1 /*
   2  * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/metadataOnStackMark.hpp"
  27 #include "classfile/symbolTable.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc/g1/concurrentMarkThread.inline.hpp"
  30 #include "gc/g1/g1CollectedHeap.inline.hpp"
  31 #include "gc/g1/g1CollectorPolicy.hpp"
  32 #include "gc/g1/g1CollectorState.hpp"
  33 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  34 #include "gc/g1/g1HeapVerifier.hpp"
  35 #include "gc/g1/g1OopClosures.inline.hpp"
  36 #include "gc/g1/g1StringDedup.hpp"
  37 #include "gc/g1/heapRegion.inline.hpp"
  38 #include "gc/g1/heapRegionRemSet.hpp"
  39 #include "gc/g1/heapRegionSet.inline.hpp"
  40 #include "gc/g1/suspendibleThreadSet.hpp"
  41 #include "gc/shared/gcId.hpp"
  42 #include "gc/shared/gcTimer.hpp"
  43 #include "gc/shared/gcTrace.hpp"
  44 #include "gc/shared/gcTraceTime.inline.hpp"
  45 #include "gc/shared/genOopClosures.inline.hpp"
  46 #include "gc/shared/referencePolicy.hpp"
  47 #include "gc/shared/strongRootsScope.hpp"
  48 #include "gc/shared/taskqueue.inline.hpp"
  49 #include "gc/shared/vmGCOperations.hpp"
  50 #include "logging/log.hpp"
  51 #include "memory/allocation.hpp"
  52 #include "memory/resourceArea.hpp"
  53 #include "oops/oop.inline.hpp"
  54 #include "runtime/atomic.inline.hpp"
  55 #include "runtime/handles.inline.hpp"
  56 #include "runtime/java.hpp"
  57 #include "runtime/prefetch.inline.hpp"
  58 #include "services/memTracker.hpp"
  59 
  60 // Concurrent marking bit map wrapper
  61 
  62 G1CMBitMapRO::G1CMBitMapRO(int shifter) :
  63   _bm(),
  64   _shifter(shifter) {
  65   _bmStartWord = 0;
  66   _bmWordSize = 0;
  67 }
  68 
  69 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr,
  70                                                  const HeapWord* limit) const {
  71   // First we must round addr *up* to a possible object boundary.
  72   addr = (HeapWord*)align_size_up((intptr_t)addr,
  73                                   HeapWordSize << _shifter);
  74   size_t addrOffset = heapWordToOffset(addr);
  75   assert(limit != NULL, "limit must not be NULL");
  76   size_t limitOffset = heapWordToOffset(limit);
  77   size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset);
  78   HeapWord* nextAddr = offsetToHeapWord(nextOffset);
  79   assert(nextAddr >= addr, "get_next_one postcondition");
  80   assert(nextAddr == limit || isMarked(nextAddr),
  81          "get_next_one postcondition");
  82   return nextAddr;
  83 }
  84 
  85 #ifndef PRODUCT
  86 bool G1CMBitMapRO::covers(MemRegion heap_rs) const {
  87   // assert(_bm.map() == _virtual_space.low(), "map inconsistency");
  88   assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize,
  89          "size inconsistency");
  90   return _bmStartWord == (HeapWord*)(heap_rs.start()) &&
  91          _bmWordSize  == heap_rs.word_size();
  92 }
  93 #endif
  94 
  95 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const {
  96   _bm.print_on_error(st, prefix);
  97 }
  98 
  99 size_t G1CMBitMap::compute_size(size_t heap_size) {
 100   return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
 101 }
 102 
 103 size_t G1CMBitMap::mark_distance() {
 104   return MinObjAlignmentInBytes * BitsPerByte;
 105 }
 106 
 107 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
 108   _bmStartWord = heap.start();
 109   _bmWordSize = heap.word_size();
 110 
 111   _bm.set_map((BitMap::bm_word_t*) storage->reserved().start());
 112   _bm.set_size(_bmWordSize >> _shifter);
 113 
 114   storage->set_mapping_changed_listener(&_listener);
 115 }
 116 
 117 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
 118   if (zero_filled) {
 119     return;
 120   }
 121   // We need to clear the bitmap on commit, removing any existing information.
 122   MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
 123   _bm->clearRange(mr);
 124 }
 125 
 126 // Closure used for clearing the given mark bitmap.
 127 class ClearBitmapHRClosure : public HeapRegionClosure {
 128  private:
 129   G1ConcurrentMark* _cm;
 130   G1CMBitMap* _bitmap;
 131   bool _may_yield;      // The closure may yield during iteration. If yielded, abort the iteration.
 132  public:
 133   ClearBitmapHRClosure(G1ConcurrentMark* cm, G1CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) {
 134     assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield.");
 135   }
 136 
 137   virtual bool doHeapRegion(HeapRegion* r) {
 138     size_t const chunk_size_in_words = M / HeapWordSize;
 139 
 140     HeapWord* cur = r->bottom();
 141     HeapWord* const end = r->end();
 142 
 143     while (cur < end) {
 144       MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 145       _bitmap->clearRange(mr);
 146 
 147       cur += chunk_size_in_words;
 148 
 149       // Abort iteration if after yielding the marking has been aborted.
 150       if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) {
 151         return true;
 152       }
 153       // Repeat the asserts from before the start of the closure. We will do them
 154       // as asserts here to minimize their overhead on the product. However, we
 155       // will have them as guarantees at the beginning / end of the bitmap
 156       // clearing to get some checking in the product.
 157       assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant");
 158       assert(!_may_yield || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
 159     }
 160 
 161     return false;
 162   }
 163 };
 164 
 165 class ParClearNextMarkBitmapTask : public AbstractGangTask {
 166   ClearBitmapHRClosure* _cl;
 167   HeapRegionClaimer     _hrclaimer;
 168   bool                  _suspendible; // If the task is suspendible, workers must join the STS.
 169 
 170 public:
 171   ParClearNextMarkBitmapTask(ClearBitmapHRClosure *cl, uint n_workers, bool suspendible) :
 172       _cl(cl), _suspendible(suspendible), AbstractGangTask("Parallel Clear Bitmap Task"), _hrclaimer(n_workers) {}
 173 
 174   void work(uint worker_id) {
 175     SuspendibleThreadSetJoiner sts_join(_suspendible);
 176     G1CollectedHeap::heap()->heap_region_par_iterate(_cl, worker_id, &_hrclaimer, true);
 177   }
 178 };
 179 
 180 void G1CMBitMap::clearAll() {
 181   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 182   ClearBitmapHRClosure cl(NULL, this, false /* may_yield */);
 183   uint n_workers = g1h->workers()->active_workers();
 184   ParClearNextMarkBitmapTask task(&cl, n_workers, false);
 185   g1h->workers()->run_task(&task);
 186   guarantee(cl.complete(), "Must have completed iteration.");
 187   return;
 188 }
 189 
 190 void G1CMBitMap::clearRange(MemRegion mr) {
 191   mr.intersection(MemRegion(_bmStartWord, _bmWordSize));
 192   assert(!mr.is_empty(), "unexpected empty region");
 193   // convert address range into offset range
 194   _bm.at_put_range(heapWordToOffset(mr.start()),
 195                    heapWordToOffset(mr.end()), false);
 196 }
 197 
 198 G1CMMarkStack::G1CMMarkStack(G1ConcurrentMark* cm) :
 199   _base(NULL), _cm(cm)
 200 {}
 201 
 202 bool G1CMMarkStack::allocate(size_t capacity) {
 203   // allocate a stack of the requisite depth
 204   ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop)));
 205   if (!rs.is_reserved()) {
 206     warning("ConcurrentMark MarkStack allocation failure");
 207     return false;
 208   }
 209   MemTracker::record_virtual_memory_type((address)rs.base(), mtGC);
 210   if (!_virtual_space.initialize(rs, rs.size())) {
 211     warning("ConcurrentMark MarkStack backing store failure");
 212     // Release the virtual memory reserved for the marking stack
 213     rs.release();
 214     return false;
 215   }
 216   assert(_virtual_space.committed_size() == rs.size(),
 217          "Didn't reserve backing store for all of G1ConcurrentMark stack?");
 218   _base = (oop*) _virtual_space.low();
 219   setEmpty();
 220   _capacity = (jint) capacity;
 221   _saved_index = -1;
 222   _should_expand = false;
 223   return true;
 224 }
 225 
 226 void G1CMMarkStack::expand() {
 227   // Called, during remark, if we've overflown the marking stack during marking.
 228   assert(isEmpty(), "stack should been emptied while handling overflow");
 229   assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted");
 230   // Clear expansion flag
 231   _should_expand = false;
 232   if (_capacity == (jint) MarkStackSizeMax) {
 233     log_trace(gc)("(benign) Can't expand marking stack capacity, at max size limit");
 234     return;
 235   }
 236   // Double capacity if possible
 237   jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax);
 238   // Do not give up existing stack until we have managed to
 239   // get the double capacity that we desired.
 240   ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity *
 241                                                            sizeof(oop)));
 242   if (rs.is_reserved()) {
 243     // Release the backing store associated with old stack
 244     _virtual_space.release();
 245     // Reinitialize virtual space for new stack
 246     if (!_virtual_space.initialize(rs, rs.size())) {
 247       fatal("Not enough swap for expanded marking stack capacity");
 248     }
 249     _base = (oop*)(_virtual_space.low());
 250     _index = 0;
 251     _capacity = new_capacity;
 252   } else {
 253     // Failed to double capacity, continue;
 254     log_trace(gc)("(benign) Failed to expand marking stack capacity from " SIZE_FORMAT "K to " SIZE_FORMAT "K",
 255                   _capacity / K, new_capacity / K);
 256   }
 257 }
 258 
 259 void G1CMMarkStack::set_should_expand() {
 260   // If we're resetting the marking state because of an
 261   // marking stack overflow, record that we should, if
 262   // possible, expand the stack.
 263   _should_expand = _cm->has_overflown();
 264 }
 265 
 266 G1CMMarkStack::~G1CMMarkStack() {
 267   if (_base != NULL) {
 268     _base = NULL;
 269     _virtual_space.release();
 270   }
 271 }
 272 
 273 void G1CMMarkStack::par_push_arr(oop* ptr_arr, int n) {
 274   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 275   jint start = _index;
 276   jint next_index = start + n;
 277   if (next_index > _capacity) {
 278     _overflow = true;
 279     return;
 280   }
 281   // Otherwise.
 282   _index = next_index;
 283   for (int i = 0; i < n; i++) {
 284     int ind = start + i;
 285     assert(ind < _capacity, "By overflow test above.");
 286     _base[ind] = ptr_arr[i];
 287   }
 288 }
 289 
 290 bool G1CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) {
 291   MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
 292   jint index = _index;
 293   if (index == 0) {
 294     *n = 0;
 295     return false;
 296   } else {
 297     int k = MIN2(max, index);
 298     jint  new_ind = index - k;
 299     for (int j = 0; j < k; j++) {
 300       ptr_arr[j] = _base[new_ind + j];
 301     }
 302     _index = new_ind;
 303     *n = k;
 304     return true;
 305   }
 306 }
 307 
 308 void G1CMMarkStack::note_start_of_gc() {
 309   assert(_saved_index == -1,
 310          "note_start_of_gc()/end_of_gc() bracketed incorrectly");
 311   _saved_index = _index;
 312 }
 313 
 314 void G1CMMarkStack::note_end_of_gc() {
 315   // This is intentionally a guarantee, instead of an assert. If we
 316   // accidentally add something to the mark stack during GC, it
 317   // will be a correctness issue so it's better if we crash. we'll
 318   // only check this once per GC anyway, so it won't be a performance
 319   // issue in any way.
 320   guarantee(_saved_index == _index,
 321             "saved index: %d index: %d", _saved_index, _index);
 322   _saved_index = -1;
 323 }
 324 
 325 G1CMRootRegions::G1CMRootRegions() :
 326   _young_list(NULL), _cm(NULL), _scan_in_progress(false),
 327   _should_abort(false),  _next_survivor(NULL) { }
 328 
 329 void G1CMRootRegions::init(G1CollectedHeap* g1h, G1ConcurrentMark* cm) {
 330   _young_list = g1h->young_list();
 331   _cm = cm;
 332 }
 333 
 334 void G1CMRootRegions::prepare_for_scan() {
 335   assert(!scan_in_progress(), "pre-condition");
 336 
 337   // Currently, only survivors can be root regions.
 338   assert(_next_survivor == NULL, "pre-condition");
 339   _next_survivor = _young_list->first_survivor_region();
 340   _scan_in_progress = (_next_survivor != NULL);
 341   _should_abort = false;
 342 }
 343 
 344 HeapRegion* G1CMRootRegions::claim_next() {
 345   if (_should_abort) {
 346     // If someone has set the should_abort flag, we return NULL to
 347     // force the caller to bail out of their loop.
 348     return NULL;
 349   }
 350 
 351   // Currently, only survivors can be root regions.
 352   HeapRegion* res = _next_survivor;
 353   if (res != NULL) {
 354     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 355     // Read it again in case it changed while we were waiting for the lock.
 356     res = _next_survivor;
 357     if (res != NULL) {
 358       if (res == _young_list->last_survivor_region()) {
 359         // We just claimed the last survivor so store NULL to indicate
 360         // that we're done.
 361         _next_survivor = NULL;
 362       } else {
 363         _next_survivor = res->get_next_young_region();
 364       }
 365     } else {
 366       // Someone else claimed the last survivor while we were trying
 367       // to take the lock so nothing else to do.
 368     }
 369   }
 370   assert(res == NULL || res->is_survivor(), "post-condition");
 371 
 372   return res;
 373 }
 374 
 375 void G1CMRootRegions::notify_scan_done() {
 376   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 377   _scan_in_progress = false;
 378   RootRegionScan_lock->notify_all();
 379 }
 380 
 381 void G1CMRootRegions::cancel_scan() {
 382   notify_scan_done();
 383 }
 384 
 385 void G1CMRootRegions::scan_finished() {
 386   assert(scan_in_progress(), "pre-condition");
 387 
 388   // Currently, only survivors can be root regions.
 389   if (!_should_abort) {
 390     assert(_next_survivor == NULL, "we should have claimed all survivors");
 391   }
 392   _next_survivor = NULL;
 393 
 394   notify_scan_done();
 395 }
 396 
 397 bool G1CMRootRegions::wait_until_scan_finished() {
 398   if (!scan_in_progress()) return false;
 399 
 400   {
 401     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 402     while (scan_in_progress()) {
 403       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 404     }
 405   }
 406   return true;
 407 }
 408 
 409 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 410   return MAX2((n_par_threads + 2) / 4, 1U);
 411 }
 412 
 413 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
 414   _g1h(g1h),
 415   _markBitMap1(),
 416   _markBitMap2(),
 417   _parallel_marking_threads(0),
 418   _max_parallel_marking_threads(0),
 419   _sleep_factor(0.0),
 420   _marking_task_overhead(1.0),
 421   _cleanup_list("Cleanup List"),
 422   _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/),
 423   _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >>
 424             CardTableModRefBS::card_shift,
 425             false /* in_resource_area*/),
 426 
 427   _prevMarkBitMap(&_markBitMap1),
 428   _nextMarkBitMap(&_markBitMap2),
 429 
 430   _markStack(this),
 431   // _finger set in set_non_marking_state
 432 
 433   _max_worker_id(ParallelGCThreads),
 434   // _active_tasks set in set_non_marking_state
 435   // _tasks set inside the constructor
 436   _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)),
 437   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 438 
 439   _has_overflown(false),
 440   _concurrent(false),
 441   _has_aborted(false),
 442   _restart_for_overflow(false),
 443   _concurrent_marking_in_progress(false),
 444   _concurrent_phase_started(false),
 445 
 446   // _verbose_level set below
 447 
 448   _init_times(),
 449   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 450   _cleanup_times(),
 451   _total_counting_time(0.0),
 452   _total_rs_scrub_time(0.0),
 453 
 454   _parallel_workers(NULL),
 455 
 456   _count_card_bitmaps(NULL),
 457   _count_marked_bytes(NULL),
 458   _completed_initialization(false) {
 459 
 460   _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 461   _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
 462 
 463   // Create & start a ConcurrentMark thread.
