1 /*
   2  * Copyright (c) 2001, 2014, 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 "code/codeCache.hpp"
  28 #include "gc_implementation/g1/concurrentMark.inline.hpp"
  29 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
  30 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
  31 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
  32 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
  33 #include "gc_implementation/g1/g1Log.hpp"
  34 #include "gc_implementation/g1/g1OopClosures.inline.hpp"
  35 #include "gc_implementation/g1/g1RemSet.hpp"
  36 #include "gc_implementation/g1/heapRegion.inline.hpp"
  37 #include "gc_implementation/g1/heapRegionManager.inline.hpp"
  38 #include "gc_implementation/g1/heapRegionRemSet.hpp"
  39 #include "gc_implementation/g1/heapRegionSet.inline.hpp"
  40 #include "gc_implementation/shared/vmGCOperations.hpp"
  41 #include "gc_implementation/shared/gcTimer.hpp"
  42 #include "gc_implementation/shared/gcTrace.hpp"
  43 #include "gc_implementation/shared/gcTraceTime.hpp"
  44 #include "memory/allocation.hpp"
  45 #include "memory/genOopClosures.inline.hpp"
  46 #include "memory/referencePolicy.hpp"
  47 #include "memory/resourceArea.hpp"
  48 #include "oops/oop.inline.hpp"
  49 #include "runtime/handles.inline.hpp"
  50 #include "runtime/java.hpp"
  51 #include "runtime/atomic.inline.hpp"
  52 #include "runtime/prefetch.inline.hpp"
  53 #include "services/memTracker.hpp"
  54 
  55 // Concurrent marking bit map wrapper
  56 
  57 CMBitMapRO::CMBitMapRO(int shifter) :
  58   _bm(),
  59   _shifter(shifter) {
  60   _bmStartWord = 0;
  61   _bmWordSize = 0;
  62 }
  63 
  64 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
  65                                                const HeapWord* limit) const {
  66   // First we must round addr *up* to a possible object boundary.
  67   addr = (HeapWord*)align_size_up((intptr_t)addr,
  68                                   HeapWordSize << _shifter);
  69   size_t addrOffset = heapWordToOffset(addr);
  70   if (limit == NULL) {
  71     limit = _bmStartWord + _bmWordSize;
  72   }
  73   size_t limitOffset = heapWordToOffset(limit);
  74   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  75   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  76   assert(nextAddr >= addr, "get_next_one postcondition");
  77   assert(nextAddr == limit || isMarked(nextAddr),
  78          "get_next_one postcondition");
  79   return nextAddr;
  80 }
  81 
  82 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr,
  83                                                  const HeapWord* limit) const {
  84   size_t addrOffset = heapWordToOffset(addr);
  85   if (limit == NULL) {
  86     limit = _bmStartWord + _bmWordSize;
  87   }
  88   size_t limitOffset = heapWordToOffset(limit);
  89   size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset);
  90   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  91   assert(nextAddr >= addr, "get_next_one postcondition");
  92   assert(nextAddr == limit || !isMarked(nextAddr),
  93          "get_next_one postcondition");
  94   return nextAddr;
  95 }
  96 
  97 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const {
  98   assert((diff & ((1 << _shifter) - 1)) == 0, "argument check");
  99   return (int) (diff >> _shifter);
 100 }
 101 
 102 #ifndef PRODUCT
 103 bool CMBitMapRO::covers(MemRegion heap_rs) const {
 104   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
 105   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
 106          "size inconsistency");
 107   return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
 108          _bmWordSize  == heap_rs.word_size();
 109 }
 110 #endif
 111 
 112 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
 113   _bm.print_on_error(st, prefix);
 114 }
 115 
 116 size_t CMBitMap::compute_size(size_t heap_size) {
 117   return heap_size / mark_distance();
 118 }
 119 
 120 size_t CMBitMap::mark_distance() {
 121   return MinObjAlignmentInBytes * BitsPerByte;
 122 }
 123 
 124 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
 125   _bmStartWord = heap.start();
 126   _bmWordSize = heap.word_size();
 127 
 128   _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
 129   _bm.set_size(_bmWordSize >> _shifter);
 130 
 131   storage->set_mapping_changed_listener(&_listener);
 132 }
 133 
 134 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions) {
 135   // We need to clear the bitmap on commit, removing any existing information.
 136   MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
 137   _bm->clearRange(mr);
 138 }
 139 
 140 // Closure used for clearing the given mark bitmap.
 141 class ClearBitmapHRClosure : public HeapRegionClosure {
 142  private:
 143   ConcurrentMark* _cm;
 144   CMBitMap* _bitmap;
 145   bool _may_yield;      // The closure may yield during iteration. If yielded, abort the iteration.
 146  public:
 147   ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
 148     assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
 149   }
 150 
 151   virtual bool doHeapRegion(HeapRegion* r) {
 152     size_t const chunk_size_in_words = M / HeapWordSize;
 153 
 154     HeapWord* cur = r->bottom();
 155     HeapWord* const end = r->end();
 156 
 157     while (cur < end) {
 158       MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 159       _bitmap->clearRange(mr);
 160 
 161       cur += chunk_size_in_words;
 162 
 163       // Abort iteration if after yielding the marking has been aborted.
 164       if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
 165         return true;
 166       }
 167       // Repeat the asserts from before the start of the closure. We will do them
 168       // as asserts here to minimize their overhead on the product. However, we
 169       // will have them as guarantees at the beginning / end of the bitmap
 170       // clearing to get some checking in the product.
 171       assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
 172       assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant");
 173     }
 174 
 175     return false;
 176   }
 177 };
 178 
 179 void CMBitMap::clearAll() {
 180   ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
 181   G1CollectedHeap::heap()->heap_region_iterate(&cl);
 182   guarantee(cl.complete(), "Must have completed iteration.");
 183   return;
 184 }
 185 
 186 void CMBitMap::markRange(MemRegion mr) {
 187   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 188   assert(!mr.is_empty(), "unexpected empty region");
 189   assert((offsetToHeapWord(heapWordToOffset(mr.end())) ==
 190           ((HeapWord *) mr.end())),
 191          "markRange memory region end is not card aligned");
 192   // convert address range into offset range
 193   _bm.at_put_range(heapWordToOffset(mr.start()),
 194                    heapWordToOffset(mr.end()), true);
 195 }
 196 
 197 void CMBitMap::clearRange(MemRegion mr) {
 198   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 199   assert(!mr.is_empty(), "unexpected empty region");
 200   // convert address range into offset range
 201   _bm.at_put_range(heapWordToOffset(mr.start()),
 202                    heapWordToOffset(mr.end()), false);
 203 }
 204 
 205 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr,
 206                                             HeapWord* end_addr) {
 207   HeapWord* start = getNextMarkedWordAddress(addr);
 208   start = MIN2(start, end_addr);
 209   HeapWord* end   = getNextUnmarkedWordAddress(start);
 210   end = MIN2(end, end_addr);
 211   assert(start <= end, "Consistency check");
 212   MemRegion mr(start, end);
 213   if (!mr.is_empty()) {
 214     clearRange(mr);
 215   }
 216   return mr;
 217 }
 218 
 219 CMMarkStack::CMMarkStack(ConcurrentMark* cm) :
 220   _base(NULL), _cm(cm)
 221 #ifdef ASSERT
 222   , _drain_in_progress(false)
 223   , _drain_in_progress_yields(false)
 224 #endif
 225 {}
 226 
 227 bool CMMarkStack::allocate(size_t capacity) {
 228   // allocate a stack of the requisite depth
 229   ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
 230   if (!rs.is_reserved()) {
 231     warning("ConcurrentMark MarkStack allocation failure");
 232     return false;
 233   }
 234   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 235   if (!_virtual_space.initialize(rs, rs.size())) {
 236     warning("ConcurrentMark MarkStack backing store failure");
 237     // Release the virtual memory reserved for the marking stack
 238     rs.release();
 239     return false;
 240   }
 241   assert(_virtual_space.committed_size() == rs.size(),
 242          "Didn't reserve backing store for all of ConcurrentMark stack?");
 243   _base = (oop*) _virtual_space.low();
 244   setEmpty();
 245   _capacity = (jint) capacity;
 246   _saved_index = -1;
 247   _should_expand = false;
 248   NOT_PRODUCT(_max_depth = 0);
 249   return true;
 250 }
 251 
 252 void CMMarkStack::expand() {
 253   // Called, during remark, if we've overflown the marking stack during marking.
 254   assert(isEmpty(), "stack should been emptied while handling overflow");
 255   assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
 256   // Clear expansion flag
 257   _should_expand = false;
 258   if (_capacity == (jint) MarkStackSizeMax) {
 259     if (PrintGCDetails && Verbose) {
 260       gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit");
 261     }
 262     return;
 263   }
 264   // Double capacity if possible
 265   jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
 266   // Do not give up existing stack until we have managed to
 267   // get the double capacity that we desired.
 268   ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
 269                                                            sizeof(oop)));
 270   if (rs.is_reserved()) {
 271     // Release the backing store associated with old stack
 272     _virtual_space.release();
 273     // Reinitialize virtual space for new stack
 274     if (!_virtual_space.initialize(rs, rs.size())) {
 275       fatal("Not enough swap for expanded marking stack capacity");
 276     }
 277     _base = (oop*)(_virtual_space.low());
 278     _index = 0;
 279     _capacity = new_capacity;
 280   } else {
 281     if (PrintGCDetails && Verbose) {
 282       // Failed to double capacity, continue;
 283       gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from "
 284                           SIZE_FORMAT"K to " SIZE_FORMAT"K",
 285                           _capacity / K, new_capacity / K);
 286     }
 287   }
 288 }
 289 
 290 void CMMarkStack::set_should_expand() {
 291   // If we're resetting the marking state because of an
 292   // marking stack overflow, record that we should, if
 293   // possible, expand the stack.
 294   _should_expand = _cm->has_overflown();
 295 }
 296 
 297 CMMarkStack::~CMMarkStack() {
 298   if (_base != NULL) {
 299     _base = NULL;
 300     _virtual_space.release();
 301   }
 302 }
 303 
 304 void CMMarkStack::par_push(oop ptr) {
 305   while (true) {
 306     if (isFull()) {
 307       _overflow = true;
 308       return;
 309     }
 310     // Otherwise...
 311     jint index = _index;
 312     jint next_index = index+1;
 313     jint res = Atomic::cmpxchg(next_index, &_index, index);
 314     if (res == index) {
 315       _base[index] = ptr;
 316       // Note that we don't maintain this atomically.  We could, but it
 317       // doesn't seem necessary.
 318       NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 319       return;
 320     }
 321     // Otherwise, we need to try again.
 322   }
 323 }
 324 
 325 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) {
 326   while (true) {
 327     if (isFull()) {
 328       _overflow = true;
 329       return;
 330     }
 331     // Otherwise...
 332     jint index = _index;
 333     jint next_index = index + n;
 334     if (next_index > _capacity) {
 335       _overflow = true;
 336       return;
 337     }
 338     jint res = Atomic::cmpxchg(next_index, &_index, index);
 339     if (res == index) {
 340       for (int i = 0; i < n; i++) {
 341         int  ind = index + i;
 342         assert(ind < _capacity, "By overflow test above.");
 343         _base[ind] = ptr_arr[i];
 344       }
 345       NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 346       return;
 347     }
 348     // Otherwise, we need to try again.
 349   }
 350 }
 351 
 352 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 353   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 354   jint start = _index;
 355   jint next_index = start + n;
 356   if (next_index > _capacity) {
 357     _overflow = true;
 358     return;
 359   }
 360   // Otherwise.
 361   _index = next_index;
 362   for (int i = 0; i < n; i++) {
 363     int ind = start + i;
 364     assert(ind < _capacity, "By overflow test above.");
 365     _base[ind] = ptr_arr[i];
 366   }
 367   NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index));
 368 }
 369 
 370 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 371   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 372   jint index = _index;
 373   if (index == 0) {
 374     *n = 0;
 375     return false;
 376   } else {
 377     int k = MIN2(max, index);
 378     jint  new_ind = index - k;
 379     for (int j = 0; j < k; j++) {
 380       ptr_arr[j] = _base[new_ind + j];
 381     }
 382     _index = new_ind;
 383     *n = k;
 384     return true;
 385   }
 386 }
 387 
 388 template<class OopClosureClass>
 389 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) {
 390   assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after
 391          || SafepointSynchronize::is_at_safepoint(),
 392          "Drain recursion must be yield-safe.");
 393   bool res = true;
 394   debug_only(_drain_in_progress = true);
 395   debug_only(_drain_in_progress_yields = yield_after);
 396   while (!isEmpty()) {
 397     oop newOop = pop();
 398     assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop");
 399     assert(newOop->is_oop(), "Expected an oop");
 400     assert(bm == NULL || bm->isMarked((HeapWord*)newOop),
 401            "only grey objects on this stack");
 402     newOop->oop_iterate(cl);
 403     if (yield_after && _cm->do_yield_check()) {
 404       res = false;
 405       break;
 406     }
 407   }
 408   debug_only(_drain_in_progress = false);
 409   return res;
 410 }
 411 
 412 void CMMarkStack::note_start_of_gc() {
 413   assert(_saved_index == -1,
 414          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 415   _saved_index = _index;
 416 }
 417 
 418 void CMMarkStack::note_end_of_gc() {
 419   // This is intentionally a guarantee, instead of an assert. If we
 420   // accidentally add something to the mark stack during GC, it
 421   // will be a correctness issue so it's better if we crash. we'll
 422   // only check this once per GC anyway, so it won't be a performance
 423   // issue in any way.
 424   guarantee(_saved_index == _index,
 425             err_msg("saved index: %d index: %d", _saved_index, _index));
 426   _saved_index = -1;
 427 }
 428 
 429 void CMMarkStack::oops_do(OopClosure* f) {
 430   assert(_saved_index == _index,
 431          err_msg("saved index: %d index: %d", _saved_index, _index));
 432   for (int i = 0; i < _index; i += 1) {
 433     f->do_oop(&_base[i]);
 434   }
 435 }
 436 
 437 CMRootRegions::CMRootRegions() :
 438   _young_list(NULL), _cm(NULL), _scan_in_progress(false),
 439   _should_abort(false),  _next_survivor(NULL) { }
 440 
 441 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) {
 442   _young_list = g1h->young_list();
 443   _cm = cm;
 444 }
 445 
 446 void CMRootRegions::prepare_for_scan() {
 447   assert(!scan_in_progress(), "pre-condition");
 448 
 449   // Currently, only survivors can be root regions.
 450   assert(_next_survivor == NULL, "pre-condition");
 451   _next_survivor = _young_list->first_survivor_region();
 452   _scan_in_progress = (_next_survivor != NULL);
 453   _should_abort = false;
 454 }
 455 
 456 HeapRegion* CMRootRegions::claim_next() {
 457   if (_should_abort) {
 458     // If someone has set the should_abort flag, we return NULL to
 459     // force the caller to bail out of their loop.
 460     return NULL;
 461   }
 462 
 463   // Currently, only survivors can be root regions.
 464   HeapRegion* res = _next_survivor;
 465   if (res != NULL) {
 466     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 467     // Read it again in case it changed while we were waiting for the lock.
 468     res = _next_survivor;
 469     if (res != NULL) {
 470       if (res == _young_list->last_survivor_region()) {
 471         // We just claimed the last survivor so store NULL to indicate
 472         // that we're done.
 473         _next_survivor = NULL;
 474       } else {
 475         _next_survivor = res->get_next_young_region();
 476       }
 477     } else {
 478       // Someone else claimed the last survivor while we were trying
 479       // to take the lock so nothing else to do.
 480     }
 481   }
 482   assert(res == NULL || res->is_survivor(), "post-condition");
 483 
 484   return res;
 485 }
 486 
 487 void CMRootRegions::scan_finished() {
 488   assert(scan_in_progress(), "pre-condition");
 489 
 490   // Currently, only survivors can be root regions.
 491   if (!_should_abort) {
 492     assert(_next_survivor == NULL, "we should have claimed all survivors");
 493   }
 494   _next_survivor = NULL;
 495 
 496   {
 497     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 498     _scan_in_progress = false;
 499     RootRegionScan_lock->notify_all();
 500   }
 501 }
 502 
 503 bool CMRootRegions::wait_until_scan_finished() {
 504   if (!scan_in_progress()) return false;
 505 
 506   {
 507     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 508     while (scan_in_progress()) {
 509       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 510     }
 511   }
 512   return true;
 513 }
 514 
 515 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
 516 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
 517 #endif // _MSC_VER
 518 
 519 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 520   return MAX2((n_par_threads + 2) / 4, 1U);
 521 }
 522 
 523 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
 524   _g1h(g1h),
 525   _markBitMap1(),
 526   _markBitMap2(),
 527   _parallel_marking_threads(0),
 528   _max_parallel_marking_threads(0),
 529   _sleep_factor(0.0),
 530   _marking_task_overhead(1.0),
 531   _cleanup_sleep_factor(0.0),
 532   _cleanup_task_overhead(1.0),
 533   _cleanup_list("Cleanup List"),
 534   _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
 535   _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
 536             CardTableModRefBS::card_shift,
 537             false /* in_resource_area*/),
 538 
 539   _prevMarkBitMap(&_markBitMap1),
 540   _nextMarkBitMap(&_markBitMap2),
 541 
 542   _markStack(this),
 543   // _finger set in set_non_marking_state
 544 
 545   _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)),
 546   // _active_tasks set in set_non_marking_state
 547   // _tasks set inside the constructor
 548   _task_queues(new CMTaskQueueSet((int) _max_worker_id)),
 549   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 550 
 551   _has_overflown(false),
 552   _concurrent(false),
 553   _has_aborted(false),
 554   _aborted_gc_id(GCId::undefined()),
 555   _restart_for_overflow(false),
 556   _concurrent_marking_in_progress(false),
 557 
 558   // _verbose_level set below
 559 
 560   _init_times(),
 561   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 562   _cleanup_times(),
 563   _total_counting_time(0.0),
 564   _total_rs_scrub_time(0.0),
 565 
 566   _parallel_workers(NULL),
 567 
 568   _count_card_bitmaps(NULL),
 569   _count_marked_bytes(NULL),
 570   _completed_initialization(false) {
 571   CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel;
 572   if (verbose_level < no_verbose) {
 573     verbose_level = no_verbose;
 574   }
 575   if (verbose_level > high_verbose) {
 576     verbose_level = high_verbose;
 577   }
 578   _verbose_level = verbose_level;
 579 
 580   if (verbose_low()) {
 581     gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", "
 582                            "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end));
 583   }
 584 
 585   _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 586   _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
 587 
 588   // Create & start a ConcurrentMark thread.
 589   _cmThread = new ConcurrentMarkThread(this);
 590   assert(cmThread() != NULL, "CM Thread should have been created");
 591   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 592   if (_cmThread->osthread() == NULL) {
 593       vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 594   }
 595 
 596   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 597   assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
 598   assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
 599 
 600   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 601   satb_qs.set_buffer_size(G1SATBBufferSize);
 602 
 603   _root_regions.init(_g1h, this);
 604 
 605   if (ConcGCThreads > ParallelGCThreads) {
 606     warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") "
 607             "than ParallelGCThreads (" UINTX_FORMAT ").",
 608             ConcGCThreads, ParallelGCThreads);
 609     return;
 610   }
 611   if (ParallelGCThreads == 0) {
 612     // if we are not running with any parallel GC threads we will not
 613     // spawn any marking threads either
 614     _parallel_marking_threads =       0;
 615     _max_parallel_marking_threads =   0;
 616     _sleep_factor             =     0.0;
 617     _marking_task_overhead    =     1.0;
 618   } else {
 619     if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
 620       // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
 621       // if both are set
 622       _sleep_factor             = 0.0;
 623       _marking_task_overhead    = 1.0;
 624     } else if (G1MarkingOverheadPercent > 0) {
 625       // We will calculate the number of parallel marking threads based
 626       // on a target overhead with respect to the soft real-time goal
 627       double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 628       double overall_cm_overhead =
 629         (double) MaxGCPauseMillis * marking_overhead /
 630         (double) GCPauseIntervalMillis;
 631       double cpu_ratio = 1.0 / (double) os::processor_count();
 632       double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 633       double marking_task_overhead =
 634         overall_cm_overhead / marking_thread_num *
 635                                                 (double) os::processor_count();
 636       double sleep_factor =
 637                          (1.0 - marking_task_overhead) / marking_task_overhead;
 638 
 639       FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num);
 640       _sleep_factor             = sleep_factor;
 641       _marking_task_overhead    = marking_task_overhead;
 642     } else {
 643       // Calculate the number of parallel marking threads by scaling
 644       // the number of parallel GC threads.
