619     _verbose = _cm->verbose_medium();
1620   }
1621 
1622   void work(uint worker_id) {
1623     assert(worker_id < _n_workers, "invariant");
1624 
1625     VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1626                                             _actual_region_bm, _actual_card_bm,
1627                                             _expected_region_bm,
1628                                             _expected_card_bm,
1629                                             _verbose);
1630 
1631     if (G1CollectedHeap::use_parallel_gc_threads()) {
1632       _g1h->heap_region_par_iterate_chunked(&verify_cl,
1633                                             worker_id,
1634                                             _n_workers,
1635                                             HeapRegion::VerifyCountClaimValue);
1636     } else {
1637       _g1h->heap_region_iterate(&verify_cl);
1638     }
1639 
1640     Atomic::add(verify_cl.failures(), &_failures);
1641   }
1642 
1643   int failures() const { return _failures; }
1644 };
1645 
1646 // Closure that finalizes the liveness counting data.
1647 // Used during the cleanup pause.
1648 // Sets the bits corresponding to the interval [NTAMS, top]
1649 // (which contains the implicitly live objects) in the
1650 // card liveness bitmap. Also sets the bit for each region,
1651 // containing live data, in the region liveness bitmap.
1652 
1653 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1654  public:
1655   FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1656                               BitMap* region_bm,
1657                               BitMap* card_bm) :
1658     CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1659 
1660   bool doHeapRegion(HeapRegion* hr) {
1661 
1662     if (hr->continuesHumongous()) {
1663       // We will ignore these here and process them when their
1664       // associated "starts humongous" region is processed (see
1665       // set_bit_for_heap_region()). Note that we cannot rely on their
1666       // associated "starts humongous" region to have their bit set to
1667       // 1 since, due to the region chunking in the parallel region
1668       // iteration, a "continues humongous" region might be visited
1669       // before its associated "starts humongous".
1670       return false;
1671     }
1672 
1673     HeapWord* ntams = hr->next_top_at_mark_start();
1674     HeapWord* top   = hr->top();
1675 
1676     assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1677 
1678     // Mark the allocated-since-marking portion...
1679     if (ntams < top) {
1680       // This definitely means the region has live objects.
1681       set_bit_for_region(hr);
1682 
1683       // Now set the bits in the card bitmap for [ntams, top)
1684       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1685       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1686 
1687       // Note: if we're looking at the last region in heap - top
1688       // could be actually just beyond the end of the heap; end_idx
1689       // will then correspond to a (non-existent) card that is also
1690       // just beyond the heap.
1691       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1692         // end of object is not card aligned - increment to cover
1693         // all the cards spanned by the object
1694         end_idx += 1;
1695       }
1696 
1697       assert(end_idx <= _card_bm->size(),
1698              err_msg("oob: end_idx=  "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1699                      end_idx, _card_bm->size()));
1700       assert(start_idx < _card_bm->size(),
1701              err_msg("oob: start_idx=  "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1702                      start_idx, _card_bm->size()));
1703 
1704       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1705     }
1706 
1707     // Set the bit for the region if it contains live data
1708     if (hr->next_marked_bytes() > 0) {
1709       set_bit_for_region(hr);
1710     }
1711 
1712     return false;
1713   }
1714 };
1715 
1716 class G1ParFinalCountTask: public AbstractGangTask {
1717 protected:
1718   G1CollectedHeap* _g1h;
1719   ConcurrentMark* _cm;
1720   BitMap* _actual_region_bm;
1721   BitMap* _actual_card_bm;
1722 
1723   uint    _n_workers;
1724 
1725 public:
1726   G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1727     : AbstractGangTask("G1 final counting"),
1728       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1729       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1730       _n_workers(0) {
1731     // Use the value already set as the number of active threads
1732     // in the call to run_task().
1733     if (G1CollectedHeap::use_parallel_gc_threads()) {
1734       assert( _g1h->workers()->active_workers() > 0,
1735         "Should have been previously set");
1736       _n_workers = _g1h->workers()->active_workers();
1737     } else {
1738       _n_workers = 1;
1739     }
1740   }
1741 
1742   void work(uint worker_id) {
1743     assert(worker_id < _n_workers, "invariant");
1744 
1745     FinalCountDataUpdateClosure final_update_cl(_g1h,
1746                                                 _actual_region_bm,
1747                                                 _actual_card_bm);
1748 
1749     if (G1CollectedHeap::use_parallel_gc_threads()) {
1750       _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1751                                             worker_id,
1752                                             _n_workers,
1753                                             HeapRegion::FinalCountClaimValue);
1754     } else {
1755       _g1h->heap_region_iterate(&final_update_cl);
1756     }
1757   }
1758 };
1759 
1760 class G1ParNoteEndTask;
1761 
1762 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1763   G1CollectedHeap* _g1;
1764   int _worker_num;
1765   size_t _max_live_bytes;
1766   uint _regions_claimed;
1767   size_t _freed_bytes;
1768   FreeRegionList* _local_cleanup_list;
1769   OldRegionSet* _old_proxy_set;
1770   HumongousRegionSet* _humongous_proxy_set;
1771   HRRSCleanupTask* _hrrs_cleanup_task;
1772   double _claimed_region_time;
1773   double _max_region_time;
1774 
1775 public:
1776   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1777                              int worker_num,
1778                              FreeRegionList* local_cleanup_list,
1779                              OldRegionSet* old_proxy_set,
1780                              HumongousRegionSet* humongous_proxy_set,
1781                              HRRSCleanupTask* hrrs_cleanup_task) :
1782     _g1(g1), _worker_num(worker_num),
1783     _max_live_bytes(0), _regions_claimed(0),
1784     _freed_bytes(0),
1785     _claimed_region_time(0.0), _max_region_time(0.0),
1786     _local_cleanup_list(local_cleanup_list),
1787     _old_proxy_set(old_proxy_set),
1788     _humongous_proxy_set(humongous_proxy_set),
1789     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1790 
1791   size_t freed_bytes() { return _freed_bytes; }
1792 
1793   bool doHeapRegion(HeapRegion *hr) {
1794     if (hr->continuesHumongous()) {
1795       return false;
1796     }
1797     // We use a claim value of zero here because all regions
1798     // were claimed with value 1 in the FinalCount task.
1799     _g1->reset_gc_time_stamps(hr);
1800     double start = os::elapsedTime();
1801     _regions_claimed++;
1802     hr->note_end_of_marking();
1803     _max_live_bytes += hr->max_live_bytes();
1804     _g1->free_region_if_empty(hr,
1805                               &_freed_bytes,
1806                               _local_cleanup_list,
1807                               _old_proxy_set,
1808                               _humongous_proxy_set,
1809                               _hrrs_cleanup_task,
1810                               true /* par */);
1811     double region_time = (os::elapsedTime() - start);
1812     _claimed_region_time += region_time;
1813     if (region_time > _max_region_time) {
1814       _max_region_time = region_time;
1815     }
1816     return false;
1817   }
1818 
1819   size_t max_live_bytes() { return _max_live_bytes; }
1820   uint regions_claimed() { return _regions_claimed; }
1821   double claimed_region_time_sec() { return _claimed_region_time; }
1822   double max_region_time_sec() { return _max_region_time; }
1823 };
1824 
1825 class G1ParNoteEndTask: public AbstractGangTask {
1826   friend class G1NoteEndOfConcMarkClosure;
1827 
1828 protected:
1829   G1CollectedHeap* _g1h;
1830   size_t _max_live_bytes;
1831   size_t _freed_bytes;
1832   FreeRegionList* _cleanup_list;
1833 
1834 public:
1835   G1ParNoteEndTask(G1CollectedHeap* g1h,
1836                    FreeRegionList* cleanup_list) :
1837     AbstractGangTask("G1 note end"), _g1h(g1h),
1838     _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1839 
1840   void work(uint worker_id) {
1841     double start = os::elapsedTime();
1842     FreeRegionList local_cleanup_list("Local Cleanup List");
1843     OldRegionSet old_proxy_set("Local Cleanup Old Proxy Set");
1844     HumongousRegionSet humongous_proxy_set("Local Cleanup Humongous Proxy Set");
1845     HRRSCleanupTask hrrs_cleanup_task;
1846     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, worker_id, &local_cleanup_list,
1847                                            &old_proxy_set,
1848                                            &humongous_proxy_set,
1849                                            &hrrs_cleanup_task);
1850     if (G1CollectedHeap::use_parallel_gc_threads()) {
1851       _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1852                                             _g1h->workers()->active_workers(),
1853                                             HeapRegion::NoteEndClaimValue);
1854     } else {
1855       _g1h->heap_region_iterate(&g1_note_end);
1856     }
1857     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1858 
1859     // Now update the lists
1860     _g1h->update_sets_after_freeing_regions(g1_note_end.freed_bytes(),
1861                                             NULL /* free_list */,
1862                                             &old_proxy_set,
1863                                             &humongous_proxy_set,
1864                                             true /* par */);
1865     {
1866       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1867       _max_live_bytes += g1_note_end.max_live_bytes();
1868       _freed_bytes += g1_note_end.freed_bytes();
1869 
1870       // If we iterate over the global cleanup list at the end of
1871       // cleanup to do this printing we will not guarantee to only
1872       // generate output for the newly-reclaimed regions (the list
1873       // might not be empty at the beginning of cleanup; we might
1874       // still be working on its previous contents). So we do the
1875       // printing here, before we append the new regions to the global
1876       // cleanup list.
1877 
1878       G1HRPrinter* hr_printer = _g1h->hr_printer();
1879       if (hr_printer->is_active()) {
1880         HeapRegionLinkedListIterator iter(&local_cleanup_list);
1881         while (iter.more_available()) {
1882           HeapRegion* hr = iter.get_next();
1883           hr_printer->cleanup(hr);
1884         }
1885       }
1886 
1887       _cleanup_list->add_as_tail(&local_cleanup_list);
1888       assert(local_cleanup_list.is_empty(), "post-condition");
1889 
1890       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1891     }
1892   }
1893   size_t max_live_bytes() { return _max_live_bytes; }
1894   size_t freed_bytes() { return _freed_bytes; }
1895 };
1896 
1897 class G1ParScrubRemSetTask: public AbstractGangTask {
1898 protected:
1899   G1RemSet* _g1rs;
1900   BitMap* _region_bm;
1901   BitMap* _card_bm;
1902 public:
1903   G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1904                        BitMap* region_bm, BitMap* card_bm) :
1905     AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1906     _region_bm(region_bm), _card_bm(card_bm) { }
1907 
1908   void work(uint worker_id) {
1909     if (G1CollectedHeap::use_parallel_gc_threads()) {
1910       _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1911                        HeapRegion::ScrubRemSetClaimValue);
1912     } else {
1913       _g1rs->scrub(_region_bm, _card_bm);
1914     }
1915   }
1916 
1917 };
1918 
1919 void ConcurrentMark::cleanup() {
1920   // world is stopped at this checkpoint
1921   assert(SafepointSynchronize::is_at_safepoint(),
1922          "world should be stopped");
1923   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1924 
1925   // If a full collection has happened, we shouldn't do this.
1926   if (has_aborted()) {
1927     g1h->set_marking_complete(); // So bitmap clearing isn't confused
1928     return;
1929   }
1930 
1931   HRSPhaseSetter x(HRSPhaseCleanup);
1932   g1h->verify_region_sets_optional();
1933 
1934   if (VerifyDuringGC) {
1935     HandleMark hm;  // handle scope
1936     gclog_or_tty->print(" VerifyDuringGC:(before)");
1937     Universe::heap()->prepare_for_verify();
1938     Universe::verify(/* silent      */ false,
1939                      /* option      */ VerifyOption_G1UsePrevMarking);
1940   }
1941 
1942   G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
1943   g1p->record_concurrent_mark_cleanup_start();
1944 
1945   double start = os::elapsedTime();
1946 
1947   HeapRegionRemSet::reset_for_cleanup_tasks();
1948 
1949   uint n_workers;
1950 
1951   // Do counting once more with the world stopped for good measure.
1952   G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1953 
1954   if (G1CollectedHeap::use_parallel_gc_threads()) {
1955    assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
1956            "sanity check");
1957 
1958     g1h->set_par_threads();
1959     n_workers = g1h->n_par_threads();
1960     assert(g1h->n_par_threads() == n_workers,
1961            "Should not have been reset");
1962     g1h->workers()->run_task(&g1_par_count_task);
1963     // Done with the parallel phase so reset to 0.
1964     g1h->set_par_threads(0);
1965 
1966     assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
1967            "sanity check");
1968   } else {
1969     n_workers = 1;
1970     g1_par_count_task.work(0);
1971   }
1972 
1973   if (VerifyDuringGC) {
1974     // Verify that the counting data accumulated during marking matches
1975     // that calculated by walking the marking bitmap.
1976 
1977     // Bitmaps to hold expected values
1978     BitMap expected_region_bm(_region_bm.size(), false);
1979     BitMap expected_card_bm(_card_bm.size(), false);
1980 
1981     G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
1982                                                  &_region_bm,
1983                                                  &_card_bm,
1984                                                  &expected_region_bm,
1985                                                  &expected_card_bm);
1986 
1987     if (G1CollectedHeap::use_parallel_gc_threads()) {
1988       g1h->set_par_threads((int)n_workers);
1989       g1h->workers()->run_task(&g1_par_verify_task);
1990       // Done with the parallel phase so reset to 0.