 464   _cmThread = new ConcurrentMarkThread(this);
 465   assert(cmThread() != NULL, "CM Thread should have been created");
 466   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 467   if (_cmThread->osthread() == NULL) {
 468       vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 469   }
 470 
 471   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 472   assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency");
 473   assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency");
 474 
 475   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 476   satb_qs.set_buffer_size(G1SATBBufferSize);
 477 
 478   _root_regions.init(_g1h, this);
 479 
 480   if (ConcGCThreads > ParallelGCThreads) {
 481     warning("Can't have more ConcGCThreads (%u) "
 482             "than ParallelGCThreads (%u).",
 483             ConcGCThreads, ParallelGCThreads);
 484     return;
 485   }
 486   if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
 487     // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
 488     // if both are set
 489     _sleep_factor             = 0.0;
 490     _marking_task_overhead    = 1.0;
 491   } else if (G1MarkingOverheadPercent > 0) {
 492     // We will calculate the number of parallel marking threads based
 493     // on a target overhead with respect to the soft real-time goal
 494     double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 495     double overall_cm_overhead =
 496       (double) MaxGCPauseMillis * marking_overhead /
 497       (double) GCPauseIntervalMillis;
 498     double cpu_ratio = 1.0 / (double) os::processor_count();
 499     double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 500     double marking_task_overhead =
 501       overall_cm_overhead / marking_thread_num *
 502                                               (double) os::processor_count();
 503     double sleep_factor =
 504                        (1.0 - marking_task_overhead) / marking_task_overhead;
 505 
 506     FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num);
 507     _sleep_factor             = sleep_factor;
 508     _marking_task_overhead    = marking_task_overhead;
 509   } else {
 510     // Calculate the number of parallel marking threads by scaling
 511     // the number of parallel GC threads.
 512     uint marking_thread_num = scale_parallel_threads(ParallelGCThreads);
 513     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 514     _sleep_factor             = 0.0;
 515     _marking_task_overhead    = 1.0;
 516   }
 517 
 518   assert(ConcGCThreads > 0, "Should have been set");
 519   _parallel_marking_threads = ConcGCThreads;
 520   _max_parallel_marking_threads = _parallel_marking_threads;
 521 
 522   _parallel_workers = new WorkGang("G1 Marker",
 523        _max_parallel_marking_threads, false, true);
 524   if (_parallel_workers == NULL) {
 525     vm_exit_during_initialization("Failed necessary allocation.");
 526   } else {
 527     _parallel_workers->initialize_workers();
 528   }
 529 
 530   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 531     size_t mark_stack_size =
 532       MIN2(MarkStackSizeMax,
 533           MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 534     // Verify that the calculated value for MarkStackSize is in range.
 535     // It would be nice to use the private utility routine from Arguments.
 536     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 537       warning("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 538               "must be between 1 and " SIZE_FORMAT,
 539               mark_stack_size, MarkStackSizeMax);
 540       return;
 541     }
 542     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 543   } else {
 544     // Verify MarkStackSize is in range.
 545     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 546       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 547         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 548           warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 549                   "must be between 1 and " SIZE_FORMAT,
 550                   MarkStackSize, MarkStackSizeMax);
 551           return;
 552         }
 553       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 554         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 555           warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 556                   " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 557                   MarkStackSize, MarkStackSizeMax);
 558           return;
 559         }
 560       }
 561     }
 562   }
 563 
 564   if (!_markStack.allocate(MarkStackSize)) {
 565     warning("Failed to allocate CM marking stack");
 566     return;
 567   }
 568 
 569   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC);
 570   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 571 
 572   _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap,  _max_worker_id, mtGC);
 573   _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC);
 574 
 575   BitMap::idx_t card_bm_size = _card_bm.size();
 576 
 577   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 578   _active_tasks = _max_worker_id;
 579 
 580   uint max_regions = _g1h->max_regions();
 581   for (uint i = 0; i < _max_worker_id; ++i) {
 582     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 583     task_queue->initialize();
 584     _task_queues->register_queue(i, task_queue);
 585 
 586     _count_card_bitmaps[i] = BitMap(card_bm_size, false);
 587     _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC);
 588 
 589     _tasks[i] = new G1CMTask(i, this,
 590                              _count_marked_bytes[i],
 591                              &_count_card_bitmaps[i],
 592                              task_queue, _task_queues);
 593 
 594     _accum_task_vtime[i] = 0.0;
 595   }
 596 
 597   // Calculate the card number for the bottom of the heap. Used
 598   // in biasing indexes into the accounting card bitmaps.
 599   _heap_bottom_card_num =
 600     intptr_t(uintptr_t(_g1h->reserved_region().start()) >>
 601                                 CardTableModRefBS::card_shift);
 602 
 603   // Clear all the liveness counting data
 604   clear_all_count_data();
 605 
 606   // so that the call below can read a sensible value
 607   _heap_start = g1h->reserved_region().start();
 608   set_non_marking_state();
 609   _completed_initialization = true;
 610 }
 611 
 612 void G1ConcurrentMark::reset() {
 613   // Starting values for these two. This should be called in a STW
 614   // phase.
 615   MemRegion reserved = _g1h->g1_reserved();
 616   _heap_start = reserved.start();
 617   _heap_end   = reserved.end();
 618 
 619   // Separated the asserts so that we know which one fires.
 620   assert(_heap_start != NULL, "heap bounds should look ok");
 621   assert(_heap_end != NULL, "heap bounds should look ok");
 622   assert(_heap_start < _heap_end, "heap bounds should look ok");
 623 
 624   // Reset all the marking data structures and any necessary flags
 625   reset_marking_state();
 626 
 627   // We do reset all of them, since different phases will use
 628   // different number of active threads. So, it's easiest to have all
 629   // of them ready.
 630   for (uint i = 0; i < _max_worker_id; ++i) {
 631     _tasks[i]->reset(_nextMarkBitMap);
 632   }
 633 
 634   // we need this to make sure that the flag is on during the evac
 635   // pause with initial mark piggy-backed
 636   set_concurrent_marking_in_progress();
 637 }
 638 
 639 
 640 void G1ConcurrentMark::reset_marking_state(bool clear_overflow) {
 641   _markStack.set_should_expand();
 642   _markStack.setEmpty();        // Also clears the _markStack overflow flag
 643   if (clear_overflow) {
 644     clear_has_overflown();
 645   } else {
 646     assert(has_overflown(), "pre-condition");
 647   }
 648   _finger = _heap_start;
 649 
 650   for (uint i = 0; i < _max_worker_id; ++i) {
 651     G1CMTaskQueue* queue = _task_queues->queue(i);
 652     queue->set_empty();
 653   }
 654 }
 655 
 656 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 657   assert(active_tasks <= _max_worker_id, "we should not have more");
 658 
 659   _active_tasks = active_tasks;
 660   // Need to update the three data structures below according to the
 661   // number of active threads for this phase.
 662   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 663   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 664   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 665 }
 666 
 667 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 668   set_concurrency(active_tasks);
 669 
 670   _concurrent = concurrent;
 671   // We propagate this to all tasks, not just the active ones.
 672   for (uint i = 0; i < _max_worker_id; ++i)
 673     _tasks[i]->set_concurrent(concurrent);
 674 
 675   if (concurrent) {
 676     set_concurrent_marking_in_progress();
 677   } else {
 678     // We currently assume that the concurrent flag has been set to
 679     // false before we start remark. At this point we should also be
 680     // in a STW phase.
 681     assert(!concurrent_marking_in_progress(), "invariant");
 682     assert(out_of_regions(),
 683            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 684            p2i(_finger), p2i(_heap_end));
 685   }
 686 }
 687 
 688 void G1ConcurrentMark::set_non_marking_state() {
 689   // We set the global marking state to some default values when we're
 690   // not doing marking.
 691   reset_marking_state();
 692   _active_tasks = 0;
 693   clear_concurrent_marking_in_progress();
 694 }
 695 
 696 G1ConcurrentMark::~G1ConcurrentMark() {
 697   // The G1ConcurrentMark instance is never freed.
 698   ShouldNotReachHere();
 699 }
 700 
 701 void G1ConcurrentMark::clearNextBitmap() {
 702   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 703 
 704   // Make sure that the concurrent mark thread looks to still be in
 705   // the current cycle.
 706   guarantee(cmThread()->during_cycle(), "invariant");
 707 
 708   // We are finishing up the current cycle by clearing the next
 709   // marking bitmap and getting it ready for the next cycle. During
 710   // this time no other cycle can start. So, let's make sure that this
 711   // is the case.
 712   guarantee(!g1h->collector_state()->mark_in_progress(), "invariant");
 713 
 714   ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */);
 715   ParClearNextMarkBitmapTask task(&cl, parallel_marking_threads(), true);
 716   _parallel_workers->run_task(&task);
 717 
 718   // Clear the liveness counting data. If the marking has been aborted, the abort()
 719   // call already did that.
 720   if (cl.complete()) {
 721     clear_all_count_data();
 722   }
 723 
 724   // Repeat the asserts from above.
 725   guarantee(cmThread()->during_cycle(), "invariant");
 726   guarantee(!g1h->collector_state()->mark_in_progress(), "invariant");
 727 }
 728 
 729 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 730   G1CMBitMap* _bitmap;
 731   bool _error;
 732  public:
 733   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 734   }
 735 
 736   virtual bool doHeapRegion(HeapRegion* r) {
 737     // This closure can be called concurrently to the mutator, so we must make sure
 738     // that the result of the getNextMarkedWordAddress() call is compared to the
 739     // value passed to it as limit to detect any found bits.
 740     // end never changes in G1.
 741     HeapWord* end = r->end();
 742     return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end;
 743   }
 744 };
 745 
 746 bool G1ConcurrentMark::nextMarkBitmapIsClear() {
 747   CheckBitmapClearHRClosure cl(_nextMarkBitMap);
 748   _g1h->heap_region_iterate(&cl);
 749   return cl.complete();
 750 }
 751 
 752 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 753 public:
 754   bool doHeapRegion(HeapRegion* r) {
 755     r->note_start_of_marking();
 756     return false;
 757   }
 758 };
 759 
 760 void G1ConcurrentMark::checkpointRootsInitialPre() {
 761   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 762   G1CollectorPolicy* g1p = g1h->g1_policy();
 763 
 764   _has_aborted = false;
 765 
 766   // Initialize marking structures. This has to be done in a STW phase.
 767   reset();
 768 
 769   // For each region note start of marking.
 770   NoteStartOfMarkHRClosure startcl;
 771   g1h->heap_region_iterate(&startcl);
 772 }
 773 
 774 
 775 void G1ConcurrentMark::checkpointRootsInitialPost() {
 776   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 777 
 778   // Start Concurrent Marking weak-reference discovery.
 779   ReferenceProcessor* rp = g1h->ref_processor_cm();
 780   // enable ("weak") refs discovery
 781   rp->enable_discovery();
 782   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 783 
 784   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 785   // This is the start of  the marking cycle, we're expected all
 786   // threads to have SATB queues with active set to false.
 787   satb_mq_set.set_active_all_threads(true, /* new active value */
 788                                      false /* expected_active */);
 789 
 790   _root_regions.prepare_for_scan();
 791 
 792   // update_g1_committed() will be called at the end of an evac pause
 793   // when marking is on. So, it's also called at the end of the
 794   // initial-mark pause to update the heap end, if the heap expands
 795   // during it. No need to call it here.
 796 }
 797 
 798 /*
 799  * Notice that in the next two methods, we actually leave the STS
 800  * during the barrier sync and join it immediately afterwards. If we
 801  * do not do this, the following deadlock can occur: one thread could
 802  * be in the barrier sync code, waiting for the other thread to also
 803  * sync up, whereas another one could be trying to yield, while also
 804  * waiting for the other threads to sync up too.
 805  *
 806  * Note, however, that this code is also used during remark and in
 807  * this case we should not attempt to leave / enter the STS, otherwise
 808  * we'll either hit an assert (debug / fastdebug) or deadlock
 809  * (product). So we should only leave / enter the STS if we are
 810  * operating concurrently.
 811  *
 812  * Because the thread that does the sync barrier has left the STS, it
 813  * is possible to be suspended for a Full GC or an evacuation pause
 814  * could occur. This is actually safe, since the entering the sync
 815  * barrier is one of the last things do_marking_step() does, and it
 816  * doesn't manipulate any data structures afterwards.
 817  */
 818 
 819 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 820   bool barrier_aborted;
 821   {
 822     SuspendibleThreadSetLeaver sts_leave(concurrent());
 823     barrier_aborted = !_first_overflow_barrier_sync.enter();
 824   }
 825 
 826   // at this point everyone should have synced up and not be doing any
 827   // more work
 828 
 829   if (barrier_aborted) {
 830     // If the barrier aborted we ignore the overflow condition and
 831     // just abort the whole marking phase as quickly as possible.
 832     return;
 833   }
 834 
 835   // If we're executing the concurrent phase of marking, reset the marking
 836   // state; otherwise the marking state is reset after reference processing,
 837   // during the remark pause.
 838   // If we reset here as a result of an overflow during the remark we will
 839   // see assertion failures from any subsequent set_concurrency_and_phase()
 840   // calls.
 841   if (concurrent()) {
 842     // let the task associated with with worker 0 do this
 843     if (worker_id == 0) {
 844       // task 0 is responsible for clearing the global data structures
 845       // We should be here because of an overflow. During STW we should
 846       // not clear the overflow flag since we rely on it being true when
 847       // we exit this method to abort the pause and restart concurrent
 848       // marking.
 849       reset_marking_state(true /* clear_overflow */);
 850 
 851       log_info(gc)("Concurrent Mark reset for overflow");
 852     }
 853   }
 854 
 855   // after this, each task should reset its own data structures then
 856   // then go into the second barrier
 857 }
 858 
 859 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 860   SuspendibleThreadSetLeaver sts_leave(concurrent());
 861   _second_overflow_barrier_sync.enter();
 862 
 863   // at this point everything should be re-initialized and ready to go
 864 }
 865 
 866 class G1CMConcurrentMarkingTask: public AbstractGangTask {
 867 private:
 868   G1ConcurrentMark*     _cm;
 869   ConcurrentMarkThread* _cmt;
 870 
 871 public:
 872   void work(uint worker_id) {
 873     assert(Thread::current()->is_ConcurrentGC_thread(),
 874            "this should only be done by a conc GC thread");
 875     ResourceMark rm;
 876 
 877     double start_vtime = os::elapsedVTime();
 878 
 879     {
 880       SuspendibleThreadSetJoiner sts_join;
 881 
 882       assert(worker_id < _cm->active_tasks(), "invariant");
 883       G1CMTask* the_task = _cm->task(worker_id);
 884       the_task->record_start_time();
 885       if (!_cm->has_aborted()) {
 886         do {
 887           double start_vtime_sec = os::elapsedVTime();
 888           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
 889 
 890           the_task->do_marking_step(mark_step_duration_ms,
 891                                     true  /* do_termination */,
 892                                     false /* is_serial*/);
 893 
 894           double end_vtime_sec = os::elapsedVTime();
 895           double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
 896           _cm->clear_has_overflown();
 897 
 898           _cm->do_yield_check(worker_id);
 899 
 900           jlong sleep_time_ms;
 901           if (!_cm->has_aborted() && the_task->has_aborted()) {
 902             sleep_time_ms =
 903               (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
 904             {
 905               SuspendibleThreadSetLeaver sts_leave;
 906               os::sleep(Thread::current(), sleep_time_ms, false);
 907             }
 908           }
 909         } while (!_cm->has_aborted() && the_task->has_aborted());
 910       }
 911       the_task->record_end_time();
 912       guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
 913     }
 914 
 915     double end_vtime = os::elapsedVTime();
 916     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 917   }
 918 
 919   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm,
 920                             ConcurrentMarkThread* cmt) :
 921       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
 922 
 923   ~G1CMConcurrentMarkingTask() { }
 924 };
 925 
 926 // Calculates the number of active workers for a concurrent
 927 // phase.
 928 uint G1ConcurrentMark::calc_parallel_marking_threads() {
 929   uint n_conc_workers = 0;
 930   if (!UseDynamicNumberOfGCThreads ||
 931       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 932        !ForceDynamicNumberOfGCThreads)) {
 933     n_conc_workers = max_parallel_marking_threads();
 934   } else {
 935     n_conc_workers =
 936       AdaptiveSizePolicy::calc_default_active_workers(
 937                                    max_parallel_marking_threads(),
 938                                    1, /* Minimum workers */
 939                                    parallel_marking_threads(),
 940                                    Threads::number_of_non_daemon_threads());
 941     // Don't scale down "n_conc_workers" by scale_parallel_threads() because
 942     // that scaling has already gone into "_max_parallel_marking_threads".