 645       uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads);
 646       FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num);
 647       _sleep_factor             = 0.0;
 648       _marking_task_overhead    = 1.0;
 649     }
 650 
 651     assert(ConcGCThreads > 0, "Should have been set");
 652     _parallel_marking_threads = (uint) ConcGCThreads;
 653     _max_parallel_marking_threads = _parallel_marking_threads;
 654 
 655     if (parallel_marking_threads() > 1) {
 656       _cleanup_task_overhead = 1.0;
 657     } else {
 658       _cleanup_task_overhead = marking_task_overhead();
 659     }
 660     _cleanup_sleep_factor =
 661                      (1.0 - cleanup_task_overhead()) / cleanup_task_overhead();
 662 
 663 #if 0
 664     gclog_or_tty->print_cr("Marking Threads          %d", parallel_marking_threads());
 665     gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead());
 666     gclog_or_tty->print_cr("CM Sleep Factor          %1.4lf", sleep_factor());
 667     gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead());
 668     gclog_or_tty->print_cr("CL Sleep Factor          %1.4lf", cleanup_sleep_factor());
 669 #endif
 670 
 671     guarantee(parallel_marking_threads() > 0, "peace of mind");
 672     _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
 673          _max_parallel_marking_threads, false, true);
 674     if (_parallel_workers == NULL) {
 675       vm_exit_during_initialization("Failed necessary allocation.");
 676     } else {
 677       _parallel_workers->initialize_workers();
 678     }
 679   }
 680 
 681   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 682     uintx mark_stack_size =
 683       MIN2(MarkStackSizeMax,
 684           MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 685     // Verify that the calculated value for MarkStackSize is in range.
 686     // It would be nice to use the private utility routine from Arguments.
 687     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 688       warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): "
 689               "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
 690               mark_stack_size, (uintx) 1, MarkStackSizeMax);
 691       return;
 692     }
 693     FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size);
 694   } else {
 695     // Verify MarkStackSize is in range.
 696     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 697       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 698         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 699           warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): "
 700                   "must be between " UINTX_FORMAT " and " UINTX_FORMAT,
 701                   MarkStackSize, (uintx) 1, MarkStackSizeMax);
 702           return;
 703         }
 704       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 705         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 706           warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")"
 707                   " or for MarkStackSizeMax (" UINTX_FORMAT ")",
 708                   MarkStackSize, MarkStackSizeMax);
 709           return;
 710         }
 711       }
 712     }
 713   }
 714 
 715   if (!_markStack.allocate(MarkStackSize)) {
 716     warning("Failed to allocate CM marking stack");
 717     return;
 718   }
 719 
 720   _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC);
 721   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 722 
 723   _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap,  _max_worker_id, mtGC);
 724   _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
 725 
 726   BitMap::idx_t card_bm_size = _card_bm.size();
 727 
 728   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 729   _active_tasks = _max_worker_id;
 730 
 731   size_t max_regions = (size_t) _g1h->max_regions();
 732   for (uint i = 0; i < _max_worker_id; ++i) {
 733     CMTaskQueue* task_queue = new CMTaskQueue();
 734     task_queue->initialize();
 735     _task_queues->register_queue(i, task_queue);
 736 
 737     _count_card_bitmaps[i] = BitMap(card_bm_size, false);
 738     _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
 739 
 740     _tasks[i] = new CMTask(i, this,
 741                            _count_marked_bytes[i],
 742                            &_count_card_bitmaps[i],
 743                            task_queue, _task_queues);
 744 
 745     _accum_task_vtime[i] = 0.0;
 746   }
 747 
 748   // Calculate the card number for the bottom of the heap. Used
 749   // in biasing indexes into the accounting card bitmaps.
 750   _heap_bottom_card_num =
 751     intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
 752                                 CardTableModRefBS::card_shift);
 753 
 754   // Clear all the liveness counting data
 755   clear_all_count_data();
 756 
 757   // so that the call below can read a sensible value
 758   _heap_start = g1h->reserved_region().start();
 759   set_non_marking_state();
 760   _completed_initialization = true;
 761 }
 762 
 763 void ConcurrentMark::reset() {
 764   // Starting values for these two. This should be called in a STW
 765   // phase.
 766   MemRegion reserved = _g1h->g1_reserved();
 767   _heap_start = reserved.start();
 768   _heap_end   = reserved.end();
 769 
 770   // Separated the asserts so that we know which one fires.
 771   assert(_heap_start != NULL, "heap bounds should look ok");
 772   assert(_heap_end != NULL, "heap bounds should look ok");
 773   assert(_heap_start < _heap_end, "heap bounds should look ok");
 774 
 775   // Reset all the marking data structures and any necessary flags
 776   reset_marking_state();
 777 
 778   if (verbose_low()) {
 779     gclog_or_tty->print_cr("[global] resetting");
 780   }
 781 
 782   // We do reset all of them, since different phases will use
 783   // different number of active threads. So, it's easiest to have all
 784   // of them ready.
 785   for (uint i = 0; i < _max_worker_id; ++i) {
 786     _tasks[i]->reset(_nextMarkBitMap);
 787   }
 788 
 789   // we need this to make sure that the flag is on during the evac
 790   // pause with initial mark piggy-backed
 791   set_concurrent_marking_in_progress();
 792 }
 793 
 794 
 795 void ConcurrentMark::reset_marking_state(bool clear_overflow) {
 796   _markStack.set_should_expand();
 797   _markStack.setEmpty();        // Also clears the _markStack overflow flag
 798   if (clear_overflow) {
 799     clear_has_overflown();
 800   } else {
 801     assert(has_overflown(), "pre-condition");
 802   }
 803   _finger = _heap_start;
 804 
 805   for (uint i = 0; i < _max_worker_id; ++i) {
 806     CMTaskQueue* queue = _task_queues->queue(i);
 807     queue->set_empty();
 808   }
 809 }
 810 
 811 void ConcurrentMark::set_concurrency(uint active_tasks) {
 812   assert(active_tasks <= _max_worker_id, "we should not have more");
 813 
 814   _active_tasks = active_tasks;
 815   // Need to update the three data structures below according to the
 816   // number of active threads for this phase.
 817   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 818   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 819   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 820 }
 821 
 822 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 823   set_concurrency(active_tasks);
 824 
 825   _concurrent = concurrent;
 826   // We propagate this to all tasks, not just the active ones.
 827   for (uint i = 0; i < _max_worker_id; ++i)
 828     _tasks[i]->set_concurrent(concurrent);
 829 
 830   if (concurrent) {
 831     set_concurrent_marking_in_progress();
 832   } else {
 833     // We currently assume that the concurrent flag has been set to
 834     // false before we start remark. At this point we should also be
 835     // in a STW phase.
 836     assert(!concurrent_marking_in_progress(), "invariant");
 837     assert(out_of_regions(),
 838            err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT,
 839                    p2i(_finger), p2i(_heap_end)));
 840   }
 841 }
 842 
 843 void ConcurrentMark::set_non_marking_state() {
 844   // We set the global marking state to some default values when we're
 845   // not doing marking.
 846   reset_marking_state();
 847   _active_tasks = 0;
 848   clear_concurrent_marking_in_progress();
 849 }
 850 
 851 ConcurrentMark::~ConcurrentMark() {
 852   // The ConcurrentMark instance is never freed.
 853   ShouldNotReachHere();
 854 }
 855 
 856 void ConcurrentMark::clearNextBitmap() {
 857   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 858 
 859   // Make sure that the concurrent mark thread looks to still be in
 860   // the current cycle.
 861   guarantee(cmThread()->during_cycle(), "invariant");
 862 
 863   // We are finishing up the current cycle by clearing the next
 864   // marking bitmap and getting it ready for the next cycle. During
 865   // this time no other cycle can start. So, let's make sure that this
 866   // is the case.
 867   guarantee(!g1h->mark_in_progress(), "invariant");
 868 
 869   ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
 870   g1h->heap_region_iterate(&cl);
 871 
 872   // Clear the liveness counting data. If the marking has been aborted, the abort()
 873   // call already did that.
 874   if (cl.complete()) {
 875     clear_all_count_data();
 876   }
 877 
 878   // Repeat the asserts from above.
 879   guarantee(cmThread()->during_cycle(), "invariant");
 880   guarantee(!g1h->mark_in_progress(), "invariant");
 881 }
 882 
 883 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 884   CMBitMap* _bitmap;
 885   bool _error;
 886  public:
 887   CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) {
 888   }
 889 
 890   virtual bool doHeapRegion(HeapRegion* r) {
 891     return _bitmap->getNextMarkedWordAddress(r->bottom(), r->end()) != r->end();
 892   }
 893 };
 894 
 895 bool ConcurrentMark::nextMarkBitmapIsClear() {
 896   CheckBitmapClearHRClosure cl(_nextMarkBitMap);
 897   _g1h->heap_region_iterate(&cl);
 898   return cl.complete();
 899 }
 900 
 901 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 902 public:
 903   bool doHeapRegion(HeapRegion* r) {
 904     if (!r->continuesHumongous()) {
 905       r->note_start_of_marking();
 906     }
 907     return false;
 908   }
 909 };
 910 
 911 void ConcurrentMark::checkpointRootsInitialPre() {
 912   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 913   G1CollectorPolicy* g1p = g1h->g1_policy();
 914 
 915   _has_aborted = false;
 916 
 917 #ifndef PRODUCT
 918   if (G1PrintReachableAtInitialMark) {
 919     print_reachable("at-cycle-start",
 920                     VerifyOption_G1UsePrevMarking, true /* all */);
 921   }
 922 #endif
 923 
 924   // Initialize marking structures. This has to be done in a STW phase.
 925   reset();
 926 
 927   // For each region note start of marking.
 928   NoteStartOfMarkHRClosure startcl;
 929   g1h->heap_region_iterate(&startcl);
 930 }
 931 
 932 
 933 void ConcurrentMark::checkpointRootsInitialPost() {
 934   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 935 
 936   // If we force an overflow during remark, the remark operation will
 937   // actually abort and we'll restart concurrent marking. If we always
 938   // force an overflow during remark we'll never actually complete the
 939   // marking phase. So, we initialize this here, at the start of the
 940   // cycle, so that at the remaining overflow number will decrease at
 941   // every remark and we'll eventually not need to cause one.
 942   force_overflow_stw()->init();
 943 
 944   // Start Concurrent Marking weak-reference discovery.
 945   ReferenceProcessor* rp = g1h->ref_processor_cm();
 946   // enable ("weak") refs discovery
 947   rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/);
 948   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 949 
 950   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 951   // This is the start of  the marking cycle, we're expected all
 952   // threads to have SATB queues with active set to false.
 953   satb_mq_set.set_active_all_threads(true, /* new active value */
 954                                      false /* expected_active */);
 955 
 956   _root_regions.prepare_for_scan();
 957 
 958   // update_g1_committed() will be called at the end of an evac pause
 959   // when marking is on. So, it's also called at the end of the
 960   // initial-mark pause to update the heap end, if the heap expands
 961   // during it. No need to call it here.
 962 }
 963 
 964 /*
 965  * Notice that in the next two methods, we actually leave the STS
 966  * during the barrier sync and join it immediately afterwards. If we
 967  * do not do this, the following deadlock can occur: one thread could
 968  * be in the barrier sync code, waiting for the other thread to also
 969  * sync up, whereas another one could be trying to yield, while also
 970  * waiting for the other threads to sync up too.
 971  *
 972  * Note, however, that this code is also used during remark and in
 973  * this case we should not attempt to leave / enter the STS, otherwise
 974  * we'll either hit an assert (debug / fastdebug) or deadlock
 975  * (product). So we should only leave / enter the STS if we are
 976  * operating concurrently.
 977  *
 978  * Because the thread that does the sync barrier has left the STS, it
 979  * is possible to be suspended for a Full GC or an evacuation pause
 980  * could occur. This is actually safe, since the entering the sync
 981  * barrier is one of the last things do_marking_step() does, and it
 982  * doesn't manipulate any data structures afterwards.
 983  */
 984 
 985 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 986   if (verbose_low()) {
 987     gclog_or_tty->print_cr("[%u] entering first barrier", worker_id);
 988   }
 989 
 990   if (concurrent()) {
 991     SuspendibleThreadSet::leave();
 992   }
 993 
 994   bool barrier_aborted = !_first_overflow_barrier_sync.enter();
 995 
 996   if (concurrent()) {
 997     SuspendibleThreadSet::join();
 998   }
 999   // at this point everyone should have synced up and not be doing any
1000   // more work
1001 
1002   if (verbose_low()) {
1003     if (barrier_aborted) {
1004       gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id);
1005     } else {
1006       gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id);
1007     }
1008   }
1009 
1010   if (barrier_aborted) {
1011     // If the barrier aborted we ignore the overflow condition and
1012     // just abort the whole marking phase as quickly as possible.
1013     return;
1014   }
1015 
1016   // If we're executing the concurrent phase of marking, reset the marking
1017   // state; otherwise the marking state is reset after reference processing,
1018   // during the remark pause.
1019   // If we reset here as a result of an overflow during the remark we will
1020   // see assertion failures from any subsequent set_concurrency_and_phase()
1021   // calls.
1022   if (concurrent()) {
1023     // let the task associated with with worker 0 do this
1024     if (worker_id == 0) {
1025       // task 0 is responsible for clearing the global data structures
1026       // We should be here because of an overflow. During STW we should
1027       // not clear the overflow flag since we rely on it being true when
1028       // we exit this method to abort the pause and restart concurrent
1029       // marking.
1030       reset_marking_state(true /* clear_overflow */);
1031       force_overflow()->update();
1032 
1033       if (G1Log::fine()) {
1034         gclog_or_tty->gclog_stamp(concurrent_gc_id());
1035         gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]");
1036       }
1037     }
1038   }
1039 
1040   // after this, each task should reset its own data structures then
1041   // then go into the second barrier
1042 }
1043 
1044 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
1045   if (verbose_low()) {
1046     gclog_or_tty->print_cr("[%u] entering second barrier", worker_id);
1047   }
1048 
1049   if (concurrent()) {
1050     SuspendibleThreadSet::leave();
1051   }
1052 
1053   bool barrier_aborted = !_second_overflow_barrier_sync.enter();
1054 
1055   if (concurrent()) {
1056     SuspendibleThreadSet::join();
1057   }
1058   // at this point everything should be re-initialized and ready to go
1059 
1060   if (verbose_low()) {
1061     if (barrier_aborted) {
1062       gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id);
1063     } else {
1064       gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id);
1065     }
1066   }
1067 }
1068 
1069 #ifndef PRODUCT
1070 void ForceOverflowSettings::init() {
1071   _num_remaining = G1ConcMarkForceOverflow;
1072   _force = false;
1073   update();
1074 }
1075 
1076 void ForceOverflowSettings::update() {
1077   if (_num_remaining > 0) {
1078     _num_remaining -= 1;
1079     _force = true;
1080   } else {
1081     _force = false;
1082   }
1083 }
1084 
1085 bool ForceOverflowSettings::should_force() {
1086   if (_force) {
1087     _force = false;
1088     return true;
1089   } else {
1090     return false;
1091   }
1092 }
1093 #endif // !PRODUCT
1094 
1095 class CMConcurrentMarkingTask: public AbstractGangTask {
1096 private:
1097   ConcurrentMark*       _cm;
1098   ConcurrentMarkThread* _cmt;
1099 
1100 public:
1101   void work(uint worker_id) {
1102     assert(Thread::current()->is_ConcurrentGC_thread(),
1103            "this should only be done by a conc GC thread");
1104     ResourceMark rm;
1105 
1106     double start_vtime = os::elapsedVTime();
1107 
1108     SuspendibleThreadSet::join();
1109 
1110     assert(worker_id < _cm->active_tasks(), "invariant");
1111     CMTask* the_task = _cm->task(worker_id);
1112     the_task->record_start_time();
1113     if (!_cm->has_aborted()) {
1114       do {
1115         double start_vtime_sec = os::elapsedVTime();
1116         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1117 
1118         the_task->do_marking_step(mark_step_duration_ms,
1119                                   true  /* do_termination */,
1120                                   false /* is_serial*/);
1121 
1122         double end_vtime_sec = os::elapsedVTime();
1123         double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
1124         _cm->clear_has_overflown();
1125 
1126         _cm->do_yield_check(worker_id);
1127 
1128         jlong sleep_time_ms;
1129         if (!_cm->has_aborted() && the_task->has_aborted()) {
1130           sleep_time_ms =
1131             (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
1132           SuspendibleThreadSet::leave();
1133           os::sleep(Thread::current(), sleep_time_ms, false);
1134           SuspendibleThreadSet::join();
1135         }
1136       } while (!_cm->has_aborted() && the_task->has_aborted());
1137     }
1138     the_task->record_end_time();
1139     guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
1140 
1141     SuspendibleThreadSet::leave();
1142 
1143     double end_vtime = os::elapsedVTime();
1144     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
1145   }
1146 
1147   CMConcurrentMarkingTask(ConcurrentMark* cm,
1148                           ConcurrentMarkThread* cmt) :
1149       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
1150 
1151   ~CMConcurrentMarkingTask() { }
1152 };
1153 
1154 // Calculates the number of active workers for a concurrent
1155 // phase.
1156 uint ConcurrentMark::calc_parallel_marking_threads() {
1157   if (G1CollectedHeap::use_parallel_gc_threads()) {
1158     uint n_conc_workers = 0;
1159     if (!UseDynamicNumberOfGCThreads ||
1160         (!FLAG_IS_DEFAULT(ConcGCThreads) &&
1161          !ForceDynamicNumberOfGCThreads)) {
1162       n_conc_workers = max_parallel_marking_threads();
1163     } else {
1164       n_conc_workers =
1165         AdaptiveSizePolicy::calc_default_active_workers(
1166                                      max_parallel_marking_threads(),
1167                                      1, /* Minimum workers */
1168                                      parallel_marking_threads(),
1169                                      Threads::number_of_non_daemon_threads());
1170       // Don't scale down "n_conc_workers" by scale_parallel_threads() because
1171       // that scaling has already gone into "_max_parallel_marking_threads".
1172     }
1173     assert(n_conc_workers > 0, "Always need at least 1");
1174     return n_conc_workers;
1175   }
1176   // If we are not running with any parallel GC threads we will not
1177   // have spawned any marking threads either. Hence the number of
1178   // concurrent workers should be 0.
1179   return 0;
1180 }
1181 
1182 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
1183   // Currently, only survivors can be root regions.
1184   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
1185   G1RootRegionScanClosure cl(_g1h, this, worker_id);
1186 
1187   const uintx interval = PrefetchScanIntervalInBytes;
1188   HeapWord* curr = hr->bottom();
1189   const HeapWord* end = hr->top();
1190   while (curr < end) {
1191     Prefetch::read(curr, interval);
1192     oop obj = oop(curr);
1193     int size = obj->oop_iterate(&cl);
1194     assert(size == obj->size(), "sanity");
1195     curr += size;
1196   }
1197 }
1198 
1199 class CMRootRegionScanTask : public AbstractGangTask {
1200 private:
1201   ConcurrentMark* _cm;
1202 
1203 public:
1204   CMRootRegionScanTask(ConcurrentMark* cm) :
1205     AbstractGangTask("Root Region Scan"), _cm(cm) { }
1206 
1207   void work(uint worker_id) {
1208     assert(Thread::current()->is_ConcurrentGC_thread(),
1209            "this should only be done by a conc GC thread");
1210 
1211     CMRootRegions* root_regions = _cm->root_regions();
1212     HeapRegion* hr = root_regions->claim_next();
1213     while (hr != NULL) {
1214       _cm->scanRootRegion(hr, worker_id);
1215       hr = root_regions->claim_next();
1216     }
1217   }
1218 };
1219 
1220 void ConcurrentMark::scanRootRegions() {
1221   // Start of concurrent marking.
1222   ClassLoaderDataGraph::clear_claimed_marks();
1223 
1224   // scan_in_progress() will have been set to true only if there was
1225   // at least one root region to scan. So, if it's false, we
1226   // should not attempt to do any further work.
1227   if (root_regions()->scan_in_progress()) {
1228     _parallel_marking_threads = calc_parallel_marking_threads();
1229     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1230            "Maximum number of marking threads exceeded");
1231     uint active_workers = MAX2(1U, parallel_marking_threads());
1232 
1233     CMRootRegionScanTask task(this);
1234     if (use_parallel_marking_threads()) {
1235       _parallel_workers->set_active_workers((int) active_workers);
1236       _parallel_workers->run_task(&task);
1237     } else {
1238       task.work(0);
1239     }
1240 
1241     // It's possible that has_aborted() is true here without actually
1242     // aborting the survivor scan earlier. This is OK as it's
1243     // mainly used for sanity checking.
1244     root_regions()->scan_finished();
1245   }
1246 }
1247 
1248 void ConcurrentMark::markFromRoots() {
1249   // we might be tempted to assert that:
1250   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1251   //        "inconsistent argument?");
1252   // However that wouldn't be right, because it's possible that
1253   // a safepoint is indeed in progress as a younger generation
1254   // stop-the-world GC happens even as we mark in this generation.
1255 
1256   _restart_for_overflow = false;
1257   force_overflow_conc()->init();
1258 
1259   // _g1h has _n_par_threads
1260   _parallel_marking_threads = calc_parallel_marking_threads();
1261   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1262     "Maximum number of marking threads exceeded");
1263 
1264   uint active_workers = MAX2(1U, parallel_marking_threads());
1265 
1266   // Parallel task terminator is set in "set_concurrency_and_phase()"
1267   set_concurrency_and_phase(active_workers, true /* concurrent */);
1268 
1269   CMConcurrentMarkingTask markingTask(this, cmThread());
1270   if (use_parallel_marking_threads()) {
1271     _parallel_workers->set_active_workers((int)active_workers);
1272     // Don't set _n_par_threads because it affects MT in process_roots()
1273     // and the decisions on that MT processing is made elsewhere.
1274     assert(_parallel_workers->active_workers() > 0, "Should have been set");
1275     _parallel_workers->run_task(&markingTask);
1276   } else {
1277     markingTask.work(0);
1278   }
1279   print_stats();
1280 }
1281 
1282 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1283   // world is stopped at this checkpoint
1284   assert(SafepointSynchronize::is_at_safepoint(),
1285          "world should be stopped");
1286 
1287   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1288 
1289   // If a full collection has happened, we shouldn't do this.