1991       g1h->set_par_threads(0);
1992 
1993       assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
1994              "sanity check");
1995     } else {
1996       g1_par_verify_task.work(0);
1997     }
1998 
1999     guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2000   }
2001 
2002   size_t start_used_bytes = g1h->used();
2003   g1h->set_marking_complete();
2004 
2005   double count_end = os::elapsedTime();
2006   double this_final_counting_time = (count_end - start);
2007   _total_counting_time += this_final_counting_time;
2008 
2009   if (G1PrintRegionLivenessInfo) {
2010     G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2011     _g1h->heap_region_iterate(&cl);
2012   }
2013 
2014   // Install newly created mark bitMap as "prev".
2015   swapMarkBitMaps();
2016 
2017   g1h->reset_gc_time_stamp();
2018 
2019   // Note end of marking in all heap regions.
2020   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2021   if (G1CollectedHeap::use_parallel_gc_threads()) {
2022     g1h->set_par_threads((int)n_workers);
2023     g1h->workers()->run_task(&g1_par_note_end_task);
2024     g1h->set_par_threads(0);
2025 
2026     assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2027            "sanity check");
2028   } else {
2029     g1_par_note_end_task.work(0);
2030   }
2031   g1h->check_gc_time_stamps();
2032 
2033   if (!cleanup_list_is_empty()) {
2034     // The cleanup list is not empty, so we'll have to process it
2035     // concurrently. Notify anyone else that might be wanting free
2036     // regions that there will be more free regions coming soon.
2037     g1h->set_free_regions_coming();
2038   }
2039 
2040   // call below, since it affects the metric by which we sort the heap
2041   // regions.
2042   if (G1ScrubRemSets) {
2043     double rs_scrub_start = os::elapsedTime();
2044     G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2045     if (G1CollectedHeap::use_parallel_gc_threads()) {
2046       g1h->set_par_threads((int)n_workers);
2047       g1h->workers()->run_task(&g1_par_scrub_rs_task);
2048       g1h->set_par_threads(0);
2049 
2050       assert(g1h->check_heap_region_claim_values(
2051                                             HeapRegion::ScrubRemSetClaimValue),
2052              "sanity check");
2053     } else {
2054       g1_par_scrub_rs_task.work(0);
2055     }
2056 
2057     double rs_scrub_end = os::elapsedTime();
2058     double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2059     _total_rs_scrub_time += this_rs_scrub_time;
2060   }
2061 
2062   // this will also free any regions totally full of garbage objects,
2063   // and sort the regions.
2064   g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2065 
2066   // Statistics.
2067   double end = os::elapsedTime();
2068   _cleanup_times.add((end - start) * 1000.0);
2069 
2070   if (G1Log::fine()) {
2071     g1h->print_size_transition(gclog_or_tty,
2072                                start_used_bytes,
2073                                g1h->used(),
2074                                g1h->capacity());
2075   }
2076 
2077   // Clean up will have freed any regions completely full of garbage.
2078   // Update the soft reference policy with the new heap occupancy.
2079   Universe::update_heap_info_at_gc();
2080 
2081   // We need to make this be a "collection" so any collection pause that
2082   // races with it goes around and waits for completeCleanup to finish.
2083   g1h->increment_total_collections();
2084 
2085   // We reclaimed old regions so we should calculate the sizes to make
2086   // sure we update the old gen/space data.
2087   g1h->g1mm()->update_sizes();
2088 
2089   if (VerifyDuringGC) {
2090     HandleMark hm;  // handle scope
2091     gclog_or_tty->print(" VerifyDuringGC:(after)");
2092     Universe::heap()->prepare_for_verify();
2093     Universe::verify(/* silent      */ false,
2094                      /* option      */ VerifyOption_G1UsePrevMarking);
2095   }
2096 
2097   g1h->verify_region_sets_optional();
2098 }
2099 
2100 void ConcurrentMark::completeCleanup() {
2101   if (has_aborted()) return;
2102 
2103   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2104 
2105   _cleanup_list.verify_optional();
2106   FreeRegionList tmp_free_list("Tmp Free List");
2107 
2108   if (G1ConcRegionFreeingVerbose) {
2109     gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2110                            "cleanup list has %u entries",
2111                            _cleanup_list.length());
2112   }
2113 
2114   // Noone else should be accessing the _cleanup_list at this point,
2115   // so it's not necessary to take any locks
2116   while (!_cleanup_list.is_empty()) {
2117     HeapRegion* hr = _cleanup_list.remove_head();
2118     assert(hr != NULL, "the list was not empty");
2119     hr->par_clear();
2120     tmp_free_list.add_as_tail(hr);
2121 
2122     // Instead of adding one region at a time to the secondary_free_list,
2123     // we accumulate them in the local list and move them a few at a
2124     // time. This also cuts down on the number of notify_all() calls
2125     // we do during this process. We'll also append the local list when
2126     // _cleanup_list is empty (which means we just removed the last
2127     // region from the _cleanup_list).
2128     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2129         _cleanup_list.is_empty()) {
2130       if (G1ConcRegionFreeingVerbose) {
2131         gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2132                                "appending %u entries to the secondary_free_list, "
2133                                "cleanup list still has %u entries",
2134                                tmp_free_list.length(),
2135                                _cleanup_list.length());
2136       }
2137 
2138       {
2139         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2140         g1h->secondary_free_list_add_as_tail(&tmp_free_list);
2141         SecondaryFreeList_lock->notify_all();
2142       }
2143 
2144       if (G1StressConcRegionFreeing) {
2145         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2146           os::sleep(Thread::current(), (jlong) 1, false);
2147         }
2148       }
2149     }
2150   }
2151   assert(tmp_free_list.is_empty(), "post-condition");
2152 }
2153 
2154 // Support closures for reference procssing in G1
2155 
2156 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2157   HeapWord* addr = (HeapWord*)obj;
2158   return addr != NULL &&
2159          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2160 }
2161 
2162 class G1CMKeepAliveClosure: public ExtendedOopClosure {
2163   G1CollectedHeap* _g1;
2164   ConcurrentMark*  _cm;
2165  public:
2166   G1CMKeepAliveClosure(G1CollectedHeap* g1, ConcurrentMark* cm) :
2167     _g1(g1), _cm(cm) {
2168     assert(Thread::current()->is_VM_thread(), "otherwise fix worker id");
2169   }
2170 
2171   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2172   virtual void do_oop(      oop* p) { do_oop_work(p); }
2173 
2174   template <class T> void do_oop_work(T* p) {
2175     oop obj = oopDesc::load_decode_heap_oop(p);
2176     HeapWord* addr = (HeapWord*)obj;
2177 
2178     if (_cm->verbose_high()) {
2179       gclog_or_tty->print_cr("\t[0] we're looking at location "
2180                              "*"PTR_FORMAT" = "PTR_FORMAT,
2181                              p, (void*) obj);
2182     }
2183 
2184     if (_g1->is_in_g1_reserved(addr) && _g1->is_obj_ill(obj)) {
2185       _cm->mark_and_count(obj);
2186       _cm->mark_stack_push(obj);
2187     }
2188   }
2189 };
2190 
2191 class G1CMDrainMarkingStackClosure: public VoidClosure {
2192   ConcurrentMark*               _cm;
2193   CMMarkStack*                  _markStack;
2194   G1CMKeepAliveClosure*         _oopClosure;
2195  public:
2196   G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMMarkStack* markStack,
2197                                G1CMKeepAliveClosure* oopClosure) :
2198     _cm(cm),
2199     _markStack(markStack),
2200     _oopClosure(oopClosure) { }
2201 
2202   void do_void() {
2203     _markStack->drain(_oopClosure, _cm->nextMarkBitMap(), false);
2204   }
2205 };
2206 
2207 // 'Keep Alive' closure used by parallel reference processing.
2208 // An instance of this closure is used in the parallel reference processing
2209 // code rather than an instance of G1CMKeepAliveClosure. We could have used
2210 // the G1CMKeepAliveClosure as it is MT-safe. Also reference objects are
2211 // placed on to discovered ref lists once so we can mark and push with no
2212 // need to check whether the object has already been marked. Using the
2213 // G1CMKeepAliveClosure would mean, however, having all the worker threads
2214 // operating on the global mark stack. This means that an individual
2215 // worker would be doing lock-free pushes while it processes its own
2216 // discovered ref list followed by drain call. If the discovered ref lists
2217 // are unbalanced then this could cause interference with the other
2218 // workers. Using a CMTask (and its embedded local data structures)
2219 // avoids that potential interference.
2220 class G1CMParKeepAliveAndDrainClosure: public OopClosure {
2221   ConcurrentMark*  _cm;
2222   CMTask*          _task;
2223   int              _ref_counter_limit;
2224   int              _ref_counter;
2225  public:
2226   G1CMParKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task) :
2227     _cm(cm), _task(task),
2228     _ref_counter_limit(G1RefProcDrainInterval) {
2229     assert(_ref_counter_limit > 0, "sanity");
2230     _ref_counter = _ref_counter_limit;
2231   }
2232 
2233   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2234   virtual void do_oop(      oop* p) { do_oop_work(p); }
2235 
2236   template <class T> void do_oop_work(T* p) {
2237     if (!_cm->has_overflown()) {
2238       oop obj = oopDesc::load_decode_heap_oop(p);
2239       if (_cm->verbose_high()) {
2240         gclog_or_tty->print_cr("\t[%u] we're looking at location "
2241                                "*"PTR_FORMAT" = "PTR_FORMAT,
2242                                _task->worker_id(), p, (void*) obj);
2243       }
2244 
2245       _task->deal_with_reference(obj);
2246       _ref_counter--;
2247 
2248       if (_ref_counter == 0) {
2249         // We have dealt with _ref_counter_limit references, pushing them and objects
2250         // reachable from them on to the local stack (and possibly the global stack).
2251         // Call do_marking_step() to process these entries. We call the routine in a
2252         // loop, which we'll exit if there's nothing more to do (i.e. we're done
2253         // with the entries that we've pushed as a result of the deal_with_reference
2254         // calls above) or we overflow.
2255         // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() flag
2256         // while there may still be some work to do. (See the comment at the
2257         // beginning of CMTask::do_marking_step() for those conditions - one of which
2258         // is reaching the specified time target.) It is only when
2259         // CMTask::do_marking_step() returns without setting the has_aborted() flag
2260         // that the marking has completed.
2261         do {
2262           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2263           _task->do_marking_step(mark_step_duration_ms,
2264                                  false /* do_stealing    */,
2265                                  false /* do_termination */);
2266         } while (_task->has_aborted() && !_cm->has_overflown());
2267         _ref_counter = _ref_counter_limit;
2268       }
2269     } else {
2270       if (_cm->verbose_high()) {
2271          gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2272       }
2273     }
2274   }
2275 };
2276 
2277 class G1CMParDrainMarkingStackClosure: public VoidClosure {
2278   ConcurrentMark* _cm;
2279   CMTask* _task;
2280  public:
2281   G1CMParDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task) :
2282     _cm(cm), _task(task) { }
2283 
2284   void do_void() {
2285     do {
2286       if (_cm->verbose_high()) {
2287         gclog_or_tty->print_cr("\t[%u] Drain: Calling do marking_step",
2288                                _task->worker_id());
2289       }
2290 
2291       // We call CMTask::do_marking_step() to completely drain the local and
2292       // global marking stacks. The routine is called in a loop, which we'll
2293       // exit if there's nothing more to do (i.e. we'completely drained the
2294       // entries that were pushed as a result of applying the
2295       // G1CMParKeepAliveAndDrainClosure to the entries on the discovered ref
2296       // lists above) or we overflow the global marking stack.
2297       // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() flag
2298       // while there may still be some work to do. (See the comment at the
2299       // beginning of CMTask::do_marking_step() for those conditions - one of which
2300       // is reaching the specified time target.) It is only when
2301       // CMTask::do_marking_step() returns without setting the has_aborted() flag
2302       // that the marking has completed.
2303 
2304       _task->do_marking_step(1000000000.0 /* something very large */,
2305                              true /* do_stealing    */,
2306                              true /* do_termination */);
2307     } while (_task->has_aborted() && !_cm->has_overflown());
2308   }
2309 };
2310 
2311 // Implementation of AbstractRefProcTaskExecutor for parallel
2312 // reference processing at the end of G1 concurrent marking
2313 
2314 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2315 private:
2316   G1CollectedHeap* _g1h;
2317   ConcurrentMark*  _cm;
2318   WorkGang*        _workers;
2319   int              _active_workers;
2320 
2321 public:
2322   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2323                         ConcurrentMark* cm,
2324                         WorkGang* workers,
2325                         int n_workers) :
2326     _g1h(g1h), _cm(cm),
2327     _workers(workers), _active_workers(n_workers) { }
2328 
2329   // Executes the given task using concurrent marking worker threads.
2330   virtual void execute(ProcessTask& task);
2331   virtual void execute(EnqueueTask& task);
2332 };
2333 
2334 class G1CMRefProcTaskProxy: public AbstractGangTask {
2335   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2336   ProcessTask&     _proc_task;
2337   G1CollectedHeap* _g1h;
2338   ConcurrentMark*  _cm;
2339 
2340 public:
2341   G1CMRefProcTaskProxy(ProcessTask& proc_task,
2342                      G1CollectedHeap* g1h,
2343                      ConcurrentMark* cm) :
2344     AbstractGangTask("Process reference objects in parallel"),
2345     _proc_task(proc_task), _g1h(g1h), _cm(cm) { }
2346 
2347   virtual void work(uint worker_id) {
2348     CMTask* marking_task = _cm->task(worker_id);
2349     G1CMIsAliveClosure g1_is_alive(_g1h);
2350     G1CMParKeepAliveAndDrainClosure g1_par_keep_alive(_cm, marking_task);
2351     G1CMParDrainMarkingStackClosure g1_par_drain(_cm, marking_task);
2352 
2353     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2354   }
2355 };
2356 
2357 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2358   assert(_workers != NULL, "Need parallel worker threads.");
2359 
2360   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2361 
2362   // We need to reset the phase for each task execution so that
2363   // the termination protocol of CMTask::do_marking_step works.