 943   }
 944   assert(n_conc_workers > 0, "Always need at least 1");
 945   return n_conc_workers;
 946 }
 947 
 948 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) {
 949   // Currently, only survivors can be root regions.
 950   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 951   G1RootRegionScanClosure cl(_g1h, this, worker_id);
 952 
 953   const uintx interval = PrefetchScanIntervalInBytes;
 954   HeapWord* curr = hr->bottom();
 955   const HeapWord* end = hr->top();
 956   while (curr < end) {
 957     Prefetch::read(curr, interval);
 958     oop obj = oop(curr);
 959     int size = obj->oop_iterate_size(&cl);
 960     assert(size == obj->size(), "sanity");
 961     curr += size;
 962   }
 963 }
 964 
 965 class G1CMRootRegionScanTask : public AbstractGangTask {
 966 private:
 967   G1ConcurrentMark* _cm;
 968 
 969 public:
 970   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 971     AbstractGangTask("Root Region Scan"), _cm(cm) { }
 972 
 973   void work(uint worker_id) {
 974     assert(Thread::current()->is_ConcurrentGC_thread(),
 975            "this should only be done by a conc GC thread");
 976 
 977     G1CMRootRegions* root_regions = _cm->root_regions();
 978     HeapRegion* hr = root_regions->claim_next();
 979     while (hr != NULL) {
 980       _cm->scanRootRegion(hr, worker_id);
 981       hr = root_regions->claim_next();
 982     }
 983   }
 984 };
 985 
 986 void G1ConcurrentMark::scanRootRegions() {
 987   // scan_in_progress() will have been set to true only if there was
 988   // at least one root region to scan. So, if it's false, we
 989   // should not attempt to do any further work.
 990   if (root_regions()->scan_in_progress()) {
 991     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 992     GCTraceConcTime(Info, gc) tt("Concurrent Root Region Scan");
 993 
 994     _parallel_marking_threads = calc_parallel_marking_threads();
 995     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
 996            "Maximum number of marking threads exceeded");
 997     uint active_workers = MAX2(1U, parallel_marking_threads());
 998 
 999     G1CMRootRegionScanTask task(this);
1000     _parallel_workers->set_active_workers(active_workers);
1001     _parallel_workers->run_task(&task);
1002 
1003     // It's possible that has_aborted() is true here without actually
1004     // aborting the survivor scan earlier. This is OK as it's
1005     // mainly used for sanity checking.
1006     root_regions()->scan_finished();
1007   }
1008 }
1009 
1010 void G1ConcurrentMark::register_concurrent_phase_start(const char* title) {
1011   assert(!_concurrent_phase_started, "Sanity");
1012   _concurrent_phase_started = true;
1013   _g1h->gc_timer_cm()->register_gc_concurrent_start(title);
1014 }
1015 
1016 void G1ConcurrentMark::register_concurrent_phase_end() {
1017   if (_concurrent_phase_started) {
1018     _concurrent_phase_started = false;
1019     _g1h->gc_timer_cm()->register_gc_concurrent_end();
1020   }
1021 }
1022 
1023 void G1ConcurrentMark::markFromRoots() {
1024   // we might be tempted to assert that:
1025   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1026   //        "inconsistent argument?");
1027   // However that wouldn't be right, because it's possible that
1028   // a safepoint is indeed in progress as a younger generation
1029   // stop-the-world GC happens even as we mark in this generation.
1030 
1031   _restart_for_overflow = false;
1032 
1033   // _g1h has _n_par_threads
1034   _parallel_marking_threads = calc_parallel_marking_threads();
1035   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1036     "Maximum number of marking threads exceeded");
1037 
1038   uint active_workers = MAX2(1U, parallel_marking_threads());
1039   assert(active_workers > 0, "Should have been set");
1040 
1041   // Parallel task terminator is set in "set_concurrency_and_phase()"
1042   set_concurrency_and_phase(active_workers, true /* concurrent */);
1043 
1044   G1CMConcurrentMarkingTask markingTask(this, cmThread());
1045   _parallel_workers->set_active_workers(active_workers);
1046   _parallel_workers->run_task(&markingTask);
1047   print_stats();
1048 }
1049 
1050 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1051   // world is stopped at this checkpoint
1052   assert(SafepointSynchronize::is_at_safepoint(),
1053          "world should be stopped");
1054 
1055   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1056 
1057   // If a full collection has happened, we shouldn't do this.
1058   if (has_aborted()) {
1059     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1060     return;
1061   }
1062 
1063   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1064 
1065   if (VerifyDuringGC) {
1066     HandleMark hm;  // handle scope
1067     g1h->prepare_for_verify();
1068     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1069   }
1070   g1h->verifier()->check_bitmaps("Remark Start");
1071 
1072   G1CollectorPolicy* g1p = g1h->g1_policy();
1073   g1p->record_concurrent_mark_remark_start();
1074 
1075   double start = os::elapsedTime();
1076 
1077   checkpointRootsFinalWork();
1078 
1079   double mark_work_end = os::elapsedTime();
1080 
1081   weakRefsWork(clear_all_soft_refs);
1082 
1083   if (has_overflown()) {
1084     // Oops.  We overflowed.  Restart concurrent marking.
1085     _restart_for_overflow = true;
1086     log_develop_trace(gc)("Remark led to restart for overflow.");
1087 
1088     // Verify the heap w.r.t. the previous marking bitmap.
1089     if (VerifyDuringGC) {
1090       HandleMark hm;  // handle scope
1091       g1h->prepare_for_verify();
1092       Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)");
1093     }
1094 
1095     // Clear the marking state because we will be restarting
1096     // marking due to overflowing the global mark stack.
1097     reset_marking_state();
1098   } else {
1099     {
1100       GCTraceTime(Debug, gc) trace("GC Aggregate Data", g1h->gc_timer_cm());
1101 
1102       // Aggregate the per-task counting data that we have accumulated
1103       // while marking.
1104       aggregate_count_data();
1105     }
1106 
1107     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1108     // We're done with marking.
1109     // This is the end of  the marking cycle, we're expected all
1110     // threads to have SATB queues with active set to true.
1111     satb_mq_set.set_active_all_threads(false, /* new active value */
1112                                        true /* expected_active */);
1113 
1114     if (VerifyDuringGC) {
1115       HandleMark hm;  // handle scope
1116       g1h->prepare_for_verify();
1117       Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)");
1118     }
1119     g1h->verifier()->check_bitmaps("Remark End");
1120     assert(!restart_for_overflow(), "sanity");
1121     // Completely reset the marking state since marking completed
1122     set_non_marking_state();
1123   }
1124 
1125   // Expand the marking stack, if we have to and if we can.
1126   if (_markStack.should_expand()) {
1127     _markStack.expand();
1128   }
1129 
1130   // Statistics
1131   double now = os::elapsedTime();
1132   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1133   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1134   _remark_times.add((now - start) * 1000.0);
1135 
1136   g1p->record_concurrent_mark_remark_end();
1137 
1138   G1CMIsAliveClosure is_alive(g1h);
1139   g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive);
1140 }
1141 
1142 // Base class of the closures that finalize and verify the
1143 // liveness counting data.
1144 class G1CMCountDataClosureBase: public HeapRegionClosure {
1145 protected:
1146   G1CollectedHeap* _g1h;
1147   G1ConcurrentMark* _cm;
1148   CardTableModRefBS* _ct_bs;
1149 
1150   BitMap* _region_bm;
1151   BitMap* _card_bm;
1152 
1153   // Takes a region that's not empty (i.e., it has at least one
1154   // live object in it and sets its corresponding bit on the region
1155   // bitmap to 1.
1156   void set_bit_for_region(HeapRegion* hr) {
1157     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1158     _region_bm->par_at_put(index, true);
1159   }
1160 
1161 public:
1162   G1CMCountDataClosureBase(G1CollectedHeap* g1h,
1163                            BitMap* region_bm, BitMap* card_bm):
1164     _g1h(g1h), _cm(g1h->concurrent_mark()),
1165     _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
1166     _region_bm(region_bm), _card_bm(card_bm) { }
1167 };
1168 
1169 // Closure that calculates the # live objects per region. Used
1170 // for verification purposes during the cleanup pause.
1171 class CalcLiveObjectsClosure: public G1CMCountDataClosureBase {
1172   G1CMBitMapRO* _bm;
1173   size_t _region_marked_bytes;
1174 
1175 public:
1176   CalcLiveObjectsClosure(G1CMBitMapRO *bm, G1CollectedHeap* g1h,
1177                          BitMap* region_bm, BitMap* card_bm) :
1178     G1CMCountDataClosureBase(g1h, region_bm, card_bm),
1179     _bm(bm), _region_marked_bytes(0) { }
1180 
1181   bool doHeapRegion(HeapRegion* hr) {
1182     HeapWord* ntams = hr->next_top_at_mark_start();
1183     HeapWord* start = hr->bottom();
1184 
1185     assert(start <= hr->end() && start <= ntams && ntams <= hr->end(),
1186            "Preconditions not met - "
1187            "start: " PTR_FORMAT ", ntams: " PTR_FORMAT ", end: " PTR_FORMAT,
1188            p2i(start), p2i(ntams), p2i(hr->end()));
1189 
1190     // Find the first marked object at or after "start".
1191     start = _bm->getNextMarkedWordAddress(start, ntams);
1192 
1193     size_t marked_bytes = 0;
1194 
1195     while (start < ntams) {
1196       oop obj = oop(start);
1197       int obj_sz = obj->size();
1198       HeapWord* obj_end = start + obj_sz;
1199 
1200       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
1201       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end);
1202 
1203       // Note: if we're looking at the last region in heap - obj_end
1204       // could be actually just beyond the end of the heap; end_idx
1205       // will then correspond to a (non-existent) card that is also
1206       // just beyond the heap.
1207       if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) {
1208         // end of object is not card aligned - increment to cover
1209         // all the cards spanned by the object
1210         end_idx += 1;
1211       }
1212 
1213       // Set the bits in the card BM for the cards spanned by this object.
1214       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1215 
1216       // Add the size of this object to the number of marked bytes.
1217       marked_bytes += (size_t)obj_sz * HeapWordSize;
1218 
1219       // This will happen if we are handling a humongous object that spans
1220       // several heap regions.
1221       if (obj_end > hr->end()) {
1222         break;
1223       }
1224       // Find the next marked object after this one.
1225       start = _bm->getNextMarkedWordAddress(obj_end, ntams);
1226     }
1227 
1228     // Mark the allocated-since-marking portion...
1229     HeapWord* top = hr->top();
1230     if (ntams < top) {
1231       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1232       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1233 
1234       // Note: if we're looking at the last region in heap - top
1235       // could be actually just beyond the end of the heap; end_idx
1236       // will then correspond to a (non-existent) card that is also
1237       // just beyond the heap.
1238       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1239         // end of object is not card aligned - increment to cover
1240         // all the cards spanned by the object
1241         end_idx += 1;
1242       }
1243       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1244 
1245       // This definitely means the region has live objects.
1246       set_bit_for_region(hr);
1247     }
1248 
1249     // Update the live region bitmap.
1250     if (marked_bytes > 0) {
1251       set_bit_for_region(hr);
1252     }
1253 
1254     // Set the marked bytes for the current region so that
1255     // it can be queried by a calling verification routine
1256     _region_marked_bytes = marked_bytes;
1257 
1258     return false;
1259   }
1260 
1261   size_t region_marked_bytes() const { return _region_marked_bytes; }
1262 };
1263 
1264 // Heap region closure used for verifying the counting data
1265 // that was accumulated concurrently and aggregated during
1266 // the remark pause. This closure is applied to the heap
1267 // regions during the STW cleanup pause.
1268 
1269 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure {
1270   G1CollectedHeap* _g1h;
1271   G1ConcurrentMark* _cm;
1272   CalcLiveObjectsClosure _calc_cl;
1273   BitMap* _region_bm;   // Region BM to be verified
1274   BitMap* _card_bm;     // Card BM to be verified
1275 
1276   BitMap* _exp_region_bm; // Expected Region BM values
1277   BitMap* _exp_card_bm;   // Expected card BM values
1278 
1279   int _failures;
1280 
1281 public:
1282   VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h,
1283                                 BitMap* region_bm,
1284                                 BitMap* card_bm,
1285                                 BitMap* exp_region_bm,
1286                                 BitMap* exp_card_bm) :
1287     _g1h(g1h), _cm(g1h->concurrent_mark()),
1288     _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm),
1289     _region_bm(region_bm), _card_bm(card_bm),
1290     _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm),
1291     _failures(0) { }
1292 
1293   int failures() const { return _failures; }
1294 
1295   bool doHeapRegion(HeapRegion* hr) {
1296     int failures = 0;
1297 
1298     // Call the CalcLiveObjectsClosure to walk the marking bitmap for
1299     // this region and set the corresponding bits in the expected region
1300     // and card bitmaps.
1301     bool res = _calc_cl.doHeapRegion(hr);
1302     assert(res == false, "should be continuing");
1303 
1304     // Verify the marked bytes for this region.
1305     size_t exp_marked_bytes = _calc_cl.region_marked_bytes();
1306     size_t act_marked_bytes = hr->next_marked_bytes();
1307 
1308     if (exp_marked_bytes > act_marked_bytes) {
1309       if (hr->is_starts_humongous()) {
1310         // For start_humongous regions, the size of the whole object will be
1311         // in exp_marked_bytes.
1312         HeapRegion* region = hr;
1313         int num_regions;
1314         for (num_regions = 0; region != NULL; num_regions++) {
1315           region = _g1h->next_region_in_humongous(region);
1316         }
1317         if ((num_regions-1) * HeapRegion::GrainBytes >= exp_marked_bytes) {
1318           failures += 1;
1319         } else if (num_regions * HeapRegion::GrainBytes < exp_marked_bytes) {
1320           failures += 1;
1321         }
1322       } else {
1323         // We're not OK if expected marked bytes > actual marked bytes. It means
1324         // we have missed accounting some objects during the actual marking.
1325         failures += 1;
1326       }
1327     }
1328 
1329     // Verify the bit, for this region, in the actual and expected
1330     // (which was just calculated) region bit maps.
1331     // We're not OK if the bit in the calculated expected region
1332     // bitmap is set and the bit in the actual region bitmap is not.
1333     BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index();
1334 
1335     bool expected = _exp_region_bm->at(index);
1336     bool actual = _region_bm->at(index);
1337     if (expected && !actual) {
1338       failures += 1;
1339     }
1340 
1341     // Verify that the card bit maps for the cards spanned by the current
1342     // region match. We have an error if we have a set bit in the expected
1343     // bit map and the corresponding bit in the actual bitmap is not set.
1344 
1345     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom());
1346     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top());
1347 
1348     for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) {
1349       expected = _exp_card_bm->at(i);
1350       actual = _card_bm->at(i);
1351 
1352       if (expected && !actual) {
1353         failures += 1;
1354       }
1355     }
1356 
1357     _failures += failures;
1358 
1359     // We could stop iteration over the heap when we
1360     // find the first violating region by returning true.
1361     return false;
1362   }
1363 };
1364 
1365 class G1ParVerifyFinalCountTask: public AbstractGangTask {
1366 protected:
1367   G1CollectedHeap* _g1h;
1368   G1ConcurrentMark* _cm;
1369   BitMap* _actual_region_bm;
1370   BitMap* _actual_card_bm;
1371 
1372   uint    _n_workers;
1373 
1374   BitMap* _expected_region_bm;
1375   BitMap* _expected_card_bm;
1376 
1377   int  _failures;
1378 
1379   HeapRegionClaimer _hrclaimer;
1380 
1381 public:
1382   G1ParVerifyFinalCountTask(G1CollectedHeap* g1h,
1383                             BitMap* region_bm, BitMap* card_bm,
1384                             BitMap* expected_region_bm, BitMap* expected_card_bm)
1385     : AbstractGangTask("G1 verify final counting"),
1386       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1387       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1388       _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm),
1389       _failures(0),
1390       _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1391     assert(VerifyDuringGC, "don't call this otherwise");
1392     assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity");
1393     assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity");
1394   }
1395 
1396   void work(uint worker_id) {
1397     assert(worker_id < _n_workers, "invariant");
1398 
1399     VerifyLiveObjectDataHRClosure verify_cl(_g1h,
1400                                             _actual_region_bm, _actual_card_bm,
1401                                             _expected_region_bm,
1402                                             _expected_card_bm);
1403 
1404     _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer);
1405 
1406     Atomic::add(verify_cl.failures(), &_failures);
1407   }
1408 
1409   int failures() const { return _failures; }
1410 };
1411 
1412 // Closure that finalizes the liveness counting data.