1290   if (has_aborted()) {
1291     g1h->set_marking_complete(); // So bitmap clearing isn't confused
1292     return;
1293   }
1294 
1295   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1296 
1297   if (VerifyDuringGC) {
1298     HandleMark hm;  // handle scope
1299     Universe::heap()->prepare_for_verify();
1300     Universe::verify(VerifyOption_G1UsePrevMarking,
1301                      " VerifyDuringGC:(before)");
1302   }
1303   g1h->check_bitmaps("Remark Start");
1304 
1305   G1CollectorPolicy* g1p = g1h->g1_policy();
1306   g1p->record_concurrent_mark_remark_start();
1307 
1308   double start = os::elapsedTime();
1309 
1310   checkpointRootsFinalWork();
1311 
1312   double mark_work_end = os::elapsedTime();
1313 
1314   weakRefsWork(clear_all_soft_refs);
1315 
1316   if (has_overflown()) {
1317     // Oops.  We overflowed.  Restart concurrent marking.
1318     _restart_for_overflow = true;
1319     if (G1TraceMarkStackOverflow) {
1320       gclog_or_tty->print_cr("\nRemark led to restart for overflow.");
1321     }
1322 
1323     // Verify the heap w.r.t. the previous marking bitmap.
1324     if (VerifyDuringGC) {
1325       HandleMark hm;  // handle scope
1326       Universe::heap()->prepare_for_verify();
1327       Universe::verify(VerifyOption_G1UsePrevMarking,
1328                        " VerifyDuringGC:(overflow)");
1329     }
1330 
1331     // Clear the marking state because we will be restarting
1332     // marking due to overflowing the global mark stack.
1333     reset_marking_state();
1334   } else {
1335     // Aggregate the per-task counting data that we have accumulated
1336     // while marking.
1337     aggregate_count_data();
1338 
1339     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1340     // We're done with marking.
1341     // This is the end of  the marking cycle, we're expected all
1342     // threads to have SATB queues with active set to true.
1343     satb_mq_set.set_active_all_threads(false, /* new active value */
1344                                        true /* expected_active */);
1345 
1346     if (VerifyDuringGC) {
1347       HandleMark hm;  // handle scope
1348       Universe::heap()->prepare_for_verify();
1349       Universe::verify(VerifyOption_G1UseNextMarking,
1350                        " VerifyDuringGC:(after)");
1351     }
1352     g1h->check_bitmaps("Remark End");
1353     assert(!restart_for_overflow(), "sanity");
1354     // Completely reset the marking state since marking completed
1355     set_non_marking_state();
1356   }
1357 
1358   // Expand the marking stack, if we have to and if we can.
1359   if (_markStack.should_expand()) {
1360     _markStack.expand();
1361   }
1362 
1363   // Statistics
1364   double now = os::elapsedTime();
1365   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1366   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1367   _remark_times.add((now - start) * 1000.0);
1368 
1369   g1p->record_concurrent_mark_remark_end();
1370 
1371   G1CMIsAliveClosure is_alive(g1h);
1372   g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1373 }
1374 
1375 // Base class of the closures that finalize and verify the
1376 // liveness counting data.
1377 class CMCountDataClosureBase: public HeapRegionClosure {
1378 protected:
1379   G1CollectedHeap* _g1h;
1380   ConcurrentMark* _cm;
1381   CardTableModRefBS* _ct_bs;
1382 
1383   BitMap* _region_bm;
1384   BitMap* _card_bm;
1385 
1386   // Takes a region that's not empty (i.e., it has at least one
1387   // live object in it and sets its corresponding bit on the region
1388   // bitmap to 1. If the region is "starts humongous" it will also set
1389   // to 1 the bits on the region bitmap that correspond to its
1390   // associated "continues humongous" regions.
1391   void set_bit_for_region(HeapRegion* hr) {
1392     assert(!hr->continuesHumongous(), "should have filtered those out");
1393 
1394     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1395     if (!hr->startsHumongous()) {
1396       // Normal (non-humongous) case: just set the bit.
1397       _region_bm->par_at_put(index, true);
1398     } else {
1399       // Starts humongous case: calculate how many regions are part of
1400       // this humongous region and then set the bit range.
1401       BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index();
1402       _region_bm->par_at_put_range(index, end_index, true);
1403     }
1404   }
1405 
1406 public:
1407   CMCountDataClosureBase(G1CollectedHeap* g1h,
1408                          BitMap* region_bm, BitMap* card_bm):
1409     _g1h(g1h), _cm(g1h->concurrent_mark()),
1410     _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
1411     _region_bm(region_bm), _card_bm(card_bm) { }
1412 };
1413 
1414 // Closure that calculates the # live objects per region. Used
1415 // for verification purposes during the cleanup pause.
1416 class CalcLiveObjectsClosure: public CMCountDataClosureBase {
1417   CMBitMapRO* _bm;
1418   size_t _region_marked_bytes;
1419 
1420 public:
1421   CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h,
1422                          BitMap* region_bm, BitMap* card_bm) :
1423     CMCountDataClosureBase(g1h, region_bm, card_bm),
1424     _bm(bm), _region_marked_bytes(0) { }
1425 
1426   bool doHeapRegion(HeapRegion* hr) {
1427 
1428     if (hr->continuesHumongous()) {
1429       // We will ignore these here and process them when their
1430       // associated "starts humongous" region is processed (see
1431       // set_bit_for_heap_region()). Note that we cannot rely on their
1432       // associated "starts humongous" region to have their bit set to
1433       // 1 since, due to the region chunking in the parallel region
1434       // iteration, a "continues humongous" region might be visited
1435       // before its associated "starts humongous".
1436       return false;
1437     }
1438 
1439     HeapWord* ntams = hr->next_top_at_mark_start();
1440     HeapWord* start = hr->bottom();
1441 
1442     assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1443            err_msg("Preconditions not met - "
1444                    "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT,
1445                    p2i(start), p2i(ntams), p2i(hr->end())));
1446 
1447     // Find the first marked object at or after "start".
1448     start = _bm->getNextMarkedWordAddress(start, ntams);
1449 
1450     size_t marked_bytes = 0;
1451 
1452     while (start < ntams) {
1453       oop obj = oop(start);
1454       int obj_sz = obj->size();
1455       HeapWord* obj_end = start + obj_sz;
1456 
1457       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1458       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1459 
1460       // Note: if we're looking at the last region in heap - obj_end
1461       // could be actually just beyond the end of the heap; end_idx
1462       // will then correspond to a (non-existent) card that is also
1463       // just beyond the heap.
1464       if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1465         // end of object is not card aligned - increment to cover
1466         // all the cards spanned by the object
1467         end_idx += 1;
1468       }
1469 
1470       // Set the bits in the card BM for the cards spanned by this object.
1471       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1472 
1473       // Add the size of this object to the number of marked bytes.
1474       marked_bytes += (size_t)obj_sz * HeapWordSize;
1475 
1476       // Find the next marked object after this one.
1477       start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1478     }
1479 
1480     // Mark the allocated-since-marking portion...
1481     HeapWord* top = hr->top();
1482     if (ntams < top) {
1483       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1484       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1485 
1486       // Note: if we're looking at the last region in heap - top
1487       // could be actually just beyond the end of the heap; end_idx
1488       // will then correspond to a (non-existent) card that is also
1489       // just beyond the heap.
1490       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1491         // end of object is not card aligned - increment to cover
1492         // all the cards spanned by the object
1493         end_idx += 1;
1494       }
1495       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1496 
1497       // This definitely means the region has live objects.
1498       set_bit_for_region(hr);
1499     }
1500 
1501     // Update the live region bitmap.
1502     if (marked_bytes > 0) {
1503       set_bit_for_region(hr);
1504     }
1505 
1506     // Set the marked bytes for the current region so that
1507     // it can be queried by a calling verification routine
1508     _region_marked_bytes = marked_bytes;
1509 
1510     return false;
1511   }
1512 
1513   size_t region_marked_bytes() const { return _region_marked_bytes; }
1514 };
1515 
1516 // Heap region closure used for verifying the counting data
1517 // that was accumulated concurrently and aggregated during
1518 // the remark pause. This closure is applied to the heap
1519 // regions during the STW cleanup pause.
1520 
1521 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1522   G1CollectedHeap* _g1h;
1523   ConcurrentMark* _cm;
1524   CalcLiveObjectsClosure _calc_cl;
1525   BitMap* _region_bm;   // Region BM to be verified
1526   BitMap* _card_bm;     // Card BM to be verified
1527   bool _verbose;        // verbose output?
1528 
1529   BitMap* _exp_region_bm; // Expected Region BM values
1530   BitMap* _exp_card_bm;   // Expected card BM values
1531 
1532   int _failures;
1533 
1534 public:
1535   VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1536                                 BitMap* region_bm,
1537                                 BitMap* card_bm,
1538                                 BitMap* exp_region_bm,
1539                                 BitMap* exp_card_bm,
1540                                 bool verbose) :
1541     _g1h(g1h), _cm(g1h->concurrent_mark()),
1542     _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1543     _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose),
1544     _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1545     _failures(0) { }
1546 
1547   int failures() const { return _failures; }
1548 
1549   bool doHeapRegion(HeapRegion* hr) {
1550     if (hr->continuesHumongous()) {
1551       // We will ignore these here and process them when their
1552       // associated "starts humongous" region is processed (see
1553       // set_bit_for_heap_region()). Note that we cannot rely on their
1554       // associated "starts humongous" region to have their bit set to
1555       // 1 since, due to the region chunking in the parallel region
1556       // iteration, a "continues humongous" region might be visited
1557       // before its associated "starts humongous".
1558       return false;
1559     }
1560 
1561     int failures = 0;
1562 
1563     // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1564     // this region and set the corresponding bits in the expected region
1565     // and card bitmaps.
1566     bool res = _calc_cl.doHeapRegion(hr);
1567     assert(res == false, "should be continuing");
1568 
1569     MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL),
1570                     Mutex::_no_safepoint_check_flag);
1571 
1572     // Verify the marked bytes for this region.
1573     size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1574     size_t act_marked_bytes = hr->next_marked_bytes();
1575 
1576     // We're not OK if expected marked bytes > actual marked bytes. It means
1577     // we have missed accounting some objects during the actual marking.
1578     if (exp_marked_bytes > act_marked_bytes) {
1579       if (_verbose) {
1580         gclog_or_tty->print_cr("Region %u: marked bytes mismatch: "
1581                                "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT,
1582                                hr->hrm_index(), exp_marked_bytes, act_marked_bytes);
1583       }
1584       failures += 1;
1585     }
1586 
1587     // Verify the bit, for this region, in the actual and expected
1588     // (which was just calculated) region bit maps.
1589     // We're not OK if the bit in the calculated expected region
1590     // bitmap is set and the bit in the actual region bitmap is not.
1591     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1592 
1593     bool expected = _exp_region_bm->at(index);
1594     bool actual = _region_bm->at(index);
1595     if (expected && !actual) {
1596       if (_verbose) {
1597         gclog_or_tty->print_cr("Region %u: region bitmap mismatch: "
1598                                "expected: %s, actual: %s",
1599                                hr->hrm_index(),
1600                                BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1601       }
1602       failures += 1;
1603     }
1604 
1605     // Verify that the card bit maps for the cards spanned by the current
1606     // region match. We have an error if we have a set bit in the expected
1607     // bit map and the corresponding bit in the actual bitmap is not set.
1608 
1609     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1610     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1611 
1612     for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1613       expected = _exp_card_bm->at(i);
1614       actual = _card_bm->at(i);
1615 
1616       if (expected && !actual) {
1617         if (_verbose) {
1618           gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": "
1619                                  "expected: %s, actual: %s",
1620                                  hr->hrm_index(), i,
1621                                  BOOL_TO_STR(expected), BOOL_TO_STR(actual));
1622         }
1623         failures += 1;
1624       }
1625     }
1626 
1627     if (failures > 0 && _verbose)  {
1628       gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", "
1629                              "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT,
1630                              HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()),
1631                              _calc_cl.region_marked_bytes(), hr->next_marked_bytes());
1632     }
1633 
1634     _failures += failures;
1635 
1636     // We could stop iteration over the heap when we
1637     // find the first violating region by returning true.
1638     return false;
1639   }
1640 };
1641 
1642 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1643 protected:
1644   G1CollectedHeap* _g1h;
1645   ConcurrentMark* _cm;
1646   BitMap* _actual_region_bm;
1647   BitMap* _actual_card_bm;
1648 
1649   uint    _n_workers;
1650 
1651   BitMap* _expected_region_bm;
1652   BitMap* _expected_card_bm;
1653 
1654   int  _failures;
1655   bool _verbose;
1656 
1657 public:
1658   G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1659                             BitMap* region_bm, BitMap* card_bm,
1660                             BitMap* expected_region_bm, BitMap* expected_card_bm)
1661     : AbstractGangTask("G1 verify final counting"),
1662       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1663       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1664       _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1665       _failures(0), _verbose(false),
1666       _n_workers(0) {
1667     assert(VerifyDuringGC, "don't call this otherwise");
1668 
1669     // Use the value already set as the number of active threads
1670     // in the call to run_task().
1671     if (G1CollectedHeap::use_parallel_gc_threads()) {
1672       assert( _g1h->workers()->active_workers() > 0,
1673         "Should have been previously set");
1674       _n_workers = _g1h->workers()->active_workers();
1675     } else {
1676       _n_workers = 1;
1677     }
1678 
1679     assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1680     assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1681 
1682     _verbose = _cm->verbose_medium();
1683   }
1684 
1685   void work(uint worker_id) {
1686     assert(worker_id < _n_workers, "invariant");
1687 
1688     VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1689                                             _actual_region_bm, _actual_card_bm,
1690                                             _expected_region_bm,
1691                                             _expected_card_bm,
1692                                             _verbose);
1693 
1694     if (G1CollectedHeap::use_parallel_gc_threads()) {
1695       _g1h->heap_region_par_iterate_chunked(&verify_cl,
1696                                             worker_id,
1697                                             _n_workers,
1698                                             HeapRegion::VerifyCountClaimValue);
1699     } else {
1700       _g1h->heap_region_iterate(&verify_cl);
1701     }
1702 
1703     Atomic::add(verify_cl.failures(), &_failures);
1704   }
1705 
1706   int failures() const { return _failures; }
1707 };
1708 
1709 // Closure that finalizes the liveness counting data.
1710 // Used during the cleanup pause.
1711 // Sets the bits corresponding to the interval [NTAMS, top]
1712 // (which contains the implicitly live objects) in the
1713 // card liveness bitmap. Also sets the bit for each region,
1714 // containing live data, in the region liveness bitmap.
1715 
1716 class FinalCountDataUpdateClosure: public CMCountDataClosureBase {
1717  public:
1718   FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1719                               BitMap* region_bm,
1720                               BitMap* card_bm) :
1721     CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1722 
1723   bool doHeapRegion(HeapRegion* hr) {
1724 
1725     if (hr->continuesHumongous()) {
1726       // We will ignore these here and process them when their
1727       // associated "starts humongous" region is processed (see
1728       // set_bit_for_heap_region()). Note that we cannot rely on their
1729       // associated "starts humongous" region to have their bit set to
1730       // 1 since, due to the region chunking in the parallel region
1731       // iteration, a "continues humongous" region might be visited
1732       // before its associated "starts humongous".
1733       return false;
1734     }
1735 
1736     HeapWord* ntams = hr->next_top_at_mark_start();
1737     HeapWord* top   = hr->top();
1738 
1739     assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1740 
1741     // Mark the allocated-since-marking portion...
1742     if (ntams < top) {
1743       // This definitely means the region has live objects.
1744       set_bit_for_region(hr);
1745 
1746       // Now set the bits in the card bitmap for [ntams, top)
1747       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1748       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1749 
1750       // Note: if we're looking at the last region in heap - top
1751       // could be actually just beyond the end of the heap; end_idx
1752       // will then correspond to a (non-existent) card that is also
1753       // just beyond the heap.
1754       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1755         // end of object is not card aligned - increment to cover
1756         // all the cards spanned by the object
1757         end_idx += 1;
1758       }
1759 
1760       assert(end_idx <= _card_bm->size(),
1761              err_msg("oob: end_idx=  "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1762                      end_idx, _card_bm->size()));
1763       assert(start_idx < _card_bm->size(),
1764              err_msg("oob: start_idx=  "SIZE_FORMAT", bitmap size= "SIZE_FORMAT,
1765                      start_idx, _card_bm->size()));
1766 
1767       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1768     }
1769 
1770     // Set the bit for the region if it contains live data
1771     if (hr->next_marked_bytes() > 0) {
1772       set_bit_for_region(hr);
1773     }
1774 
1775     return false;
1776   }
1777 };
1778 
1779 class G1ParFinalCountTask: public AbstractGangTask {
1780 protected:
1781   G1CollectedHeap* _g1h;
1782   ConcurrentMark* _cm;
1783   BitMap* _actual_region_bm;
1784   BitMap* _actual_card_bm;
1785 
1786   uint    _n_workers;
1787 
1788 public:
1789   G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1790     : AbstractGangTask("G1 final counting"),
1791       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1792       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1793       _n_workers(0) {
1794     // Use the value already set as the number of active threads
1795     // in the call to run_task().
1796     if (G1CollectedHeap::use_parallel_gc_threads()) {
1797       assert( _g1h->workers()->active_workers() > 0,
1798         "Should have been previously set");
1799       _n_workers = _g1h->workers()->active_workers();
1800     } else {
1801       _n_workers = 1;
1802     }
1803   }
1804 
1805   void work(uint worker_id) {
1806     assert(worker_id < _n_workers, "invariant");
1807 
1808     FinalCountDataUpdateClosure final_update_cl(_g1h,
1809                                                 _actual_region_bm,
1810                                                 _actual_card_bm);
1811 
1812     if (G1CollectedHeap::use_parallel_gc_threads()) {
1813       _g1h->heap_region_par_iterate_chunked(&final_update_cl,
1814                                             worker_id,
1815                                             _n_workers,
1816                                             HeapRegion::FinalCountClaimValue);
1817     } else {
1818       _g1h->heap_region_iterate(&final_update_cl);
1819     }
1820   }
1821 };
1822 
1823 class G1ParNoteEndTask;
1824 
1825 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1826   G1CollectedHeap* _g1;
1827   size_t _max_live_bytes;
1828   uint _regions_claimed;
1829   size_t _freed_bytes;
1830   FreeRegionList* _local_cleanup_list;
1831   HeapRegionSetCount _old_regions_removed;
1832   HeapRegionSetCount _humongous_regions_removed;
1833   HRRSCleanupTask* _hrrs_cleanup_task;
1834   double _claimed_region_time;
1835   double _max_region_time;
1836 
1837 public:
1838   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1839                              FreeRegionList* local_cleanup_list,
1840                              HRRSCleanupTask* hrrs_cleanup_task) :
1841     _g1(g1),
1842     _max_live_bytes(0), _regions_claimed(0),
1843     _freed_bytes(0),
1844     _claimed_region_time(0.0), _max_region_time(0.0),
1845     _local_cleanup_list(local_cleanup_list),
1846     _old_regions_removed(),
1847     _humongous_regions_removed(),
1848     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1849 
1850   size_t freed_bytes() { return _freed_bytes; }
1851   const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; }
1852   const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; }
1853 
1854   bool doHeapRegion(HeapRegion *hr) {
1855     if (hr->continuesHumongous()) {
1856       return false;
1857     }
1858     // We use a claim value of zero here because all regions
1859     // were claimed with value 1 in the FinalCount task.
1860     _g1->reset_gc_time_stamps(hr);
1861     double start = os::elapsedTime();
1862     _regions_claimed++;
1863     hr->note_end_of_marking();
1864     _max_live_bytes += hr->max_live_bytes();
1865 
1866     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1867       _freed_bytes += hr->used();
1868       hr->set_containing_set(NULL);
1869       if (hr->isHumongous()) {
1870         assert(hr->startsHumongous(), "we should only see starts humongous");
1871         _humongous_regions_removed.increment(1u, hr->capacity());
1872         _g1->free_humongous_region(hr, _local_cleanup_list, true);
1873       } else {
1874         _old_regions_removed.increment(1u, hr->capacity());
1875         _g1->free_region(hr, _local_cleanup_list, true);
1876       }
1877     } else {
1878       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1879     }
1880 
1881     double region_time = (os::elapsedTime() - start);
1882     _claimed_region_time += region_time;
1883     if (region_time > _max_region_time) {
1884       _max_region_time = region_time;
1885     }
1886     return false;
1887   }
1888 
1889   size_t max_live_bytes() { return _max_live_bytes; }
1890   uint regions_claimed() { return _regions_claimed; }
1891   double claimed_region_time_sec() { return _claimed_region_time; }
1892   double max_region_time_sec() { return _max_region_time; }
1893 };
1894 
1895 class G1ParNoteEndTask: public AbstractGangTask {
1896   friend class G1NoteEndOfConcMarkClosure;
1897 
1898 protected:
1899   G1CollectedHeap* _g1h;
1900   size_t _max_live_bytes;
1901   size_t _freed_bytes;
1902   FreeRegionList* _cleanup_list;
1903 
1904 public:
1905   G1ParNoteEndTask(G1CollectedHeap* g1h,
1906                    FreeRegionList* cleanup_list) :
1907     AbstractGangTask("G1 note end"), _g1h(g1h),
1908     _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { }
1909 
1910   void work(uint worker_id) {
1911     double start = os::elapsedTime();
1912     FreeRegionList local_cleanup_list("Local Cleanup List");
1913     HRRSCleanupTask hrrs_cleanup_task;
1914     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1915                                            &hrrs_cleanup_task);
1916     if (G1CollectedHeap::use_parallel_gc_threads()) {
1917       _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id,
1918                                             _g1h->workers()->active_workers(),
1919                                             HeapRegion::NoteEndClaimValue);
1920     } else {
1921       _g1h->heap_region_iterate(&g1_note_end);
1922     }
1923     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1924 
1925     // Now update the lists
1926     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1927     {
1928       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1929       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1930       _max_live_bytes += g1_note_end.max_live_bytes();
1931       _freed_bytes += g1_note_end.freed_bytes();
1932 
1933       // If we iterate over the global cleanup list at the end of
1934       // cleanup to do this printing we will not guarantee to only
1935       // generate output for the newly-reclaimed regions (the list
1936       // might not be empty at the beginning of cleanup; we might
1937       // still be working on its previous contents). So we do the
1938       // printing here, before we append the new regions to the global
1939       // cleanup list.