2364   _cm->set_phase(_active_workers, false /* concurrent */);
2365   _g1h->set_par_threads(_active_workers);
2366   _workers->run_task(&proc_task_proxy);
2367   _g1h->set_par_threads(0);
2368 }
2369 
2370 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2371   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2372   EnqueueTask& _enq_task;
2373 
2374 public:
2375   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2376     AbstractGangTask("Enqueue reference objects in parallel"),
2377     _enq_task(enq_task) { }
2378 
2379   virtual void work(uint worker_id) {
2380     _enq_task.work(worker_id);
2381   }
2382 };
2383 
2384 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2385   assert(_workers != NULL, "Need parallel worker threads.");
2386 
2387   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2388 
2389   _g1h->set_par_threads(_active_workers);
2390   _workers->run_task(&enq_task_proxy);
2391   _g1h->set_par_threads(0);
2392 }
2393 
2394 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2395   ResourceMark rm;
2396   HandleMark   hm;
2397 
2398   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2399 
2400   // Is alive closure.
2401   G1CMIsAliveClosure g1_is_alive(g1h);
2402 
2403   // Inner scope to exclude the cleaning of the string and symbol
2404   // tables from the displayed time.
2405   {
2406     if (G1Log::finer()) {
2407       gclog_or_tty->put(' ');
2408     }
2409     TraceTime t("GC ref-proc", G1Log::finer(), false, gclog_or_tty);
2410 
2411     ReferenceProcessor* rp = g1h->ref_processor_cm();
2412 
2413     // See the comment in G1CollectedHeap::ref_processing_init()
2414     // about how reference processing currently works in G1.
2415 
2416     // Process weak references.
2417     rp->setup_policy(clear_all_soft_refs);
2418     assert(_markStack.isEmpty(), "mark stack should be empty");
2419 
2420     G1CMKeepAliveClosure g1_keep_alive(g1h, this);
2421     G1CMDrainMarkingStackClosure
2422       g1_drain_mark_stack(this, &_markStack, &g1_keep_alive);
2423 
2424     // We use the work gang from the G1CollectedHeap and we utilize all
2425     // the worker threads.
2426     uint active_workers = g1h->workers() ? g1h->workers()->active_workers() : 1U;
2427     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2428 
2429     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2430                                               g1h->workers(), active_workers);
2431 
2432     if (rp->processing_is_mt()) {
2433       // Set the degree of MT here.  If the discovery is done MT, there
2434       // may have been a different number of threads doing the discovery
2435       // and a different number of discovered lists may have Ref objects.
2436       // That is OK as long as the Reference lists are balanced (see
2437       // balance_all_queues() and balance_queues()).
2438       rp->set_active_mt_degree(active_workers);
2439 
2440       rp->process_discovered_references(&g1_is_alive,
2441                                       &g1_keep_alive,
2442                                       &g1_drain_mark_stack,
2443                                       &par_task_executor);
2444 
2445       // The work routines of the parallel keep_alive and drain_marking_stack
2446       // will set the has_overflown flag if we overflow the global marking
2447       // stack.
2448     } else {
2449       rp->process_discovered_references(&g1_is_alive,
2450                                         &g1_keep_alive,
2451                                         &g1_drain_mark_stack,
2452                                         NULL);
2453     }
2454 
2455     assert(_markStack.overflow() || _markStack.isEmpty(),
2456             "mark stack should be empty (unless it overflowed)");
2457     if (_markStack.overflow()) {
2458       // Should have been done already when we tried to push an
2459       // entry on to the global mark stack. But let's do it again.
2460       set_has_overflown();
2461     }
2462 
2463     if (rp->processing_is_mt()) {
2464       assert(rp->num_q() == active_workers, "why not");
2465       rp->enqueue_discovered_references(&par_task_executor);
2466     } else {
2467       rp->enqueue_discovered_references();
2468     }
2469 
2470     rp->verify_no_references_recorded();
2471     assert(!rp->discovery_enabled(), "Post condition");
2472   }
2473 
2474   // Now clean up stale oops in StringTable
2475   StringTable::unlink(&g1_is_alive);
2476   // Clean up unreferenced symbols in symbol table.
2477   SymbolTable::unlink();
2478 }
2479 
2480 void ConcurrentMark::swapMarkBitMaps() {
2481   CMBitMapRO* temp = _prevMarkBitMap;
2482   _prevMarkBitMap  = (CMBitMapRO*)_nextMarkBitMap;
2483   _nextMarkBitMap  = (CMBitMap*)  temp;
2484 }
2485 
2486 class CMRemarkTask: public AbstractGangTask {
2487 private:
2488   ConcurrentMark *_cm;
2489 
2490 public:
2491   void work(uint worker_id) {
2492     // Since all available tasks are actually started, we should
2493     // only proceed if we're supposed to be actived.
2494     if (worker_id < _cm->active_tasks()) {
2495       CMTask* task = _cm->task(worker_id);
2496       task->record_start_time();
2497       do {
2498         task->do_marking_step(1000000000.0 /* something very large */,
2499                               true /* do_stealing    */,
2500                               true /* do_termination */);
2501       } while (task->has_aborted() && !_cm->has_overflown());
2502       // If we overflow, then we do not want to restart. We instead
2503       // want to abort remark and do concurrent marking again.
2504       task->record_end_time();
2505     }
2506   }
2507 
2508   CMRemarkTask(ConcurrentMark* cm, int active_workers) :
2509     AbstractGangTask("Par Remark"), _cm(cm) {
2510     _cm->terminator()->reset_for_reuse(active_workers);
2511   }
2512 };
2513 
2514 void ConcurrentMark::checkpointRootsFinalWork() {
2515   ResourceMark rm;
2516   HandleMark   hm;
2517   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2518 
2519   g1h->ensure_parsability(false);
2520 
2521   if (G1CollectedHeap::use_parallel_gc_threads()) {
2522     G1CollectedHeap::StrongRootsScope srs(g1h);
2523     // this is remark, so we'll use up all active threads
2524     uint active_workers = g1h->workers()->active_workers();
2525     if (active_workers == 0) {
2526       assert(active_workers > 0, "Should have been set earlier");
2527       active_workers = (uint) ParallelGCThreads;
2528       g1h->workers()->set_active_workers(active_workers);
2529     }
2530     set_phase(active_workers, false /* concurrent */);
2531     // Leave _parallel_marking_threads at it's
2532     // value originally calculated in the ConcurrentMark
2533     // constructor and pass values of the active workers
2534     // through the gang in the task.
2535 
2536     CMRemarkTask remarkTask(this, active_workers);
2537     g1h->set_par_threads(active_workers);
2538     g1h->workers()->run_task(&remarkTask);
2539     g1h->set_par_threads(0);
2540   } else {
2541     G1CollectedHeap::StrongRootsScope srs(g1h);
2542     // this is remark, so we'll use up all available threads
2543     uint active_workers = 1;
2544     set_phase(active_workers, false /* concurrent */);
2545 
2546     CMRemarkTask remarkTask(this, active_workers);
2547     // We will start all available threads, even if we decide that the
2548     // active_workers will be fewer. The extra ones will just bail out
2549     // immediately.
2550     remarkTask.work(0);
2551   }
2552   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2553   guarantee(satb_mq_set.completed_buffers_num() == 0, "invariant");
2554 
2555   print_stats();
2556 
2557 #if VERIFY_OBJS_PROCESSED
2558   if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) {
2559     gclog_or_tty->print_cr("Processed = %d, enqueued = %d.",
2560                            _scan_obj_cl.objs_processed,
2561                            ThreadLocalObjQueue::objs_enqueued);
2562     guarantee(_scan_obj_cl.objs_processed ==
2563               ThreadLocalObjQueue::objs_enqueued,
2564               "Different number of objs processed and enqueued.");
2565   }
2566 #endif
2567 }
2568 
2569 #ifndef PRODUCT
2570 
2571 class PrintReachableOopClosure: public OopClosure {
2572 private:
2573   G1CollectedHeap* _g1h;
2574   outputStream*    _out;
2575   VerifyOption     _vo;
2576   bool             _all;
2577 
2578 public:
2579   PrintReachableOopClosure(outputStream* out,
2580                            VerifyOption  vo,
2581                            bool          all) :
2582     _g1h(G1CollectedHeap::heap()),
2583     _out(out), _vo(vo), _all(all) { }
2584 
2585   void do_oop(narrowOop* p) { do_oop_work(p); }
2586   void do_oop(      oop* p) { do_oop_work(p); }
2587 
2588   template <class T> void do_oop_work(T* p) {
2589     oop         obj = oopDesc::load_decode_heap_oop(p);
2590     const char* str = NULL;
2591     const char* str2 = "";
2592 
2593     if (obj == NULL) {
2594       str = "";
2595     } else if (!_g1h->is_in_g1_reserved(obj)) {
2596       str = " O";
2597     } else {
2598       HeapRegion* hr  = _g1h->heap_region_containing(obj);
2599       guarantee(hr != NULL, "invariant");
2600       bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2601       bool marked = _g1h->is_marked(obj, _vo);
2602 
2603       if (over_tams) {
2604         str = " >";
2605         if (marked) {
2606           str2 = " AND MARKED";
2607         }
2608       } else if (marked) {
2609         str = " M";
2610       } else {
2611         str = " NOT";
2612       }
2613     }
2614 
2615     _out->print_cr("  "PTR_FORMAT": "PTR_FORMAT"%s%s",
2616                    p, (void*) obj, str, str2);
2617   }
2618 };
2619 
2620 class PrintReachableObjectClosure : public ObjectClosure {
2621 private:
2622   G1CollectedHeap* _g1h;
2623   outputStream*    _out;
2624   VerifyOption     _vo;
2625   bool             _all;
2626   HeapRegion*      _hr;
2627 
2628 public:
2629   PrintReachableObjectClosure(outputStream* out,
2630                               VerifyOption  vo,
2631                               bool          all,
2632                               HeapRegion*   hr) :
2633     _g1h(G1CollectedHeap::heap()),
2634     _out(out), _vo(vo), _all(all), _hr(hr) { }
2635 
2636   void do_object(oop o) {
2637     bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2638     bool marked = _g1h->is_marked(o, _vo);
2639     bool print_it = _all || over_tams || marked;
2640 
2641     if (print_it) {
2642       _out->print_cr(" "PTR_FORMAT"%s",
2643                      o, (over_tams) ? " >" : (marked) ? " M" : "");
2644       PrintReachableOopClosure oopCl(_out, _vo, _all);
2645       o->oop_iterate_no_header(&oopCl);
2646     }
2647   }
2648 };
2649 
2650 class PrintReachableRegionClosure : public HeapRegionClosure {
2651 private:
2652   G1CollectedHeap* _g1h;
2653   outputStream*    _out;
2654   VerifyOption     _vo;
2655   bool             _all;
2656 
2657 public:
2658   bool doHeapRegion(HeapRegion* hr) {
2659     HeapWord* b = hr->bottom();
2660     HeapWord* e = hr->end();
2661     HeapWord* t = hr->top();
2662     HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2663     _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2664                    "TAMS: "PTR_FORMAT, b, e, t, p);
2665     _out->cr();
2666 
2667     HeapWord* from = b;
2668     HeapWord* to   = t;
2669 
2670     if (to > from) {
2671       _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to);
2672       _out->cr();
2673       PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2674       hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2675       _out->cr();
2676     }
2677 
2678     return false;
2679   }
2680 
2681   PrintReachableRegionClosure(outputStream* out,
2682                               VerifyOption  vo,
2683                               bool          all) :
2684     _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2685 };
2686 
2687 void ConcurrentMark::print_reachable(const char* str,
2688                                      VerifyOption vo,
2689                                      bool all) {
2690   gclog_or_tty->cr();
2691   gclog_or_tty->print_cr("== Doing heap dump... ");
2692 
2693   if (G1PrintReachableBaseFile == NULL) {
2694     gclog_or_tty->print_cr("  #### error: no base file defined");
2695     return;
2696   }
2697 
2698   if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2699       (JVM_MAXPATHLEN - 1)) {
2700     gclog_or_tty->print_cr("  #### error: file name too long");
2701     return;
2702   }
2703 
2704   char file_name[JVM_MAXPATHLEN];
2705   sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2706   gclog_or_tty->print_cr("  dumping to file %s", file_name);
2707 
2708   fileStream fout(file_name);
2709   if (!fout.is_open()) {
2710     gclog_or_tty->print_cr("  #### error: could not open file");
2711     return;
2712   }
2713 
2714   outputStream* out = &fout;
2715   out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2716   out->cr();
2717 
2718   out->print_cr("--- ITERATING OVER REGIONS");
2719   out->cr();
2720   PrintReachableRegionClosure rcl(out, vo, all);
2721   _g1h->heap_region_iterate(&rcl);
2722   out->cr();
2723 
2724   gclog_or_tty->print_cr("  done");
2725   gclog_or_tty->flush();
2726 }
2727 
2728 #endif // PRODUCT
2729 
2730 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2731   // Note we are overriding the read-only view of the prev map here, via
2732   // the cast.
2733   ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2734 }
2735 
2736 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2737   _nextMarkBitMap->clearRange(mr);
2738 }
2739 
2740 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) {
2741   clearRangePrevBitmap(mr);
2742   clearRangeNextBitmap(mr);
2743 }
2744 
2745 HeapRegion*
2746 ConcurrentMark::claim_region(uint worker_id) {
2747   // "checkpoint" the finger
2748   HeapWord* finger = _finger;
2749 
2750   // _heap_end will not change underneath our feet; it only changes at
2751   // yield points.