1413 // Used during the cleanup pause.
1414 // Sets the bits corresponding to the interval [NTAMS, top]
1415 // (which contains the implicitly live objects) in the
1416 // card liveness bitmap. Also sets the bit for each region,
1417 // containing live data, in the region liveness bitmap.
1418 
1419 class FinalCountDataUpdateClosure: public G1CMCountDataClosureBase {
1420  public:
1421   FinalCountDataUpdateClosure(G1CollectedHeap* g1h,
1422                               BitMap* region_bm,
1423                               BitMap* card_bm) :
1424     G1CMCountDataClosureBase(g1h, region_bm, card_bm) { }
1425 
1426   bool doHeapRegion(HeapRegion* hr) {
1427     HeapWord* ntams = hr->next_top_at_mark_start();
1428     HeapWord* top   = hr->top();
1429 
1430     assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions.");
1431 
1432     // Mark the allocated-since-marking portion...
1433     if (ntams < top) {
1434       // This definitely means the region has live objects.
1435       set_bit_for_region(hr);
1436 
1437       // Now set the bits in the card bitmap for [ntams, top)
1438       BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams);
1439       BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top);
1440 
1441       // Note: if we're looking at the last region in heap - top
1442       // could be actually just beyond the end of the heap; end_idx
1443       // will then correspond to a (non-existent) card that is also
1444       // just beyond the heap.
1445       if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) {
1446         // end of object is not card aligned - increment to cover
1447         // all the cards spanned by the object
1448         end_idx += 1;
1449       }
1450 
1451       assert(end_idx <= _card_bm->size(),
1452              "oob: end_idx=  " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT,
1453              end_idx, _card_bm->size());
1454       assert(start_idx < _card_bm->size(),
1455              "oob: start_idx=  " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT,
1456              start_idx, _card_bm->size());
1457 
1458       _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */);
1459     }
1460 
1461     // Set the bit for the region if it contains live data
1462     if (hr->next_marked_bytes() > 0) {
1463       set_bit_for_region(hr);
1464     }
1465 
1466     return false;
1467   }
1468 };
1469 
1470 class G1ParFinalCountTask: public AbstractGangTask {
1471 protected:
1472   G1CollectedHeap* _g1h;
1473   G1ConcurrentMark* _cm;
1474   BitMap* _actual_region_bm;
1475   BitMap* _actual_card_bm;
1476 
1477   uint    _n_workers;
1478   HeapRegionClaimer _hrclaimer;
1479 
1480 public:
1481   G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm)
1482     : AbstractGangTask("G1 final counting"),
1483       _g1h(g1h), _cm(_g1h->concurrent_mark()),
1484       _actual_region_bm(region_bm), _actual_card_bm(card_bm),
1485       _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) {
1486   }
1487 
1488   void work(uint worker_id) {
1489     assert(worker_id < _n_workers, "invariant");
1490 
1491     FinalCountDataUpdateClosure final_update_cl(_g1h,
1492                                                 _actual_region_bm,
1493                                                 _actual_card_bm);
1494 
1495     _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer);
1496   }
1497 };
1498 
1499 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1500   G1CollectedHeap* _g1;
1501   size_t _freed_bytes;
1502   FreeRegionList* _local_cleanup_list;
1503   uint _old_regions_removed;
1504   uint _humongous_regions_removed;
1505   HRRSCleanupTask* _hrrs_cleanup_task;
1506 
1507 public:
1508   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1509                              FreeRegionList* local_cleanup_list,
1510                              HRRSCleanupTask* hrrs_cleanup_task) :
1511     _g1(g1),
1512     _freed_bytes(0),
1513     _local_cleanup_list(local_cleanup_list),
1514     _old_regions_removed(0),
1515     _humongous_regions_removed(0),
1516     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1517 
1518   size_t freed_bytes() { return _freed_bytes; }
1519   const uint old_regions_removed() { return _old_regions_removed; }
1520   const uint humongous_regions_removed() { return _humongous_regions_removed; }
1521 
1522   bool doHeapRegion(HeapRegion *hr) {
1523     if (hr->is_archive()) {
1524       return false;
1525     }
1526     // We use a claim value of zero here because all regions
1527     // were claimed with value 1 in the FinalCount task.
1528     _g1->reset_gc_time_stamps(hr);
1529     hr->note_end_of_marking();
1530 
1531     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1532       _freed_bytes += hr->used();
1533       hr->set_containing_set(NULL);
1534       if (hr->is_humongous()) {
1535         _humongous_regions_removed++;
1536         _g1->free_humongous_region(hr, _local_cleanup_list, true);
1537       } else {
1538         _old_regions_removed++;
1539         _g1->free_region(hr, _local_cleanup_list, true);
1540       }
1541     } else {
1542       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1543     }
1544 
1545     return false;
1546   }
1547 };
1548 
1549 class G1ParNoteEndTask: public AbstractGangTask {
1550   friend class G1NoteEndOfConcMarkClosure;
1551 
1552 protected:
1553   G1CollectedHeap* _g1h;
1554   FreeRegionList* _cleanup_list;
1555   HeapRegionClaimer _hrclaimer;
1556 
1557 public:
1558   G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1559       AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1560   }
1561 
1562   void work(uint worker_id) {
1563     FreeRegionList local_cleanup_list("Local Cleanup List");
1564     HRRSCleanupTask hrrs_cleanup_task;
1565     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1566                                            &hrrs_cleanup_task);
1567     _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1568     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1569 
1570     // Now update the lists
1571     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1572     {
1573       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1574       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1575 
1576       // If we iterate over the global cleanup list at the end of
1577       // cleanup to do this printing we will not guarantee to only
1578       // generate output for the newly-reclaimed regions (the list
1579       // might not be empty at the beginning of cleanup; we might
1580       // still be working on its previous contents). So we do the
1581       // printing here, before we append the new regions to the global
1582       // cleanup list.
1583 
1584       G1HRPrinter* hr_printer = _g1h->hr_printer();
1585       if (hr_printer->is_active()) {
1586         FreeRegionListIterator iter(&local_cleanup_list);
1587         while (iter.more_available()) {
1588           HeapRegion* hr = iter.get_next();
1589           hr_printer->cleanup(hr);
1590         }
1591       }
1592 
1593       _cleanup_list->add_ordered(&local_cleanup_list);
1594       assert(local_cleanup_list.is_empty(), "post-condition");
1595 
1596       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1597     }
1598   }
1599 };
1600 
1601 void G1ConcurrentMark::cleanup() {
1602   // world is stopped at this checkpoint
1603   assert(SafepointSynchronize::is_at_safepoint(),
1604          "world should be stopped");
1605   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1606 
1607   // If a full collection has happened, we shouldn't do this.
1608   if (has_aborted()) {
1609     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1610     return;
1611   }
1612 
1613   g1h->verifier()->verify_region_sets_optional();
1614 
1615   if (VerifyDuringGC) {
1616     HandleMark hm;  // handle scope
1617     g1h->prepare_for_verify();
1618     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1619   }
1620   g1h->verifier()->check_bitmaps("Cleanup Start");
1621 
1622   G1CollectorPolicy* g1p = g1h->g1_policy();
1623   g1p->record_concurrent_mark_cleanup_start();
1624 
1625   double start = os::elapsedTime();
1626 
1627   HeapRegionRemSet::reset_for_cleanup_tasks();
1628 
1629   // Do counting once more with the world stopped for good measure.
1630   G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm);
1631 
1632   g1h->workers()->run_task(&g1_par_count_task);
1633 
1634   if (VerifyDuringGC) {
1635     // Verify that the counting data accumulated during marking matches
1636     // that calculated by walking the marking bitmap.
1637 
1638     // Bitmaps to hold expected values
1639     BitMap expected_region_bm(_region_bm.size(), true);
1640     BitMap expected_card_bm(_card_bm.size(), true);
1641 
1642     G1ParVerifyFinalCountTask g1_par_verify_task(g1h,
1643                                                  &_region_bm,
1644                                                  &_card_bm,
1645                                                  &expected_region_bm,
1646                                                  &expected_card_bm);
1647 
1648     g1h->workers()->run_task(&g1_par_verify_task);
1649 
1650     guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures");
1651   }
1652 
1653   size_t start_used_bytes = g1h->used();
1654   g1h->collector_state()->set_mark_in_progress(false);
1655 
1656   double count_end = os::elapsedTime();
1657   double this_final_counting_time = (count_end - start);
1658   _total_counting_time += this_final_counting_time;
1659 
1660   if (log_is_enabled(Trace, gc, liveness)) {
1661     G1PrintRegionLivenessInfoClosure cl("Post-Marking");
1662     _g1h->heap_region_iterate(&cl);
1663   }
1664 
1665   // Install newly created mark bitMap as "prev".
1666   swapMarkBitMaps();
1667 
1668   g1h->reset_gc_time_stamp();
1669 
1670   uint n_workers = _g1h->workers()->active_workers();
1671 
1672   // Note end of marking in all heap regions.
1673   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1674   g1h->workers()->run_task(&g1_par_note_end_task);
1675   g1h->check_gc_time_stamps();
1676 
1677   if (!cleanup_list_is_empty()) {
1678     // The cleanup list is not empty, so we'll have to process it
1679     // concurrently. Notify anyone else that might be wanting free
1680     // regions that there will be more free regions coming soon.
1681     g1h->set_free_regions_coming();
1682   }
1683 
1684   // call below, since it affects the metric by which we sort the heap
1685   // regions.
1686   if (G1ScrubRemSets) {
1687     double rs_scrub_start = os::elapsedTime();
1688     g1h->scrub_rem_set(&_region_bm, &_card_bm);
1689     _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start);
1690   }
1691 
1692   // this will also free any regions totally full of garbage objects,
1693   // and sort the regions.
1694   g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1695 
1696   // Statistics.
1697   double end = os::elapsedTime();
1698   _cleanup_times.add((end - start) * 1000.0);
1699 
1700   // Clean up will have freed any regions completely full of garbage.
1701   // Update the soft reference policy with the new heap occupancy.
1702   Universe::update_heap_info_at_gc();
1703 
1704   if (VerifyDuringGC) {
1705     HandleMark hm;  // handle scope
1706     g1h->prepare_for_verify();
1707     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)");
1708   }
1709 
1710   g1h->verifier()->check_bitmaps("Cleanup End");
1711 
1712   g1h->verifier()->verify_region_sets_optional();
1713 
1714   // We need to make this be a "collection" so any collection pause that
1715   // races with it goes around and waits for completeCleanup to finish.
1716   g1h->increment_total_collections();
1717 
1718   // Clean out dead classes and update Metaspace sizes.
1719   if (ClassUnloadingWithConcurrentMark) {
1720     ClassLoaderDataGraph::purge();
1721   }
1722   MetaspaceGC::compute_new_size();
1723 
1724   // We reclaimed old regions so we should calculate the sizes to make
1725   // sure we update the old gen/space data.
1726   g1h->g1mm()->update_sizes();
1727   g1h->allocation_context_stats().update_after_mark();
1728 
1729   g1h->trace_heap_after_concurrent_cycle();
1730 }
1731 
1732 void G1ConcurrentMark::completeCleanup() {
1733   if (has_aborted()) return;
1734 
1735   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1736 
1737   _cleanup_list.verify_optional();
1738   FreeRegionList tmp_free_list("Tmp Free List");
1739 
1740   log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1741                                   "cleanup list has %u entries",
1742                                   _cleanup_list.length());
1743 
1744   // No one else should be accessing the _cleanup_list at this point,
1745   // so it is not necessary to take any locks
1746   while (!_cleanup_list.is_empty()) {
1747     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
1748     assert(hr != NULL, "Got NULL from a non-empty list");
1749     hr->par_clear();
1750     tmp_free_list.add_ordered(hr);
1751 
1752     // Instead of adding one region at a time to the secondary_free_list,
1753     // we accumulate them in the local list and move them a few at a
1754     // time. This also cuts down on the number of notify_all() calls
1755     // we do during this process. We'll also append the local list when
1756     // _cleanup_list is empty (which means we just removed the last
1757     // region from the _cleanup_list).
1758     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
1759         _cleanup_list.is_empty()) {
1760       log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1761                                       "appending %u entries to the secondary_free_list, "
1762                                       "cleanup list still has %u entries",
1763                                       tmp_free_list.length(),
1764                                       _cleanup_list.length());
1765 
1766       {
1767         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
1768         g1h->secondary_free_list_add(&tmp_free_list);
1769         SecondaryFreeList_lock->notify_all();
1770       }
1771 #ifndef PRODUCT
1772       if (G1StressConcRegionFreeing) {
1773         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
1774           os::sleep(Thread::current(), (jlong) 1, false);
1775         }
1776       }
1777 #endif
1778     }
1779   }
1780   assert(tmp_free_list.is_empty(), "post-condition");
1781 }
1782 
1783 // Supporting Object and Oop closures for reference discovery
1784 // and processing in during marking
1785 
1786 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1787   HeapWord* addr = (HeapWord*)obj;
1788   return addr != NULL &&
1789          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
1790 }
1791 
1792 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1793 // Uses the G1CMTask associated with a worker thread (for serial reference
1794 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1795 // trace referent objects.
1796 //
1797 // Using the G1CMTask and embedded local queues avoids having the worker
1798 // threads operating on the global mark stack. This reduces the risk
1799 // of overflowing the stack - which we would rather avoid at this late
1800 // state. Also using the tasks' local queues removes the potential
1801 // of the workers interfering with each other that could occur if
1802 // operating on the global stack.
1803 
1804 class G1CMKeepAliveAndDrainClosure: public OopClosure {
1805   G1ConcurrentMark* _cm;
1806   G1CMTask*         _task;
1807   int               _ref_counter_limit;
1808   int               _ref_counter;
1809   bool              _is_serial;
1810  public:
1811   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1812     _cm(cm), _task(task), _is_serial(is_serial),
1813     _ref_counter_limit(G1RefProcDrainInterval) {
1814     assert(_ref_counter_limit > 0, "sanity");
1815     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1816     _ref_counter = _ref_counter_limit;
1817   }
1818 
1819   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1820   virtual void do_oop(      oop* p) { do_oop_work(p); }
1821 
1822   template <class T> void do_oop_work(T* p) {
1823     if (!_cm->has_overflown()) {
1824       oop obj = oopDesc::load_decode_heap_oop(p);
1825       _task->deal_with_reference(obj);
1826       _ref_counter--;
1827 
1828       if (_ref_counter == 0) {
1829         // We have dealt with _ref_counter_limit references, pushing them
1830         // and objects reachable from them on to the local stack (and
1831         // possibly the global stack). Call G1CMTask::do_marking_step() to
1832         // process these entries.
1833         //
1834         // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1835         // there's nothing more to do (i.e. we're done with the entries that
1836         // were pushed as a result of the G1CMTask::deal_with_reference() calls
1837         // above) or we overflow.
1838         //
1839         // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1840         // flag while there may still be some work to do. (See the comment at
1841         // the beginning of G1CMTask::do_marking_step() for those conditions -
1842         // one of which is reaching the specified time target.) It is only
1843         // when G1CMTask::do_marking_step() returns without setting the
1844         // has_aborted() flag that the marking step has completed.
1845         do {
1846           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1847           _task->do_marking_step(mark_step_duration_ms,
1848                                  false      /* do_termination */,
1849                                  _is_serial);
1850         } while (_task->has_aborted() && !_cm->has_overflown());
1851         _ref_counter = _ref_counter_limit;
1852       }
1853     }
1854   }
1855 };
1856 
1857 // 'Drain' oop closure used by both serial and parallel reference processing.
1858 // Uses the G1CMTask associated with a given worker thread (for serial
1859 // reference processing the G1CMtask for worker 0 is used). Calls the
1860 // do_marking_step routine, with an unbelievably large timeout value,
1861 // to drain the marking data structures of the remaining entries
1862 // added by the 'keep alive' oop closure above.
1863 
1864 class G1CMDrainMarkingStackClosure: public VoidClosure {
1865   G1ConcurrentMark* _cm;
1866   G1CMTask*         _task;
1867   bool              _is_serial;
1868  public:
1869   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1870     _cm(cm), _task(task), _is_serial(is_serial) {
1871     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1872   }
1873 
1874   void do_void() {
1875     do {
1876       // We call G1CMTask::do_marking_step() to completely drain the local
1877       // and global marking stacks of entries pushed by the 'keep alive'
1878       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1879       //
1880       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1881       // if there's nothing more to do (i.e. we've completely drained the
1882       // entries that were pushed as a a result of applying the 'keep alive'
1883       // closure to the entries on the discovered ref lists) or we overflow
1884       // the global marking stack.