1940 
1941       G1HRPrinter* hr_printer = _g1h->hr_printer();
1942       if (hr_printer->is_active()) {
1943         FreeRegionListIterator iter(&local_cleanup_list);
1944         while (iter.more_available()) {
1945           HeapRegion* hr = iter.get_next();
1946           hr_printer->cleanup(hr);
1947         }
1948       }
1949 
1950       _cleanup_list->add_ordered(&local_cleanup_list);
1951       assert(local_cleanup_list.is_empty(), "post-condition");
1952 
1953       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1954     }
1955   }
1956   size_t max_live_bytes() { return _max_live_bytes; }
1957   size_t freed_bytes() { return _freed_bytes; }
1958 };
1959 
1960 class G1ParScrubRemSetTask: public AbstractGangTask {
1961 protected:
1962   G1RemSet* _g1rs;
1963   BitMap* _region_bm;
1964   BitMap* _card_bm;
1965 public:
1966   G1ParScrubRemSetTask(G1CollectedHeap* g1h,
1967                        BitMap* region_bm, BitMap* card_bm) :
1968     AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()),
1969     _region_bm(region_bm), _card_bm(card_bm) { }
1970 
1971   void work(uint worker_id) {
1972     if (G1CollectedHeap::use_parallel_gc_threads()) {
1973       _g1rs->scrub_par(_region_bm, _card_bm, worker_id,
1974                        HeapRegion::ScrubRemSetClaimValue);
1975     } else {
1976       _g1rs->scrub(_region_bm, _card_bm);
1977     }
1978   }
1979 
1980 };
1981 
1982 void ConcurrentMark::cleanup() {
1983   // world is stopped at this checkpoint
1984   assert(SafepointSynchronize::is_at_safepoint(),
1985          "world should be stopped");
1986   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1987 
1988   // If a full collection has happened, we shouldn't do this.
1989   if (has_aborted()) {
1990     g1h->set_marking_complete(); // So bitmap clearing isn't confused
1991     return;
1992   }
1993 
1994   g1h->verify_region_sets_optional();
1995 
1996   if (VerifyDuringGC) {
1997     HandleMark hm;  // handle scope
1998     Universe::heap()->prepare_for_verify();
1999     Universe::verify(VerifyOption_G1UsePrevMarking,
2000                      " VerifyDuringGC:(before)");
2001   }
2002   g1h->check_bitmaps("Cleanup Start");
2003 
2004   G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy();
2005   g1p->record_concurrent_mark_cleanup_start();
2006 
2007   double start = os::elapsedTime();
2008 
2009   HeapRegionRemSet::reset_for_cleanup_tasks();
2010 
2011   uint n_workers;
2012 
2013   // Do counting once more with the world stopped for good measure.
2014   G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
2015 
2016   if (G1CollectedHeap::use_parallel_gc_threads()) {
2017    assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2018            "sanity check");
2019 
2020     g1h->set_par_threads();
2021     n_workers = g1h->n_par_threads();
2022     assert(g1h->n_par_threads() == n_workers,
2023            "Should not have been reset");
2024     g1h->workers()->run_task(&g1_par_count_task);
2025     // Done with the parallel phase so reset to 0.
2026     g1h->set_par_threads(0);
2027 
2028     assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue),
2029            "sanity check");
2030   } else {
2031     n_workers = 1;
2032     g1_par_count_task.work(0);
2033   }
2034 
2035   if (VerifyDuringGC) {
2036     // Verify that the counting data accumulated during marking matches
2037     // that calculated by walking the marking bitmap.
2038 
2039     // Bitmaps to hold expected values
2040     BitMap expected_region_bm(_region_bm.size(), true);
2041     BitMap expected_card_bm(_card_bm.size(), true);
2042 
2043     G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
2044                                                  &_region_bm,
2045                                                  &_card_bm,
2046                                                  &expected_region_bm,
2047                                                  &expected_card_bm);
2048 
2049     if (G1CollectedHeap::use_parallel_gc_threads()) {
2050       g1h->set_par_threads((int)n_workers);
2051       g1h->workers()->run_task(&g1_par_verify_task);
2052       // Done with the parallel phase so reset to 0.
2053       g1h->set_par_threads(0);
2054 
2055       assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue),
2056              "sanity check");
2057     } else {
2058       g1_par_verify_task.work(0);
2059     }
2060 
2061     guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
2062   }
2063 
2064   size_t start_used_bytes = g1h->used();
2065   g1h->set_marking_complete();
2066 
2067   double count_end = os::elapsedTime();
2068   double this_final_counting_time = (count_end - start);
2069   _total_counting_time += this_final_counting_time;
2070 
2071   if (G1PrintRegionLivenessInfo) {
2072     G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking");
2073     _g1h->heap_region_iterate(&cl);
2074   }
2075 
2076   // Install newly created mark bitMap as "prev".
2077   swapMarkBitMaps();
2078 
2079   g1h->reset_gc_time_stamp();
2080 
2081   // Note end of marking in all heap regions.
2082   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list);
2083   if (G1CollectedHeap::use_parallel_gc_threads()) {
2084     g1h->set_par_threads((int)n_workers);
2085     g1h->workers()->run_task(&g1_par_note_end_task);
2086     g1h->set_par_threads(0);
2087 
2088     assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue),
2089            "sanity check");
2090   } else {
2091     g1_par_note_end_task.work(0);
2092   }
2093   g1h->check_gc_time_stamps();
2094 
2095   if (!cleanup_list_is_empty()) {
2096     // The cleanup list is not empty, so we'll have to process it
2097     // concurrently. Notify anyone else that might be wanting free
2098     // regions that there will be more free regions coming soon.
2099     g1h->set_free_regions_coming();
2100   }
2101 
2102   // call below, since it affects the metric by which we sort the heap
2103   // regions.
2104   if (G1ScrubRemSets) {
2105     double rs_scrub_start = os::elapsedTime();
2106     G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
2107     if (G1CollectedHeap::use_parallel_gc_threads()) {
2108       g1h->set_par_threads((int)n_workers);
2109       g1h->workers()->run_task(&g1_par_scrub_rs_task);
2110       g1h->set_par_threads(0);
2111 
2112       assert(g1h->check_heap_region_claim_values(
2113                                             HeapRegion::ScrubRemSetClaimValue),
2114              "sanity check");
2115     } else {
2116       g1_par_scrub_rs_task.work(0);
2117     }
2118 
2119     double rs_scrub_end = os::elapsedTime();
2120     double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start);
2121     _total_rs_scrub_time += this_rs_scrub_time;
2122   }
2123 
2124   // this will also free any regions totally full of garbage objects,
2125   // and sort the regions.
2126   g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers);
2127 
2128   // Statistics.
2129   double end = os::elapsedTime();
2130   _cleanup_times.add((end - start) * 1000.0);
2131 
2132   if (G1Log::fine()) {
2133     g1h->print_size_transition(gclog_or_tty,
2134                                start_used_bytes,
2135                                g1h->used(),
2136                                g1h->capacity());
2137   }
2138 
2139   // Clean up will have freed any regions completely full of garbage.
2140   // Update the soft reference policy with the new heap occupancy.
2141   Universe::update_heap_info_at_gc();
2142 
2143   if (VerifyDuringGC) {
2144     HandleMark hm;  // handle scope
2145     Universe::heap()->prepare_for_verify();
2146     Universe::verify(VerifyOption_G1UsePrevMarking,
2147                      " VerifyDuringGC:(after)");
2148   }
2149 
2150   g1h->check_bitmaps("Cleanup End");
2151 
2152   g1h->verify_region_sets_optional();
2153 
2154   // We need to make this be a "collection" so any collection pause that
2155   // races with it goes around and waits for completeCleanup to finish.
2156   g1h->increment_total_collections();
2157 
2158   // Clean out dead classes and update Metaspace sizes.
2159   if (ClassUnloadingWithConcurrentMark) {
2160     ClassLoaderDataGraph::purge();
2161   }
2162   MetaspaceGC::compute_new_size();
2163 
2164   // We reclaimed old regions so we should calculate the sizes to make
2165   // sure we update the old gen/space data.
2166   g1h->g1mm()->update_sizes();
2167 
2168   g1h->trace_heap_after_concurrent_cycle();
2169 }
2170 
2171 void ConcurrentMark::completeCleanup() {
2172   if (has_aborted()) return;
2173 
2174   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2175 
2176   _cleanup_list.verify_optional();
2177   FreeRegionList tmp_free_list("Tmp Free List");
2178 
2179   if (G1ConcRegionFreeingVerbose) {
2180     gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2181                            "cleanup list has %u entries",
2182                            _cleanup_list.length());
2183   }
2184 
2185   // No one else should be accessing the _cleanup_list at this point,
2186   // so it is not necessary to take any locks
2187   while (!_cleanup_list.is_empty()) {
2188     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
2189     assert(hr != NULL, "Got NULL from a non-empty list");
2190     hr->par_clear();
2191     tmp_free_list.add_ordered(hr);
2192 
2193     // Instead of adding one region at a time to the secondary_free_list,
2194     // we accumulate them in the local list and move them a few at a
2195     // time. This also cuts down on the number of notify_all() calls
2196     // we do during this process. We'll also append the local list when
2197     // _cleanup_list is empty (which means we just removed the last
2198     // region from the _cleanup_list).
2199     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
2200         _cleanup_list.is_empty()) {
2201       if (G1ConcRegionFreeingVerbose) {
2202         gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : "
2203                                "appending %u entries to the secondary_free_list, "
2204                                "cleanup list still has %u entries",
2205                                tmp_free_list.length(),
2206                                _cleanup_list.length());
2207       }
2208 
2209       {
2210         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
2211         g1h->secondary_free_list_add(&tmp_free_list);
2212         SecondaryFreeList_lock->notify_all();
2213       }
2214 
2215       if (G1StressConcRegionFreeing) {
2216         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
2217           os::sleep(Thread::current(), (jlong) 1, false);
2218         }
2219       }
2220     }
2221   }
2222   assert(tmp_free_list.is_empty(), "post-condition");
2223 }
2224 
2225 // Supporting Object and Oop closures for reference discovery
2226 // and processing in during marking
2227 
2228 bool G1CMIsAliveClosure::do_object_b(oop obj) {
2229   HeapWord* addr = (HeapWord*)obj;
2230   return addr != NULL &&
2231          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
2232 }
2233 
2234 // 'Keep Alive' oop closure used by both serial parallel reference processing.
2235 // Uses the CMTask associated with a worker thread (for serial reference
2236 // processing the CMTask for worker 0 is used) to preserve (mark) and
2237 // trace referent objects.
2238 //
2239 // Using the CMTask and embedded local queues avoids having the worker
2240 // threads operating on the global mark stack. This reduces the risk
2241 // of overflowing the stack - which we would rather avoid at this late
2242 // state. Also using the tasks' local queues removes the potential
2243 // of the workers interfering with each other that could occur if
2244 // operating on the global stack.
2245 
2246 class G1CMKeepAliveAndDrainClosure: public OopClosure {
2247   ConcurrentMark* _cm;
2248   CMTask*         _task;
2249   int             _ref_counter_limit;
2250   int             _ref_counter;
2251   bool            _is_serial;
2252  public:
2253   G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2254     _cm(cm), _task(task), _is_serial(is_serial),
2255     _ref_counter_limit(G1RefProcDrainInterval) {
2256     assert(_ref_counter_limit > 0, "sanity");
2257     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2258     _ref_counter = _ref_counter_limit;
2259   }
2260 
2261   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
2262   virtual void do_oop(      oop* p) { do_oop_work(p); }
2263 
2264   template <class T> void do_oop_work(T* p) {
2265     if (!_cm->has_overflown()) {
2266       oop obj = oopDesc::load_decode_heap_oop(p);
2267       if (_cm->verbose_high()) {
2268         gclog_or_tty->print_cr("\t[%u] we're looking at location "
2269                                "*"PTR_FORMAT" = "PTR_FORMAT,
2270                                _task->worker_id(), p2i(p), p2i((void*) obj));
2271       }
2272 
2273       _task->deal_with_reference(obj);
2274       _ref_counter--;
2275 
2276       if (_ref_counter == 0) {
2277         // We have dealt with _ref_counter_limit references, pushing them
2278         // and objects reachable from them on to the local stack (and
2279         // possibly the global stack). Call CMTask::do_marking_step() to
2280         // process these entries.
2281         //
2282         // We call CMTask::do_marking_step() in a loop, which we'll exit if
2283         // there's nothing more to do (i.e. we're done with the entries that
2284         // were pushed as a result of the CMTask::deal_with_reference() calls
2285         // above) or we overflow.
2286         //
2287         // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2288         // flag while there may still be some work to do. (See the comment at
2289         // the beginning of CMTask::do_marking_step() for those conditions -
2290         // one of which is reaching the specified time target.) It is only
2291         // when CMTask::do_marking_step() returns without setting the
2292         // has_aborted() flag that the marking step has completed.
2293         do {
2294           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
2295           _task->do_marking_step(mark_step_duration_ms,
2296                                  false      /* do_termination */,
2297                                  _is_serial);
2298         } while (_task->has_aborted() && !_cm->has_overflown());
2299         _ref_counter = _ref_counter_limit;
2300       }
2301     } else {
2302       if (_cm->verbose_high()) {
2303          gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id());
2304       }
2305     }
2306   }
2307 };
2308 
2309 // 'Drain' oop closure used by both serial and parallel reference processing.
2310 // Uses the CMTask associated with a given worker thread (for serial
2311 // reference processing the CMtask for worker 0 is used). Calls the
2312 // do_marking_step routine, with an unbelievably large timeout value,
2313 // to drain the marking data structures of the remaining entries
2314 // added by the 'keep alive' oop closure above.
2315 
2316 class G1CMDrainMarkingStackClosure: public VoidClosure {
2317   ConcurrentMark* _cm;
2318   CMTask*         _task;
2319   bool            _is_serial;
2320  public:
2321   G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) :
2322     _cm(cm), _task(task), _is_serial(is_serial) {
2323     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
2324   }
2325 
2326   void do_void() {
2327     do {
2328       if (_cm->verbose_high()) {
2329         gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s",
2330                                _task->worker_id(), BOOL_TO_STR(_is_serial));
2331       }
2332 
2333       // We call CMTask::do_marking_step() to completely drain the local
2334       // and global marking stacks of entries pushed by the 'keep alive'
2335       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
2336       //
2337       // CMTask::do_marking_step() is called in a loop, which we'll exit
2338       // if there's nothing more to do (i.e. we've completely drained the
2339       // entries that were pushed as a a result of applying the 'keep alive'
2340       // closure to the entries on the discovered ref lists) or we overflow
2341       // the global marking stack.
2342       //
2343       // Note: CMTask::do_marking_step() can set the CMTask::has_aborted()
2344       // flag while there may still be some work to do. (See the comment at
2345       // the beginning of CMTask::do_marking_step() for those conditions -
2346       // one of which is reaching the specified time target.) It is only
2347       // when CMTask::do_marking_step() returns without setting the
2348       // has_aborted() flag that the marking step has completed.
2349 
2350       _task->do_marking_step(1000000000.0 /* something very large */,
2351                              true         /* do_termination */,
2352                              _is_serial);
2353     } while (_task->has_aborted() && !_cm->has_overflown());
2354   }
2355 };
2356 
2357 // Implementation of AbstractRefProcTaskExecutor for parallel
2358 // reference processing at the end of G1 concurrent marking
2359 
2360 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
2361 private:
2362   G1CollectedHeap* _g1h;
2363   ConcurrentMark*  _cm;
2364   WorkGang*        _workers;
2365   int              _active_workers;
2366 
2367 public:
2368   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
2369                         ConcurrentMark* cm,
2370                         WorkGang* workers,
2371                         int n_workers) :
2372     _g1h(g1h), _cm(cm),
2373     _workers(workers), _active_workers(n_workers) { }
2374 
2375   // Executes the given task using concurrent marking worker threads.
2376   virtual void execute(ProcessTask& task);
2377   virtual void execute(EnqueueTask& task);
2378 };
2379 
2380 class G1CMRefProcTaskProxy: public AbstractGangTask {
2381   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
2382   ProcessTask&     _proc_task;
2383   G1CollectedHeap* _g1h;
2384   ConcurrentMark*  _cm;
2385 
2386 public:
2387   G1CMRefProcTaskProxy(ProcessTask& proc_task,
2388                      G1CollectedHeap* g1h,
2389                      ConcurrentMark* cm) :
2390     AbstractGangTask("Process reference objects in parallel"),
2391     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
2392     ReferenceProcessor* rp = _g1h->ref_processor_cm();
2393     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
2394   }
2395 
2396   virtual void work(uint worker_id) {
2397     ResourceMark rm;
2398     HandleMark hm;
2399     CMTask* task = _cm->task(worker_id);
2400     G1CMIsAliveClosure g1_is_alive(_g1h);
2401     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
2402     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
2403 
2404     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
2405   }
2406 };
2407 
2408 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
2409   assert(_workers != NULL, "Need parallel worker threads.");
2410   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2411 
2412   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
2413 
2414   // We need to reset the concurrency level before each
2415   // proxy task execution, so that the termination protocol
2416   // and overflow handling in CMTask::do_marking_step() knows
2417   // how many workers to wait for.
2418   _cm->set_concurrency(_active_workers);
2419   _g1h->set_par_threads(_active_workers);
2420   _workers->run_task(&proc_task_proxy);
2421   _g1h->set_par_threads(0);
2422 }
2423 
2424 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
2425   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
2426   EnqueueTask& _enq_task;
2427 
2428 public:
2429   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
2430     AbstractGangTask("Enqueue reference objects in parallel"),
2431     _enq_task(enq_task) { }
2432 
2433   virtual void work(uint worker_id) {
2434     _enq_task.work(worker_id);
2435   }
2436 };
2437 
2438 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
2439   assert(_workers != NULL, "Need parallel worker threads.");
2440   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
2441 
2442   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
2443 
2444   // Not strictly necessary but...
2445   //
2446   // We need to reset the concurrency level before each
2447   // proxy task execution, so that the termination protocol
2448   // and overflow handling in CMTask::do_marking_step() knows
2449   // how many workers to wait for.
2450   _cm->set_concurrency(_active_workers);
2451   _g1h->set_par_threads(_active_workers);
2452   _workers->run_task(&enq_task_proxy);
2453   _g1h->set_par_threads(0);
2454 }
2455 
2456 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
2457   G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
2458 }
2459 
2460 // Helper class to get rid of some boilerplate code.
2461 class G1RemarkGCTraceTime : public GCTraceTime {
2462   static bool doit_and_prepend(bool doit) {
2463     if (doit) {
2464       gclog_or_tty->put(' ');
2465     }
2466     return doit;
2467   }
2468 
2469  public:
2470   G1RemarkGCTraceTime(const char* title, bool doit)
2471     : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(),
2472         G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) {
2473   }
2474 };
2475 
2476 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2477   if (has_overflown()) {
2478     // Skip processing the discovered references if we have
2479     // overflown the global marking stack. Reference objects
2480     // only get discovered once so it is OK to not
2481     // de-populate the discovered reference lists. We could have,
2482     // but the only benefit would be that, when marking restarts,
2483     // less reference objects are discovered.
2484     return;
2485   }
2486 
2487   ResourceMark rm;
2488   HandleMark   hm;
2489 
2490   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2491 
2492   // Is alive closure.
2493   G1CMIsAliveClosure g1_is_alive(g1h);
2494 
2495   // Inner scope to exclude the cleaning of the string and symbol
2496   // tables from the displayed time.
2497   {
2498     if (G1Log::finer()) {
2499       gclog_or_tty->put(' ');
2500     }
2501     GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm(), concurrent_gc_id());
2502 
2503     ReferenceProcessor* rp = g1h->ref_processor_cm();
2504 
2505     // See the comment in G1CollectedHeap::ref_processing_init()
2506     // about how reference processing currently works in G1.
2507 
2508     // Set the soft reference policy
2509     rp->setup_policy(clear_all_soft_refs);
2510     assert(_markStack.isEmpty(), "mark stack should be empty");
2511 
2512     // Instances of the 'Keep Alive' and 'Complete GC' closures used
2513     // in serial reference processing. Note these closures are also
2514     // used for serially processing (by the the current thread) the
2515     // JNI references during parallel reference processing.