2752   while (finger < _heap_end) {
2753     assert(_g1h->is_in_g1_reserved(finger), "invariant");
2754 
2755     // Note on how this code handles humongous regions. In the
2756     // normal case the finger will reach the start of a "starts
2757     // humongous" (SH) region. Its end will either be the end of the
2758     // last "continues humongous" (CH) region in the sequence, or the
2759     // standard end of the SH region (if the SH is the only region in
2760     // the sequence). That way claim_region() will skip over the CH
2761     // regions. However, there is a subtle race between a CM thread
2762     // executing this method and a mutator thread doing a humongous
2763     // object allocation. The two are not mutually exclusive as the CM
2764     // thread does not need to hold the Heap_lock when it gets
2765     // here. So there is a chance that claim_region() will come across
2766     // a free region that's in the progress of becoming a SH or a CH
2767     // region. In the former case, it will either
2768     //   a) Miss the update to the region's end, in which case it will
2769     //      visit every subsequent CH region, will find their bitmaps
2770     //      empty, and do nothing, or
2771     //   b) Will observe the update of the region's end (in which case
2772     //      it will skip the subsequent CH regions).
2773     // If it comes across a region that suddenly becomes CH, the
2774     // scenario will be similar to b). So, the race between
2775     // claim_region() and a humongous object allocation might force us
2776     // to do a bit of unnecessary work (due to some unnecessary bitmap
2777     // iterations) but it should not introduce and correctness issues.
2778     HeapRegion* curr_region   = _g1h->heap_region_containing_raw(finger);
2779     HeapWord*   bottom        = curr_region->bottom();
2780     HeapWord*   end           = curr_region->end();
2781     HeapWord*   limit         = curr_region->next_top_at_mark_start();
2782 
2783     if (verbose_low()) {
2784       gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2785                              "["PTR_FORMAT", "PTR_FORMAT"), "
2786                              "limit = "PTR_FORMAT,
2787                              worker_id, curr_region, bottom, end, limit);
2788     }
2789 
2790     // Is the gap between reading the finger and doing the CAS too long?
2791     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2792     if (res == finger) {
2793       // we succeeded
2794 
2795       // notice that _finger == end cannot be guaranteed here since,
2796       // someone else might have moved the finger even further
2797       assert(_finger >= end, "the finger should have moved forward");
2798 
2799       if (verbose_low()) {
2800         gclog_or_tty->print_cr("[%u] we were successful with region = "
2801                                PTR_FORMAT, worker_id, curr_region);
2802       }
2803 
2804       if (limit > bottom) {
2805         if (verbose_low()) {
2806           gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2807                                  "returning it ", worker_id, curr_region);
2808         }
2809         return curr_region;
2810       } else {
2811         assert(limit == bottom,
2812                "the region limit should be at bottom");
2813         if (verbose_low()) {
2814           gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
2815                                  "returning NULL", worker_id, curr_region);
2816         }
2817         // we return NULL and the caller should try calling
2818         // claim_region() again.
2819         return NULL;
2820       }
2821     } else {
2822       assert(_finger > finger, "the finger should have moved forward");
2823       if (verbose_low()) {
2824         gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
2825                                "global finger = "PTR_FORMAT", "
2826                                "our finger = "PTR_FORMAT,
2827                                worker_id, _finger, finger);
2828       }
2829 
2830       // read it again
2831       finger = _finger;
2832     }
2833   }
2834 
2835   return NULL;
2836 }
2837 
2838 #ifndef PRODUCT
2839 enum VerifyNoCSetOopsPhase {
2840   VerifyNoCSetOopsStack,
2841   VerifyNoCSetOopsQueues,
2842   VerifyNoCSetOopsSATBCompleted,
2843   VerifyNoCSetOopsSATBThread
2844 };
2845 
2846 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure  {
2847 private:
2848   G1CollectedHeap* _g1h;
2849   VerifyNoCSetOopsPhase _phase;
2850   int _info;
2851 
2852   const char* phase_str() {
2853     switch (_phase) {
2854     case VerifyNoCSetOopsStack:         return "Stack";
2855     case VerifyNoCSetOopsQueues:        return "Queue";
2856     case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
2857     case VerifyNoCSetOopsSATBThread:    return "Thread SATB Buffers";
2858     default:                            ShouldNotReachHere();
2859     }
2860     return NULL;
2861   }
2862 
2863   void do_object_work(oop obj) {
2864     guarantee(!_g1h->obj_in_cs(obj),
2865               err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
2866                       (void*) obj, phase_str(), _info));
2867   }
2868 
2869 public:
2870   VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
2871 
2872   void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
2873     _phase = phase;
2874     _info = info;
2875   }
2876 
2877   virtual void do_oop(oop* p) {
2878     oop obj = oopDesc::load_decode_heap_oop(p);
2879     do_object_work(obj);
2880   }
2881 
2882   virtual void do_oop(narrowOop* p) {
2883     // We should not come across narrow oops while scanning marking
2884     // stacks and SATB buffers.
2885     ShouldNotReachHere();
2886   }
2887 
2888   virtual void do_object(oop obj) {
2889     do_object_work(obj);
2890   }
2891 };
2892 
2893 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
2894                                          bool verify_enqueued_buffers,
2895                                          bool verify_thread_buffers,
2896                                          bool verify_fingers) {
2897   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2898   if (!G1CollectedHeap::heap()->mark_in_progress()) {
2899     return;
2900   }
2901 
2902   VerifyNoCSetOopsClosure cl;
2903 
2904   if (verify_stacks) {
2905     // Verify entries on the global mark stack
2906     cl.set_phase(VerifyNoCSetOopsStack);
2907     _markStack.oops_do(&cl);
2908 
2909     // Verify entries on the task queues
2910     for (uint i = 0; i < _max_worker_id; i += 1) {
2911       cl.set_phase(VerifyNoCSetOopsQueues, i);
2912       CMTaskQueue* queue = _task_queues->queue(i);
2913       queue->oops_do(&cl);
2914     }
2915   }
2916 
2917   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
2918 
2919   // Verify entries on the enqueued SATB buffers
2920   if (verify_enqueued_buffers) {
2921     cl.set_phase(VerifyNoCSetOopsSATBCompleted);
2922     satb_qs.iterate_completed_buffers_read_only(&cl);
2923   }
2924 
2925   // Verify entries on the per-thread SATB buffers
2926   if (verify_thread_buffers) {
2927     cl.set_phase(VerifyNoCSetOopsSATBThread);
2928     satb_qs.iterate_thread_buffers_read_only(&cl);
2929   }
2930 
2931   if (verify_fingers) {
2932     // Verify the global finger
2933     HeapWord* global_finger = finger();
2934     if (global_finger != NULL && global_finger < _heap_end) {
2935       // The global finger always points to a heap region boundary. We
2936       // use heap_region_containing_raw() to get the containing region
2937       // given that the global finger could be pointing to a free region
2938       // which subsequently becomes continues humongous. If that
2939       // happens, heap_region_containing() will return the bottom of the
2940       // corresponding starts humongous region and the check below will
2941       // not hold any more.
2942       HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
2943       guarantee(global_finger == global_hr->bottom(),
2944                 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
2945                         global_finger, HR_FORMAT_PARAMS(global_hr)));
2946     }
2947 
2948     // Verify the task fingers
2949     assert(parallel_marking_threads() <= _max_worker_id, "sanity");
2950     for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
2951       CMTask* task = _tasks[i];
2952       HeapWord* task_finger = task->finger();
2953       if (task_finger != NULL && task_finger < _heap_end) {
2954         // See above note on the global finger verification.
2955         HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
2956         guarantee(task_finger == task_hr->bottom() ||
2957                   !task_hr->in_collection_set(),
2958                   err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
2959                           task_finger, HR_FORMAT_PARAMS(task_hr)));
2960       }
2961     }
2962   }
2963 }
2964 #endif // PRODUCT
2965 
2966 void ConcurrentMark::clear_marking_state(bool clear_overflow) {
2967   _markStack.set_should_expand();
2968   _markStack.setEmpty();        // Also clears the _markStack overflow flag
2969   if (clear_overflow) {
2970     clear_has_overflown();
2971   } else {
2972     assert(has_overflown(), "pre-condition");
2973   }
2974   _finger = _heap_start;
2975 
2976   for (uint i = 0; i < _max_worker_id; ++i) {
2977     CMTaskQueue* queue = _task_queues->queue(i);
2978     queue->set_empty();
2979   }
2980 }
2981 
2982 // Aggregate the counting data that was constructed concurrently
2983 // with marking.
2984 class AggregateCountDataHRClosure: public HeapRegionClosure {
2985   G1CollectedHeap* _g1h;
2986   ConcurrentMark* _cm;
2987   CardTableModRefBS* _ct_bs;
2988   BitMap* _cm_card_bm;
2989   uint _max_worker_id;
2990 
2991  public:
2992   AggregateCountDataHRClosure(G1CollectedHeap* g1h,
2993                               BitMap* cm_card_bm,
2994                               uint max_worker_id) :
2995     _g1h(g1h), _cm(g1h->concurrent_mark()),
2996     _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
2997     _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
2998 
2999   bool doHeapRegion(HeapRegion* hr) {
3000     if (hr->continuesHumongous()) {
3001       // We will ignore these here and process them when their
3002       // associated "starts humongous" region is processed.
3003       // Note that we cannot rely on their associated
3004       // "starts humongous" region to have their bit set to 1
3005       // since, due to the region chunking in the parallel region
3006       // iteration, a "continues humongous" region might be visited
3007       // before its associated "starts humongous".
3008       return false;
3009     }
3010 
3011     HeapWord* start = hr->bottom();
3012     HeapWord* limit = hr->next_top_at_mark_start();
3013     HeapWord* end = hr->end();
3014 
3015     assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3016            err_msg("Preconditions not met - "
3017                    "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3018                    "top: "PTR_FORMAT", end: "PTR_FORMAT,
3019                    start, limit, hr->top(), hr->end()));
3020 
3021     assert(hr->next_marked_bytes() == 0, "Precondition");
3022 
3023     if (start == limit) {
3024       // NTAMS of this region has not been set so nothing to do.
3025       return false;
3026     }
3027 
3028     // 'start' should be in the heap.
3029     assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3030     // 'end' *may* be just beyone the end of the heap (if hr is the last region)
3031     assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3032 
3033     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3034     BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3035     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3036 
3037     // If ntams is not card aligned then we bump card bitmap index
3038     // for limit so that we get the all the cards spanned by
3039     // the object ending at ntams.
3040     // Note: if this is the last region in the heap then ntams
3041     // could be actually just beyond the end of the the heap;
3042     // limit_idx will then  correspond to a (non-existent) card
3043     // that is also outside the heap.
3044     if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3045       limit_idx += 1;
3046     }
3047 
3048     assert(limit_idx <= end_idx, "or else use atomics");
3049 
3050     // Aggregate the "stripe" in the count data associated with hr.
3051     uint hrs_index = hr->hrs_index();
3052     size_t marked_bytes = 0;
3053 
3054     for (uint i = 0; i < _max_worker_id; i += 1) {
3055       size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3056       BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3057 
3058       // Fetch the marked_bytes in this region for task i and
3059       // add it to the running total for this region.
3060       marked_bytes += marked_bytes_array[hrs_index];
3061 
3062       // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3063       // into the global card bitmap.
3064       BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3065 
3066       while (scan_idx < limit_idx) {
3067         assert(task_card_bm->at(scan_idx) == true, "should be");
3068         _cm_card_bm->set_bit(scan_idx);
3069         assert(_cm_card_bm->at(scan_idx) == true, "should be");
3070 
3071         // BitMap::get_next_one_offset() can handle the case when
3072         // its left_offset parameter is greater than its right_offset
3073         // parameter. It does, however, have an early exit if
3074         // left_offset == right_offset. So let's limit the value
3075         // passed in for left offset here.
3076         BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3077         scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3078       }
3079     }
3080 
3081     // Update the marked bytes for this region.
3082     hr->add_to_marked_bytes(marked_bytes);
3083 
3084     // Next heap region
3085     return false;
3086   }
3087 };
3088 
3089 class G1AggregateCountDataTask: public AbstractGangTask {
3090 protected:
3091   G1CollectedHeap* _g1h;
3092   ConcurrentMark* _cm;
3093   BitMap* _cm_card_bm;
3094   uint _max_worker_id;
3095   int _active_workers;
3096 
3097 public:
3098   G1AggregateCountDataTask(G1CollectedHeap* g1h,
3099                            ConcurrentMark* cm,
3100                            BitMap* cm_card_bm,
3101                            uint max_worker_id,
3102                            int n_workers) :
3103     AbstractGangTask("Count Aggregation"),
3104     _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3105     _max_worker_id(max_worker_id),
3106     _active_workers(n_workers) { }
3107 
3108   void work(uint worker_id) {
3109     AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3110 
3111     if (G1CollectedHeap::use_parallel_gc_threads()) {
3112       _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3113                                             _active_workers,
3114                                             HeapRegion::AggregateCountClaimValue);
3115     } else {
3116       _g1h->heap_region_iterate(&cl);
3117     }
3118   }
3119 };
3120 
3121 
3122 void ConcurrentMark::aggregate_count_data() {
3123   int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3124                         _g1h->workers()->active_workers() :
3125                         1);
3126 
3127   G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3128                                            _max_worker_id, n_workers);
3129 
3130   if (G1CollectedHeap::use_parallel_gc_threads()) {
3131     assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3132            "sanity check");
3133     _g1h->set_par_threads(n_workers);
3134     _g1h->workers()->run_task(&g1_par_agg_task);
3135     _g1h->set_par_threads(0);
3136 
3137     assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3138            "sanity check");
3139     _g1h->reset_heap_region_claim_values();
3140   } else {
3141     g1_par_agg_task.work(0);
3142   }
3143 }
3144 
3145 // Clear the per-worker arrays used to store the per-region counting data
3146 void ConcurrentMark::clear_all_count_data() {
3147   // Clear the global card bitmap - it will be filled during
3148   // liveness count aggregation (during remark) and the
3149   // final counting task.