1885       //
1886       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1887       // flag while there may still be some work to do. (See the comment at
1888       // the beginning of G1CMTask::do_marking_step() for those conditions -
1889       // one of which is reaching the specified time target.) It is only
1890       // when G1CMTask::do_marking_step() returns without setting the
1891       // has_aborted() flag that the marking step has completed.
1892 
1893       _task->do_marking_step(1000000000.0 /* something very large */,
1894                              true         /* do_termination */,
1895                              _is_serial);
1896     } while (_task->has_aborted() && !_cm->has_overflown());
1897   }
1898 };
1899 
1900 // Implementation of AbstractRefProcTaskExecutor for parallel
1901 // reference processing at the end of G1 concurrent marking
1902 
1903 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
1904 private:
1905   G1CollectedHeap*  _g1h;
1906   G1ConcurrentMark* _cm;
1907   WorkGang*         _workers;
1908   uint              _active_workers;
1909 
1910 public:
1911   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1912                           G1ConcurrentMark* cm,
1913                           WorkGang* workers,
1914                           uint n_workers) :
1915     _g1h(g1h), _cm(cm),
1916     _workers(workers), _active_workers(n_workers) { }
1917 
1918   // Executes the given task using concurrent marking worker threads.
1919   virtual void execute(ProcessTask& task);
1920   virtual void execute(EnqueueTask& task);
1921 };
1922 
1923 class G1CMRefProcTaskProxy: public AbstractGangTask {
1924   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1925   ProcessTask&      _proc_task;
1926   G1CollectedHeap*  _g1h;
1927   G1ConcurrentMark* _cm;
1928 
1929 public:
1930   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1931                        G1CollectedHeap* g1h,
1932                        G1ConcurrentMark* cm) :
1933     AbstractGangTask("Process reference objects in parallel"),
1934     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1935     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1936     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1937   }
1938 
1939   virtual void work(uint worker_id) {
1940     ResourceMark rm;
1941     HandleMark hm;
1942     G1CMTask* task = _cm->task(worker_id);
1943     G1CMIsAliveClosure g1_is_alive(_g1h);
1944     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1945     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1946 
1947     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1948   }
1949 };
1950 
1951 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1952   assert(_workers != NULL, "Need parallel worker threads.");
1953   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1954 
1955   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1956 
1957   // We need to reset the concurrency level before each
1958   // proxy task execution, so that the termination protocol
1959   // and overflow handling in G1CMTask::do_marking_step() knows
1960   // how many workers to wait for.
1961   _cm->set_concurrency(_active_workers);
1962   _workers->run_task(&proc_task_proxy);
1963 }
1964 
1965 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
1966   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1967   EnqueueTask& _enq_task;
1968 
1969 public:
1970   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1971     AbstractGangTask("Enqueue reference objects in parallel"),
1972     _enq_task(enq_task) { }
1973 
1974   virtual void work(uint worker_id) {
1975     _enq_task.work(worker_id);
1976   }
1977 };
1978 
1979 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1980   assert(_workers != NULL, "Need parallel worker threads.");
1981   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1982 
1983   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1984 
1985   // Not strictly necessary but...
1986   //
1987   // We need to reset the concurrency level before each
1988   // proxy task execution, so that the termination protocol
1989   // and overflow handling in G1CMTask::do_marking_step() knows
1990   // how many workers to wait for.
1991   _cm->set_concurrency(_active_workers);
1992   _workers->run_task(&enq_task_proxy);
1993 }
1994 
1995 void G1ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) {
1996   G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes);
1997 }
1998 
1999 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
2000   if (has_overflown()) {
2001     // Skip processing the discovered references if we have
2002     // overflown the global marking stack. Reference objects
2003     // only get discovered once so it is OK to not
2004     // de-populate the discovered reference lists. We could have,
2005     // but the only benefit would be that, when marking restarts,
2006     // less reference objects are discovered.
2007     return;
2008   }
2009 
2010   ResourceMark rm;
2011   HandleMark   hm;
2012 
2013   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2014 
2015   // Is alive closure.
2016   G1CMIsAliveClosure g1_is_alive(g1h);
2017 
2018   // Inner scope to exclude the cleaning of the string and symbol
2019   // tables from the displayed time.
2020   {
2021     GCTraceTime(Debug, gc) trace("GC Ref Proc", g1h->gc_timer_cm());
2022 
2023     ReferenceProcessor* rp = g1h->ref_processor_cm();
2024 
2025     // See the comment in G1CollectedHeap::ref_processing_init()
2026     // about how reference processing currently works in G1.
2027 
2028     // Set the soft reference policy
2029     rp->setup_policy(clear_all_soft_refs);
2030     assert(_markStack.isEmpty(), "mark stack should be empty");
2031 
2032     // Instances of the 'Keep Alive' and 'Complete GC' closures used
2033     // in serial reference processing. Note these closures are also
2034     // used for serially processing (by the the current thread) the
2035     // JNI references during parallel reference processing.
2036     //
2037     // These closures do not need to synchronize with the worker
2038     // threads involved in parallel reference processing as these
2039     // instances are executed serially by the current thread (e.g.
2040     // reference processing is not multi-threaded and is thus
2041     // performed by the current thread instead of a gang worker).
2042     //
2043     // The gang tasks involved in parallel reference processing create
2044     // their own instances of these closures, which do their own
2045     // synchronization among themselves.
2046     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
2047     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
2048 
2049     // We need at least one active thread. If reference processing
2050     // is not multi-threaded we use the current (VMThread) thread,
2051     // otherwise we use the work gang from the G1CollectedHeap and
2052     // we utilize all the worker threads we can.
2053     bool processing_is_mt = rp->processing_is_mt();
2054     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
2055     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
2056 
2057     // Parallel processing task executor.
2058     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
2059                                               g1h->workers(), active_workers);
2060     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
2061 
2062     // Set the concurrency level. The phase was already set prior to
2063     // executing the remark task.
2064     set_concurrency(active_workers);
2065 
2066     // Set the degree of MT processing here.  If the discovery was done MT,
2067     // the number of threads involved during discovery could differ from
2068     // the number of active workers.  This is OK as long as the discovered
2069     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
2070     rp->set_active_mt_degree(active_workers);
2071 
2072     // Process the weak references.
2073     const ReferenceProcessorStats& stats =
2074         rp->process_discovered_references(&g1_is_alive,
2075                                           &g1_keep_alive,
2076                                           &g1_drain_mark_stack,
2077                                           executor,
2078                                           g1h->gc_timer_cm());
2079     g1h->gc_tracer_cm()->report_gc_reference_stats(stats);
2080 
2081     // The do_oop work routines of the keep_alive and drain_marking_stack
2082     // oop closures will set the has_overflown flag if we overflow the
2083     // global marking stack.
2084 
2085     assert(_markStack.overflow() || _markStack.isEmpty(),
2086             "mark stack should be empty (unless it overflowed)");
2087 
2088     if (_markStack.overflow()) {
2089       // This should have been done already when we tried to push an
2090       // entry on to the global mark stack. But let's do it again.
2091       set_has_overflown();
2092     }
2093 
2094     assert(rp->num_q() == active_workers, "why not");
2095 
2096     rp->enqueue_discovered_references(executor);
2097 
2098     rp->verify_no_references_recorded();
2099     assert(!rp->discovery_enabled(), "Post condition");
2100   }
2101 
2102   if (has_overflown()) {
2103     // We can not trust g1_is_alive if the marking stack overflowed
2104     return;
2105   }
2106 
2107   assert(_markStack.isEmpty(), "Marking should have completed");
2108 
2109   // Unload Klasses, String, Symbols, Code Cache, etc.
2110   {
2111     GCTraceTime(Debug, gc) trace("Unloading", g1h->gc_timer_cm());
2112 
2113     if (ClassUnloadingWithConcurrentMark) {
2114       bool purged_classes;
2115 
2116       {
2117         GCTraceTime(Trace, gc) trace("System Dictionary Unloading", g1h->gc_timer_cm());
2118         purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */);
2119       }
2120 
2121       {
2122         GCTraceTime(Trace, gc) trace("Parallel Unloading", g1h->gc_timer_cm());
2123         weakRefsWorkParallelPart(&g1_is_alive, purged_classes);
2124       }
2125     }
2126 
2127     if (G1StringDedup::is_enabled()) {
2128       GCTraceTime(Trace, gc) trace("String Deduplication Unlink", g1h->gc_timer_cm());
2129       G1StringDedup::unlink(&g1_is_alive);
2130     }
2131   }
2132 }
2133 
2134 void G1ConcurrentMark::swapMarkBitMaps() {
2135   G1CMBitMapRO* temp = _prevMarkBitMap;
2136   _prevMarkBitMap    = (G1CMBitMapRO*)_nextMarkBitMap;
2137   _nextMarkBitMap    = (G1CMBitMap*)  temp;
2138 }
2139 
2140 // Closure for marking entries in SATB buffers.
2141 class G1CMSATBBufferClosure : public SATBBufferClosure {
2142 private:
2143   G1CMTask* _task;
2144   G1CollectedHeap* _g1h;
2145 
2146   // This is very similar to G1CMTask::deal_with_reference, but with
2147   // more relaxed requirements for the argument, so this must be more
2148   // circumspect about treating the argument as an object.
2149   void do_entry(void* entry) const {
2150     _task->increment_refs_reached();
2151     HeapRegion* hr = _g1h->heap_region_containing(entry);
2152     if (entry < hr->next_top_at_mark_start()) {
2153       // Until we get here, we don't know whether entry refers to a valid
2154       // object; it could instead have been a stale reference.
2155       oop obj = static_cast<oop>(entry);
2156       assert(obj->is_oop(true /* ignore mark word */),
2157              "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj));
2158       _task->make_reference_grey(obj, hr);
2159     }
2160   }
2161 
2162 public:
2163   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
2164     : _task(task), _g1h(g1h) { }
2165 
2166   virtual void do_buffer(void** buffer, size_t size) {
2167     for (size_t i = 0; i < size; ++i) {
2168       do_entry(buffer[i]);
2169     }
2170   }
2171 };
2172 
2173 class G1RemarkThreadsClosure : public ThreadClosure {
2174   G1CMSATBBufferClosure _cm_satb_cl;
2175   G1CMOopClosure _cm_cl;
2176   MarkingCodeBlobClosure _code_cl;
2177   int _thread_parity;
2178 
2179  public:
2180   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
2181     _cm_satb_cl(task, g1h),
2182     _cm_cl(g1h, g1h->concurrent_mark(), task),
2183     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
2184     _thread_parity(Threads::thread_claim_parity()) {}
2185 
2186   void do_thread(Thread* thread) {
2187     if (thread->is_Java_thread()) {
2188       if (thread->claim_oops_do(true, _thread_parity)) {
2189         JavaThread* jt = (JavaThread*)thread;
2190 
2191         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
2192         // however the liveness of oops reachable from nmethods have very complex lifecycles:
2193         // * Alive if on the stack of an executing method
2194         // * Weakly reachable otherwise
2195         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
2196         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
2197         jt->nmethods_do(&_code_cl);
2198 
2199         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
2200       }
2201     } else if (thread->is_VM_thread()) {
2202       if (thread->claim_oops_do(true, _thread_parity)) {
2203         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
2204       }
2205     }
2206   }
2207 };
2208 
2209 class G1CMRemarkTask: public AbstractGangTask {
2210 private:
2211   G1ConcurrentMark* _cm;
2212 public:
2213   void work(uint worker_id) {
2214     // Since all available tasks are actually started, we should
2215     // only proceed if we're supposed to be active.
2216     if (worker_id < _cm->active_tasks()) {
2217       G1CMTask* task = _cm->task(worker_id);
2218       task->record_start_time();
2219       {
2220         ResourceMark rm;
2221         HandleMark hm;
2222 
2223         G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
2224         Threads::threads_do(&threads_f);
2225       }
2226 
2227       do {
2228         task->do_marking_step(1000000000.0 /* something very large */,
2229                               true         /* do_termination       */,
2230                               false        /* is_serial            */);
2231       } while (task->has_aborted() && !_cm->has_overflown());
2232       // If we overflow, then we do not want to restart. We instead
2233       // want to abort remark and do concurrent marking again.
2234       task->record_end_time();
2235     }
2236   }
2237 
2238   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
2239     AbstractGangTask("Par Remark"), _cm(cm) {
2240     _cm->terminator()->reset_for_reuse(active_workers);
2241   }
2242 };
2243 
2244 void G1ConcurrentMark::checkpointRootsFinalWork() {
2245   ResourceMark rm;
2246   HandleMark   hm;
2247   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2248 
2249   GCTraceTime(Debug, gc) trace("Finalize Marking", g1h->gc_timer_cm());
2250 
2251   g1h->ensure_parsability(false);
2252 
2253   // this is remark, so we'll use up all active threads
2254   uint active_workers = g1h->workers()->active_workers();
2255   set_concurrency_and_phase(active_workers, false /* concurrent */);
2256   // Leave _parallel_marking_threads at it's
2257   // value originally calculated in the G1ConcurrentMark
2258   // constructor and pass values of the active workers
2259   // through the gang in the task.
2260 
2261   {
2262     StrongRootsScope srs(active_workers);
2263 
2264     G1CMRemarkTask remarkTask(this, active_workers);
2265     // We will start all available threads, even if we decide that the
2266     // active_workers will be fewer. The extra ones will just bail out
2267     // immediately.
2268     g1h->workers()->run_task(&remarkTask);
2269   }
2270 
2271   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2272   guarantee(has_overflown() ||
2273             satb_mq_set.completed_buffers_num() == 0,
2274             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
2275             BOOL_TO_STR(has_overflown()),
2276             satb_mq_set.completed_buffers_num());
2277 
2278   print_stats();
2279 }
2280 
2281 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
2282   // Note we are overriding the read-only view of the prev map here, via
2283   // the cast.
2284   ((G1CMBitMap*)_prevMarkBitMap)->clearRange(mr);
2285 }
2286 
2287 HeapRegion*
2288 G1ConcurrentMark::claim_region(uint worker_id) {
2289   // "checkpoint" the finger
2290   HeapWord* finger = _finger;
2291 
2292   // _heap_end will not change underneath our feet; it only changes at
2293   // yield points.
2294   while (finger < _heap_end) {
2295     assert(_g1h->is_in_g1_reserved(finger), "invariant");
2296 
2297     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
2298 
2299     // Above heap_region_containing may return NULL as we always scan claim
2300     // until the end of the heap. In this case, just jump to the next region.
2301     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
2302 
2303     // Is the gap between reading the finger and doing the CAS too long?
2304     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
2305     if (res == finger && curr_region != NULL) {
2306       // we succeeded
2307       HeapWord*   bottom        = curr_region->bottom();
2308       HeapWord*   limit         = curr_region->next_top_at_mark_start();
2309 
2310       // notice that _finger == end cannot be guaranteed here since,
2311       // someone else might have moved the finger even further
2312       assert(_finger >= end, "the finger should have moved forward");
2313 
2314       if (limit > bottom) {
2315         return curr_region;
2316       } else {
2317         assert(limit == bottom,
2318                "the region limit should be at bottom");
2319         // we return NULL and the caller should try calling
2320         // claim_region() again.
2321         return NULL;
2322       }
2323     } else {
2324       assert(_finger > finger, "the finger should have moved forward");
2325       // read it again
2326       finger = _finger;
2327     }
2328   }
2329 
2330   return NULL;
2331 }
2332 
2333 #ifndef PRODUCT
2334 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
2335 private:
2336   G1CollectedHeap* _g1h;
2337   const char* _phase;
2338   int _info;
2339 
2340 public:
2341   VerifyNoCSetOops(const char* phase, int info = -1) :
2342     _g1h(G1CollectedHeap::heap()),
2343     _phase(phase),
2344     _info(info)
2345   { }
2346 
2347   void operator()(oop obj) const {
2348     guarantee(obj->is_oop(),
2349               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
2350               p2i(obj), _phase, _info);
2351     guarantee(!_g1h->obj_in_cs(obj),
2352               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
2353               p2i(obj), _phase, _info);
2354   }
2355 };
2356 
2357 void G1ConcurrentMark::verify_no_cset_oops() {
2358   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
2359   if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
2360     return;
2361   }
2362 
2363   // Verify entries on the global mark stack
2364   _markStack.iterate(VerifyNoCSetOops("Stack"));
2365 
2366   // Verify entries on the task queues
2367   for (uint i = 0; i < _max_worker_id; ++i) {
2368     G1CMTaskQueue* queue = _task_queues->queue(i);
2369     queue->iterate(VerifyNoCSetOops("Queue", i));
2370   }
2371 
2372   // Verify the global finger
2373   HeapWord* global_finger = finger();
2374   if (global_finger != NULL && global_finger < _heap_end) {
2375     // Since we always iterate over all regions, we might get a NULL HeapRegion
2376     // here.