2516     //
2517     // These closures do not need to synchronize with the worker
2518     // threads involved in parallel reference processing as these
2519     // instances are executed serially by the current thread (e.g.
2520     // reference processing is not multi-threaded and is thus
2521     // performed by the current thread instead of a gang worker).
2522     //
2523     // The gang tasks involved in parallel reference processing create
2524     // their own instances of these closures, which do their own
2525     // synchronization among themselves.
2526     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2527     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2528 
2529     // We need at least one active thread. If reference processing
2530     // is not multi-threaded we use the current (VMThread) thread,
2531     // otherwise we use the work gang from the G1CollectedHeap and
2532     // we utilize all the worker threads we can.
2533     bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL;
2534     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2535     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2536 
2537     // Parallel processing task executor.
2538     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2539                                               g1h->workers(), active_workers);
2540     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2541 
2542     // Set the concurrency level. The phase was already set prior to
2543     // executing the remark task.
2544     set_concurrency(active_workers);
2545 
2546     // Set the degree of MT processing here.  If the discovery was done MT,
2547     // the number of threads involved during discovery could differ from
2548     // the number of active workers.  This is OK as long as the discovered
2549     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2550     rp->set_active_mt_degree(active_workers);
2551 
2552     // Process the weak references.
2553     const ReferenceProcessorStats& stats =
2554         rp->process_discovered_references(&g1_is_alive,
2555                                           &g1_keep_alive,
2556                                           &g1_drain_mark_stack,
2557                                           executor,
2558                                           g1h->gc_timer_cm(),
2559                                           concurrent_gc_id());
2560     g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2561 
2562     // The do_oop work routines of the keep_alive and drain_marking_stack
2563     // oop closures will set the has_overflown flag if we overflow the
2564     // global marking stack.
2565 
2566     assert(_markStack.overflow() || _markStack.isEmpty(),
2567             "mark stack should be empty (unless it overflowed)");
2568 
2569     if (_markStack.overflow()) {
2570       // This should have been done already when we tried to push an
2571       // entry on to the global mark stack. But let's do it again.
2572       set_has_overflown();
2573     }
2574 
2575     assert(rp->num_q() == active_workers, "why not");
2576 
2577     rp->enqueue_discovered_references(executor);
2578 
2579     rp->verify_no_references_recorded();
2580     assert(!rp->discovery_enabled(), "Post condition");
2581   }
2582 
2583   if (has_overflown()) {
2584     // We can not trust g1_is_alive if the marking stack overflowed
2585     return;
2586   }
2587 
2588   assert(_markStack.isEmpty(), "Marking should have completed");
2589 
2590   // Unload Klasses, String, Symbols, Code Cache, etc.
2591   {
2592     G1RemarkGCTraceTime trace("Unloading", G1Log::finer());
2593 
2594     if (ClassUnloadingWithConcurrentMark) {
2595       bool purged_classes;
2596 
2597       {
2598         G1RemarkGCTraceTime trace("System Dictionary Unloading", G1Log::finest());
2599         purged_classes = SystemDictionary::do_unloading(&g1_is_alive);
2600       }
2601 
2602       {
2603         G1RemarkGCTraceTime trace("Parallel Unloading", G1Log::finest());
2604         weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2605       }
2606     }
2607 
2608     if (G1StringDedup::is_enabled()) {
2609       G1RemarkGCTraceTime trace("String Deduplication Unlink", G1Log::finest());
2610       G1StringDedup::unlink(&g1_is_alive);
2611     }
2612   }
2613 }
2614 
2615 void ConcurrentMark::swapMarkBitMaps() {
2616   CMBitMapRO* temp = _prevMarkBitMap;
2617   _prevMarkBitMap  = (CMBitMapRO*)_nextMarkBitMap;
2618   _nextMarkBitMap  = (CMBitMap*)  temp;
2619 }
2620 
2621 class CMObjectClosure;
2622 
2623 // Closure for iterating over objects, currently only used for
2624 // processing SATB buffers.
2625 class CMObjectClosure : public ObjectClosure {
2626 private:
2627   CMTask* _task;
2628 
2629 public:
2630   void do_object(oop obj) {
2631     _task->deal_with_reference(obj);
2632   }
2633 
2634   CMObjectClosure(CMTask* task) : _task(task) { }
2635 };
2636 
2637 class G1RemarkThreadsClosure : public ThreadClosure {
2638   CMObjectClosure _cm_obj;
2639   G1CMOopClosure _cm_cl;
2640   MarkingCodeBlobClosure _code_cl;
2641   int _thread_parity;
2642   bool _is_par;
2643 
2644  public:
2645   G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) :
2646     _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2647     _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {}
2648 
2649   void do_thread(Thread* thread) {
2650     if (thread->is_Java_thread()) {
2651       if (thread->claim_oops_do(_is_par, _thread_parity)) {
2652         JavaThread* jt = (JavaThread*)thread;
2653 
2654         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2655         // however the liveness of oops reachable from nmethods have very complex lifecycles:
2656         // * Alive if on the stack of an executing method
2657         // * Weakly reachable otherwise
2658         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2659         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2660         jt->nmethods_do(&_code_cl);
2661 
2662         jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj);
2663       }
2664     } else if (thread->is_VM_thread()) {
2665       if (thread->claim_oops_do(_is_par, _thread_parity)) {
2666         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj);
2667       }
2668     }
2669   }
2670 };
2671 
2672 class CMRemarkTask: public AbstractGangTask {
2673 private:
2674   ConcurrentMark* _cm;
2675   bool            _is_serial;
2676 public:
2677   void work(uint worker_id) {
2678     // Since all available tasks are actually started, we should
2679     // only proceed if we're supposed to be active.
2680     if (worker_id < _cm->active_tasks()) {
2681       CMTask* task = _cm->task(worker_id);
2682       task->record_start_time();
2683       {
2684         ResourceMark rm;
2685         HandleMark hm;
2686 
2687         G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial);
2688         Threads::threads_do(&threads_f);
2689       }
2690 
2691       do {
2692         task->do_marking_step(1000000000.0 /* something very large */,
2693                               true         /* do_termination       */,
2694                               _is_serial);
2695       } while (task->has_aborted() && !_cm->has_overflown());
2696       // If we overflow, then we do not want to restart. We instead
2697       // want to abort remark and do concurrent marking again.
2698       task->record_end_time();
2699     }
2700   }
2701 
2702   CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) :
2703     AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) {
2704     _cm->terminator()->reset_for_reuse(active_workers);
2705   }
2706 };
2707 
2708 void ConcurrentMark::checkpointRootsFinalWork() {
2709   ResourceMark rm;
2710   HandleMark   hm;
2711   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2712 
2713   G1RemarkGCTraceTime trace("Finalize Marking", G1Log::finer());
2714 
2715   g1h->ensure_parsability(false);
2716 
2717   if (G1CollectedHeap::use_parallel_gc_threads()) {
2718     G1CollectedHeap::StrongRootsScope srs(g1h);
2719     // this is remark, so we'll use up all active threads
2720     uint active_workers = g1h->workers()->active_workers();
2721     if (active_workers == 0) {
2722       assert(active_workers > 0, "Should have been set earlier");
2723       active_workers = (uint) ParallelGCThreads;
2724       g1h->workers()->set_active_workers(active_workers);
2725     }
2726     set_concurrency_and_phase(active_workers, false /* concurrent */);
2727     // Leave _parallel_marking_threads at it's
2728     // value originally calculated in the ConcurrentMark
2729     // constructor and pass values of the active workers
2730     // through the gang in the task.
2731 
2732     CMRemarkTask remarkTask(this, active_workers, false /* is_serial */);
2733     // We will start all available threads, even if we decide that the
2734     // active_workers will be fewer. The extra ones will just bail out
2735     // immediately.
2736     g1h->set_par_threads(active_workers);
2737     g1h->workers()->run_task(&remarkTask);
2738     g1h->set_par_threads(0);
2739   } else {
2740     G1CollectedHeap::StrongRootsScope srs(g1h);
2741     uint active_workers = 1;
2742     set_concurrency_and_phase(active_workers, false /* concurrent */);
2743 
2744     // Note - if there's no work gang then the VMThread will be
2745     // the thread to execute the remark - serially. We have
2746     // to pass true for the is_serial parameter so that
2747     // CMTask::do_marking_step() doesn't enter the sync
2748     // barriers in the event of an overflow. Doing so will
2749     // cause an assert that the current thread is not a
2750     // concurrent GC thread.
2751     CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/);
2752     remarkTask.work(0);
2753   }
2754   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2755   guarantee(has_overflown() ||
2756             satb_mq_set.completed_buffers_num() == 0,
2757             err_msg("Invariant: has_overflown = %s, num buffers = %d",
2758                     BOOL_TO_STR(has_overflown()),
2759                     satb_mq_set.completed_buffers_num()));
2760 
2761   print_stats();
2762 }
2763 
2764 #ifndef PRODUCT
2765 
2766 class PrintReachableOopClosure: public OopClosure {
2767 private:
2768   G1CollectedHeap* _g1h;
2769   outputStream*    _out;
2770   VerifyOption     _vo;
2771   bool             _all;
2772 
2773 public:
2774   PrintReachableOopClosure(outputStream* out,
2775                            VerifyOption  vo,
2776                            bool          all) :
2777     _g1h(G1CollectedHeap::heap()),
2778     _out(out), _vo(vo), _all(all) { }
2779 
2780   void do_oop(narrowOop* p) { do_oop_work(p); }
2781   void do_oop(      oop* p) { do_oop_work(p); }
2782 
2783   template <class T> void do_oop_work(T* p) {
2784     oop         obj = oopDesc::load_decode_heap_oop(p);
2785     const char* str = NULL;
2786     const char* str2 = "";
2787 
2788     if (obj == NULL) {
2789       str = "";
2790     } else if (!_g1h->is_in_g1_reserved(obj)) {
2791       str = " O";
2792     } else {
2793       HeapRegion* hr  = _g1h->heap_region_containing(obj);
2794       bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo);
2795       bool marked = _g1h->is_marked(obj, _vo);
2796 
2797       if (over_tams) {
2798         str = " >";
2799         if (marked) {
2800           str2 = " AND MARKED";
2801         }
2802       } else if (marked) {
2803         str = " M";
2804       } else {
2805         str = " NOT";
2806       }
2807     }
2808 
2809     _out->print_cr("  "PTR_FORMAT": "PTR_FORMAT"%s%s",
2810                    p2i(p), p2i((void*) obj), str, str2);
2811   }
2812 };
2813 
2814 class PrintReachableObjectClosure : public ObjectClosure {
2815 private:
2816   G1CollectedHeap* _g1h;
2817   outputStream*    _out;
2818   VerifyOption     _vo;
2819   bool             _all;
2820   HeapRegion*      _hr;
2821 
2822 public:
2823   PrintReachableObjectClosure(outputStream* out,
2824                               VerifyOption  vo,
2825                               bool          all,
2826                               HeapRegion*   hr) :
2827     _g1h(G1CollectedHeap::heap()),
2828     _out(out), _vo(vo), _all(all), _hr(hr) { }
2829 
2830   void do_object(oop o) {
2831     bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo);
2832     bool marked = _g1h->is_marked(o, _vo);
2833     bool print_it = _all || over_tams || marked;
2834 
2835     if (print_it) {
2836       _out->print_cr(" "PTR_FORMAT"%s",
2837                      p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : "");
2838       PrintReachableOopClosure oopCl(_out, _vo, _all);
2839       o->oop_iterate_no_header(&oopCl);
2840     }
2841   }
2842 };
2843 
2844 class PrintReachableRegionClosure : public HeapRegionClosure {
2845 private:
2846   G1CollectedHeap* _g1h;
2847   outputStream*    _out;
2848   VerifyOption     _vo;
2849   bool             _all;
2850 
2851 public:
2852   bool doHeapRegion(HeapRegion* hr) {
2853     HeapWord* b = hr->bottom();
2854     HeapWord* e = hr->end();
2855     HeapWord* t = hr->top();
2856     HeapWord* p = _g1h->top_at_mark_start(hr, _vo);
2857     _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" "
2858                    "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p));
2859     _out->cr();
2860 
2861     HeapWord* from = b;
2862     HeapWord* to   = t;
2863 
2864     if (to > from) {
2865       _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to));
2866       _out->cr();
2867       PrintReachableObjectClosure ocl(_out, _vo, _all, hr);
2868       hr->object_iterate_mem_careful(MemRegion(from, to), &ocl);
2869       _out->cr();
2870     }
2871 
2872     return false;
2873   }
2874 
2875   PrintReachableRegionClosure(outputStream* out,
2876                               VerifyOption  vo,
2877                               bool          all) :
2878     _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { }
2879 };
2880 
2881 void ConcurrentMark::print_reachable(const char* str,
2882                                      VerifyOption vo,
2883                                      bool all) {
2884   gclog_or_tty->cr();
2885   gclog_or_tty->print_cr("== Doing heap dump... ");
2886 
2887   if (G1PrintReachableBaseFile == NULL) {
2888     gclog_or_tty->print_cr("  #### error: no base file defined");
2889     return;
2890   }
2891 
2892   if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) >
2893       (JVM_MAXPATHLEN - 1)) {
2894     gclog_or_tty->print_cr("  #### error: file name too long");
2895     return;
2896   }
2897 
2898   char file_name[JVM_MAXPATHLEN];
2899   sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str);
2900   gclog_or_tty->print_cr("  dumping to file %s", file_name);
2901 
2902   fileStream fout(file_name);
2903   if (!fout.is_open()) {
2904     gclog_or_tty->print_cr("  #### error: could not open file");
2905     return;
2906   }
2907 
2908   outputStream* out = &fout;
2909   out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo));
2910   out->cr();
2911 
2912   out->print_cr("--- ITERATING OVER REGIONS");
2913   out->cr();
2914   PrintReachableRegionClosure rcl(out, vo, all);
2915   _g1h->heap_region_iterate(&rcl);
2916   out->cr();
2917 
2918   gclog_or_tty->print_cr("  done");
2919   gclog_or_tty->flush();
2920 }
2921 
2922 #endif // PRODUCT
2923 
2924 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2925   // Note we are overriding the read-only view of the prev map here, via
2926   // the cast.
2927   ((CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2928 }
2929 
2930 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) {
2931   _nextMarkBitMap->clearRange(mr);
2932 }
2933 
2934 HeapRegion*
2935 ConcurrentMark::claim_region(uint worker_id) {
2936   // "checkpoint" the finger
2937   HeapWord* finger = _finger;
2938 
2939   // _heap_end will not change underneath our feet; it only changes at
2940   // yield points.
2941   while (finger < _heap_end) {
2942     assert(_g1h->is_in_g1_reserved(finger), "invariant");
2943 
2944     // Note on how this code handles humongous regions. In the
2945     // normal case the finger will reach the start of a "starts
2946     // humongous" (SH) region. Its end will either be the end of the
2947     // last "continues humongous" (CH) region in the sequence, or the
2948     // standard end of the SH region (if the SH is the only region in
2949     // the sequence). That way claim_region() will skip over the CH
2950     // regions. However, there is a subtle race between a CM thread
2951     // executing this method and a mutator thread doing a humongous
2952     // object allocation. The two are not mutually exclusive as the CM
2953     // thread does not need to hold the Heap_lock when it gets
2954     // here. So there is a chance that claim_region() will come across
2955     // a free region that's in the progress of becoming a SH or a CH
2956     // region. In the former case, it will either
2957     //   a) Miss the update to the region's end, in which case it will
2958     //      visit every subsequent CH region, will find their bitmaps
2959     //      empty, and do nothing, or
2960     //   b) Will observe the update of the region's end (in which case
2961     //      it will skip the subsequent CH regions).
2962     // If it comes across a region that suddenly becomes CH, the
2963     // scenario will be similar to b). So, the race between
2964     // claim_region() and a humongous object allocation might force us
2965     // to do a bit of unnecessary work (due to some unnecessary bitmap
2966     // iterations) but it should not introduce and correctness issues.
2967     HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger);
2968 
2969     // Above heap_region_containing_raw may return NULL as we always scan claim
2970     // until the end of the heap. In this case, just jump to the next region.
2971     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2972 
2973     // Is the gap between reading the finger and doing the CAS too long?
2974     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2975     if (res == finger && curr_region != NULL) {
2976       // we succeeded
2977       HeapWord*   bottom        = curr_region->bottom();
2978       HeapWord*   limit         = curr_region->next_top_at_mark_start();
2979 
2980       if (verbose_low()) {
2981         gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" "
2982                                "["PTR_FORMAT", "PTR_FORMAT"), "
2983                                "limit = "PTR_FORMAT,
2984                                worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit));
2985       }
2986 
2987       // notice that _finger == end cannot be guaranteed here since,
2988       // someone else might have moved the finger even further
2989       assert(_finger >= end, "the finger should have moved forward");
2990 
2991       if (verbose_low()) {
2992         gclog_or_tty->print_cr("[%u] we were successful with region = "
2993                                PTR_FORMAT, worker_id, p2i(curr_region));
2994       }
2995 
2996       if (limit > bottom) {
2997         if (verbose_low()) {
2998           gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, "
2999                                  "returning it ", worker_id, p2i(curr_region));
3000         }
3001         return curr_region;
3002       } else {
3003         assert(limit == bottom,
3004                "the region limit should be at bottom");
3005         if (verbose_low()) {
3006           gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, "
3007                                  "returning NULL", worker_id, p2i(curr_region));
3008         }
3009         // we return NULL and the caller should try calling
3010         // claim_region() again.
3011         return NULL;
3012       }
3013     } else {
3014       assert(_finger > finger, "the finger should have moved forward");
3015       if (verbose_low()) {
3016         if (curr_region == NULL) {
3017           gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, "
3018                                  "global finger = "PTR_FORMAT", "
3019                                  "our finger = "PTR_FORMAT,
3020                                  worker_id, p2i(_finger), p2i(finger));
3021         } else {
3022           gclog_or_tty->print_cr("[%u] somebody else moved the finger, "
3023                                  "global finger = "PTR_FORMAT", "
3024                                  "our finger = "PTR_FORMAT,
3025                                  worker_id, p2i(_finger), p2i(finger));
3026         }
3027       }
3028 
3029       // read it again
3030       finger = _finger;
3031     }
3032   }
3033 
3034   return NULL;
3035 }
3036 
3037 #ifndef PRODUCT
3038 enum VerifyNoCSetOopsPhase {
3039   VerifyNoCSetOopsStack,
3040   VerifyNoCSetOopsQueues,
3041   VerifyNoCSetOopsSATBCompleted,
3042   VerifyNoCSetOopsSATBThread
3043 };
3044 
3045 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure  {
3046 private:
3047   G1CollectedHeap* _g1h;
3048   VerifyNoCSetOopsPhase _phase;
3049   int _info;
3050 
3051   const char* phase_str() {
3052     switch (_phase) {
3053     case VerifyNoCSetOopsStack:         return "Stack";
3054     case VerifyNoCSetOopsQueues:        return "Queue";
3055     case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers";
3056     case VerifyNoCSetOopsSATBThread:    return "Thread SATB Buffers";
3057     default:                            ShouldNotReachHere();
3058     }
3059     return NULL;
3060   }
3061 
3062   void do_object_work(oop obj) {
3063     guarantee(!_g1h->obj_in_cs(obj),
3064               err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d",
3065                       p2i((void*) obj), phase_str(), _info));
3066   }
3067 
3068 public:
3069   VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { }
3070 
3071   void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) {
3072     _phase = phase;
3073     _info = info;
3074   }
3075 
3076   virtual void do_oop(oop* p) {
3077     oop obj = oopDesc::load_decode_heap_oop(p);
3078     do_object_work(obj);
3079   }
3080 
3081   virtual void do_oop(narrowOop* p) {
3082     // We should not come across narrow oops while scanning marking
3083     // stacks and SATB buffers.
3084     ShouldNotReachHere();
3085   }
3086 
3087   virtual void do_object(oop obj) {
3088     do_object_work(obj);
3089   }
3090 };
3091 
3092 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks,
3093                                          bool verify_enqueued_buffers,
3094                                          bool verify_thread_buffers,
3095                                          bool verify_fingers) {
3096   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
3097   if (!G1CollectedHeap::heap()->mark_in_progress()) {
3098     return;
3099   }
3100 
3101   VerifyNoCSetOopsClosure cl;
3102 
3103   if (verify_stacks) {
3104     // Verify entries on the global mark stack
3105     cl.set_phase(VerifyNoCSetOopsStack);
3106     _markStack.oops_do(&cl);
3107 
3108     // Verify entries on the task queues
3109     for (uint i = 0; i < _max_worker_id; i += 1) {
3110       cl.set_phase(VerifyNoCSetOopsQueues, i);
3111       CMTaskQueue* queue = _task_queues->queue(i);
3112       queue->oops_do(&cl);
3113     }
3114   }
3115 
3116   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
3117 
3118   // Verify entries on the enqueued SATB buffers
3119   if (verify_enqueued_buffers) {
3120     cl.set_phase(VerifyNoCSetOopsSATBCompleted);
3121     satb_qs.iterate_completed_buffers_read_only(&cl);
3122   }
3123 
3124   // Verify entries on the per-thread SATB buffers
3125   if (verify_thread_buffers) {
3126     cl.set_phase(VerifyNoCSetOopsSATBThread);
3127     satb_qs.iterate_thread_buffers_read_only(&cl);
3128   }
3129 
3130   if (verify_fingers) {
3131     // Verify the global finger
3132     HeapWord* global_finger = finger();
3133     if (global_finger != NULL && global_finger < _heap_end) {
3134       // The global finger always points to a heap region boundary. We
3135       // use heap_region_containing_raw() to get the containing region
3136       // given that the global finger could be pointing to a free region
3137       // which subsequently becomes continues humongous. If that
3138       // happens, heap_region_containing() will return the bottom of the
3139       // corresponding starts humongous region and the check below will
3140       // not hold any more.
