1 /*
   2  * Copyright (c) 2001, 2021, 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/classLoaderDataGraph.hpp"
  27 #include "classfile/systemDictionary.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc/g1/g1BarrierSet.hpp"
  30 #include "gc/g1/g1BatchedTask.hpp"
  31 #include "gc/g1/g1CardSetMemory.hpp"
  32 #include "gc/g1/g1CollectedHeap.inline.hpp"
  33 #include "gc/g1/g1CollectorState.hpp"
  34 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  35 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp"
  36 #include "gc/g1/g1DirtyCardQueue.hpp"
  37 #include "gc/g1/g1HeapVerifier.hpp"
  38 #include "gc/g1/g1OopClosures.inline.hpp"
  39 #include "gc/g1/g1Policy.hpp"
  40 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
  41 #include "gc/g1/g1ThreadLocalData.hpp"
  42 #include "gc/g1/g1Trace.hpp"
  43 #include "gc/g1/heapRegion.inline.hpp"
  44 #include "gc/g1/heapRegionManager.hpp"
  45 #include "gc/g1/heapRegionRemSet.inline.hpp"
  46 #include "gc/g1/heapRegionSet.inline.hpp"
  47 #include "gc/shared/gcId.hpp"
  48 #include "gc/shared/gcTimer.hpp"
  49 #include "gc/shared/gcTraceTime.inline.hpp"
  50 #include "gc/shared/gcVMOperations.hpp"
  51 #include "gc/shared/genOopClosures.inline.hpp"
  52 #include "gc/shared/referencePolicy.hpp"
  53 #include "gc/shared/strongRootsScope.hpp"
  54 #include "gc/shared/suspendibleThreadSet.hpp"
  55 #include "gc/shared/taskTerminator.hpp"
  56 #include "gc/shared/taskqueue.inline.hpp"
  57 #include "gc/shared/weakProcessor.inline.hpp"
  58 #include "gc/shared/workerPolicy.hpp"
  59 #include "include/jvm.h"
  60 #include "logging/log.hpp"
  61 #include "memory/allocation.hpp"
  62 #include "memory/iterator.hpp"
  63 #include "memory/metaspaceUtils.hpp"
  64 #include "memory/resourceArea.hpp"
  65 #include "memory/universe.hpp"
  66 #include "oops/access.inline.hpp"
  67 #include "oops/oop.inline.hpp"
  68 #include "runtime/atomic.hpp"
  69 #include "runtime/globals_extension.hpp"
  70 #include "runtime/handles.inline.hpp"
  71 #include "runtime/java.hpp"
  72 #include "runtime/orderAccess.hpp"
  73 #include "runtime/prefetch.inline.hpp"
  74 #include "services/memTracker.hpp"
  75 #include "utilities/align.hpp"
  76 #include "utilities/formatBuffer.hpp"
  77 #include "utilities/growableArray.hpp"
  78 
  79 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
  80   assert(addr < _cm->finger(), "invariant");
  81   assert(addr >= _task->finger(), "invariant");
  82 
  83   // We move that task's local finger along.
  84   _task->move_finger_to(addr);
  85 
  86   _task->scan_task_entry(G1TaskQueueEntry::from_oop(cast_to_oop(addr)));
  87   // we only partially drain the local queue and global stack
  88   _task->drain_local_queue(true);
  89   _task->drain_global_stack(true);
  90 
  91   // if the has_aborted flag has been raised, we need to bail out of
  92   // the iteration
  93   return !_task->has_aborted();
  94 }
  95 
  96 G1CMMarkStack::G1CMMarkStack() :
  97   _max_chunk_capacity(0),
  98   _base(NULL),
  99   _chunk_capacity(0) {
 100   set_empty();
 101 }
 102 
 103 bool G1CMMarkStack::resize(size_t new_capacity) {
 104   assert(is_empty(), "Only resize when stack is empty.");
 105   assert(new_capacity <= _max_chunk_capacity,
 106          "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
 107 
 108   TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
 109 
 110   if (new_base == NULL) {
 111     log_warning(gc)("Failed to reserve memory for new overflow mark stack with " SIZE_FORMAT " chunks and size " SIZE_FORMAT "B.", new_capacity, new_capacity * sizeof(TaskQueueEntryChunk));
 112     return false;
 113   }
 114   // Release old mapping.