2377     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
2378     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
2379               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
2380               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
2381   }
2382 
2383   // Verify the task fingers
2384   assert(parallel_marking_threads() <= _max_worker_id, "sanity");
2385   for (uint i = 0; i < parallel_marking_threads(); ++i) {
2386     G1CMTask* task = _tasks[i];
2387     HeapWord* task_finger = task->finger();
2388     if (task_finger != NULL && task_finger < _heap_end) {
2389       // See above note on the global finger verification.
2390       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
2391       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2392                 !task_hr->in_collection_set(),
2393                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2394                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
2395     }
2396   }
2397 }
2398 #endif // PRODUCT
2399 
2400 // Aggregate the counting data that was constructed concurrently
2401 // with marking.
2402 class AggregateCountDataHRClosure: public HeapRegionClosure {
2403   G1CollectedHeap* _g1h;
2404   G1ConcurrentMark* _cm;
2405   CardTableModRefBS* _ct_bs;
2406   BitMap* _cm_card_bm;
2407   uint _max_worker_id;
2408 
2409  public:
2410   AggregateCountDataHRClosure(G1CollectedHeap* g1h,
2411                               BitMap* cm_card_bm,
2412                               uint max_worker_id) :
2413     _g1h(g1h), _cm(g1h->concurrent_mark()),
2414     _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())),
2415     _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { }
2416 
2417   bool doHeapRegion(HeapRegion* hr) {
2418     HeapWord* start = hr->bottom();
2419     HeapWord* limit = hr->next_top_at_mark_start();
2420     HeapWord* end = hr->end();
2421 
2422     assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(),
2423            "Preconditions not met - "
2424            "start: " PTR_FORMAT ", limit: " PTR_FORMAT ", "
2425            "top: " PTR_FORMAT ", end: " PTR_FORMAT,
2426            p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end()));
2427 
2428     assert(hr->next_marked_bytes() == 0, "Precondition");
2429 
2430     if (start == limit) {
2431       // NTAMS of this region has not been set so nothing to do.
2432       return false;
2433     }
2434 
2435     // 'start' should be in the heap.
2436     assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity");
2437     // 'end' *may* be just beyond the end of the heap (if hr is the last region)
2438     assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity");
2439 
2440     BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start);
2441     BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit);
2442     BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end);
2443 
2444     // If ntams is not card aligned then we bump card bitmap index
2445     // for limit so that we get the all the cards spanned by
2446     // the object ending at ntams.
2447     // Note: if this is the last region in the heap then ntams
2448     // could be actually just beyond the end of the the heap;
2449     // limit_idx will then  correspond to a (non-existent) card
2450     // that is also outside the heap.
2451     if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) {
2452       limit_idx += 1;
2453     }
2454 
2455     assert(limit_idx <= end_idx, "or else use atomics");
2456 
2457     // Aggregate the "stripe" in the count data associated with hr.
2458     uint hrm_index = hr->hrm_index();
2459     size_t marked_bytes = 0;
2460 
2461     for (uint i = 0; i < _max_worker_id; i += 1) {
2462       size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i);
2463       BitMap* task_card_bm = _cm->count_card_bitmap_for(i);
2464 
2465       // Fetch the marked_bytes in this region for task i and
2466       // add it to the running total for this region.
2467       marked_bytes += marked_bytes_array[hrm_index];
2468 
2469       // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx)
2470       // into the global card bitmap.
2471       BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx);
2472 
2473       while (scan_idx < limit_idx) {
2474         assert(task_card_bm->at(scan_idx) == true, "should be");
2475         _cm_card_bm->set_bit(scan_idx);
2476         assert(_cm_card_bm->at(scan_idx) == true, "should be");
2477 
2478         // BitMap::get_next_one_offset() can handle the case when
2479         // its left_offset parameter is greater than its right_offset
2480         // parameter. It does, however, have an early exit if
2481         // left_offset == right_offset. So let's limit the value
2482         // passed in for left offset here.
2483         BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx);
2484         scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx);
2485       }
2486     }
2487 
2488     // Update the marked bytes for this region.
2489     hr->add_to_marked_bytes(marked_bytes);
2490 
2491     // Next heap region
2492     return false;
2493   }
2494 };
2495 
2496 class G1AggregateCountDataTask: public AbstractGangTask {
2497 protected:
2498   G1CollectedHeap* _g1h;
2499   G1ConcurrentMark* _cm;
2500   BitMap* _cm_card_bm;
2501   uint _max_worker_id;
2502   uint _active_workers;
2503   HeapRegionClaimer _hrclaimer;
2504 
2505 public:
2506   G1AggregateCountDataTask(G1CollectedHeap* g1h,
2507                            G1ConcurrentMark* cm,
2508                            BitMap* cm_card_bm,
2509                            uint max_worker_id,
2510                            uint n_workers) :
2511       AbstractGangTask("Count Aggregation"),
2512       _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm),
2513       _max_worker_id(max_worker_id),
2514       _active_workers(n_workers),
2515       _hrclaimer(_active_workers) {
2516   }
2517 
2518   void work(uint worker_id) {
2519     AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id);
2520 
2521     _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer);
2522   }
2523 };
2524 
2525 
2526 void G1ConcurrentMark::aggregate_count_data() {
2527   uint n_workers = _g1h->workers()->active_workers();
2528 
2529   G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm,
2530                                            _max_worker_id, n_workers);
2531 
2532   _g1h->workers()->run_task(&g1_par_agg_task);
2533 }
2534 
2535 // Clear the per-worker arrays used to store the per-region counting data
2536 void G1ConcurrentMark::clear_all_count_data() {
2537   // Clear the global card bitmap - it will be filled during
2538   // liveness count aggregation (during remark) and the
2539   // final counting task.
2540   _card_bm.clear();
2541 
2542   // Clear the global region bitmap - it will be filled as part
2543   // of the final counting task.
2544   _region_bm.clear();
2545 
2546   uint max_regions = _g1h->max_regions();
2547   assert(_max_worker_id > 0, "uninitialized");
2548 
2549   for (uint i = 0; i < _max_worker_id; i += 1) {
2550     BitMap* task_card_bm = count_card_bitmap_for(i);
2551     size_t* marked_bytes_array = count_marked_bytes_array_for(i);
2552 
2553     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
2554     assert(marked_bytes_array != NULL, "uninitialized");
2555 
2556     memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t));
2557     task_card_bm->clear();
2558   }
2559 }
2560 
2561 void G1ConcurrentMark::print_stats() {
2562   if (!log_is_enabled(Debug, gc, stats)) {
2563     return;
2564   }
2565   log_debug(gc, stats)("---------------------------------------------------------------------");
2566   for (size_t i = 0; i < _active_tasks; ++i) {
2567     _tasks[i]->print_stats();
2568     log_debug(gc, stats)("---------------------------------------------------------------------");
2569   }
2570 }
2571 
2572 // abandon current marking iteration due to a Full GC
2573 void G1ConcurrentMark::abort() {
2574   if (!cmThread()->during_cycle() || _has_aborted) {
2575     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2576     return;
2577   }
2578 
2579   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2580   // concurrent bitmap clearing.
2581   _nextMarkBitMap->clearAll();
2582 
2583   // Note we cannot clear the previous marking bitmap here
2584   // since VerifyDuringGC verifies the objects marked during
2585   // a full GC against the previous bitmap.
2586 
2587   // Clear the liveness counting data
2588   clear_all_count_data();
2589   // Empty mark stack
2590   reset_marking_state();
2591   for (uint i = 0; i < _max_worker_id; ++i) {
2592     _tasks[i]->clear_region_fields();
2593   }
2594   _first_overflow_barrier_sync.abort();
2595   _second_overflow_barrier_sync.abort();
2596   _has_aborted = true;
2597 
2598   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2599   satb_mq_set.abandon_partial_marking();
2600   // This can be called either during or outside marking, we'll read
2601   // the expected_active value from the SATB queue set.
2602   satb_mq_set.set_active_all_threads(
2603                                  false, /* new active value */
2604                                  satb_mq_set.is_active() /* expected_active */);
2605 
2606   _g1h->trace_heap_after_concurrent_cycle();
2607 
2608   // Close any open concurrent phase timing
2609   register_concurrent_phase_end();
2610 
2611   _g1h->register_concurrent_cycle_end();
2612 }
2613 
2614 static void print_ms_time_info(const char* prefix, const char* name,
2615                                NumberSeq& ns) {
2616   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2617                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2618   if (ns.num() > 0) {
2619     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2620                            prefix, ns.sd(), ns.maximum());
2621   }
2622 }
2623 
2624 void G1ConcurrentMark::print_summary_info() {
2625   LogHandle(gc, marking) log;
2626   if (!log.is_trace()) {
2627     return;
2628   }
2629 
2630   log.trace(" Concurrent marking:");
2631   print_ms_time_info("  ", "init marks", _init_times);
2632   print_ms_time_info("  ", "remarks", _remark_times);
2633   {
2634     print_ms_time_info("     ", "final marks", _remark_mark_times);
2635     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2636 
2637   }
2638   print_ms_time_info("  ", "cleanups", _cleanup_times);
2639   log.trace("    Final counting total time = %8.2f s (avg = %8.2f ms).",
2640             _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2641   if (G1ScrubRemSets) {
2642     log.trace("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
2643               _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2644   }
2645   log.trace("  Total stop_world time = %8.2f s.",
2646             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2647   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2648             cmThread()->vtime_accum(), cmThread()->vtime_mark_accum());
2649 }
2650 
2651 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2652   _parallel_workers->print_worker_threads_on(st);
2653 }
2654 
2655 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2656   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2657       p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
2658   _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
2659   _nextMarkBitMap->print_on_error(st, " Next Bits: ");
2660 }
2661 
2662 // We take a break if someone is trying to stop the world.
2663 bool G1ConcurrentMark::do_yield_check(uint worker_id) {
2664   if (SuspendibleThreadSet::should_yield()) {
2665     if (worker_id == 0) {
2666       _g1h->g1_policy()->record_concurrent_pause();
2667     }
2668     SuspendibleThreadSet::yield();
2669     return true;
2670   } else {
2671     return false;
2672   }
2673 }
2674 
2675 // Closure for iteration over bitmaps
2676 class G1CMBitMapClosure : public BitMapClosure {
2677 private:
2678   // the bitmap that is being iterated over
2679   G1CMBitMap*                 _nextMarkBitMap;
2680   G1ConcurrentMark*           _cm;
2681   G1CMTask*                   _task;
2682 
2683 public:
2684   G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm, G1CMBitMap* nextMarkBitMap) :
2685     _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { }
2686 
2687   bool do_bit(size_t offset) {
2688     HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset);
2689     assert(_nextMarkBitMap->isMarked(addr), "invariant");
2690     assert( addr < _cm->finger(), "invariant");
2691     assert(addr >= _task->finger(), "invariant");
2692 
2693     // We move that task's local finger along.
2694     _task->move_finger_to(addr);
2695 
2696     _task->scan_object(oop(addr));
2697     // we only partially drain the local queue and global stack
2698     _task->drain_local_queue(true);
2699     _task->drain_global_stack(true);
2700 
2701     // if the has_aborted flag has been raised, we need to bail out of
2702     // the iteration
2703     return !_task->has_aborted();
2704   }
2705 };
2706 
2707 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2708   ReferenceProcessor* result = NULL;
2709   if (G1UseConcMarkReferenceProcessing) {
2710     result = g1h->ref_processor_cm();
2711     assert(result != NULL, "should not be NULL");
2712   }
2713   return result;
2714 }
2715 
2716 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2717                                G1ConcurrentMark* cm,
2718                                G1CMTask* task)
2719   : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2720     _g1h(g1h), _cm(cm), _task(task)
2721 { }
2722 
2723 void G1CMTask::setup_for_region(HeapRegion* hr) {
2724   assert(hr != NULL,
2725         "claim_region() should have filtered out NULL regions");
2726   _curr_region  = hr;
2727   _finger       = hr->bottom();
2728   update_region_limit();
2729 }
2730 
2731 void G1CMTask::update_region_limit() {
2732   HeapRegion* hr            = _curr_region;
2733   HeapWord* bottom          = hr->bottom();
2734   HeapWord* limit           = hr->next_top_at_mark_start();
2735 
2736   if (limit == bottom) {
2737     // The region was collected underneath our feet.
2738     // We set the finger to bottom to ensure that the bitmap
2739     // iteration that will follow this will not do anything.
2740     // (this is not a condition that holds when we set the region up,
2741     // as the region is not supposed to be empty in the first place)
2742     _finger = bottom;
2743   } else if (limit >= _region_limit) {
2744     assert(limit >= _finger, "peace of mind");
2745   } else {
2746     assert(limit < _region_limit, "only way to get here");
2747     // This can happen under some pretty unusual circumstances.  An
2748     // evacuation pause empties the region underneath our feet (NTAMS
2749     // at bottom). We then do some allocation in the region (NTAMS
2750     // stays at bottom), followed by the region being used as a GC
2751     // alloc region (NTAMS will move to top() and the objects
2752     // originally below it will be grayed). All objects now marked in
2753     // the region are explicitly grayed, if below the global finger,
2754     // and we do not need in fact to scan anything else. So, we simply
2755     // set _finger to be limit to ensure that the bitmap iteration
2756     // doesn't do anything.
2757     _finger = limit;
2758   }
2759 
2760   _region_limit = limit;
2761 }
2762 
2763 void G1CMTask::giveup_current_region() {
2764   assert(_curr_region != NULL, "invariant");
2765   clear_region_fields();
2766 }
2767 
2768 void G1CMTask::clear_region_fields() {
2769   // Values for these three fields that indicate that we're not
2770   // holding on to a region.
2771   _curr_region   = NULL;
2772   _finger        = NULL;
2773   _region_limit  = NULL;
2774 }
2775 
2776 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2777   if (cm_oop_closure == NULL) {
2778     assert(_cm_oop_closure != NULL, "invariant");
2779   } else {
2780     assert(_cm_oop_closure == NULL, "invariant");
2781   }
2782   _cm_oop_closure = cm_oop_closure;
2783 }
2784 
2785 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) {
2786   guarantee(nextMarkBitMap != NULL, "invariant");
2787   _nextMarkBitMap                = nextMarkBitMap;
2788   clear_region_fields();
2789 
2790   _calls                         = 0;
2791   _elapsed_time_ms               = 0.0;
2792   _termination_time_ms           = 0.0;
2793   _termination_start_time_ms     = 0.0;
2794 }
2795 
2796 bool G1CMTask::should_exit_termination() {
2797   regular_clock_call();
2798   // This is called when we are in the termination protocol. We should
2799   // quit if, for some reason, this task wants to abort or the global
2800   // stack is not empty (this means that we can get work from it).
2801   return !_cm->mark_stack_empty() || has_aborted();
2802 }
2803 
2804 void G1CMTask::reached_limit() {
2805   assert(_words_scanned >= _words_scanned_limit ||
2806          _refs_reached >= _refs_reached_limit ,
2807          "shouldn't have been called otherwise");
2808   regular_clock_call();
2809 }
2810 
2811 void G1CMTask::regular_clock_call() {
2812   if (has_aborted()) return;
2813 
2814   // First, we need to recalculate the words scanned and refs reached
2815   // limits for the next clock call.
2816   recalculate_limits();
2817 
2818   // During the regular clock call we do the following
2819 
2820   // (1) If an overflow has been flagged, then we abort.
2821   if (_cm->has_overflown()) {
2822     set_has_aborted();
2823     return;
2824   }
2825 
2826   // If we are not concurrent (i.e. we're doing remark) we don't need
2827   // to check anything else. The other steps are only needed during
2828   // the concurrent marking phase.
2829   if (!concurrent()) return;
2830 
2831   // (2) If marking has been aborted for Full GC, then we also abort.
2832   if (_cm->has_aborted()) {
2833     set_has_aborted();
2834     return;
2835   }
2836 
2837   double curr_time_ms = os::elapsedVTime() * 1000.0;
2838 
2839   // (4) We check whether we should yield. If we have to, then we abort.