3141       // Since we always iterate over all regions, we might get a NULL HeapRegion
3142       // here.
3143       HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger);
3144       guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
3145                 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT,
3146                         p2i(global_finger), HR_FORMAT_PARAMS(global_hr)));
3147     }
3148 
3149     // Verify the task fingers
3150     assert(parallel_marking_threads() <= _max_worker_id, "sanity");
3151     for (int i = 0; i < (int) parallel_marking_threads(); i += 1) {
3152       CMTask* task = _tasks[i];
3153       HeapWord* task_finger = task->finger();
3154       if (task_finger != NULL && task_finger < _heap_end) {
3155         // See above note on the global finger verification.
3156         HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger);
3157         guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
3158                   !task_hr->in_collection_set(),
3159                   err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT,
3160                           p2i(task_finger), HR_FORMAT_PARAMS(task_hr)));
3161       }
3162     }
3163   }
3164 }
3165 #endif // PRODUCT
3166 
3167 // Aggregate the counting data that was constructed concurrently
3168 // with marking.
3169 class AggregateCountDataHRClosure: public HeapRegionClosure {
3170   G1CollectedHeap* _g1h;
3171   ConcurrentMark* _cm;
3172   CardTableModRefBS* _ct_bs;
3173   BitMap* _cm_card_bm;
3174   uint _max_worker_id;
3175 
3176  public:
3177   AggregateCountDataHRClosure(G1CollectedHeap* g1h,
3178                               BitMap* cm_card_bm,
3179                               uint max_worker_id) :
3180     _g1h(g1h), _cm(g1h->concurrent_mark()),
3181     _ct_bs((CardTableModRefBS*) (g1h->barrier_set())),
3182     _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
3183 
3184   bool doHeapRegion(HeapRegion* hr) {
3185     if (hr->continuesHumongous()) {
3186       // We will ignore these here and process them when their
3187       // associated "starts humongous" region is processed.
3188       // Note that we cannot rely on their associated
3189       // "starts humongous" region to have their bit set to 1
3190       // since, due to the region chunking in the parallel region
3191       // iteration, a "continues humongous" region might be visited
3192       // before its associated "starts humongous".
3193       return false;
3194     }
3195 
3196     HeapWord* start = hr->bottom();
3197     HeapWord* limit = hr->next_top_at_mark_start();
3198     HeapWord* end = hr->end();
3199 
3200     assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
3201            err_msg("Preconditions not met - "
3202                    "start: "PTR_FORMAT", limit: "PTR_FORMAT", "
3203                    "top: "PTR_FORMAT", end: "PTR_FORMAT,
3204                    p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())));
3205 
3206     assert(hr->next_marked_bytes() == 0, "Precondition");
3207 
3208     if (start == limit) {
3209       // NTAMS of this region has not been set so nothing to do.
3210       return false;
3211     }
3212 
3213     // 'start' should be in the heap.
3214     assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
3215     // 'end' *may* be just beyond the end of the heap (if hr is the last region)
3216     assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
3217 
3218     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
3219     BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
3220     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
3221 
3222     // If ntams is not card aligned then we bump card bitmap index
3223     // for limit so that we get the all the cards spanned by
3224     // the object ending at ntams.
3225     // Note: if this is the last region in the heap then ntams
3226     // could be actually just beyond the end of the the heap;
3227     // limit_idx will then  correspond to a (non-existent) card
3228     // that is also outside the heap.
3229     if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
3230       limit_idx += 1;
3231     }
3232 
3233     assert(limit_idx <= end_idx, "or else use atomics");
3234 
3235     // Aggregate the "stripe" in the count data associated with hr.
3236     uint hrm_index = hr->hrm_index();
3237     size_t marked_bytes = 0;
3238 
3239     for (uint i = 0; i < _max_worker_id; i += 1) {
3240       size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
3241       BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
3242 
3243       // Fetch the marked_bytes in this region for task i and
3244       // add it to the running total for this region.
3245       marked_bytes += marked_bytes_array[hrm_index];
3246 
3247       // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
3248       // into the global card bitmap.
3249       BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
3250 
3251       while (scan_idx < limit_idx) {
3252         assert(task_card_bm->at(scan_idx) == true, "should be");
3253         _cm_card_bm->set_bit(scan_idx);
3254         assert(_cm_card_bm->at(scan_idx) == true, "should be");
3255 
3256         // BitMap::get_next_one_offset() can handle the case when
3257         // its left_offset parameter is greater than its right_offset
3258         // parameter. It does, however, have an early exit if
3259         // left_offset == right_offset. So let's limit the value
3260         // passed in for left offset here.
3261         BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
3262         scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
3263       }
3264     }
3265 
3266     // Update the marked bytes for this region.
3267     hr->add_to_marked_bytes(marked_bytes);
3268 
3269     // Next heap region
3270     return false;
3271   }
3272 };
3273 
3274 class G1AggregateCountDataTask: public AbstractGangTask {
3275 protected:
3276   G1CollectedHeap* _g1h;
3277   ConcurrentMark* _cm;
3278   BitMap* _cm_card_bm;
3279   uint _max_worker_id;
3280   int _active_workers;
3281 
3282 public:
3283   G1AggregateCountDataTask(G1CollectedHeap* g1h,
3284                            ConcurrentMark* cm,
3285                            BitMap* cm_card_bm,
3286                            uint max_worker_id,
3287                            int n_workers) :
3288     AbstractGangTask("Count Aggregation"),
3289     _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
3290     _max_worker_id(max_worker_id),
3291     _active_workers(n_workers) { }
3292 
3293   void work(uint worker_id) {
3294     AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
3295 
3296     if (G1CollectedHeap::use_parallel_gc_threads()) {
3297       _g1h->heap_region_par_iterate_chunked(&cl, worker_id,
3298                                             _active_workers,
3299                                             HeapRegion::AggregateCountClaimValue);
3300     } else {
3301       _g1h->heap_region_iterate(&cl);
3302     }
3303   }
3304 };
3305 
3306 
3307 void ConcurrentMark::aggregate_count_data() {
3308   int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ?
3309                         _g1h->workers()->active_workers() :
3310                         1);
3311 
3312   G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
3313                                            _max_worker_id, n_workers);
3314 
3315   if (G1CollectedHeap::use_parallel_gc_threads()) {
3316     assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
3317            "sanity check");
3318     _g1h->set_par_threads(n_workers);
3319     _g1h->workers()->run_task(&g1_par_agg_task);
3320     _g1h->set_par_threads(0);
3321 
3322     assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue),
3323            "sanity check");
3324     _g1h->reset_heap_region_claim_values();
3325   } else {
3326     g1_par_agg_task.work(0);
3327   }
3328 }
3329 
3330 // Clear the per-worker arrays used to store the per-region counting data
3331 void ConcurrentMark::clear_all_count_data() {
3332   // Clear the global card bitmap - it will be filled during
3333   // liveness count aggregation (during remark) and the
3334   // final counting task.
3335   _card_bm.clear();
3336 
3337   // Clear the global region bitmap - it will be filled as part
3338   // of the final counting task.
3339   _region_bm.clear();
3340 
3341   uint max_regions = _g1h->max_regions();
3342   assert(_max_worker_id > 0, "uninitialized");
3343 
3344   for (uint i = 0; i < _max_worker_id; i += 1) {
3345     BitMap* task_card_bm = count_card_bitmap_for(i);
3346     size_t* marked_bytes_array = count_marked_bytes_array_for(i);
3347 
3348     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
3349     assert(marked_bytes_array != NULL, "uninitialized");
3350 
3351     memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
3352     task_card_bm->clear();
3353   }
3354 }
3355 
3356 void ConcurrentMark::print_stats() {
3357   if (verbose_stats()) {
3358     gclog_or_tty->print_cr("---------------------------------------------------------------------");
3359     for (size_t i = 0; i < _active_tasks; ++i) {
3360       _tasks[i]->print_stats();
3361       gclog_or_tty->print_cr("---------------------------------------------------------------------");
3362     }
3363   }
3364 }
3365 
3366 // abandon current marking iteration due to a Full GC
3367 void ConcurrentMark::abort() {
3368   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
3369   // concurrent bitmap clearing.
3370   _nextMarkBitMap->clearAll();
3371 
3372   // Note we cannot clear the previous marking bitmap here
3373   // since VerifyDuringGC verifies the objects marked during
3374   // a full GC against the previous bitmap.
3375 
3376   // Clear the liveness counting data
3377   clear_all_count_data();
3378   // Empty mark stack
3379   reset_marking_state();
3380   for (uint i = 0; i < _max_worker_id; ++i) {
3381     _tasks[i]->clear_region_fields();
3382   }
3383   _first_overflow_barrier_sync.abort();
3384   _second_overflow_barrier_sync.abort();
3385   const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id();
3386   if (!gc_id.is_undefined()) {
3387     // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance
3388     // to detect that it was aborted. Only keep track of the first GC id that we aborted.
3389     _aborted_gc_id = gc_id;
3390    }
3391   _has_aborted = true;
3392 
3393   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3394   satb_mq_set.abandon_partial_marking();
3395   // This can be called either during or outside marking, we'll read
3396   // the expected_active value from the SATB queue set.
3397   satb_mq_set.set_active_all_threads(
3398                                  false, /* new active value */
3399                                  satb_mq_set.is_active() /* expected_active */);
3400 
3401   _g1h->trace_heap_after_concurrent_cycle();
3402   _g1h->register_concurrent_cycle_end();
3403 }
3404 
3405 const GCId& ConcurrentMark::concurrent_gc_id() {
3406   if (has_aborted()) {
3407     return _aborted_gc_id;
3408   }
3409   return _g1h->gc_tracer_cm()->gc_id();
3410 }
3411 
3412 static void print_ms_time_info(const char* prefix, const char* name,
3413                                NumberSeq& ns) {
3414   gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
3415                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
3416   if (ns.num() > 0) {
3417     gclog_or_tty->print_cr("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
3418                            prefix, ns.sd(), ns.maximum());
3419   }
3420 }
3421 
3422 void ConcurrentMark::print_summary_info() {
3423   gclog_or_tty->print_cr(" Concurrent marking:");
3424   print_ms_time_info("  ", "init marks", _init_times);
3425   print_ms_time_info("  ", "remarks", _remark_times);
3426   {
3427     print_ms_time_info("     ", "final marks", _remark_mark_times);
3428     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
3429 
3430   }
3431   print_ms_time_info("  ", "cleanups", _cleanup_times);
3432   gclog_or_tty->print_cr("    Final counting total time = %8.2f s (avg = %8.2f ms).",
3433                          _total_counting_time,
3434                          (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 /
3435                           (double)_cleanup_times.num()
3436                          : 0.0));
3437   if (G1ScrubRemSets) {
3438     gclog_or_tty->print_cr("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
3439                            _total_rs_scrub_time,
3440                            (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 /
3441                             (double)_cleanup_times.num()
3442                            : 0.0));
3443   }
3444   gclog_or_tty->print_cr("  Total stop_world time = %8.2f s.",
3445                          (_init_times.sum() + _remark_times.sum() +
3446                           _cleanup_times.sum())/1000.0);
3447   gclog_or_tty->print_cr("  Total concurrent time = %8.2f s "
3448                 "(%8.2f s marking).",
3449                 cmThread()->vtime_accum(),
3450                 cmThread()->vtime_mark_accum());
3451 }
3452 
3453 void ConcurrentMark::print_worker_threads_on(outputStream* st) const {
3454   if (use_parallel_marking_threads()) {
3455     _parallel_workers->print_worker_threads_on(st);
3456   }
3457 }
3458 
3459 void ConcurrentMark::print_on_error(outputStream* st) const {
3460   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
3461       p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
3462   _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
3463   _nextMarkBitMap->print_on_error(st, " Next Bits: ");
3464 }
3465 
3466 // We take a break if someone is trying to stop the world.
3467 bool ConcurrentMark::do_yield_check(uint worker_id) {
3468   if (SuspendibleThreadSet::should_yield()) {
3469     if (worker_id == 0) {
3470       _g1h->g1_policy()->record_concurrent_pause();
3471     }
3472     SuspendibleThreadSet::yield();
3473     return true;
3474   } else {
3475     return false;
3476   }
3477 }
3478 
3479 #ifndef PRODUCT
3480 // for debugging purposes
3481 void ConcurrentMark::print_finger() {
3482   gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT,
3483                          p2i(_heap_start), p2i(_heap_end), p2i(_finger));
3484   for (uint i = 0; i < _max_worker_id; ++i) {
3485     gclog_or_tty->print("   %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger()));
3486   }
3487   gclog_or_tty->cr();
3488 }
3489 #endif
3490 
3491 void CMTask::scan_object(oop obj) {
3492   assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
3493 
3494   if (_cm->verbose_high()) {
3495     gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
3496                            _worker_id, p2i((void*) obj));
3497   }
3498 
3499   size_t obj_size = obj->size();
3500   _words_scanned += obj_size;
3501 
3502   obj->oop_iterate(_cm_oop_closure);
3503   statsOnly( ++_objs_scanned );
3504   check_limits();
3505 }
3506 
3507 // Closure for iteration over bitmaps
3508 class CMBitMapClosure : public BitMapClosure {
3509 private:
3510   // the bitmap that is being iterated over
3511   CMBitMap*                   _nextMarkBitMap;
3512   ConcurrentMark*             _cm;
3513   CMTask*                     _task;
3514 
3515 public:
3516   CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) :
3517     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
3518 
3519   bool do_bit(size_t offset) {
3520     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
3521     assert(_nextMarkBitMap->isMarked(addr), "invariant");
3522     assert( addr < _cm->finger(), "invariant");
3523 
3524     statsOnly( _task->increase_objs_found_on_bitmap() );
3525     assert(addr >= _task->finger(), "invariant");
3526 
3527     // We move that task's local finger along.
3528     _task->move_finger_to(addr);
3529 
3530     _task->scan_object(oop(addr));
3531     // we only partially drain the local queue and global stack
3532     _task->drain_local_queue(true);
3533     _task->drain_global_stack(true);
3534 
3535     // if the has_aborted flag has been raised, we need to bail out of
3536     // the iteration
3537     return !_task->has_aborted();
3538   }
3539 };
3540 
3541 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
3542                                ConcurrentMark* cm,
3543                                CMTask* task)
3544   : _g1h(g1h), _cm(cm), _task(task) {
3545   assert(_ref_processor == NULL, "should be initialized to NULL");
3546 
3547   if (G1UseConcMarkReferenceProcessing) {
3548     _ref_processor = g1h->ref_processor_cm();
3549     assert(_ref_processor != NULL, "should not be NULL");
3550   }
3551 }
3552 
3553 void CMTask::setup_for_region(HeapRegion* hr) {
3554   assert(hr != NULL,
3555         "claim_region() should have filtered out NULL regions");
3556   assert(!hr->continuesHumongous(),
3557         "claim_region() should have filtered out continues humongous regions");
3558 
3559   if (_cm->verbose_low()) {
3560     gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT,
3561                            _worker_id, p2i(hr));
3562   }
3563 
3564   _curr_region  = hr;
3565   _finger       = hr->bottom();
3566   update_region_limit();
3567 }
3568 
3569 void CMTask::update_region_limit() {
3570   HeapRegion* hr            = _curr_region;
3571   HeapWord* bottom          = hr->bottom();
3572   HeapWord* limit           = hr->next_top_at_mark_start();
3573 
3574   if (limit == bottom) {
3575     if (_cm->verbose_low()) {
3576       gclog_or_tty->print_cr("[%u] found an empty region "
3577                              "["PTR_FORMAT", "PTR_FORMAT")",
3578                              _worker_id, p2i(bottom), p2i(limit));
3579     }
3580     // The region was collected underneath our feet.
3581     // We set the finger to bottom to ensure that the bitmap
3582     // iteration that will follow this will not do anything.
3583     // (this is not a condition that holds when we set the region up,
3584     // as the region is not supposed to be empty in the first place)
3585     _finger = bottom;
3586   } else if (limit >= _region_limit) {
3587     assert(limit >= _finger, "peace of mind");
3588   } else {
3589     assert(limit < _region_limit, "only way to get here");
3590     // This can happen under some pretty unusual circumstances.  An
3591     // evacuation pause empties the region underneath our feet (NTAMS
3592     // at bottom). We then do some allocation in the region (NTAMS
3593     // stays at bottom), followed by the region being used as a GC
3594     // alloc region (NTAMS will move to top() and the objects
3595     // originally below it will be grayed). All objects now marked in
3596     // the region are explicitly grayed, if below the global finger,
3597     // and we do not need in fact to scan anything else. So, we simply
3598     // set _finger to be limit to ensure that the bitmap iteration
3599     // doesn't do anything.
3600     _finger = limit;
3601   }
3602 
3603   _region_limit = limit;
3604 }
3605 
3606 void CMTask::giveup_current_region() {
3607   assert(_curr_region != NULL, "invariant");
3608   if (_cm->verbose_low()) {
3609     gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT,
3610                            _worker_id, p2i(_curr_region));
3611   }
3612   clear_region_fields();
3613 }
3614 
3615 void CMTask::clear_region_fields() {
3616   // Values for these three fields that indicate that we're not
3617   // holding on to a region.
3618   _curr_region   = NULL;
3619   _finger        = NULL;
3620   _region_limit  = NULL;
3621 }
3622 
3623 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
3624   if (cm_oop_closure == NULL) {
3625     assert(_cm_oop_closure != NULL, "invariant");
3626   } else {
3627     assert(_cm_oop_closure == NULL, "invariant");
3628   }
3629   _cm_oop_closure = cm_oop_closure;
3630 }
3631 
3632 void CMTask::reset(CMBitMap* nextMarkBitMap) {
3633   guarantee(nextMarkBitMap != NULL, "invariant");
3634 
3635   if (_cm->verbose_low()) {
3636     gclog_or_tty->print_cr("[%u] resetting", _worker_id);
3637   }
3638 
3639   _nextMarkBitMap                = nextMarkBitMap;
3640   clear_region_fields();
3641 
3642   _calls                         = 0;
3643   _elapsed_time_ms               = 0.0;
3644   _termination_time_ms           = 0.0;
3645   _termination_start_time_ms     = 0.0;
3646 
3647 #if _MARKING_STATS_
3648   _local_pushes                  = 0;
3649   _local_pops                    = 0;
3650   _local_max_size                = 0;
3651   _objs_scanned                  = 0;
3652   _global_pushes                 = 0;
3653   _global_pops                   = 0;
3654   _global_max_size               = 0;
3655   _global_transfers_to           = 0;
3656   _global_transfers_from         = 0;
3657   _regions_claimed               = 0;
3658   _objs_found_on_bitmap          = 0;
3659   _satb_buffers_processed        = 0;
3660   _steal_attempts                = 0;
3661   _steals                        = 0;
3662   _aborted                       = 0;
3663   _aborted_overflow              = 0;
3664   _aborted_cm_aborted            = 0;
3665   _aborted_yield                 = 0;
3666   _aborted_timed_out             = 0;
3667   _aborted_satb                  = 0;
3668   _aborted_termination           = 0;
3669 #endif // _MARKING_STATS_
3670 }
3671 
3672 bool CMTask::should_exit_termination() {
3673   regular_clock_call();
3674   // This is called when we are in the termination protocol. We should
3675   // quit if, for some reason, this task wants to abort or the global
3676   // stack is not empty (this means that we can get work from it).
3677   return !_cm->mark_stack_empty() || has_aborted();
3678 }
3679 
3680 void CMTask::reached_limit() {
3681   assert(_words_scanned >= _words_scanned_limit ||
3682          _refs_reached >= _refs_reached_limit ,
3683          "shouldn't have been called otherwise");
3684   regular_clock_call();
3685 }
3686 
3687 void CMTask::regular_clock_call() {
3688   if (has_aborted()) return;
3689 
3690   // First, we need to recalculate the words scanned and refs reached
3691   // limits for the next clock call.
3692   recalculate_limits();
3693 
3694   // During the regular clock call we do the following
3695 
3696   // (1) If an overflow has been flagged, then we abort.
3697   if (_cm->has_overflown()) {
3698     set_has_aborted();
3699     return;
3700   }
3701 
3702   // If we are not concurrent (i.e. we're doing remark) we don't need
3703   // to check anything else. The other steps are only needed during
3704   // the concurrent marking phase.
3705   if (!concurrent()) return;
3706 
3707   // (2) If marking has been aborted for Full GC, then we also abort.
3708   if (_cm->has_aborted()) {
3709     set_has_aborted();
3710     statsOnly( ++_aborted_cm_aborted );
3711     return;
3712   }
3713 
3714   double curr_time_ms = os::elapsedVTime() * 1000.0;
3715 
3716   // (3) If marking stats are enabled, then we update the step history.