3150   _card_bm.clear();
3151 
3152   // Clear the global region bitmap - it will be filled as part
3153   // of the final counting task.
3154   _region_bm.clear();
3155 
3156   uint max_regions = _g1h->max_regions();
3157   assert(_max_worker_id > 0, "uninitialized");
3158 
3159   for (uint i = 0; i < _max_worker_id; i += 1) {
3160     BitMap* task_card_bm = count_card_bitmap_for(i);
3161     size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3162 
3163     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3164     assert(marked_bytes_array != NULL, "uninitialized");
3165 
3166     memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3167     task_card_bm->clear();
3168   }
3169 }
3170 
3171 void ConcurrentMark::print_stats() {
3172   if (verbose_stats()) {
3173     gclog_or_tty->print_cr("---------------------------------------------------------------------");
3174     for (size_t i = 0; i < _active_tasks; ++i) {
3175       _tasks[i]->print_stats();
3176       gclog_or_tty->print_cr("---------------------------------------------------------------------");
3177     }
3178   }
3179 }
3180 
3181 // abandon current marking iteration due to a Full GC
3182 void ConcurrentMark::abort() {
3183   // Clear all marks to force marking thread to do nothing
3184   _nextMarkBitMap->clearAll();
3185   // Clear the liveness counting data
3186   clear_all_count_data();
3187   // Empty mark stack
3188   clear_marking_state();
3189   for (uint i = 0; i < _max_worker_id; ++i) {
3190     _tasks[i]->clear_region_fields();
3191   }
3192   _has_aborted = true;
3193 
3194   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3195   satb_mq_set.abandon_partial_marking();
3196   // This can be called either during or outside marking, we'll read
3197   // the expected_active value from the SATB queue set.
3198   satb_mq_set.set_active_all_threads(
3199                                  false, /* new active value */
3200                                  satb_mq_set.is_active() /* expected_active */);
3201 }
3202 
3203 static void print_ms_time_info(const char* prefix, const char* name,
3204                                NumberSeq& ns) {
3205   gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3206                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3207   if (ns.num() > 0) {
3208     gclog_or_tty->print_cr("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
3209                            prefix, ns.sd(), ns.maximum());
3210   }
3211 }
3212 
3213 void ConcurrentMark::print_summary_info() {
3214   gclog_or_tty->print_cr(" Concurrent marking:");
3215   print_ms_time_info("  ", "init marks", _init_times);
3216   print_ms_time_info("  ", "remarks", _remark_times);
3217   {
3218     print_ms_time_info("     ", "final marks", _remark_mark_times);
3219     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
3220 
3221   }
3222   print_ms_time_info("  ", "cleanups", _cleanup_times);
3223   gclog_or_tty->print_cr("    Final counting total time = %8.2f s (avg = %8.2f ms).",
3224                          _total_counting_time,
3225                          (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3226                           (double)_cleanup_times.num()
3227                          : 0.0));
3228   if (G1ScrubRemSets) {
3229     gclog_or_tty->print_cr("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
3230                            _total_rs_scrub_time,
3231                            (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3232                             (double)_cleanup_times.num()
3233                            : 0.0));
3234   }
3235   gclog_or_tty->print_cr("  Total stop_world time = %8.2f s.",
3236                          (_init_times.sum() + _remark_times.sum() +
3237                           _cleanup_times.sum())/1000.0);
3238   gclog_or_tty->print_cr("  Total concurrent time = %8.2f s "
3239                 "(%8.2f s marking).",
3240                 cmThread()->vtime_accum(),
3241                 cmThread()->vtime_mark_accum());
3242 }
3243 
3244 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3245   _parallel_workers->print_worker_threads_on(st);
3246 }
3247 
3248 // We take a break if someone is trying to stop the world.
3249 bool ConcurrentMark::do_yield_check(uint worker_id) {
3250   if (should_yield()) {
3251     if (worker_id == 0) {
3252       _g1h->g1_policy()->record_concurrent_pause();
3253     }
3254     cmThread()->yield();
3255     return true;
3256   } else {
3257     return false;
3258   }
3259 }
3260 
3261 bool ConcurrentMark::should_yield() {
3262   return cmThread()->should_yield();
3263 }
3264 
3265 bool ConcurrentMark::containing_card_is_marked(void* p) {
3266   size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1);
3267   return _card_bm.at(offset >> CardTableModRefBS::card_shift);
3268 }
3269 
3270 bool ConcurrentMark::containing_cards_are_marked(void* start,
3271                                                  void* last) {
3272   return containing_card_is_marked(start) &&
3273          containing_card_is_marked(last);
3274 }
3275 
3276 #ifndef PRODUCT
3277 // for debugging purposes
3278 void ConcurrentMark::print_finger() {
3279   gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3280                          _heap_start, _heap_end, _finger);
3281   for (uint i = 0; i < _max_worker_id; ++i) {
3282     gclog_or_tty->print("   %u: "PTR_FORMAT, i, _tasks[i]->finger());
3283   }
3284   gclog_or_tty->print_cr("");
3285 }
3286 #endif
3287 
3288 void CMTask::scan_object(oop obj) {
3289   assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3290 
3291   if (_cm->verbose_high()) {
3292     gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3293                            _worker_id, (void*) obj);
3294   }
3295 
3296   size_t obj_size = obj->size();
3297   _words_scanned += obj_size;
3298 
3299   obj->oop_iterate(_cm_oop_closure);
3300   statsOnly( ++_objs_scanned );
3301   check_limits();
3302 }
3303 
3304 // Closure for iteration over bitmaps
3305 class CMBitMapClosure : public BitMapClosure {
3306 private:
3307   // the bitmap that is being iterated over
3308   CMBitMap*                   _nextMarkBitMap;
3309   ConcurrentMark*             _cm;
3310   CMTask*                     _task;
3311 
3312 public:
3313   CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3314     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3315 
3316   bool do_bit(size_t offset) {
3317     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3318     assert(_nextMarkBitMap->isMarked(addr), "invariant");
3319     assert( addr < _cm->finger(), "invariant");
3320 
3321     statsOnly( _task->increase_objs_found_on_bitmap() );
3322     assert(addr >= _task->finger(), "invariant");
3323 
3324     // We move that task's local finger along.
3325     _task->move_finger_to(addr);
3326 
3327     _task->scan_object(oop(addr));
3328     // we only partially drain the local queue and global stack
3329     _task->drain_local_queue(true);
3330     _task->drain_global_stack(true);
3331 
3332     // if the has_aborted flag has been raised, we need to bail out of
3333     // the iteration
3334     return !_task->has_aborted();
3335   }
3336 };
3337 
3338 // Closure for iterating over objects, currently only used for
3339 // processing SATB buffers.
3340 class CMObjectClosure : public ObjectClosure {
3341 private:
3342   CMTask* _task;
3343 
3344 public:
3345   void do_object(oop obj) {
3346     _task->deal_with_reference(obj);
3347   }
3348 
3349   CMObjectClosure(CMTask* task) : _task(task) { }
3350 };
3351 
3352 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3353                                ConcurrentMark* cm,
3354                                CMTask* task)
3355   : _g1h(g1h), _cm(cm), _task(task) {
3356   assert(_ref_processor == NULL, "should be initialized to NULL");
3357 
3358   if (G1UseConcMarkReferenceProcessing) {
3359     _ref_processor = g1h->ref_processor_cm();
3360     assert(_ref_processor != NULL, "should not be NULL");
3361   }
3362 }
3363 
3364 void CMTask::setup_for_region(HeapRegion* hr) {
3365   // Separated the asserts so that we know which one fires.
3366   assert(hr != NULL,
3367         "claim_region() should have filtered out continues humongous regions");
3368   assert(!hr->continuesHumongous(),
3369         "claim_region() should have filtered out continues humongous regions");
3370 
3371   if (_cm->verbose_low()) {
3372     gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3373                            _worker_id, hr);
3374   }
3375 
3376   _curr_region  = hr;
3377   _finger       = hr->bottom();
3378   update_region_limit();
3379 }
3380 
3381 void CMTask::update_region_limit() {
3382   HeapRegion* hr            = _curr_region;
3383   HeapWord* bottom          = hr->bottom();
3384   HeapWord* limit           = hr->next_top_at_mark_start();
3385 
3386   if (limit == bottom) {
3387     if (_cm->verbose_low()) {
3388       gclog_or_tty->print_cr("[%u] found an empty region "
3389                              "["PTR_FORMAT", "PTR_FORMAT")",
3390                              _worker_id, bottom, limit);
3391     }
3392     // The region was collected underneath our feet.
3393     // We set the finger to bottom to ensure that the bitmap
3394     // iteration that will follow this will not do anything.
3395     // (this is not a condition that holds when we set the region up,
3396     // as the region is not supposed to be empty in the first place)
3397     _finger = bottom;
3398   } else if (limit >= _region_limit) {
3399     assert(limit >= _finger, "peace of mind");
3400   } else {
3401     assert(limit < _region_limit, "only way to get here");
3402     // This can happen under some pretty unusual circumstances.  An
3403     // evacuation pause empties the region underneath our feet (NTAMS
3404     // at bottom). We then do some allocation in the region (NTAMS
3405     // stays at bottom), followed by the region being used as a GC
3406     // alloc region (NTAMS will move to top() and the objects
3407     // originally below it will be grayed). All objects now marked in
3408     // the region are explicitly grayed, if below the global finger,
3409     // and we do not need in fact to scan anything else. So, we simply
3410     // set _finger to be limit to ensure that the bitmap iteration
3411     // doesn't do anything.
3412     _finger = limit;
3413   }
3414 
3415   _region_limit = limit;
3416 }
3417 
3418 void CMTask::giveup_current_region() {
3419   assert(_curr_region != NULL, "invariant");
3420   if (_cm->verbose_low()) {
3421     gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3422                            _worker_id, _curr_region);
3423   }
3424   clear_region_fields();
3425 }
3426 
3427 void CMTask::clear_region_fields() {
3428   // Values for these three fields that indicate that we're not
3429   // holding on to a region.
3430   _curr_region   = NULL;
3431   _finger        = NULL;
3432   _region_limit  = NULL;
3433 }
3434 
3435 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3436   if (cm_oop_closure == NULL) {
3437     assert(_cm_oop_closure != NULL, "invariant");
3438   } else {
3439     assert(_cm_oop_closure == NULL, "invariant");
3440   }
3441   _cm_oop_closure = cm_oop_closure;
3442 }
3443 
3444 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3445   guarantee(nextMarkBitMap != NULL, "invariant");
3446 
3447   if (_cm->verbose_low()) {
3448     gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3449   }
3450 
3451   _nextMarkBitMap                = nextMarkBitMap;
3452   clear_region_fields();
3453 
3454   _calls                         = 0;
3455   _elapsed_time_ms               = 0.0;
3456   _termination_time_ms           = 0.0;
3457   _termination_start_time_ms     = 0.0;
3458 
3459 #if _MARKING_STATS_
3460   _local_pushes                  = 0;
3461   _local_pops                    = 0;
3462   _local_max_size                = 0;
3463   _objs_scanned                  = 0;
3464   _global_pushes                 = 0;
3465   _global_pops                   = 0;
3466   _global_max_size               = 0;
3467   _global_transfers_to           = 0;
3468   _global_transfers_from         = 0;
3469   _regions_claimed               = 0;
3470   _objs_found_on_bitmap          = 0;
3471   _satb_buffers_processed        = 0;
3472   _steal_attempts                = 0;
3473   _steals                        = 0;
3474   _aborted                       = 0;
3475   _aborted_overflow              = 0;
3476   _aborted_cm_aborted            = 0;
3477   _aborted_yield                 = 0;
3478   _aborted_timed_out             = 0;
3479   _aborted_satb                  = 0;
3480   _aborted_termination           = 0;
3481 #endif // _MARKING_STATS_
3482 }
3483 
3484 bool CMTask::should_exit_termination() {
3485   regular_clock_call();
3486   // This is called when we are in the termination protocol. We should
3487   // quit if, for some reason, this task wants to abort or the global
3488   // stack is not empty (this means that we can get work from it).
3489   return !_cm->mark_stack_empty() || has_aborted();
3490 }
3491 
3492 void CMTask::reached_limit() {
3493   assert(_words_scanned >= _words_scanned_limit ||
3494          _refs_reached >= _refs_reached_limit ,
3495          "shouldn't have been called otherwise");
3496   regular_clock_call();
3497 }
3498 
3499 void CMTask::regular_clock_call() {
3500   if (has_aborted()) return;
3501 
3502   // First, we need to recalculate the words scanned and refs reached
3503   // limits for the next clock call.
3504   recalculate_limits();
3505 
3506   // During the regular clock call we do the following
3507 
3508   // (1) If an overflow has been flagged, then we abort.
3509   if (_cm->has_overflown()) {
3510     set_has_aborted();
3511     return;
3512   }
3513 
3514   // If we are not concurrent (i.e. we're doing remark) we don't need
3515   // to check anything else. The other steps are only needed during
3516   // the concurrent marking phase.
3517   if (!concurrent()) return;
3518 
3519   // (2) If marking has been aborted for Full GC, then we also abort.