2840   if (SuspendibleThreadSet::should_yield()) {
2841     // We should yield. To do this we abort the task. The caller is
2842     // responsible for yielding.
2843     set_has_aborted();
2844     return;
2845   }
2846 
2847   // (5) We check whether we've reached our time quota. If we have,
2848   // then we abort.
2849   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2850   if (elapsed_time_ms > _time_target_ms) {
2851     set_has_aborted();
2852     _has_timed_out = true;
2853     return;
2854   }
2855 
2856   // (6) Finally, we check whether there are enough completed STAB
2857   // buffers available for processing. If there are, we abort.
2858   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2859   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2860     // we do need to process SATB buffers, we'll abort and restart
2861     // the marking task to do so
2862     set_has_aborted();
2863     return;
2864   }
2865 }
2866 
2867 void G1CMTask::recalculate_limits() {
2868   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2869   _words_scanned_limit      = _real_words_scanned_limit;
2870 
2871   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2872   _refs_reached_limit       = _real_refs_reached_limit;
2873 }
2874 
2875 void G1CMTask::decrease_limits() {
2876   // This is called when we believe that we're going to do an infrequent
2877   // operation which will increase the per byte scanned cost (i.e. move
2878   // entries to/from the global stack). It basically tries to decrease the
2879   // scanning limit so that the clock is called earlier.
2880 
2881   _words_scanned_limit = _real_words_scanned_limit -
2882     3 * words_scanned_period / 4;
2883   _refs_reached_limit  = _real_refs_reached_limit -
2884     3 * refs_reached_period / 4;
2885 }
2886 
2887 void G1CMTask::move_entries_to_global_stack() {
2888   // local array where we'll store the entries that will be popped
2889   // from the local queue
2890   oop buffer[global_stack_transfer_size];
2891 
2892   int n = 0;
2893   oop obj;
2894   while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) {
2895     buffer[n] = obj;
2896     ++n;
2897   }
2898 
2899   if (n > 0) {
2900     // we popped at least one entry from the local queue
2901 
2902     if (!_cm->mark_stack_push(buffer, n)) {
2903       set_has_aborted();
2904     }
2905   }
2906 
2907   // this operation was quite expensive, so decrease the limits
2908   decrease_limits();
2909 }
2910 
2911 void G1CMTask::get_entries_from_global_stack() {
2912   // local array where we'll store the entries that will be popped
2913   // from the global stack.
2914   oop buffer[global_stack_transfer_size];
2915   int n;
2916   _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n);
2917   assert(n <= global_stack_transfer_size,
2918          "we should not pop more than the given limit");
2919   if (n > 0) {
2920     // yes, we did actually pop at least one entry
2921     for (int i = 0; i < n; ++i) {
2922       bool success = _task_queue->push(buffer[i]);
2923       // We only call this when the local queue is empty or under a
2924       // given target limit. So, we do not expect this push to fail.
2925       assert(success, "invariant");
2926     }
2927   }
2928 
2929   // this operation was quite expensive, so decrease the limits
2930   decrease_limits();
2931 }
2932 
2933 void G1CMTask::drain_local_queue(bool partially) {
2934   if (has_aborted()) return;
2935 
2936   // Decide what the target size is, depending whether we're going to
2937   // drain it partially (so that other tasks can steal if they run out
2938   // of things to do) or totally (at the very end).
2939   size_t target_size;
2940   if (partially) {
2941     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2942   } else {
2943     target_size = 0;
2944   }
2945 
2946   if (_task_queue->size() > target_size) {
2947     oop obj;
2948     bool ret = _task_queue->pop_local(obj);
2949     while (ret) {
2950       assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" );
2951       assert(!_g1h->is_on_master_free_list(
2952                   _g1h->heap_region_containing((HeapWord*) obj)), "invariant");
2953 
2954       scan_object(obj);
2955 
2956       if (_task_queue->size() <= target_size || has_aborted()) {
2957         ret = false;
2958       } else {
2959         ret = _task_queue->pop_local(obj);
2960       }
2961     }
2962   }
2963 }
2964 
2965 void G1CMTask::drain_global_stack(bool partially) {
2966   if (has_aborted()) return;
2967 
2968   // We have a policy to drain the local queue before we attempt to
2969   // drain the global stack.
2970   assert(partially || _task_queue->size() == 0, "invariant");
2971 
2972   // Decide what the target size is, depending whether we're going to
2973   // drain it partially (so that other tasks can steal if they run out
2974   // of things to do) or totally (at the very end).  Notice that,
2975   // because we move entries from the global stack in chunks or
2976   // because another task might be doing the same, we might in fact
2977   // drop below the target. But, this is not a problem.
2978   size_t target_size;
2979   if (partially) {
2980     target_size = _cm->partial_mark_stack_size_target();
2981   } else {
2982     target_size = 0;
2983   }
2984 
2985   if (_cm->mark_stack_size() > target_size) {
2986     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2987       get_entries_from_global_stack();
2988       drain_local_queue(partially);
2989     }
2990   }
2991 }
2992 
2993 // SATB Queue has several assumptions on whether to call the par or
2994 // non-par versions of the methods. this is why some of the code is
2995 // replicated. We should really get rid of the single-threaded version
2996 // of the code to simplify things.
2997 void G1CMTask::drain_satb_buffers() {
2998   if (has_aborted()) return;
2999 
3000   // We set this so that the regular clock knows that we're in the
3001   // middle of draining buffers and doesn't set the abort flag when it
3002   // notices that SATB buffers are available for draining. It'd be
3003   // very counter productive if it did that. :-)
3004   _draining_satb_buffers = true;
3005 
3006   G1CMSATBBufferClosure satb_cl(this, _g1h);
3007   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
3008 
3009   // This keeps claiming and applying the closure to completed buffers
3010   // until we run out of buffers or we need to abort.
3011   while (!has_aborted() &&
3012          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
3013     regular_clock_call();
3014   }
3015 
3016   _draining_satb_buffers = false;
3017 
3018   assert(has_aborted() ||
3019          concurrent() ||
3020          satb_mq_set.completed_buffers_num() == 0, "invariant");
3021 
3022   // again, this was a potentially expensive operation, decrease the
3023   // limits to get the regular clock call early
3024   decrease_limits();
3025 }
3026 
3027 void G1CMTask::print_stats() {
3028   log_debug(gc, stats)("Marking Stats, task = %u, calls = %d",
3029                        _worker_id, _calls);
3030   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
3031                        _elapsed_time_ms, _termination_time_ms);
3032   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
3033                        _step_times_ms.num(), _step_times_ms.avg(),
3034                        _step_times_ms.sd());
3035   log_debug(gc, stats)("                    max = %1.2lfms, total = %1.2lfms",
3036                        _step_times_ms.maximum(), _step_times_ms.sum());
3037 }
3038 
3039 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) {
3040   return _task_queues->steal(worker_id, hash_seed, obj);
3041 }
3042 
3043 /*****************************************************************************
3044 
3045     The do_marking_step(time_target_ms, ...) method is the building
3046     block of the parallel marking framework. It can be called in parallel
3047     with other invocations of do_marking_step() on different tasks
3048     (but only one per task, obviously) and concurrently with the
3049     mutator threads, or during remark, hence it eliminates the need
3050     for two versions of the code. When called during remark, it will
3051     pick up from where the task left off during the concurrent marking
3052     phase. Interestingly, tasks are also claimable during evacuation
3053     pauses too, since do_marking_step() ensures that it aborts before
3054     it needs to yield.
3055 
3056     The data structures that it uses to do marking work are the
3057     following:
3058 
3059       (1) Marking Bitmap. If there are gray objects that appear only
3060       on the bitmap (this happens either when dealing with an overflow
3061       or when the initial marking phase has simply marked the roots
3062       and didn't push them on the stack), then tasks claim heap
3063       regions whose bitmap they then scan to find gray objects. A
3064       global finger indicates where the end of the last claimed region
3065       is. A local finger indicates how far into the region a task has
3066       scanned. The two fingers are used to determine how to gray an
3067       object (i.e. whether simply marking it is OK, as it will be
3068       visited by a task in the future, or whether it needs to be also
3069       pushed on a stack).
3070 
3071       (2) Local Queue. The local queue of the task which is accessed
3072       reasonably efficiently by the task. Other tasks can steal from
3073       it when they run out of work. Throughout the marking phase, a
3074       task attempts to keep its local queue short but not totally
3075       empty, so that entries are available for stealing by other
3076       tasks. Only when there is no more work, a task will totally
3077       drain its local queue.
3078 
3079       (3) Global Mark Stack. This handles local queue overflow. During
3080       marking only sets of entries are moved between it and the local
3081       queues, as access to it requires a mutex and more fine-grain
3082       interaction with it which might cause contention. If it
3083       overflows, then the marking phase should restart and iterate
3084       over the bitmap to identify gray objects. Throughout the marking
3085       phase, tasks attempt to keep the global mark stack at a small
3086       length but not totally empty, so that entries are available for
3087       popping by other tasks. Only when there is no more work, tasks
3088       will totally drain the global mark stack.
3089 
3090       (4) SATB Buffer Queue. This is where completed SATB buffers are
3091       made available. Buffers are regularly removed from this queue
3092       and scanned for roots, so that the queue doesn't get too
3093       long. During remark, all completed buffers are processed, as
3094       well as the filled in parts of any uncompleted buffers.
3095 
3096     The do_marking_step() method tries to abort when the time target
3097     has been reached. There are a few other cases when the
3098     do_marking_step() method also aborts:
3099 
3100       (1) When the marking phase has been aborted (after a Full GC).
3101 
3102       (2) When a global overflow (on the global stack) has been
3103       triggered. Before the task aborts, it will actually sync up with
3104       the other tasks to ensure that all the marking data structures
3105       (local queues, stacks, fingers etc.)  are re-initialized so that
3106       when do_marking_step() completes, the marking phase can
3107       immediately restart.
3108 
3109       (3) When enough completed SATB buffers are available. The
3110       do_marking_step() method only tries to drain SATB buffers right
3111       at the beginning. So, if enough buffers are available, the
3112       marking step aborts and the SATB buffers are processed at
3113       the beginning of the next invocation.
3114 
3115       (4) To yield. when we have to yield then we abort and yield
3116       right at the end of do_marking_step(). This saves us from a lot
3117       of hassle as, by yielding we might allow a Full GC. If this
3118       happens then objects will be compacted underneath our feet, the
3119       heap might shrink, etc. We save checking for this by just
3120       aborting and doing the yield right at the end.
3121 
3122     From the above it follows that the do_marking_step() method should
3123     be called in a loop (or, otherwise, regularly) until it completes.
3124 
3125     If a marking step completes without its has_aborted() flag being
3126     true, it means it has completed the current marking phase (and
3127     also all other marking tasks have done so and have all synced up).
3128 
3129     A method called regular_clock_call() is invoked "regularly" (in
3130     sub ms intervals) throughout marking. It is this clock method that
3131     checks all the abort conditions which were mentioned above and
3132     decides when the task should abort. A work-based scheme is used to
3133     trigger this clock method: when the number of object words the
3134     marking phase has scanned or the number of references the marking
3135     phase has visited reach a given limit. Additional invocations to
3136     the method clock have been planted in a few other strategic places
3137     too. The initial reason for the clock method was to avoid calling
3138     vtime too regularly, as it is quite expensive. So, once it was in
3139     place, it was natural to piggy-back all the other conditions on it
3140     too and not constantly check them throughout the code.
3141 
3142     If do_termination is true then do_marking_step will enter its
3143     termination protocol.
3144 
3145     The value of is_serial must be true when do_marking_step is being
3146     called serially (i.e. by the VMThread) and do_marking_step should
3147     skip any synchronization in the termination and overflow code.
3148     Examples include the serial remark code and the serial reference
3149     processing closures.
3150 
3151     The value of is_serial must be false when do_marking_step is
3152     being called by any of the worker threads in a work gang.
3153     Examples include the concurrent marking code (CMMarkingTask),
3154     the MT remark code, and the MT reference processing closures.
3155 
3156  *****************************************************************************/
3157 
3158 void G1CMTask::do_marking_step(double time_target_ms,
3159                                bool do_termination,
3160                                bool is_serial) {
3161   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
3162   assert(concurrent() == _cm->concurrent(), "they should be the same");
3163 
3164   G1CollectorPolicy* g1_policy = _g1h->g1_policy();
3165   assert(_task_queues != NULL, "invariant");
3166   assert(_task_queue != NULL, "invariant");
3167   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
3168 
3169   assert(!_claimed,
3170          "only one thread should claim this task at any one time");
3171 
3172   // OK, this doesn't safeguard again all possible scenarios, as it is
3173   // possible for two threads to set the _claimed flag at the same
3174   // time. But it is only for debugging purposes anyway and it will
3175   // catch most problems.
3176   _claimed = true;
3177 
3178   _start_time_ms = os::elapsedVTime() * 1000.0;
3179 
3180   // If do_stealing is true then do_marking_step will attempt to
3181   // steal work from the other G1CMTasks. It only makes sense to
3182   // enable stealing when the termination protocol is enabled
3183   // and do_marking_step() is not being called serially.
3184   bool do_stealing = do_termination && !is_serial;
3185 
3186   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
3187   _time_target_ms = time_target_ms - diff_prediction_ms;
3188 
3189   // set up the variables that are used in the work-based scheme to
3190   // call the regular clock method
3191   _words_scanned = 0;
3192   _refs_reached  = 0;
3193   recalculate_limits();
3194 
3195   // clear all flags
3196   clear_has_aborted();
3197   _has_timed_out = false;
3198   _draining_satb_buffers = false;
3199 
3200   ++_calls;
3201 
3202   // Set up the bitmap and oop closures. Anything that uses them is
3203   // eventually called from this method, so it is OK to allocate these
3204   // statically.
3205   G1CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap);
3206   G1CMOopClosure    cm_oop_closure(_g1h, _cm, this);
3207   set_cm_oop_closure(&cm_oop_closure);
3208 
3209   if (_cm->has_overflown()) {
3210     // This can happen if the mark stack overflows during a GC pause
3211     // and this task, after a yield point, restarts. We have to abort
3212     // as we need to get into the overflow protocol which happens
3213     // right at the end of this task.
3214     set_has_aborted();
3215   }
3216 
3217   // First drain any available SATB buffers. After this, we will not
3218   // look at SATB buffers before the next invocation of this method.
3219   // If enough completed SATB buffers are queued up, the regular clock
3220   // will abort this task so that it restarts.
3221   drain_satb_buffers();
3222   // ...then partially drain the local queue and the global stack
3223   drain_local_queue(true);
3224   drain_global_stack(true);
3225 
3226   do {
3227     if (!has_aborted() && _curr_region != NULL) {
3228       // This means that we're already holding on to a region.
3229       assert(_finger != NULL, "if region is not NULL, then the finger "
3230              "should not be NULL either");
3231 
3232       // We might have restarted this task after an evacuation pause
3233       // which might have evacuated the region we're holding on to
3234       // underneath our feet. Let's read its limit again to make sure
3235       // that we do not iterate over a region of the heap that
3236       // contains garbage (update_region_limit() will also move
3237       // _finger to the start of the region if it is found empty).
3238       update_region_limit();
3239       // We will start from _finger not from the start of the region,
3240       // as we might be restarting this task after aborting half-way
3241       // through scanning this region. In this case, _finger points to
3242       // the address where we last found a marked object. If this is a
3243       // fresh region, _finger points to start().
3244       MemRegion mr = MemRegion(_finger, _region_limit);
3245 
3246       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
3247              "humongous regions should go around loop once only");
3248 
3249       // Some special cases:
3250       // If the memory region is empty, we can just give up the region.
3251       // If the current region is humongous then we only need to check
3252       // the bitmap for the bit associated with the start of the object,
3253       // scan the object if it's live, and give up the region.
3254       // Otherwise, let's iterate over the bitmap of the part of the region
3255       // that is left.
3256       // If the iteration is successful, give up the region.
3257       if (mr.is_empty()) {
3258         giveup_current_region();
3259         regular_clock_call();
3260       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
3261         if (_nextMarkBitMap->isMarked(mr.start())) {
3262           // The object is marked - apply the closure
3263           BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start());
3264           bitmap_closure.do_bit(offset);
3265         }
3266         // Even if this task aborted while scanning the humongous object
3267         // we can (and should) give up the current region.