3717 #if _MARKING_STATS_
3718   if (_words_scanned >= _words_scanned_limit) {
3719     ++_clock_due_to_scanning;
3720   }
3721   if (_refs_reached >= _refs_reached_limit) {
3722     ++_clock_due_to_marking;
3723   }
3724 
3725   double last_interval_ms = curr_time_ms - _interval_start_time_ms;
3726   _interval_start_time_ms = curr_time_ms;
3727   _all_clock_intervals_ms.add(last_interval_ms);
3728 
3729   if (_cm->verbose_medium()) {
3730       gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, "
3731                         "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s",
3732                         _worker_id, last_interval_ms,
3733                         _words_scanned,
3734                         (_words_scanned >= _words_scanned_limit) ? " (*)" : "",
3735                         _refs_reached,
3736                         (_refs_reached >= _refs_reached_limit) ? " (*)" : "");
3737   }
3738 #endif // _MARKING_STATS_
3739 
3740   // (4) We check whether we should yield. If we have to, then we abort.
3741   if (SuspendibleThreadSet::should_yield()) {
3742     // We should yield. To do this we abort the task. The caller is
3743     // responsible for yielding.
3744     set_has_aborted();
3745     statsOnly( ++_aborted_yield );
3746     return;
3747   }
3748 
3749   // (5) We check whether we've reached our time quota. If we have,
3750   // then we abort.
3751   double elapsed_time_ms = curr_time_ms - _start_time_ms;
3752   if (elapsed_time_ms > _time_target_ms) {
3753     set_has_aborted();
3754     _has_timed_out = true;
3755     statsOnly( ++_aborted_timed_out );
3756     return;
3757   }
3758 
3759   // (6) Finally, we check whether there are enough completed STAB
3760   // buffers available for processing. If there are, we abort.
3761   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3762   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
3763     if (_cm->verbose_low()) {
3764       gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers",
3765                              _worker_id);
3766     }
3767     // we do need to process SATB buffers, we'll abort and restart
3768     // the marking task to do so
3769     set_has_aborted();
3770     statsOnly( ++_aborted_satb );
3771     return;
3772   }
3773 }
3774 
3775 void CMTask::recalculate_limits() {
3776   _real_words_scanned_limit = _words_scanned + words_scanned_period;
3777   _words_scanned_limit      = _real_words_scanned_limit;
3778 
3779   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
3780   _refs_reached_limit       = _real_refs_reached_limit;
3781 }
3782 
3783 void CMTask::decrease_limits() {
3784   // This is called when we believe that we're going to do an infrequent
3785   // operation which will increase the per byte scanned cost (i.e. move
3786   // entries to/from the global stack). It basically tries to decrease the
3787   // scanning limit so that the clock is called earlier.
3788 
3789   if (_cm->verbose_medium()) {
3790     gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id);
3791   }
3792 
3793   _words_scanned_limit = _real_words_scanned_limit -
3794     3 * words_scanned_period / 4;
3795   _refs_reached_limit  = _real_refs_reached_limit -
3796     3 * refs_reached_period / 4;
3797 }
3798 
3799 void CMTask::move_entries_to_global_stack() {
3800   // local array where we'll store the entries that will be popped
3801   // from the local queue
3802   oop buffer[global_stack_transfer_size];
3803 
3804   int n = 0;
3805   oop obj;
3806   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
3807     buffer[n] = obj;
3808     ++n;
3809   }
3810 
3811   if (n > 0) {
3812     // we popped at least one entry from the local queue
3813 
3814     statsOnly( ++_global_transfers_to; _local_pops += n );
3815 
3816     if (!_cm->mark_stack_push(buffer, n)) {
3817       if (_cm->verbose_low()) {
3818         gclog_or_tty->print_cr("[%u] aborting due to global stack overflow",
3819                                _worker_id);
3820       }
3821       set_has_aborted();
3822     } else {
3823       // the transfer was successful
3824 
3825       if (_cm->verbose_medium()) {
3826         gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack",
3827                                _worker_id, n);
3828       }
3829       statsOnly( int tmp_size = _cm->mark_stack_size();
3830                  if (tmp_size > _global_max_size) {
3831                    _global_max_size = tmp_size;
3832                  }
3833                  _global_pushes += n );
3834     }
3835   }
3836 
3837   // this operation was quite expensive, so decrease the limits
3838   decrease_limits();
3839 }
3840 
3841 void CMTask::get_entries_from_global_stack() {
3842   // local array where we'll store the entries that will be popped
3843   // from the global stack.
3844   oop buffer[global_stack_transfer_size];
3845   int n;
3846   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
3847   assert(n <= global_stack_transfer_size,
3848          "we should not pop more than the given limit");
3849   if (n > 0) {
3850     // yes, we did actually pop at least one entry
3851 
3852     statsOnly( ++_global_transfers_from; _global_pops += n );
3853     if (_cm->verbose_medium()) {
3854       gclog_or_tty->print_cr("[%u] popped %d entries from the global stack",
3855                              _worker_id, n);
3856     }
3857     for (int i = 0; i < n; ++i) {
3858       bool success = _task_queue->push(buffer[i]);
3859       // We only call this when the local queue is empty or under a
3860       // given target limit. So, we do not expect this push to fail.
3861       assert(success, "invariant");
3862     }
3863 
3864     statsOnly( int tmp_size = _task_queue->size();
3865                if (tmp_size > _local_max_size) {
3866                  _local_max_size = tmp_size;
3867                }
3868                _local_pushes += n );
3869   }
3870 
3871   // this operation was quite expensive, so decrease the limits
3872   decrease_limits();
3873 }
3874 
3875 void CMTask::drain_local_queue(bool partially) {
3876   if (has_aborted()) return;
3877 
3878   // Decide what the target size is, depending whether we're going to
3879   // drain it partially (so that other tasks can steal if they run out
3880   // of things to do) or totally (at the very end).
3881   size_t target_size;
3882   if (partially) {
3883     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
3884   } else {
3885     target_size = 0;
3886   }
3887 
3888   if (_task_queue->size() > target_size) {
3889     if (_cm->verbose_high()) {
3890       gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT,
3891                              _worker_id, target_size);
3892     }
3893 
3894     oop obj;
3895     bool ret = _task_queue->pop_local(obj);
3896     while (ret) {
3897       statsOnly( ++_local_pops );
3898 
3899       if (_cm->verbose_high()) {
3900         gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id,
3901                                p2i((void*) obj));
3902       }
3903 
3904       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
3905       assert(!_g1h->is_on_master_free_list(
3906                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
3907 
3908       scan_object(obj);
3909 
3910       if (_task_queue->size() <= target_size || has_aborted()) {
3911         ret = false;
3912       } else {
3913         ret = _task_queue->pop_local(obj);
3914       }
3915     }
3916 
3917     if (_cm->verbose_high()) {
3918       gclog_or_tty->print_cr("[%u] drained local queue, size = %u",
3919                              _worker_id, _task_queue->size());
3920     }
3921   }
3922 }
3923 
3924 void CMTask::drain_global_stack(bool partially) {
3925   if (has_aborted()) return;
3926 
3927   // We have a policy to drain the local queue before we attempt to
3928   // drain the global stack.
3929   assert(partially || _task_queue->size() == 0, "invariant");
3930 
3931   // Decide what the target size is, depending whether we're going to
3932   // drain it partially (so that other tasks can steal if they run out
3933   // of things to do) or totally (at the very end).  Notice that,
3934   // because we move entries from the global stack in chunks or
3935   // because another task might be doing the same, we might in fact
3936   // drop below the target. But, this is not a problem.
3937   size_t target_size;
3938   if (partially) {
3939     target_size = _cm->partial_mark_stack_size_target();
3940   } else {
3941     target_size = 0;
3942   }
3943 
3944   if (_cm->mark_stack_size() > target_size) {
3945     if (_cm->verbose_low()) {
3946       gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT,
3947                              _worker_id, target_size);
3948     }
3949 
3950     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
3951       get_entries_from_global_stack();
3952       drain_local_queue(partially);
3953     }
3954 
3955     if (_cm->verbose_low()) {
3956       gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT,
3957                              _worker_id, _cm->mark_stack_size());
3958     }
3959   }
3960 }
3961 
3962 // SATB Queue has several assumptions on whether to call the par or
3963 // non-par versions of the methods. this is why some of the code is
3964 // replicated. We should really get rid of the single-threaded version
3965 // of the code to simplify things.
3966 void CMTask::drain_satb_buffers() {
3967   if (has_aborted()) return;
3968 
3969   // We set this so that the regular clock knows that we're in the
3970   // middle of draining buffers and doesn't set the abort flag when it
3971   // notices that SATB buffers are available for draining. It'd be
3972   // very counter productive if it did that. :-)
3973   _draining_satb_buffers = true;
3974 
3975   CMObjectClosure oc(this);
3976   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3977   if (G1CollectedHeap::use_parallel_gc_threads()) {
3978     satb_mq_set.set_par_closure(_worker_id, &oc);
3979   } else {
3980     satb_mq_set.set_closure(&oc);
3981   }
3982 
3983   // This keeps claiming and applying the closure to completed buffers
3984   // until we run out of buffers or we need to abort.
3985   if (G1CollectedHeap::use_parallel_gc_threads()) {
3986     while (!has_aborted() &&
3987            satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) {
3988       if (_cm->verbose_medium()) {
3989         gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3990       }
3991       statsOnly( ++_satb_buffers_processed );
3992       regular_clock_call();
3993     }
3994   } else {
3995     while (!has_aborted() &&
3996            satb_mq_set.apply_closure_to_completed_buffer()) {
3997       if (_cm->verbose_medium()) {
3998         gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id);
3999       }
4000       statsOnly( ++_satb_buffers_processed );
4001       regular_clock_call();
4002     }
4003   }
4004 
4005   _draining_satb_buffers = false;
4006 
4007   assert(has_aborted() ||
4008          concurrent() ||
4009          satb_mq_set.completed_buffers_num() == 0, "invariant");
4010 
4011   if (G1CollectedHeap::use_parallel_gc_threads()) {
4012     satb_mq_set.set_par_closure(_worker_id, NULL);
4013   } else {
4014     satb_mq_set.set_closure(NULL);
4015   }
4016 
4017   // again, this was a potentially expensive operation, decrease the
4018   // limits to get the regular clock call early
4019   decrease_limits();
4020 }
4021 
4022 void CMTask::print_stats() {
4023   gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d",
4024                          _worker_id, _calls);
4025   gclog_or_tty->print_cr("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
4026                          _elapsed_time_ms, _termination_time_ms);
4027   gclog_or_tty->print_cr("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4028                          _step_times_ms.num(), _step_times_ms.avg(),
4029                          _step_times_ms.sd());
4030   gclog_or_tty->print_cr("                    max = %1.2lfms, total = %1.2lfms",
4031                          _step_times_ms.maximum(), _step_times_ms.sum());
4032 
4033 #if _MARKING_STATS_
4034   gclog_or_tty->print_cr("  Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
4035                          _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(),
4036                          _all_clock_intervals_ms.sd());
4037   gclog_or_tty->print_cr("                         max = %1.2lfms, total = %1.2lfms",
4038                          _all_clock_intervals_ms.maximum(),
4039                          _all_clock_intervals_ms.sum());
4040   gclog_or_tty->print_cr("  Clock Causes (cum): scanning = %d, marking = %d",
4041                          _clock_due_to_scanning, _clock_due_to_marking);
4042   gclog_or_tty->print_cr("  Objects: scanned = %d, found on the bitmap = %d",
4043                          _objs_scanned, _objs_found_on_bitmap);
4044   gclog_or_tty->print_cr("  Local Queue:  pushes = %d, pops = %d, max size = %d",
4045                          _local_pushes, _local_pops, _local_max_size);
4046   gclog_or_tty->print_cr("  Global Stack: pushes = %d, pops = %d, max size = %d",
4047                          _global_pushes, _global_pops, _global_max_size);
4048   gclog_or_tty->print_cr("                transfers to = %d, transfers from = %d",
4049                          _global_transfers_to,_global_transfers_from);
4050   gclog_or_tty->print_cr("  Regions: claimed = %d", _regions_claimed);
4051   gclog_or_tty->print_cr("  SATB buffers: processed = %d", _satb_buffers_processed);
4052   gclog_or_tty->print_cr("  Steals: attempts = %d, successes = %d",
4053                          _steal_attempts, _steals);
4054   gclog_or_tty->print_cr("  Aborted: %d, due to", _aborted);
4055   gclog_or_tty->print_cr("    overflow: %d, global abort: %d, yield: %d",
4056                          _aborted_overflow, _aborted_cm_aborted, _aborted_yield);
4057   gclog_or_tty->print_cr("    time out: %d, SATB: %d, termination: %d",
4058                          _aborted_timed_out, _aborted_satb, _aborted_termination);
4059 #endif // _MARKING_STATS_
4060 }
4061 
4062 /*****************************************************************************
4063 
4064     The do_marking_step(time_target_ms, ...) method is the building
4065     block of the parallel marking framework. It can be called in parallel
4066     with other invocations of do_marking_step() on different tasks
4067     (but only one per task, obviously) and concurrently with the
4068     mutator threads, or during remark, hence it eliminates the need
4069     for two versions of the code. When called during remark, it will
4070     pick up from where the task left off during the concurrent marking
4071     phase. Interestingly, tasks are also claimable during evacuation
4072     pauses too, since do_marking_step() ensures that it aborts before
4073     it needs to yield.
4074 
4075     The data structures that it uses to do marking work are the
4076     following:
4077 
4078       (1) Marking Bitmap. If there are gray objects that appear only
4079       on the bitmap (this happens either when dealing with an overflow
4080       or when the initial marking phase has simply marked the roots
4081       and didn't push them on the stack), then tasks claim heap
4082       regions whose bitmap they then scan to find gray objects. A
4083       global finger indicates where the end of the last claimed region
4084       is. A local finger indicates how far into the region a task has
4085       scanned. The two fingers are used to determine how to gray an
4086       object (i.e. whether simply marking it is OK, as it will be
4087       visited by a task in the future, or whether it needs to be also
4088       pushed on a stack).
4089 
4090       (2) Local Queue. The local queue of the task which is accessed
4091       reasonably efficiently by the task. Other tasks can steal from
4092       it when they run out of work. Throughout the marking phase, a
4093       task attempts to keep its local queue short but not totally
4094       empty, so that entries are available for stealing by other
4095       tasks. Only when there is no more work, a task will totally
4096       drain its local queue.
4097 
4098       (3) Global Mark Stack. This handles local queue overflow. During
4099       marking only sets of entries are moved between it and the local
4100       queues, as access to it requires a mutex and more fine-grain
4101       interaction with it which might cause contention. If it
4102       overflows, then the marking phase should restart and iterate
4103       over the bitmap to identify gray objects. Throughout the marking
4104       phase, tasks attempt to keep the global mark stack at a small
4105       length but not totally empty, so that entries are available for
4106       popping by other tasks. Only when there is no more work, tasks
4107       will totally drain the global mark stack.
4108 
4109       (4) SATB Buffer Queue. This is where completed SATB buffers are
4110       made available. Buffers are regularly removed from this queue
4111       and scanned for roots, so that the queue doesn't get too
4112       long. During remark, all completed buffers are processed, as
4113       well as the filled in parts of any uncompleted buffers.
4114 
4115     The do_marking_step() method tries to abort when the time target
4116     has been reached. There are a few other cases when the
4117     do_marking_step() method also aborts:
4118 
4119       (1) When the marking phase has been aborted (after a Full GC).
4120 
4121       (2) When a global overflow (on the global stack) has been
4122       triggered. Before the task aborts, it will actually sync up with
4123       the other tasks to ensure that all the marking data structures
4124       (local queues, stacks, fingers etc.)  are re-initialized so that
4125       when do_marking_step() completes, the marking phase can
4126       immediately restart.
4127 
4128       (3) When enough completed SATB buffers are available. The
4129       do_marking_step() method only tries to drain SATB buffers right
4130       at the beginning. So, if enough buffers are available, the
4131       marking step aborts and the SATB buffers are processed at
4132       the beginning of the next invocation.
4133 
4134       (4) To yield. when we have to yield then we abort and yield
4135       right at the end of do_marking_step(). This saves us from a lot
4136       of hassle as, by yielding we might allow a Full GC. If this
4137       happens then objects will be compacted underneath our feet, the
4138       heap might shrink, etc. We save checking for this by just
4139       aborting and doing the yield right at the end.
4140 
4141     From the above it follows that the do_marking_step() method should
4142     be called in a loop (or, otherwise, regularly) until it completes.
4143 
4144     If a marking step completes without its has_aborted() flag being
4145     true, it means it has completed the current marking phase (and
4146     also all other marking tasks have done so and have all synced up).
4147 
4148     A method called regular_clock_call() is invoked "regularly" (in
4149     sub ms intervals) throughout marking. It is this clock method that
4150     checks all the abort conditions which were mentioned above and
4151     decides when the task should abort. A work-based scheme is used to
4152     trigger this clock method: when the number of object words the
4153     marking phase has scanned or the number of references the marking
4154     phase has visited reach a given limit. Additional invocations to
4155     the method clock have been planted in a few other strategic places
4156     too. The initial reason for the clock method was to avoid calling
4157     vtime too regularly, as it is quite expensive. So, once it was in
4158     place, it was natural to piggy-back all the other conditions on it
4159     too and not constantly check them throughout the code.
4160 
4161     If do_termination is true then do_marking_step will enter its
4162     termination protocol.
4163 
4164     The value of is_serial must be true when do_marking_step is being
4165     called serially (i.e. by the VMThread) and do_marking_step should
4166     skip any synchronization in the termination and overflow code.
4167     Examples include the serial remark code and the serial reference
4168     processing closures.
4169 
4170     The value of is_serial must be false when do_marking_step is
4171     being called by any of the worker threads in a work gang.
4172     Examples include the concurrent marking code (CMMarkingTask),
4173     the MT remark code, and the MT reference processing closures.
4174 
4175  *****************************************************************************/
4176 
4177 void CMTask::do_marking_step(double time_target_ms,
4178                              bool do_termination,
4179                              bool is_serial) {
4180   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
4181   assert(concurrent() == _cm->concurrent(), "they should be the same");
4182 
4183   G1CollectorPolicy* g1_policy = _g1h->g1_policy();
4184   assert(_task_queues != NULL, "invariant");
4185   assert(_task_queue != NULL, "invariant");
4186   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
4187 
4188   assert(!_claimed,
4189          "only one thread should claim this task at any one time");
4190 
4191   // OK, this doesn't safeguard again all possible scenarios, as it is
4192   // possible for two threads to set the _claimed flag at the same
4193   // time. But it is only for debugging purposes anyway and it will
4194   // catch most problems.
4195   _claimed = true;
4196 
4197   _start_time_ms = os::elapsedVTime() * 1000.0;
4198   statsOnly( _interval_start_time_ms = _start_time_ms );
4199 
4200   // If do_stealing is true then do_marking_step will attempt to
4201   // steal work from the other CMTasks. It only makes sense to
4202   // enable stealing when the termination protocol is enabled
4203   // and do_marking_step() is not being called serially.
4204   bool do_stealing = do_termination && !is_serial;
4205 
4206   double diff_prediction_ms =
4207     g1_policy->get_new_prediction(&_marking_step_diffs_ms);
4208   _time_target_ms = time_target_ms - diff_prediction_ms;
4209 
4210   // set up the variables that are used in the work-based scheme to
4211   // call the regular clock method
4212   _words_scanned = 0;
4213   _refs_reached  = 0;
4214   recalculate_limits();
4215 
4216   // clear all flags
4217   clear_has_aborted();
4218   _has_timed_out = false;
4219   _draining_satb_buffers = false;
4220 
4221   ++_calls;
4222 
4223   if (_cm->verbose_low()) {
4224     gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, "
4225                            "target = %1.2lfms >>>>>>>>>>",
4226                            _worker_id, _calls, _time_target_ms);
4227   }
4228 
4229   // Set up the bitmap and oop closures. Anything that uses them is
4230   // eventually called from this method, so it is OK to allocate these
4231   // statically.
4232   CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
4233   G1CMOopClosure  cm_oop_closure(_g1h, _cm, this);
4234   set_cm_oop_closure(&cm_oop_closure);
4235 
4236   if (_cm->has_overflown()) {
4237     // This can happen if the mark stack overflows during a GC pause
4238     // and this task, after a yield point, restarts. We have to abort
4239     // as we need to get into the overflow protocol which happens
4240     // right at the end of this task.
4241     set_has_aborted();
4242   }
4243 
4244   // First drain any available SATB buffers. After this, we will not
4245   // look at SATB buffers before the next invocation of this method.
4246   // If enough completed SATB buffers are queued up, the regular clock
4247   // will abort this task so that it restarts.
4248   drain_satb_buffers();
4249   // ...then partially drain the local queue and the global stack
4250   drain_local_queue(true);
4251   drain_global_stack(true);
4252 
4253   do {
4254     if (!has_aborted() && _curr_region != NULL) {
4255       // This means that we're already holding on to a region.
4256       assert(_finger != NULL, "if region is not NULL, then the finger "
4257              "should not be NULL either");
4258 
4259       // We might have restarted this task after an evacuation pause
4260       // which might have evacuated the region we're holding on to
4261       // underneath our feet. Let's read its limit again to make sure
4262       // that we do not iterate over a region of the heap that
4263       // contains garbage (update_region_limit() will also move
4264       // _finger to the start of the region if it is found empty).
4265       update_region_limit();
4266       // We will start from _finger not from the start of the region,
4267       // as we might be restarting this task after aborting half-way
4268       // through scanning this region. In this case, _finger points to
4269       // the address where we last found a marked object. If this is a
4270       // fresh region, _finger points to start().