3520   if (_cm->has_aborted()) {
3521     set_has_aborted();
3522     statsOnly( ++_aborted_cm_aborted );
3523     return;
3524   }
3525 
3526   double curr_time_ms = os::elapsedVTime() * 1000.0;
3527 
3528   // (3) If marking stats are enabled, then we update the step history.
3529 #if _MARKING_STATS_
3530   if (_words_scanned >= _words_scanned_limit) {
3531     ++_clock_due_to_scanning;
3532   }
3533   if (_refs_reached >= _refs_reached_limit) {
3534     ++_clock_due_to_marking;
3535   }
3536 
3537   double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3538   _interval_start_time_ms = curr_time_ms;
3539   _all_clock_intervals_ms.add(last_interval_ms);
3540 
3541   if (_cm->verbose_medium()) {
3542       gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3543                         "scanned = %d%s, refs reached = %d%s",
3544                         _worker_id, last_interval_ms,
3545                         _words_scanned,
3546                         (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3547                         _refs_reached,
3548                         (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3549   }
3550 #endif // _MARKING_STATS_
3551 
3552   // (4) We check whether we should yield. If we have to, then we abort.
3553   if (_cm->should_yield()) {
3554     // We should yield. To do this we abort the task. The caller is
3555     // responsible for yielding.
3556     set_has_aborted();
3557     statsOnly( ++_aborted_yield );
3558     return;
3559   }
3560 
3561   // (5) We check whether we've reached our time quota. If we have,
3562   // then we abort.
3563   double elapsed_time_ms = curr_time_ms - _start_time_ms;
3564   if (elapsed_time_ms > _time_target_ms) {
3565     set_has_aborted();
3566     _has_timed_out = true;
3567     statsOnly( ++_aborted_timed_out );
3568     return;
3569   }
3570 
3571   // (6) Finally, we check whether there are enough completed STAB
3572   // buffers available for processing. If there are, we abort.
3573   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3574   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3575     if (_cm->verbose_low()) {
3576       gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3577                              _worker_id);
3578     }
3579     // we do need to process SATB buffers, we'll abort and restart
3580     // the marking task to do so
3581     set_has_aborted();
3582     statsOnly( ++_aborted_satb );
3583     return;
3584   }
3585 }
3586 
3587 void CMTask::recalculate_limits() {
3588   _real_words_scanned_limit = _words_scanned + words_scanned_period;
3589   _words_scanned_limit      = _real_words_scanned_limit;
3590 
3591   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
3592   _refs_reached_limit       = _real_refs_reached_limit;
3593 }
3594 
3595 void CMTask::decrease_limits() {
3596   // This is called when we believe that we're going to do an infrequent
3597   // operation which will increase the per byte scanned cost (i.e. move
3598   // entries to/from the global stack). It basically tries to decrease the
3599   // scanning limit so that the clock is called earlier.
3600 
3601   if (_cm->verbose_medium()) {
3602     gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3603   }
3604 
3605   _words_scanned_limit = _real_words_scanned_limit -
3606     3 * words_scanned_period / 4;
3607   _refs_reached_limit  = _real_refs_reached_limit -
3608     3 * refs_reached_period / 4;
3609 }
3610 
3611 void CMTask::move_entries_to_global_stack() {
3612   // local array where we'll store the entries that will be popped
3613   // from the local queue
3614   oop buffer[global_stack_transfer_size];
3615 
3616   int n = 0;
3617   oop obj;
3618   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3619     buffer[n] = obj;
3620     ++n;
3621   }
3622 
3623   if (n > 0) {
3624     // we popped at least one entry from the local queue
3625 
3626     statsOnly( ++_global_transfers_to; _local_pops += n );
3627 
3628     if (!_cm->mark_stack_push(buffer, n)) {
3629       if (_cm->verbose_low()) {
3630         gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3631                                _worker_id);
3632       }
3633       set_has_aborted();
3634     } else {
3635       // the transfer was successful
3636 
3637       if (_cm->verbose_medium()) {
3638         gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3639                                _worker_id, n);
3640       }
3641       statsOnly( int tmp_size = _cm->mark_stack_size();
3642                  if (tmp_size > _global_max_size) {
3643                    _global_max_size = tmp_size;
3644                  }
3645                  _global_pushes += n );
3646     }
3647   }
3648 
3649   // this operation was quite expensive, so decrease the limits
3650   decrease_limits();
3651 }
3652 
3653 void CMTask::get_entries_from_global_stack() {
3654   // local array where we'll store the entries that will be popped
3655   // from the global stack.
3656   oop buffer[global_stack_transfer_size];
3657   int n;
3658   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3659   assert(n <= global_stack_transfer_size,
3660          "we should not pop more than the given limit");
3661   if (n > 0) {
3662     // yes, we did actually pop at least one entry
3663 
3664     statsOnly( ++_global_transfers_from; _global_pops += n );
3665     if (_cm->verbose_medium()) {
3666       gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3667                              _worker_id, n);
3668     }
3669     for (int i = 0; i < n; ++i) {
3670       bool success = _task_queue->push(buffer[i]);
3671       // We only call this when the local queue is empty or under a
3672       // given target limit. So, we do not expect this push to fail.
3673       assert(success, "invariant");
3674     }
3675 
3676     statsOnly( int tmp_size = _task_queue->size();
3677                if (tmp_size > _local_max_size) {
3678                  _local_max_size = tmp_size;
3679                }
3680                _local_pushes += n );
3681   }
3682 
3683   // this operation was quite expensive, so decrease the limits
3684   decrease_limits();
3685 }
3686 
3687 void CMTask::drain_local_queue(bool partially) {
3688   if (has_aborted()) return;
3689 
3690   // Decide what the target size is, depending whether we're going to
3691   // drain it partially (so that other tasks can steal if they run out
3692   // of things to do) or totally (at the very end).
3693   size_t target_size;
3694   if (partially) {
3695     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3696   } else {
3697     target_size = 0;
3698   }
3699 
3700   if (_task_queue->size() > target_size) {
3701     if (_cm->verbose_high()) {
3702       gclog_or_tty->print_cr("[%u] draining local queue, target size = %d",
3703                              _worker_id, target_size);
3704     }
3705 
3706     oop obj;
3707     bool ret = _task_queue->pop_local(obj);
3708     while (ret) {
3709       statsOnly( ++_local_pops );
3710 
3711       if (_cm->verbose_high()) {
3712         gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3713                                (void*) obj);
3714       }
3715 
3716       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3717       assert(!_g1h->is_on_master_free_list(
3718                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3719 
3720       scan_object(obj);
3721 
3722       if (_task_queue->size() <= target_size || has_aborted()) {
3723         ret = false;
3724       } else {
3725         ret = _task_queue->pop_local(obj);
3726       }
3727     }
3728 
3729     if (_cm->verbose_high()) {
3730       gclog_or_tty->print_cr("[%u] drained local queue, size = %d",
3731                              _worker_id, _task_queue->size());
3732     }
3733   }
3734 }
3735 
3736 void CMTask::drain_global_stack(bool partially) {
3737   if (has_aborted()) return;
3738 
3739   // We have a policy to drain the local queue before we attempt to
3740   // drain the global stack.
3741   assert(partially || _task_queue->size() == 0, "invariant");
3742 
3743   // Decide what the target size is, depending whether we're going to
3744   // drain it partially (so that other tasks can steal if they run out
3745   // of things to do) or totally (at the very end).  Notice that,
3746   // because we move entries from the global stack in chunks or
3747   // because another task might be doing the same, we might in fact
3748   // drop below the target. But, this is not a problem.
3749   size_t target_size;
3750   if (partially) {
3751     target_size = _cm->partial_mark_stack_size_target();
3752   } else {
3753     target_size = 0;
3754   }
3755 
3756   if (_cm->mark_stack_size() > target_size) {
3757     if (_cm->verbose_low()) {
3758       gclog_or_tty->print_cr("[%u] draining global_stack, target size %d",
3759                              _worker_id, target_size);
3760     }
3761 
3762     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3763       get_entries_from_global_stack();
3764       drain_local_queue(partially);
3765     }
3766 
3767     if (_cm->verbose_low()) {
3768       gclog_or_tty->print_cr("[%u] drained global stack, size = %d",
3769                              _worker_id, _cm->mark_stack_size());
3770     }
3771   }
3772 }
3773 
3774 // SATB Queue has several assumptions on whether to call the par or
3775 // non-par versions of the methods. this is why some of the code is
3776 // replicated. We should really get rid of the single-threaded version
3777 // of the code to simplify things.
3778 void CMTask::drain_satb_buffers() {
3779   if (has_aborted()) return;
3780 
3781   // We set this so that the regular clock knows that we're in the
3782   // middle of draining buffers and doesn't set the abort flag when it
3783   // notices that SATB buffers are available for draining. It'd be
3784   // very counter productive if it did that. :-)
3785   _draining_satb_buffers = true;
3786 
3787   CMObjectClosure oc(this);
3788   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3789   if (G1CollectedHeap::use_parallel_gc_threads()) {
3790     satb_mq_set.set_par_closure(_worker_id, &oc);
3791   } else {
3792     satb_mq_set.set_closure(&oc);
3793   }
3794 
3795   // This keeps claiming and applying the closure to completed buffers
3796   // until we run out of buffers or we need to abort.
3797   if (G1CollectedHeap::use_parallel_gc_threads()) {
3798     while (!has_aborted() &&
3799            satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3800       if (_cm->verbose_medium()) {
3801         gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3802       }
3803       statsOnly( ++_satb_buffers_processed );
3804       regular_clock_call();
3805     }
3806   } else {
3807     while (!has_aborted() &&
3808            satb_mq_set.apply_closure_to_completed_buffer()) {
3809       if (_cm->verbose_medium()) {
3810         gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3811       }
3812       statsOnly( ++_satb_buffers_processed );
3813       regular_clock_call();
3814     }
3815   }
3816 
3817   if (!concurrent() && !has_aborted()) {
3818     // We should only do this during remark.
3819     if (G1CollectedHeap::use_parallel_gc_threads()) {
3820       satb_mq_set.par_iterate_closure_all_threads(_worker_id);
3821     } else {
3822       satb_mq_set.iterate_closure_all_threads();
3823     }
3824   }
3825 
3826   _draining_satb_buffers = false;
3827 
3828   assert(has_aborted() ||
3829          concurrent() ||
3830          satb_mq_set.completed_buffers_num() == 0, "invariant");
3831 
3832   if (G1CollectedHeap::use_parallel_gc_threads()) {
3833     satb_mq_set.set_par_closure(_worker_id, NULL);
3834   } else {
3835     satb_mq_set.set_closure(NULL);
3836   }
3837 
3838   // again, this was a potentially expensive operation, decrease the
3839   // limits to get the regular clock call early
3840   decrease_limits();
3841 }
3842 
3843 void CMTask::print_stats() {
3844   gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
3845                          _worker_id, _calls);
3846   gclog_or_tty->print_cr("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3847                          _elapsed_time_ms, _termination_time_ms);
3848   gclog_or_tty->print_cr("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3849                          _step_times_ms.num(), _step_times_ms.avg(),
3850                          _step_times_ms.sd());
3851   gclog_or_tty->print_cr("                    max = %1.2lfms, total = %1.2lfms",
3852                          _step_times_ms.maximum(), _step_times_ms.sum());
3853 
3854 #if _MARKING_STATS_
3855   gclog_or_tty->print_cr("  Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3856                          _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
3857                          _all_clock_intervals_ms.sd());
3858   gclog_or_tty->print_cr("                         max = %1.2lfms, total = %1.2lfms",
3859                          _all_clock_intervals_ms.maximum(),
3860                          _all_clock_intervals_ms.sum());
3861   gclog_or_tty->print_cr("  Clock Causes (cum): scanning = %d, marking = %d",
3862                          _clock_due_to_scanning, _clock_due_to_marking);
3863   gclog_or_tty->print_cr("  Objects: scanned = %d, found on the bitmap = %d",
3864                          _objs_scanned, _objs_found_on_bitmap);
3865   gclog_or_tty->print_cr("  Local Queue:  pushes = %d, pops = %d, max size = %d",
3866                          _local_pushes, _local_pops, _local_max_size);
3867   gclog_or_tty->print_cr("  Global Stack: pushes = %d, pops = %d, max size = %d",
3868                          _global_pushes, _global_pops, _global_max_size);
3869   gclog_or_tty->print_cr("                transfers to = %d, transfers from = %d",
3870                          _global_transfers_to,_global_transfers_from);
3871   gclog_or_tty->print_cr("  Regions: claimed = %d", _regions_claimed);
3872   gclog_or_tty->print_cr("  SATB buffers: processed = %d", _satb_buffers_processed);
3873   gclog_or_tty->print_cr("  Steals: attempts = %d, successes = %d",
3874                          _steal_attempts, _steals);
3875   gclog_or_tty->print_cr("  Aborted: %d, due to", _aborted);
3876   gclog_or_tty->print_cr("    overflow: %d, global abort: %d, yield: %d",
3877                          _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
3878   gclog_or_tty->print_cr("    time out: %d, SATB: %d, termination: %d",
3879                          _aborted_timed_out, _aborted_satb, _aborted_termination);
3880 #endif // _MARKING_STATS_
3881 }
3882 
3883 /*****************************************************************************
3884 
3885     The do_marking_step(time_target_ms) method is the building block
3886     of the parallel marking framework. It can be called in parallel
3887     with other invocations of do_marking_step() on different tasks
3888     (but only one per task, obviously) and concurrently with the
3889     mutator threads, or during remark, hence it eliminates the need
3890     for two versions of the code. When called during remark, it will
3891     pick up from where the task left off during the concurrent marking
3892     phase. Interestingly, tasks are also claimable during evacuation
3893     pauses too, since do_marking_step() ensures that it aborts before
3894     it needs to yield.