3268         giveup_current_region();
3269         regular_clock_call();
3270       } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
3271         giveup_current_region();
3272         regular_clock_call();
3273       } else {
3274         assert(has_aborted(), "currently the only way to do so");
3275         // The only way to abort the bitmap iteration is to return
3276         // false from the do_bit() method. However, inside the
3277         // do_bit() method we move the _finger to point to the
3278         // object currently being looked at. So, if we bail out, we
3279         // have definitely set _finger to something non-null.
3280         assert(_finger != NULL, "invariant");
3281 
3282         // Region iteration was actually aborted. So now _finger
3283         // points to the address of the object we last scanned. If we
3284         // leave it there, when we restart this task, we will rescan
3285         // the object. It is easy to avoid this. We move the finger by
3286         // enough to point to the next possible object header (the
3287         // bitmap knows by how much we need to move it as it knows its
3288         // granularity).
3289         assert(_finger < _region_limit, "invariant");
3290         HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger);
3291         // Check if bitmap iteration was aborted while scanning the last object
3292         if (new_finger >= _region_limit) {
3293           giveup_current_region();
3294         } else {
3295           move_finger_to(new_finger);
3296         }
3297       }
3298     }
3299     // At this point we have either completed iterating over the
3300     // region we were holding on to, or we have aborted.
3301 
3302     // We then partially drain the local queue and the global stack.
3303     // (Do we really need this?)
3304     drain_local_queue(true);
3305     drain_global_stack(true);
3306 
3307     // Read the note on the claim_region() method on why it might
3308     // return NULL with potentially more regions available for
3309     // claiming and why we have to check out_of_regions() to determine
3310     // whether we're done or not.
3311     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
3312       // We are going to try to claim a new region. We should have
3313       // given up on the previous one.
3314       // Separated the asserts so that we know which one fires.
3315       assert(_curr_region  == NULL, "invariant");
3316       assert(_finger       == NULL, "invariant");
3317       assert(_region_limit == NULL, "invariant");
3318       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
3319       if (claimed_region != NULL) {
3320         // Yes, we managed to claim one
3321         setup_for_region(claimed_region);
3322         assert(_curr_region == claimed_region, "invariant");
3323       }
3324       // It is important to call the regular clock here. It might take
3325       // a while to claim a region if, for example, we hit a large
3326       // block of empty regions. So we need to call the regular clock
3327       // method once round the loop to make sure it's called
3328       // frequently enough.
3329       regular_clock_call();
3330     }
3331 
3332     if (!has_aborted() && _curr_region == NULL) {
3333       assert(_cm->out_of_regions(),
3334              "at this point we should be out of regions");
3335     }
3336   } while ( _curr_region != NULL && !has_aborted());
3337 
3338   if (!has_aborted()) {
3339     // We cannot check whether the global stack is empty, since other
3340     // tasks might be pushing objects to it concurrently.
3341     assert(_cm->out_of_regions(),
3342            "at this point we should be out of regions");
3343     // Try to reduce the number of available SATB buffers so that
3344     // remark has less work to do.
3345     drain_satb_buffers();
3346   }
3347 
3348   // Since we've done everything else, we can now totally drain the
3349   // local queue and global stack.
3350   drain_local_queue(false);
3351   drain_global_stack(false);
3352 
3353   // Attempt at work stealing from other task's queues.
3354   if (do_stealing && !has_aborted()) {
3355     // We have not aborted. This means that we have finished all that
3356     // we could. Let's try to do some stealing...
3357 
3358     // We cannot check whether the global stack is empty, since other
3359     // tasks might be pushing objects to it concurrently.
3360     assert(_cm->out_of_regions() && _task_queue->size() == 0,
3361            "only way to reach here");
3362     while (!has_aborted()) {
3363       oop obj;
3364       if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
3365         assert(_nextMarkBitMap->isMarked((HeapWord*) obj),
3366                "any stolen object should be marked");
3367         scan_object(obj);
3368 
3369         // And since we're towards the end, let's totally drain the
3370         // local queue and global stack.
3371         drain_local_queue(false);
3372         drain_global_stack(false);
3373       } else {
3374         break;
3375       }
3376     }
3377   }
3378 
3379   // We still haven't aborted. Now, let's try to get into the
3380   // termination protocol.
3381   if (do_termination && !has_aborted()) {
3382     // We cannot check whether the global stack is empty, since other
3383     // tasks might be concurrently pushing objects on it.
3384     // Separated the asserts so that we know which one fires.
3385     assert(_cm->out_of_regions(), "only way to reach here");
3386     assert(_task_queue->size() == 0, "only way to reach here");
3387     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
3388 
3389     // The G1CMTask class also extends the TerminatorTerminator class,
3390     // hence its should_exit_termination() method will also decide
3391     // whether to exit the termination protocol or not.
3392     bool finished = (is_serial ||
3393                      _cm->terminator()->offer_termination(this));
3394     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
3395     _termination_time_ms +=
3396       termination_end_time_ms - _termination_start_time_ms;
3397 
3398     if (finished) {
3399       // We're all done.
3400 
3401       if (_worker_id == 0) {
3402         // let's allow task 0 to do this
3403         if (concurrent()) {
3404           assert(_cm->concurrent_marking_in_progress(), "invariant");
3405           // we need to set this to false before the next
3406           // safepoint. This way we ensure that the marking phase
3407           // doesn't observe any more heap expansions.
3408           _cm->clear_concurrent_marking_in_progress();
3409         }
3410       }
3411 
3412       // We can now guarantee that the global stack is empty, since
3413       // all other tasks have finished. We separated the guarantees so
3414       // that, if a condition is false, we can immediately find out
3415       // which one.
3416       guarantee(_cm->out_of_regions(), "only way to reach here");
3417       guarantee(_cm->mark_stack_empty(), "only way to reach here");
3418       guarantee(_task_queue->size() == 0, "only way to reach here");
3419       guarantee(!_cm->has_overflown(), "only way to reach here");
3420       guarantee(!_cm->mark_stack_overflow(), "only way to reach here");
3421     } else {
3422       // Apparently there's more work to do. Let's abort this task. It
3423       // will restart it and we can hopefully find more things to do.
3424       set_has_aborted();
3425     }
3426   }
3427 
3428   // Mainly for debugging purposes to make sure that a pointer to the
3429   // closure which was statically allocated in this frame doesn't
3430   // escape it by accident.
3431   set_cm_oop_closure(NULL);
3432   double end_time_ms = os::elapsedVTime() * 1000.0;
3433   double elapsed_time_ms = end_time_ms - _start_time_ms;
3434   // Update the step history.
3435   _step_times_ms.add(elapsed_time_ms);
3436 
3437   if (has_aborted()) {
3438     // The task was aborted for some reason.
3439     if (_has_timed_out) {
3440       double diff_ms = elapsed_time_ms - _time_target_ms;
3441       // Keep statistics of how well we did with respect to hitting
3442       // our target only if we actually timed out (if we aborted for
3443       // other reasons, then the results might get skewed).
3444       _marking_step_diffs_ms.add(diff_ms);
3445     }
3446 
3447     if (_cm->has_overflown()) {
3448       // This is the interesting one. We aborted because a global
3449       // overflow was raised. This means we have to restart the
3450       // marking phase and start iterating over regions. However, in
3451       // order to do this we have to make sure that all tasks stop
3452       // what they are doing and re-initialize in a safe manner. We
3453       // will achieve this with the use of two barrier sync points.
3454 
3455       if (!is_serial) {
3456         // We only need to enter the sync barrier if being called
3457         // from a parallel context
3458         _cm->enter_first_sync_barrier(_worker_id);
3459 
3460         // When we exit this sync barrier we know that all tasks have
3461         // stopped doing marking work. So, it's now safe to
3462         // re-initialize our data structures. At the end of this method,
3463         // task 0 will clear the global data structures.
3464       }
3465 
3466       // We clear the local state of this task...
3467       clear_region_fields();
3468 
3469       if (!is_serial) {
3470         // ...and enter the second barrier.
3471         _cm->enter_second_sync_barrier(_worker_id);
3472       }
3473       // At this point, if we're during the concurrent phase of
3474       // marking, everything has been re-initialized and we're
3475       // ready to restart.
3476     }
3477   }
3478 
3479   _claimed = false;
3480 }
3481 
3482 G1CMTask::G1CMTask(uint worker_id,
3483                    G1ConcurrentMark* cm,
3484                    size_t* marked_bytes,
3485                    BitMap* card_bm,
3486                    G1CMTaskQueue* task_queue,
3487                    G1CMTaskQueueSet* task_queues)
3488   : _g1h(G1CollectedHeap::heap()),
3489     _worker_id(worker_id), _cm(cm),
3490     _claimed(false),
3491     _nextMarkBitMap(NULL), _hash_seed(17),
3492     _task_queue(task_queue),
3493     _task_queues(task_queues),
3494     _cm_oop_closure(NULL),
3495     _marked_bytes_array(marked_bytes),
3496     _card_bm(card_bm) {
3497   guarantee(task_queue != NULL, "invariant");
3498   guarantee(task_queues != NULL, "invariant");
3499 
3500   _marking_step_diffs_ms.add(0.5);
3501 }
3502 
3503 // These are formatting macros that are used below to ensure
3504 // consistent formatting. The *_H_* versions are used to format the
3505 // header for a particular value and they should be kept consistent
3506 // with the corresponding macro. Also note that most of the macros add
3507 // the necessary white space (as a prefix) which makes them a bit
3508 // easier to compose.
3509 
3510 // All the output lines are prefixed with this string to be able to
3511 // identify them easily in a large log file.
3512 #define G1PPRL_LINE_PREFIX            "###"
3513 
3514 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
3515 #ifdef _LP64
3516 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
3517 #else // _LP64
3518 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
3519 #endif // _LP64
3520 
3521 // For per-region info
3522 #define G1PPRL_TYPE_FORMAT            "   %-4s"
3523 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
3524 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
3525 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
3526 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
3527 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
3528 
3529 // For summary info
3530 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
3531 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
3532 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
3533 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
3534 
3535 G1PrintRegionLivenessInfoClosure::
3536 G1PrintRegionLivenessInfoClosure(const char* phase_name)
3537   : _total_used_bytes(0), _total_capacity_bytes(0),
3538     _total_prev_live_bytes(0), _total_next_live_bytes(0),
3539     _hum_used_bytes(0), _hum_capacity_bytes(0),
3540     _hum_prev_live_bytes(0), _hum_next_live_bytes(0),
3541     _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
3542   G1CollectedHeap* g1h = G1CollectedHeap::heap();
3543   MemRegion g1_reserved = g1h->g1_reserved();
3544   double now = os::elapsedTime();
3545 
3546   // Print the header of the output.
3547   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
3548   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
3549                           G1PPRL_SUM_ADDR_FORMAT("reserved")
3550                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
3551                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
3552                           HeapRegion::GrainBytes);
3553   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3554   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3555                           G1PPRL_TYPE_H_FORMAT
3556                           G1PPRL_ADDR_BASE_H_FORMAT
3557                           G1PPRL_BYTE_H_FORMAT
3558                           G1PPRL_BYTE_H_FORMAT
3559                           G1PPRL_BYTE_H_FORMAT
3560                           G1PPRL_DOUBLE_H_FORMAT
3561                           G1PPRL_BYTE_H_FORMAT
3562                           G1PPRL_BYTE_H_FORMAT,
3563                           "type", "address-range",
3564                           "used", "prev-live", "next-live", "gc-eff",
3565                           "remset", "code-roots");
3566   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3567                           G1PPRL_TYPE_H_FORMAT
3568                           G1PPRL_ADDR_BASE_H_FORMAT
3569                           G1PPRL_BYTE_H_FORMAT
3570                           G1PPRL_BYTE_H_FORMAT
3571                           G1PPRL_BYTE_H_FORMAT
3572                           G1PPRL_DOUBLE_H_FORMAT
3573                           G1PPRL_BYTE_H_FORMAT
3574                           G1PPRL_BYTE_H_FORMAT,
3575                           "", "",
3576                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3577                           "(bytes)", "(bytes)");
3578 }
3579 
3580 // It takes as a parameter a reference to one of the _hum_* fields, it
3581 // deduces the corresponding value for a region in a humongous region
3582 // series (either the region size, or what's left if the _hum_* field
3583 // is < the region size), and updates the _hum_* field accordingly.
3584 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) {
3585   size_t bytes = 0;
3586   // The > 0 check is to deal with the prev and next live bytes which
3587   // could be 0.
3588   if (*hum_bytes > 0) {
3589     bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes);
3590     *hum_bytes -= bytes;
3591   }
3592   return bytes;
3593 }
3594 
3595 // It deduces the values for a region in a humongous region series
3596 // from the _hum_* fields and updates those accordingly. It assumes
3597 // that that _hum_* fields have already been set up from the "starts
3598 // humongous" region and we visit the regions in address order.
3599 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes,
3600                                                      size_t* capacity_bytes,
3601                                                      size_t* prev_live_bytes,
3602                                                      size_t* next_live_bytes) {
3603   assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition");
3604   *used_bytes      = get_hum_bytes(&_hum_used_bytes);
3605   *capacity_bytes  = get_hum_bytes(&_hum_capacity_bytes);
3606   *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes);
3607   *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes);
3608 }
3609 
3610 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
3611   const char* type       = r->get_type_str();
3612   HeapWord* bottom       = r->bottom();
3613   HeapWord* end          = r->end();
3614   size_t capacity_bytes  = r->capacity();
3615   size_t used_bytes      = r->used();
3616   size_t prev_live_bytes = r->live_bytes();
3617   size_t next_live_bytes = r->next_live_bytes();
3618   double gc_eff          = r->gc_efficiency();
3619   size_t remset_bytes    = r->rem_set()->mem_size();
3620   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3621 
3622   if (r->is_starts_humongous()) {
3623     assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 &&
3624            _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0,
3625            "they should have been zeroed after the last time we used them");
3626     // Set up the _hum_* fields.
3627     _hum_capacity_bytes  = capacity_bytes;
3628     _hum_used_bytes      = used_bytes;
3629     _hum_prev_live_bytes = prev_live_bytes;
3630     _hum_next_live_bytes = next_live_bytes;
3631     get_hum_bytes(&used_bytes, &capacity_bytes,
3632                   &prev_live_bytes, &next_live_bytes);
3633     end = bottom + HeapRegion::GrainWords;
3634   } else if (r->is_continues_humongous()) {
3635     get_hum_bytes(&used_bytes, &capacity_bytes,
3636                   &prev_live_bytes, &next_live_bytes);
3637     assert(end == bottom + HeapRegion::GrainWords, "invariant");
3638   }
3639 
3640   _total_used_bytes      += used_bytes;
3641   _total_capacity_bytes  += capacity_bytes;
3642   _total_prev_live_bytes += prev_live_bytes;
3643   _total_next_live_bytes += next_live_bytes;
3644   _total_remset_bytes    += remset_bytes;
3645   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3646 
3647   // Print a line for this particular region.
3648   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3649                           G1PPRL_TYPE_FORMAT
3650                           G1PPRL_ADDR_BASE_FORMAT
3651                           G1PPRL_BYTE_FORMAT
3652                           G1PPRL_BYTE_FORMAT
3653                           G1PPRL_BYTE_FORMAT
3654                           G1PPRL_DOUBLE_FORMAT
3655                           G1PPRL_BYTE_FORMAT
3656                           G1PPRL_BYTE_FORMAT,
3657                           type, p2i(bottom), p2i(end),
3658                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3659                           remset_bytes, strong_code_roots_bytes);
3660 
3661   return false;
3662 }
3663 
3664 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3665   // add static memory usages to remembered set sizes
3666   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3667   // Print the footer of the output.
3668   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3669   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3670                          " SUMMARY"
3671                          G1PPRL_SUM_MB_FORMAT("capacity")
3672                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3673                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3674                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3675                          G1PPRL_SUM_MB_FORMAT("remset")
3676                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3677                          bytes_to_mb(_total_capacity_bytes),
3678                          bytes_to_mb(_total_used_bytes),
3679                          perc(_total_used_bytes, _total_capacity_bytes),
3680                          bytes_to_mb(_total_prev_live_bytes),
3681                          perc(_total_prev_live_bytes, _total_capacity_bytes),
3682                          bytes_to_mb(_total_next_live_bytes),
3683                          perc(_total_next_live_bytes, _total_capacity_bytes),
3684                          bytes_to_mb(_total_remset_bytes),
3685                          bytes_to_mb(_total_strong_code_roots_bytes));
3686 }
--- EOF ---