4271       MemRegion mr = MemRegion(_finger, _region_limit);
4272 
4273       if (_cm->verbose_low()) {
4274         gclog_or_tty->print_cr("[%u] we're scanning part "
4275                                "["PTR_FORMAT", "PTR_FORMAT") "
4276                                "of region "HR_FORMAT,
4277                                _worker_id, p2i(_finger), p2i(_region_limit),
4278                                HR_FORMAT_PARAMS(_curr_region));
4279       }
4280 
4281       assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(),
4282              "humongous regions should go around loop once only");
4283 
4284       // Some special cases:
4285       // If the memory region is empty, we can just give up the region.
4286       // If the current region is humongous then we only need to check
4287       // the bitmap for the bit associated with the start of the object,
4288       // scan the object if it's live, and give up the region.
4289       // Otherwise, let's iterate over the bitmap of the part of the region
4290       // that is left.
4291       // If the iteration is successful, give up the region.
4292       if (mr.is_empty()) {
4293         giveup_current_region();
4294         regular_clock_call();
4295       } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) {
4296         if (_nextMarkBitMap->isMarked(mr.start())) {
4297           // The object is marked - apply the closure
4298           BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
4299           bitmap_closure.do_bit(offset);
4300         }
4301         // Even if this task aborted while scanning the humongous object
4302         // we can (and should) give up the current region.
4303         giveup_current_region();
4304         regular_clock_call();
4305       } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
4306         giveup_current_region();
4307         regular_clock_call();
4308       } else {
4309         assert(has_aborted(), "currently the only way to do so");
4310         // The only way to abort the bitmap iteration is to return
4311         // false from the do_bit() method. However, inside the
4312         // do_bit() method we move the _finger to point to the
4313         // object currently being looked at. So, if we bail out, we
4314         // have definitely set _finger to something non-null.
4315         assert(_finger != NULL, "invariant");
4316 
4317         // Region iteration was actually aborted. So now _finger
4318         // points to the address of the object we last scanned. If we
4319         // leave it there, when we restart this task, we will rescan
4320         // the object. It is easy to avoid this. We move the finger by
4321         // enough to point to the next possible object header (the
4322         // bitmap knows by how much we need to move it as it knows its
4323         // granularity).
4324         assert(_finger < _region_limit, "invariant");
4325         HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
4326         // Check if bitmap iteration was aborted while scanning the last object
4327         if (new_finger >= _region_limit) {
4328           giveup_current_region();
4329         } else {
4330           move_finger_to(new_finger);
4331         }
4332       }
4333     }
4334     // At this point we have either completed iterating over the
4335     // region we were holding on to, or we have aborted.
4336 
4337     // We then partially drain the local queue and the global stack.
4338     // (Do we really need this?)
4339     drain_local_queue(true);
4340     drain_global_stack(true);
4341 
4342     // Read the note on the claim_region() method on why it might
4343     // return NULL with potentially more regions available for
4344     // claiming and why we have to check out_of_regions() to determine
4345     // whether we're done or not.
4346     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
4347       // We are going to try to claim a new region. We should have
4348       // given up on the previous one.
4349       // Separated the asserts so that we know which one fires.
4350       assert(_curr_region  == NULL, "invariant");
4351       assert(_finger       == NULL, "invariant");
4352       assert(_region_limit == NULL, "invariant");
4353       if (_cm->verbose_low()) {
4354         gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id);
4355       }
4356       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
4357       if (claimed_region != NULL) {
4358         // Yes, we managed to claim one
4359         statsOnly( ++_regions_claimed );
4360 
4361         if (_cm->verbose_low()) {
4362           gclog_or_tty->print_cr("[%u] we successfully claimed "
4363                                  "region "PTR_FORMAT,
4364                                  _worker_id, p2i(claimed_region));
4365         }
4366 
4367         setup_for_region(claimed_region);
4368         assert(_curr_region == claimed_region, "invariant");
4369       }
4370       // It is important to call the regular clock here. It might take
4371       // a while to claim a region if, for example, we hit a large
4372       // block of empty regions. So we need to call the regular clock
4373       // method once round the loop to make sure it's called
4374       // frequently enough.
4375       regular_clock_call();
4376     }
4377 
4378     if (!has_aborted() && _curr_region == NULL) {
4379       assert(_cm->out_of_regions(),
4380              "at this point we should be out of regions");
4381     }
4382   } while ( _curr_region != NULL && !has_aborted());
4383 
4384   if (!has_aborted()) {
4385     // We cannot check whether the global stack is empty, since other
4386     // tasks might be pushing objects to it concurrently.
4387     assert(_cm->out_of_regions(),
4388            "at this point we should be out of regions");
4389 
4390     if (_cm->verbose_low()) {
4391       gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id);
4392     }
4393 
4394     // Try to reduce the number of available SATB buffers so that
4395     // remark has less work to do.
4396     drain_satb_buffers();
4397   }
4398 
4399   // Since we've done everything else, we can now totally drain the
4400   // local queue and global stack.
4401   drain_local_queue(false);
4402   drain_global_stack(false);
4403 
4404   // Attempt at work stealing from other task's queues.
4405   if (do_stealing && !has_aborted()) {
4406     // We have not aborted. This means that we have finished all that
4407     // we could. Let's try to do some stealing...
4408 
4409     // We cannot check whether the global stack is empty, since other
4410     // tasks might be pushing objects to it concurrently.
4411     assert(_cm->out_of_regions() && _task_queue->size() == 0,
4412            "only way to reach here");
4413 
4414     if (_cm->verbose_low()) {
4415       gclog_or_tty->print_cr("[%u] starting to steal", _worker_id);
4416     }
4417 
4418     while (!has_aborted()) {
4419       oop obj;
4420       statsOnly( ++_steal_attempts );
4421 
4422       if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
4423         if (_cm->verbose_medium()) {
4424           gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully",
4425                                  _worker_id, p2i((void*) obj));
4426         }
4427 
4428         statsOnly( ++_steals );
4429 
4430         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
4431                "any stolen object should be marked");
4432         scan_object(obj);
4433 
4434         // And since we're towards the end, let's totally drain the
4435         // local queue and global stack.
4436         drain_local_queue(false);
4437         drain_global_stack(false);
4438       } else {
4439         break;
4440       }
4441     }
4442   }
4443 
4444   // If we are about to wrap up and go into termination, check if we
4445   // should raise the overflow flag.
4446   if (do_termination && !has_aborted()) {
4447     if (_cm->force_overflow()->should_force()) {
4448       _cm->set_has_overflown();
4449       regular_clock_call();
4450     }
4451   }
4452 
4453   // We still haven't aborted. Now, let's try to get into the
4454   // termination protocol.
4455   if (do_termination && !has_aborted()) {
4456     // We cannot check whether the global stack is empty, since other
4457     // tasks might be concurrently pushing objects on it.
4458     // Separated the asserts so that we know which one fires.
4459     assert(_cm->out_of_regions(), "only way to reach here");
4460     assert(_task_queue->size() == 0, "only way to reach here");
4461 
4462     if (_cm->verbose_low()) {
4463       gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id);
4464     }
4465 
4466     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
4467 
4468     // The CMTask class also extends the TerminatorTerminator class,
4469     // hence its should_exit_termination() method will also decide
4470     // whether to exit the termination protocol or not.
4471     bool finished = (is_serial ||
4472                      _cm->terminator()->offer_termination(this));
4473     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
4474     _termination_time_ms +=
4475       termination_end_time_ms - _termination_start_time_ms;
4476 
4477     if (finished) {
4478       // We're all done.
4479 
4480       if (_worker_id == 0) {
4481         // let's allow task 0 to do this
4482         if (concurrent()) {
4483           assert(_cm->concurrent_marking_in_progress(), "invariant");
4484           // we need to set this to false before the next
4485           // safepoint. This way we ensure that the marking phase
4486           // doesn't observe any more heap expansions.
4487           _cm->clear_concurrent_marking_in_progress();
4488         }
4489       }
4490 
4491       // We can now guarantee that the global stack is empty, since
4492       // all other tasks have finished. We separated the guarantees so
4493       // that, if a condition is false, we can immediately find out
4494       // which one.
4495       guarantee(_cm->out_of_regions(), "only way to reach here");
4496       guarantee(_cm->mark_stack_empty(), "only way to reach here");
4497       guarantee(_task_queue->size() == 0, "only way to reach here");
4498       guarantee(!_cm->has_overflown(), "only way to reach here");
4499       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
4500 
4501       if (_cm->verbose_low()) {
4502         gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id);
4503       }
4504     } else {
4505       // Apparently there's more work to do. Let's abort this task. It
4506       // will restart it and we can hopefully find more things to do.
4507 
4508       if (_cm->verbose_low()) {
4509         gclog_or_tty->print_cr("[%u] apparently there is more work to do",
4510                                _worker_id);
4511       }
4512 
4513       set_has_aborted();
4514       statsOnly( ++_aborted_termination );
4515     }
4516   }
4517 
4518   // Mainly for debugging purposes to make sure that a pointer to the
4519   // closure which was statically allocated in this frame doesn't
4520   // escape it by accident.
4521   set_cm_oop_closure(NULL);
4522   double end_time_ms = os::elapsedVTime() * 1000.0;
4523   double elapsed_time_ms = end_time_ms - _start_time_ms;
4524   // Update the step history.
4525   _step_times_ms.add(elapsed_time_ms);
4526 
4527   if (has_aborted()) {
4528     // The task was aborted for some reason.
4529 
4530     statsOnly( ++_aborted );
4531 
4532     if (_has_timed_out) {
4533       double diff_ms = elapsed_time_ms - _time_target_ms;
4534       // Keep statistics of how well we did with respect to hitting
4535       // our target only if we actually timed out (if we aborted for
4536       // other reasons, then the results might get skewed).
4537       _marking_step_diffs_ms.add(diff_ms);
4538     }
4539 
4540     if (_cm->has_overflown()) {
4541       // This is the interesting one. We aborted because a global
4542       // overflow was raised. This means we have to restart the
4543       // marking phase and start iterating over regions. However, in
4544       // order to do this we have to make sure that all tasks stop
4545       // what they are doing and re-initialize in a safe manner. We
4546       // will achieve this with the use of two barrier sync points.
4547 
4548       if (_cm->verbose_low()) {
4549         gclog_or_tty->print_cr("[%u] detected overflow", _worker_id);
4550       }
4551 
4552       if (!is_serial) {
4553         // We only need to enter the sync barrier if being called
4554         // from a parallel context
4555         _cm->enter_first_sync_barrier(_worker_id);
4556 
4557         // When we exit this sync barrier we know that all tasks have
4558         // stopped doing marking work. So, it's now safe to
4559         // re-initialize our data structures. At the end of this method,
4560         // task 0 will clear the global data structures.
4561       }
4562 
4563       statsOnly( ++_aborted_overflow );
4564 
4565       // We clear the local state of this task...
4566       clear_region_fields();
4567 
4568       if (!is_serial) {
4569         // ...and enter the second barrier.
4570         _cm->enter_second_sync_barrier(_worker_id);
4571       }
4572       // At this point, if we're during the concurrent phase of
4573       // marking, everything has been re-initialized and we're
4574       // ready to restart.
4575     }
4576 
4577     if (_cm->verbose_low()) {
4578       gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, "
4579                              "elapsed = %1.2lfms <<<<<<<<<<",
4580                              _worker_id, _time_target_ms, elapsed_time_ms);
4581       if (_cm->has_aborted()) {
4582         gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========",
4583                                _worker_id);
4584       }
4585     }
4586   } else {
4587     if (_cm->verbose_low()) {
4588       gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, "
4589                              "elapsed = %1.2lfms <<<<<<<<<<",
4590                              _worker_id, _time_target_ms, elapsed_time_ms);
4591     }
4592   }
4593 
4594   _claimed = false;
4595 }
4596 
4597 CMTask::CMTask(uint worker_id,
4598                ConcurrentMark* cm,
4599                size_t* marked_bytes,
4600                BitMap* card_bm,
4601                CMTaskQueue* task_queue,
4602                CMTaskQueueSet* task_queues)
4603   : _g1h(G1CollectedHeap::heap()),
4604     _worker_id(worker_id), _cm(cm),
4605     _claimed(false),
4606     _nextMarkBitMap(NULL), _hash_seed(17),
4607     _task_queue(task_queue),
4608     _task_queues(task_queues),
4609     _cm_oop_closure(NULL),
4610     _marked_bytes_array(marked_bytes),
4611     _card_bm(card_bm) {
4612   guarantee(task_queue != NULL, "invariant");
4613   guarantee(task_queues != NULL, "invariant");
4614 
4615   statsOnly( _clock_due_to_scanning = 0;
4616              _clock_due_to_marking  = 0 );
4617 
4618   _marking_step_diffs_ms.add(0.5);
4619 }
4620 
4621 // These are formatting macros that are used below to ensure
4622 // consistent formatting. The *_H_* versions are used to format the
4623 // header for a particular value and they should be kept consistent
4624 // with the corresponding macro. Also note that most of the macros add
4625 // the necessary white space (as a prefix) which makes them a bit
4626 // easier to compose.
4627 
4628 // All the output lines are prefixed with this string to be able to
4629 // identify them easily in a large log file.
4630 #define G1PPRL_LINE_PREFIX            "###"
4631 
4632 #define G1PPRL_ADDR_BASE_FORMAT    " "PTR_FORMAT"-"PTR_FORMAT
4633 #ifdef _LP64
4634 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
4635 #else // _LP64
4636 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
4637 #endif // _LP64
4638 
4639 // For per-region info
4640 #define G1PPRL_TYPE_FORMAT            "   %-4s"
4641 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
4642 #define G1PPRL_BYTE_FORMAT            "  "SIZE_FORMAT_W(9)
4643 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
4644 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
4645 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
4646 
4647 // For summary info
4648 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  "tag":"G1PPRL_ADDR_BASE_FORMAT
4649 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  "tag": "SIZE_FORMAT
4650 #define G1PPRL_SUM_MB_FORMAT(tag)      "  "tag": %1.2f MB"
4651 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%"
4652 
4653 G1PrintRegionLivenessInfoClosure::
4654 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name)
4655   : _out(out),
4656     _total_used_bytes(0), _total_capacity_bytes(0),
4657     _total_prev_live_bytes(0), _total_next_live_bytes(0),
4658     _hum_used_bytes(0), _hum_capacity_bytes(0),
4659     _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
4660     _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
4661   G1CollectedHeap* g1h = G1CollectedHeap::heap();
4662   MemRegion g1_reserved = g1h->g1_reserved();
4663   double now = os::elapsedTime();
4664 
4665   // Print the header of the output.
4666   _out->cr();
4667   _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
4668   _out->print_cr(G1PPRL_LINE_PREFIX" HEAP"
4669                  G1PPRL_SUM_ADDR_FORMAT("reserved")
4670                  G1PPRL_SUM_BYTE_FORMAT("region-size"),
4671                  p2i(g1_reserved.start()), p2i(g1_reserved.end()),
4672                  HeapRegion::GrainBytes);
4673   _out->print_cr(G1PPRL_LINE_PREFIX);
4674   _out->print_cr(G1PPRL_LINE_PREFIX
4675                 G1PPRL_TYPE_H_FORMAT
4676                 G1PPRL_ADDR_BASE_H_FORMAT
4677                 G1PPRL_BYTE_H_FORMAT
4678                 G1PPRL_BYTE_H_FORMAT
4679                 G1PPRL_BYTE_H_FORMAT
4680                 G1PPRL_DOUBLE_H_FORMAT
4681                 G1PPRL_BYTE_H_FORMAT
4682                 G1PPRL_BYTE_H_FORMAT,
4683                 "type", "address-range",
4684                 "used", "prev-live", "next-live", "gc-eff",
4685                 "remset", "code-roots");
4686   _out->print_cr(G1PPRL_LINE_PREFIX
4687                 G1PPRL_TYPE_H_FORMAT
4688                 G1PPRL_ADDR_BASE_H_FORMAT
4689                 G1PPRL_BYTE_H_FORMAT
4690                 G1PPRL_BYTE_H_FORMAT
4691                 G1PPRL_BYTE_H_FORMAT
4692                 G1PPRL_DOUBLE_H_FORMAT
4693                 G1PPRL_BYTE_H_FORMAT
4694                 G1PPRL_BYTE_H_FORMAT,
4695                 "", "",
4696                 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
4697                 "(bytes)", "(bytes)");
4698 }
4699 
4700 // It takes as a parameter a reference to one of the _hum_* fields, it
4701 // deduces the corresponding value for a region in a humongous region
4702 // series (either the region size, or what's left if the _hum_* field
4703 // is < the region size), and updates the _hum_* field accordingly.
4704 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
4705   size_t bytes = 0;
4706   // The > 0 check is to deal with the prev and next live bytes which
4707   // could be 0.
4708   if (*hum_bytes > 0) {
4709     bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
4710     *hum_bytes -= bytes;
4711   }
4712   return bytes;
4713 }
4714 
4715 // It deduces the values for a region in a humongous region series
4716 // from the _hum_* fields and updates those accordingly. It assumes
4717 // that that _hum_* fields have already been set up from the "starts
4718 // humongous" region and we visit the regions in address order.
4719 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
4720                                                      size_t* capacity_bytes,
4721                                                      size_t* prev_live_bytes,
4722                                                      size_t* next_live_bytes) {
4723   assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
4724   *used_bytes      = get_hum_bytes(&_hum_used_bytes);
4725   *capacity_bytes  = get_hum_bytes(&_hum_capacity_bytes);
4726   *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
4727   *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
4728 }
4729 
4730 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
4731   const char* type = "";
4732   HeapWord* bottom       = r->bottom();
4733   HeapWord* end          = r->end();
4734   size_t capacity_bytes  = r->capacity();
4735   size_t used_bytes      = r->used();
4736   size_t prev_live_bytes = r->live_bytes();
4737   size_t next_live_bytes = r->next_live_bytes();
4738   double gc_eff          = r->gc_efficiency();
4739   size_t remset_bytes    = r->rem_set()->mem_size();
4740   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
4741 
4742   if (r->used() == 0) {
4743     type = "FREE";
4744   } else if (r->is_survivor()) {
4745     type = "SURV";
4746   } else if (r->is_young()) {
4747     type = "EDEN";
4748   } else if (r->startsHumongous()) {
4749     type = "HUMS";
4750 
4751     assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
4752            _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
4753            "they should have been zeroed after the last time we used them");
4754     // Set up the _hum_* fields.
4755     _hum_capacity_bytes  = capacity_bytes;
4756     _hum_used_bytes      = used_bytes;
4757     _hum_prev_live_bytes = prev_live_bytes;
4758     _hum_next_live_bytes = next_live_bytes;
4759     get_hum_bytes(&used_bytes, &capacity_bytes,
4760                   &prev_live_bytes, &next_live_bytes);
4761     end = bottom + HeapRegion::GrainWords;
4762   } else if (r->continuesHumongous()) {
4763     type = "HUMC";
4764     get_hum_bytes(&used_bytes, &capacity_bytes,
4765                   &prev_live_bytes, &next_live_bytes);
4766     assert(end == bottom + HeapRegion::GrainWords, "invariant");
4767   } else {
4768     type = "OLD";
4769   }
4770 
4771   _total_used_bytes      += used_bytes;
4772   _total_capacity_bytes  += capacity_bytes;
4773   _total_prev_live_bytes += prev_live_bytes;
4774   _total_next_live_bytes += next_live_bytes;
4775   _total_remset_bytes    += remset_bytes;
4776   _total_strong_code_roots_bytes += strong_code_roots_bytes;
4777 
4778   // Print a line for this particular region.
4779   _out->print_cr(G1PPRL_LINE_PREFIX
4780                  G1PPRL_TYPE_FORMAT
4781                  G1PPRL_ADDR_BASE_FORMAT
4782                  G1PPRL_BYTE_FORMAT
4783                  G1PPRL_BYTE_FORMAT
4784                  G1PPRL_BYTE_FORMAT
4785                  G1PPRL_DOUBLE_FORMAT
4786                  G1PPRL_BYTE_FORMAT
4787                  G1PPRL_BYTE_FORMAT,
4788                  type, p2i(bottom), p2i(end),
4789                  used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
4790                  remset_bytes, strong_code_roots_bytes);
4791 
4792   return false;
4793 }
4794 
4795 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
4796   // add static memory usages to remembered set sizes
4797   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
4798   // Print the footer of the output.
4799   _out->print_cr(G1PPRL_LINE_PREFIX);
4800   _out->print_cr(G1PPRL_LINE_PREFIX
4801                  " SUMMARY"
4802                  G1PPRL_SUM_MB_FORMAT("capacity")
4803                  G1PPRL_SUM_MB_PERC_FORMAT("used")
4804                  G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
4805                  G1PPRL_SUM_MB_PERC_FORMAT("next-live")
4806                  G1PPRL_SUM_MB_FORMAT("remset")
4807                  G1PPRL_SUM_MB_FORMAT("code-roots"),
4808                  bytes_to_mb(_total_capacity_bytes),
4809                  bytes_to_mb(_total_used_bytes),
4810                  perc(_total_used_bytes, _total_capacity_bytes),
4811                  bytes_to_mb(_total_prev_live_bytes),
4812                  perc(_total_prev_live_bytes, _total_capacity_bytes),
4813                  bytes_to_mb(_total_next_live_bytes),
4814                  perc(_total_next_live_bytes, _total_capacity_bytes),
4815                  bytes_to_mb(_total_remset_bytes),
4816                  bytes_to_mb(_total_strong_code_roots_bytes));
4817   _out->cr();
4818 }