3895 
3896     The data structures that is uses to do marking work are the
3897     following:
3898 
3899       (1) Marking Bitmap. If there are gray objects that appear only
3900       on the bitmap (this happens either when dealing with an overflow
3901       or when the initial marking phase has simply marked the roots
3902       and didn't push them on the stack), then tasks claim heap
3903       regions whose bitmap they then scan to find gray objects. A
3904       global finger indicates where the end of the last claimed region
3905       is. A local finger indicates how far into the region a task has
3906       scanned. The two fingers are used to determine how to gray an
3907       object (i.e. whether simply marking it is OK, as it will be
3908       visited by a task in the future, or whether it needs to be also
3909       pushed on a stack).
3910 
3911       (2) Local Queue. The local queue of the task which is accessed
3912       reasonably efficiently by the task. Other tasks can steal from
3913       it when they run out of work. Throughout the marking phase, a
3914       task attempts to keep its local queue short but not totally
3915       empty, so that entries are available for stealing by other
3916       tasks. Only when there is no more work, a task will totally
3917       drain its local queue.
3918 
3919       (3) Global Mark Stack. This handles local queue overflow. During
3920       marking only sets of entries are moved between it and the local
3921       queues, as access to it requires a mutex and more fine-grain
3922       interaction with it which might cause contention. If it
3923       overflows, then the marking phase should restart and iterate
3924       over the bitmap to identify gray objects. Throughout the marking
3925       phase, tasks attempt to keep the global mark stack at a small
3926       length but not totally empty, so that entries are available for
3927       popping by other tasks. Only when there is no more work, tasks
3928       will totally drain the global mark stack.
3929 
3930       (4) SATB Buffer Queue. This is where completed SATB buffers are
3931       made available. Buffers are regularly removed from this queue
3932       and scanned for roots, so that the queue doesn't get too
3933       long. During remark, all completed buffers are processed, as
3934       well as the filled in parts of any uncompleted buffers.
3935 
3936     The do_marking_step() method tries to abort when the time target
3937     has been reached. There are a few other cases when the
3938     do_marking_step() method also aborts:
3939 
3940       (1) When the marking phase has been aborted (after a Full GC).
3941 
3942       (2) When a global overflow (on the global stack) has been
3943       triggered. Before the task aborts, it will actually sync up with
3944       the other tasks to ensure that all the marking data structures
3945       (local queues, stacks, fingers etc.)  are re-initialised so that
3946       when do_marking_step() completes, the marking phase can
3947       immediately restart.
3948 
3949       (3) When enough completed SATB buffers are available. The
3950       do_marking_step() method only tries to drain SATB buffers right
3951       at the beginning. So, if enough buffers are available, the
3952       marking step aborts and the SATB buffers are processed at
3953       the beginning of the next invocation.
3954 
3955       (4) To yield. when we have to yield then we abort and yield
3956       right at the end of do_marking_step(). This saves us from a lot
3957       of hassle as, by yielding we might allow a Full GC. If this
3958       happens then objects will be compacted underneath our feet, the
3959       heap might shrink, etc. We save checking for this by just
3960       aborting and doing the yield right at the end.
3961 
3962     From the above it follows that the do_marking_step() method should
3963     be called in a loop (or, otherwise, regularly) until it completes.
3964 
3965     If a marking step completes without its has_aborted() flag being
3966     true, it means it has completed the current marking phase (and
3967     also all other marking tasks have done so and have all synced up).
3968 
3969     A method called regular_clock_call() is invoked "regularly" (in
3970     sub ms intervals) throughout marking. It is this clock method that
3971     checks all the abort conditions which were mentioned above and
3972     decides when the task should abort. A work-based scheme is used to
3973     trigger this clock method: when the number of object words the
3974     marking phase has scanned or the number of references the marking
3975     phase has visited reach a given limit. Additional invocations to
3976     the method clock have been planted in a few other strategic places
3977     too. The initial reason for the clock method was to avoid calling
3978     vtime too regularly, as it is quite expensive. So, once it was in
3979     place, it was natural to piggy-back all the other conditions on it
3980     too and not constantly check them throughout the code.
3981 
3982  *****************************************************************************/
3983 
3984 void CMTask::do_marking_step(double time_target_ms,
3985                              bool do_stealing,
3986                              bool do_termination) {
3987   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
3988   assert(concurrent() == _cm->concurrent(), "they should be the same");
3989 
3990   G1CollectorPolicy* g1_policy = _g1h->g1_policy();
3991   assert(_task_queues != NULL, "invariant");
3992   assert(_task_queue != NULL, "invariant");
3993   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
3994 
3995   assert(!_claimed,
3996          "only one thread should claim this task at any one time");
3997 
3998   // OK, this doesn't safeguard again all possible scenarios, as it is
3999   // possible for two threads to set the _claimed flag at the same
4000   // time. But it is only for debugging purposes anyway and it will
4001   // catch most problems.
4002   _claimed = true;
4003 
4004   _start_time_ms = os::elapsedVTime() * 1000.0;
4005   statsOnly( _interval_start_time_ms = _start_time_ms );
4006 
4007   double diff_prediction_ms =
4008     g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4009   _time_target_ms = time_target_ms - diff_prediction_ms;
4010 
4011   // set up the variables that are used in the work-based scheme to
4012   // call the regular clock method
4013   _words_scanned = 0;
4014   _refs_reached  = 0;
4015   recalculate_limits();
4016 
4017   // clear all flags
4018   clear_has_aborted();
4019   _has_timed_out = false;
4020   _draining_satb_buffers = false;
4021 
4022   ++_calls;
4023 
4024   if (_cm->verbose_low()) {
4025     gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4026                            "target = %1.2lfms >>>>>>>>>>",
4027                            _worker_id, _calls, _time_target_ms);
4028   }
4029 
4030   // Set up the bitmap and oop closures. Anything that uses them is
4031   // eventually called from this method, so it is OK to allocate these
4032   // statically.
4033   CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4034   G1CMOopClosure  cm_oop_closure(_g1h, _cm, this);
4035   set_cm_oop_closure(&cm_oop_closure);
4036 
4037   if (_cm->has_overflown()) {
4038     // This can happen if the mark stack overflows during a GC pause
4039     // and this task, after a yield point, restarts. We have to abort
4040     // as we need to get into the overflow protocol which happens
4041     // right at the end of this task.
4042     set_has_aborted();
4043   }
4044 
4045   // First drain any available SATB buffers. After this, we will not
4046   // look at SATB buffers before the next invocation of this method.
4047   // If enough completed SATB buffers are queued up, the regular clock
4048   // will abort this task so that it restarts.
4049   drain_satb_buffers();
4050   // ...then partially drain the local queue and the global stack
4051   drain_local_queue(true);
4052   drain_global_stack(true);
4053 
4054   do {
4055     if (!has_aborted() && _curr_region != NULL) {
4056       // This means that we're already holding on to a region.
4057       assert(_finger != NULL, "if region is not NULL, then the finger "
4058              "should not be NULL either");
4059 
4060       // We might have restarted this task after an evacuation pause
4061       // which might have evacuated the region we're holding on to
4062       // underneath our feet. Let's read its limit again to make sure
4063       // that we do not iterate over a region of the heap that
4064       // contains garbage (update_region_limit() will also move
4065       // _finger to the start of the region if it is found empty).
4066       update_region_limit();
4067       // We will start from _finger not from the start of the region,
4068       // as we might be restarting this task after aborting half-way
4069       // through scanning this region. In this case, _finger points to
4070       // the address where we last found a marked object. If this is a
4071       // fresh region, _finger points to start().
4072       MemRegion mr = MemRegion(_finger, _region_limit);
4073 
4074       if (_cm->verbose_low()) {
4075         gclog_or_tty->print_cr("[%u] we're scanning part "
4076                                "["PTR_FORMAT", "PTR_FORMAT") "
4077                                "of region "PTR_FORMAT,
4078                                _worker_id, _finger, _region_limit, _curr_region);
4079       }
4080 
4081       // Let's iterate over the bitmap of the part of the
4082       // region that is left.
4083       if (mr.is_empty() || _nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4084         // We successfully completed iterating over the region. Now,
4085         // let's give up the region.
4086         giveup_current_region();
4087         regular_clock_call();
4088       } else {
4089         assert(has_aborted(), "currently the only way to do so");
4090         // The only way to abort the bitmap iteration is to return
4091         // false from the do_bit() method. However, inside the
4092         // do_bit() method we move the _finger to point to the
4093         // object currently being looked at. So, if we bail out, we
4094         // have definitely set _finger to something non-null.
4095         assert(_finger != NULL, "invariant");
4096 
4097         // Region iteration was actually aborted. So now _finger
4098         // points to the address of the object we last scanned. If we
4099         // leave it there, when we restart this task, we will rescan
4100         // the object. It is easy to avoid this. We move the finger by
4101         // enough to point to the next possible object header (the
4102         // bitmap knows by how much we need to move it as it knows its
4103         // granularity).
4104         assert(_finger < _region_limit, "invariant");
4105         HeapWord* new_finger = _nextMarkBitMap->nextWord(_finger);
4106         // Check if bitmap iteration was aborted while scanning the last object
4107         if (new_finger >= _region_limit) {
4108           giveup_current_region();
4109         } else {
4110           move_finger_to(new_finger);
4111         }
4112       }
4113     }
4114     // At this point we have either completed iterating over the
4115     // region we were holding on to, or we have aborted.
4116 
4117     // We then partially drain the local queue and the global stack.
4118     // (Do we really need this?)
4119     drain_local_queue(true);
4120     drain_global_stack(true);
4121 
4122     // Read the note on the claim_region() method on why it might
4123     // return NULL with potentially more regions available for
4124     // claiming and why we have to check out_of_regions() to determine
4125     // whether we're done or not.
4126     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4127       // We are going to try to claim a new region. We should have
4128       // given up on the previous one.
4129       // Separated the asserts so that we know which one fires.
4130       assert(_curr_region  == NULL, "invariant");
4131       assert(_finger       == NULL, "invariant");
4132       assert(_region_limit == NULL, "invariant");
4133       if (_cm->verbose_low()) {
4134         gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4135       }
4136       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4137       if (claimed_region != NULL) {
4138         // Yes, we managed to claim one
4139         statsOnly( ++_regions_claimed );
4140 
4141         if (_cm->verbose_low()) {
4142           gclog_or_tty->print_cr("[%u] we successfully claimed "
4143                                  "region "PTR_FORMAT,
4144                                  _worker_id, claimed_region);
4145         }
4146 
4147         setup_for_region(claimed_region);
4148         assert(_curr_region == claimed_region, "invariant");
4149       }
4150       // It is important to call the regular clock here. It might take
4151       // a while to claim a region if, for example, we hit a large
4152       // block of empty regions. So we need to call the regular clock
4153       // method once round the loop to make sure it's called
4154       // frequently enough.
4155       regular_clock_call();
4156     }
4157 
4158     if (!has_aborted() && _curr_region == NULL) {
4159       assert(_cm->out_of_regions(),
4160              "at this point we should be out of regions");
4161     }
4162   } while ( _curr_region != NULL && !has_aborted());
4163 
4164   if (!has_aborted()) {
4165     // We cannot check whether the global stack is empty, since other
4166     // tasks might be pushing objects to it concurrently.
4167     assert(_cm->out_of_regions(),
4168            "at this point we should be out of regions");
4169 
4170     if (_cm->verbose_low()) {
4171       gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4172     }
4173 
4174     // Try to reduce the number of available SATB buffers so that
4175     // remark has less work to do.
4176     drain_satb_buffers();
4177   }
4178 
4179   // Since we've done everything else, we can now totally drain the
4180   // local queue and global stack.
4181   drain_local_queue(false);
4182   drain_global_stack(false);
4183 
4184   // Attempt at work stealing from other task's queues.
4185   if (do_stealing && !has_aborted()) {
4186     // We have not aborted. This means that we have finished all that
4187     // we could. Let's try to do some stealing...
4188 
4189     // We cannot check whether the global stack is empty, since other
4190     // tasks might be pushing objects to it concurrently.
4191     assert(_cm->out_of_regions() && _task_queue->size() == 0,
4192            "only way to reach here");
4193 
4194     if (_cm->verbose_low()) {
4195       gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4196     }
4197 
4198     while (!has_aborted()) {
4199       oop obj;
4200       statsOnly( ++_steal_attempts );
4201 
4202       if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4203         if (_cm->verbose_medium()) {
4204           gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4205                                  _worker_id, (void*) obj);
4206         }
4207 
4208         statsOnly( ++_steals );
4209 
4210         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4211                "any stolen object should be marked");
4212         scan_object(obj);
4213 
4214         // And since we're towards the end, let's totally drain the
4215         // local queue and global stack.
4216         drain_local_queue(false);
4217         drain_global_stack(false);
4218       } else {
4219         break;
4220       }
4221     }
4222   }
4223 
4224   // If we are about to wrap up and go into termination, check if we
4225   // should raise the overflow flag.
4226   if (do_termination && !has_aborted()) {
4227     if (_cm->force_overflow()->should_force()) {
4228       _cm->set_has_overflown();
4229       regular_clock_call();
4230     }
4231   }
4232 
4233   // We still haven't aborted. Now, let's try to get into the
4234   // termination protocol.
4235   if (do_termination && !has_aborted()) {
4236     // We cannot check whether the global stack is empty, since other
4237     // tasks might be concurrently pushing objects on it.
4238     // Separated the asserts so that we know which one fires.
4239     assert(_cm->out_of_regions(), "only way to reach here");
4240     assert(_task_queue->size() == 0, "only way to reach here");
4241 
4242     if (_cm->verbose_low()) {
4243       gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4244     }
4245 
4246     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4247     // The CMTask class also extends the TerminatorTerminator class,
4248     // hence its should_exit_termination() method will also decide
4249     // whether to exit the termination protocol or not.
4250     bool finished = _cm->terminator()->offer_termination(this);
4251     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4252     _termination_time_ms +=
4253       termination_end_time_ms - _termination_start_time_ms;
4254 
4255     if (finished) {
4256       // We're all done.
4257 
4258       if (_worker_id == 0) {
4259         // let's allow task 0 to do this
4260         if (concurrent()) {
4261           assert(_cm->concurrent_marking_in_progress(), "invariant");
4262           // we need to set this to false before the next
4263           // safepoint. This way we ensure that the marking phase
4264           // doesn't observe any more heap expansions.
4265           _cm->clear_concurrent_marking_in_progress();
4266         }
4267       }
4268 
4269       // We can now guarantee that the global stack is empty, since
4270       // all other tasks have finished. We separated the guarantees so
4271       // that, if a condition is false, we can immediately find out
4272       // which one.
4273       guarantee(_cm->out_of_regions(), "only way to reach here");
4274       guarantee(_cm->mark_stack_empty(), "only way to reach here");
4275       guarantee(_task_queue->size() == 0, "only way to reach here");
4276       guarantee(!_cm->has_overflown(), "only way to reach here");
4277       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4278 
4279       if (_cm->verbose_low()) {
4280         gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4281       }
4282     } else {
4283       // Apparently there's more work to do. Let's abort this task. It
4284       // will restart it and we can hopefully find more things to do.
4285 
4286       if (_cm->verbose_low()) {
4287         gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4288                                _worker_id);
4289       }
4290 
4291       set_has_aborted();
4292       statsOnly( ++_aborted_termination );
4293     }
4294   }
4295 
4296   // Mainly for debugging purposes to make sure that a pointer to the
4297   // closure which was statically allocated in this frame doesn't
4298   // escape it by accident.
4299   set_cm_oop_closure(NULL);
4300   double end_time_ms = os::elapsedVTime() * 1000.0;
4301   double elapsed_time_ms = end_time_ms - _start_time_ms;
4302   // Update the step history.
4303   _step_times_ms.add(elapsed_time_ms);
4304 
4305   if (has_aborted()) {
4306     // The task was aborted for some reason.
4307 
4308     statsOnly( ++_aborted );
4309 
4310     if (_has_timed_out) {
4311       double diff_ms = elapsed_time_ms - _time_target_ms;
4312       // Keep statistics of how well we did with respect to hitting
4313       // our target only if we actually timed out (if we aborted for
4314       // other reasons, then the results might get skewed).
4315       _marking_step_diffs_ms.add(diff_ms);
4316     }
4317 
4318     if (_cm->has_overflown()) {
4319       // This is the interesting one. We aborted because a global
4320       // overflow was raised. This means we have to restart the
4321       // marking phase and start iterating over regions. However, in
4322       // order to do this we have to make sure that all tasks stop
4323       // what they are doing and re-initialise in a safe manner. We
4324       // will achieve this with the use of two barrier sync points.
4325 
4326       if (_cm->verbose_low()) {
4327         gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4328       }
4329 
4330       _cm->enter_first_sync_barrier(_worker_id);
4331       // When we exit this sync barrier we know that all tasks have
4332       // stopped doing marking work. So, it's now safe to
4333       // re-initialise our data structures. At the end of this method,
4334       // task 0 will clear the global data structures.
4335 
4336       statsOnly( ++_aborted_overflow );
4337 
4338       // We clear the local state of this task...
4339       clear_region_fields();
4340 
4341       // ...and enter the second barrier.
4342       _cm->enter_second_sync_barrier(_worker_id);
4343       // At this point everything has bee re-initialised and we're
4344       // ready to restart.
4345     }
4346 
4347     if (_cm->verbose_low()) {
4348       gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4349                              "elapsed = %1.2lfms <<<<<<<<<<",
4350                              _worker_id, _time_target_ms, elapsed_time_ms);
4351       if (_cm->has_aborted()) {
4352         gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4353                                _worker_id);
4354       }
4355     }
4356   } else {
4357     if (_cm->verbose_low()) {
4358       gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4359                              "elapsed = %1.2lfms <<<<<<<<<<",
4360                              _worker_id, _time_target_ms, elapsed_time_ms);
4361     }
4362   }
4363 
4364   _claimed = false;
4365 }
4366 
4367 CMTask::CMTask(uint worker_id,
4368                ConcurrentMark* cm,
4369                size_t* marked_bytes,
4370                BitMap* card_bm,
4371                CMTaskQueue* task_queue,
4372                CMTaskQueueSet* task_queues)
4373   : _g1h(G1CollectedHeap::heap()),
4374     _worker_id(worker_id), _cm(cm),
4375     _claimed(false),
4376     _nextMarkBitMap(NULL), _hash_seed(17),
4377     _task_queue(task_queue),
4378     _task_queues(task_queues),
4379     _cm_oop_closure(NULL),
4380     _marked_bytes_array(marked_bytes),
4381     _card_bm(card_bm) {
4382   guarantee(task_queue != NULL, "invariant");
4383   guarantee(task_queues != NULL, "invariant");
4384 
4385   statsOnly( _clock_due_to_scanning = 0;
4386              _clock_due_to_marking  = 0 );
4387 
4388   _marking_step_diffs_ms.add(0.5);
4389 }
4390 
4391 // These are formatting macros that are used below to ensure
4392 // consistent formatting. The *_H_* versions are used to format the
4393 // header for a particular value and they should be kept consistent
4394 // with the corresponding macro. Also note that most of the macros add
4395 // the necessary white space (as a prefix) which makes them a bit
4396 // easier to compose.
4397 
4398 // All the output lines are prefixed with this string to be able to
4399 // identify them easily in a large log file.
4400 #define G1PPRL_LINE_PREFIX            "###"
4401 
4402 #define G1PPRL_ADDR_BASE_FORMAT    " "PTR_FORMAT"-"PTR_FORMAT
4403 #ifdef _LP64
4404 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
4405 #else // _LP64
4406 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
4407 #endif // _LP64
4408 
4409 // For per-region info
4410 #define G1PPRL_TYPE_FORMAT            "   %-4s"
4411 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
4412 #define G1PPRL_BYTE_FORMAT            "  "SIZE_FORMAT_W(9)
4413 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
4414 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
4415 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
4416 
4417 // For summary info
4418 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  "tag":"G1PPRL_ADDR_BASE_FORMAT
4419 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  "tag": "SIZE_FORMAT
4420 #define G1PPRL_SUM_MB_FORMAT(tag)      "  "tag": %1.2f MB"
4421 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4422 
4423 G1PrintRegionLivenessInfoClosure::
4424 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4425   : _out(out),
4426     _total_used_bytes(0), _total_capacity_bytes(0),
4427     _total_prev_live_bytes(0), _total_next_live_bytes(0),
4428     _hum_used_bytes(0), _hum_capacity_bytes(0),
4429     _hum_prev_live_bytes(0), _hum_next_live_bytes(0) {
4430   G1CollectedHeap* g1h = G1CollectedHeap::heap();
4431   MemRegion g1_committed = g1h->g1_committed();
4432   MemRegion g1_reserved = g1h->g1_reserved();
4433   double now = os::elapsedTime();
4434 
4435   // Print the header of the output.
4436   _out->cr();
4437   _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4438   _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4439                  G1PPRL_SUM_ADDR_FORMAT("committed")
4440                  G1PPRL_SUM_ADDR_FORMAT("reserved")
4441                  G1PPRL_SUM_BYTE_FORMAT("region-size"),
4442                  g1_committed.start(), g1_committed.end(),
4443                  g1_reserved.start(), g1_reserved.end(),
4444                  HeapRegion::GrainBytes);
4445   _out->print_cr(G1PPRL_LINE_PREFIX);
4446   _out->print_cr(G1PPRL_LINE_PREFIX
4447                  G1PPRL_TYPE_H_FORMAT
4448                  G1PPRL_ADDR_BASE_H_FORMAT
4449                  G1PPRL_BYTE_H_FORMAT
4450                  G1PPRL_BYTE_H_FORMAT
4451                  G1PPRL_BYTE_H_FORMAT
4452                  G1PPRL_DOUBLE_H_FORMAT,
4453                  "type", "address-range",
4454                  "used", "prev-live", "next-live", "gc-eff");
4455   _out->print_cr(G1PPRL_LINE_PREFIX
4456                  G1PPRL_TYPE_H_FORMAT
4457                  G1PPRL_ADDR_BASE_H_FORMAT
4458                  G1PPRL_BYTE_H_FORMAT
4459                  G1PPRL_BYTE_H_FORMAT
4460                  G1PPRL_BYTE_H_FORMAT
4461                  G1PPRL_DOUBLE_H_FORMAT,
4462                  "", "",
4463                  "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)");
4464 }
4465 
4466 // It takes as a parameter a reference to one of the _hum_* fields, it
4467 // deduces the corresponding value for a region in a humongous region
4468 // series (either the region size, or what's left if the _hum_* field
4469 // is < the region size), and updates the _hum_* field accordingly.
4470 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4471   size_t bytes = 0;
4472   // The > 0 check is to deal with the prev and next live bytes which
4473   // could be 0.
4474   if (*hum_bytes > 0) {
4475     bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4476     *hum_bytes -= bytes;
4477   }
4478   return bytes;
4479 }
4480 
4481 // It deduces the values for a region in a humongous region series
4482 // from the _hum_* fields and updates those accordingly. It assumes
4483 // that that _hum_* fields have already been set up from the "starts
4484 // humongous" region and we visit the regions in address order.
4485 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4486                                                      size_t* capacity_bytes,
4487                                                      size_t* prev_live_bytes,
4488                                                      size_t* next_live_bytes) {
4489   assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4490   *used_bytes      = get_hum_bytes(&_hum_used_bytes);
4491   *capacity_bytes  = get_hum_bytes(&_hum_capacity_bytes);
4492   *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4493   *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4494 }
4495 
4496 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4497   const char* type = "";
4498   HeapWord* bottom       = r->bottom();
4499   HeapWord* end          = r->end();
4500   size_t capacity_bytes  = r->capacity();
4501   size_t used_bytes      = r->used();
4502   size_t prev_live_bytes = r->live_bytes();
4503   size_t next_live_bytes = r->next_live_bytes();
4504   double gc_eff          = r->gc_efficiency();
4505   if (r->used() == 0) {
4506     type = "FREE";
4507   } else if (r->is_survivor()) {
4508     type = "SURV";
4509   } else if (r->is_young()) {
4510     type = "EDEN";
4511   } else if (r->startsHumongous()) {
4512     type = "HUMS";
4513 
4514     assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4515            _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4516            "they should have been zeroed after the last time we used them");
4517     // Set up the _hum_* fields.
4518     _hum_capacity_bytes  = capacity_bytes;
4519     _hum_used_bytes      = used_bytes;
4520     _hum_prev_live_bytes = prev_live_bytes;
4521     _hum_next_live_bytes = next_live_bytes;
4522     get_hum_bytes(&used_bytes, &capacity_bytes,
4523                   &prev_live_bytes, &next_live_bytes);
4524     end = bottom + HeapRegion::GrainWords;
4525   } else if (r->continuesHumongous()) {
4526     type = "HUMC";
4527     get_hum_bytes(&used_bytes, &capacity_bytes,
4528                   &prev_live_bytes, &next_live_bytes);
4529     assert(end == bottom + HeapRegion::GrainWords, "invariant");
4530   } else {
4531     type = "OLD";
4532   }
4533 
4534   _total_used_bytes      += used_bytes;
4535   _total_capacity_bytes  += capacity_bytes;
4536   _total_prev_live_bytes += prev_live_bytes;
4537   _total_next_live_bytes += next_live_bytes;
4538 
4539   // Print a line for this particular region.
4540   _out->print_cr(G1PPRL_LINE_PREFIX
4541                  G1PPRL_TYPE_FORMAT
4542                  G1PPRL_ADDR_BASE_FORMAT
4543                  G1PPRL_BYTE_FORMAT
4544                  G1PPRL_BYTE_FORMAT
4545                  G1PPRL_BYTE_FORMAT
4546                  G1PPRL_DOUBLE_FORMAT,
4547                  type, bottom, end,
4548                  used_bytes, prev_live_bytes, next_live_bytes, gc_eff);
4549 
4550   return false;
4551 }
4552 
4553 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4554   // Print the footer of the output.
4555   _out->print_cr(G1PPRL_LINE_PREFIX);
4556   _out->print_cr(G1PPRL_LINE_PREFIX
4557                  " SUMMARY"
4558                  G1PPRL_SUM_MB_FORMAT("capacity")
4559                  G1PPRL_SUM_MB_PERC_FORMAT("used")
4560                  G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4561                  G1PPRL_SUM_MB_PERC_FORMAT("next-live"),
4562                  bytes_to_mb(_total_capacity_bytes),
4563                  bytes_to_mb(_total_used_bytes),
4564                  perc(_total_used_bytes, _total_capacity_bytes),
4565                  bytes_to_mb(_total_prev_live_bytes),
4566                  perc(_total_prev_live_bytes, _total_capacity_bytes),
4567                  bytes_to_mb(_total_next_live_bytes),
4568                  perc(_total_next_live_bytes, _total_capacity_bytes));
4569   _out->cr();
4570 }