 115   if (_base != NULL) {
 116     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 117   }
 118 
 119   _base = new_base;
 120   _chunk_capacity = new_capacity;
 121   set_empty();
 122 
 123   return true;
 124 }
 125 
 126 size_t G1CMMarkStack::capacity_alignment() {
 127   return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
 128 }
 129 
 130 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
 131   guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
 132 
 133   size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
 134 
 135   _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 136   size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 137 
 138   guarantee(initial_chunk_capacity <= _max_chunk_capacity,
 139             "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
 140             _max_chunk_capacity,
 141             initial_chunk_capacity);
 142 
 143   log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
 144                 initial_chunk_capacity, _max_chunk_capacity);
 145 
 146   return resize(initial_chunk_capacity);
 147 }
 148 
 149 void G1CMMarkStack::expand() {
 150   if (_chunk_capacity == _max_chunk_capacity) {
 151     log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
 152     return;
 153   }
 154   size_t old_capacity = _chunk_capacity;
 155   // Double capacity if possible
 156   size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
 157 
 158   if (resize(new_capacity)) {
 159     log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 160                   old_capacity, new_capacity);
 161   } else {
 162     log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 163                     old_capacity, new_capacity);
 164   }
 165 }
 166 
 167 G1CMMarkStack::~G1CMMarkStack() {
 168   if (_base != NULL) {
 169     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 170   }
 171 }
 172 
 173 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
 174   elem->next = *list;
 175   *list = elem;
 176 }
 177 
 178 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
 179   MutexLocker x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 180   add_chunk_to_list(&_chunk_list, elem);
 181   _chunks_in_chunk_list++;
 182 }
 183 
 184 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
 185   MutexLocker x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 186   add_chunk_to_list(&_free_list, elem);
 187 }
 188 
 189 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
 190   TaskQueueEntryChunk* result = *list;
 191   if (result != NULL) {
 192     *list = (*list)->next;
 193   }
 194   return result;
 195 }
 196 
 197 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
 198   MutexLocker x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 199   TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
 200   if (result != NULL) {
 201     _chunks_in_chunk_list--;
 202   }
 203   return result;
 204 }
 205 
 206 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
 207   MutexLocker x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 208   return remove_chunk_from_list(&_free_list);
 209 }
 210 
 211 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
 212   // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
 213   // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
 214   // wraparound of _hwm.
 215   if (_hwm >= _chunk_capacity) {
 216     return NULL;
 217   }
 218 
 219   size_t cur_idx = Atomic::fetch_and_add(&_hwm, 1u);
 220   if (cur_idx >= _chunk_capacity) {
 221     return NULL;
 222   }
 223 
 224   TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
 225   result->next = NULL;
 226   return result;
 227 }
 228 
 229 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
 230   // Get a new chunk.
 231   TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
 232 
 233   if (new_chunk == NULL) {
 234     // Did not get a chunk from the free list. Allocate from backing memory.
 235     new_chunk = allocate_new_chunk();
 236 
 237     if (new_chunk == NULL) {
 238       return false;
 239     }
 240   }
 241 
 242   Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 243 
 244   add_chunk_to_chunk_list(new_chunk);
 245 
 246   return true;
 247 }
 248 
 249 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
 250   TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
 251 
 252   if (cur == NULL) {
 253     return false;
 254   }
 255 
 256   Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 257 
 258   add_chunk_to_free_list(cur);
 259   return true;
 260 }
 261 
 262 void G1CMMarkStack::set_empty() {
 263   _chunks_in_chunk_list = 0;
 264   _hwm = 0;
 265   _chunk_list = NULL;
 266   _free_list = NULL;
 267 }
 268 
 269 G1CMRootMemRegions::G1CMRootMemRegions(uint const max_regions) :
 270     _root_regions(MemRegion::create_array(max_regions, mtGC)),
 271     _max_regions(max_regions),
 272     _num_root_regions(0),
 273     _claimed_root_regions(0),
 274     _scan_in_progress(false),
 275     _should_abort(false) { }
 276 
 277 G1CMRootMemRegions::~G1CMRootMemRegions() {
 278   MemRegion::destroy_array(_root_regions, _max_regions);
 279 }
 280 
 281 void G1CMRootMemRegions::reset() {
 282   _num_root_regions = 0;
 283 }
 284 
 285 void G1CMRootMemRegions::add(HeapWord* start, HeapWord* end) {
 286   assert_at_safepoint();
 287   size_t idx = Atomic::fetch_and_add(&_num_root_regions, 1u);
 288   assert(idx < _max_regions, "Trying to add more root MemRegions than there is space " SIZE_FORMAT, _max_regions);
 289   assert(start != NULL && end != NULL && start <= end, "Start (" PTR_FORMAT ") should be less or equal to "
 290          "end (" PTR_FORMAT ")", p2i(start), p2i(end));
 291   _root_regions[idx].set_start(start);
 292   _root_regions[idx].set_end(end);
 293 }
 294 
 295 void G1CMRootMemRegions::prepare_for_scan() {
 296   assert(!scan_in_progress(), "pre-condition");
 297 
 298   _scan_in_progress = _num_root_regions > 0;
 299 
 300   _claimed_root_regions = 0;
 301   _should_abort = false;
 302 }
 303 
 304 const MemRegion* G1CMRootMemRegions::claim_next() {
 305   if (_should_abort) {
 306     // If someone has set the should_abort flag, we return NULL to
 307     // force the caller to bail out of their loop.
 308     return NULL;
 309   }
 310 
 311   if (_claimed_root_regions >= _num_root_regions) {
 312     return NULL;
 313   }
 314 
 315   size_t claimed_index = Atomic::fetch_and_add(&_claimed_root_regions, 1u);
 316   if (claimed_index < _num_root_regions) {
 317     return &_root_regions[claimed_index];
 318   }
 319   return NULL;
 320 }
 321 
 322 uint G1CMRootMemRegions::num_root_regions() const {
 323   return (uint)_num_root_regions;
 324 }
 325 
 326 void G1CMRootMemRegions::notify_scan_done() {
 327   MutexLocker x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 328   _scan_in_progress = false;
 329   RootRegionScan_lock->notify_all();
 330 }
 331 
 332 void G1CMRootMemRegions::cancel_scan() {
 333   notify_scan_done();
 334 }
 335 
 336 void G1CMRootMemRegions::scan_finished() {
 337   assert(scan_in_progress(), "pre-condition");
 338 
 339   if (!_should_abort) {
 340     assert(_claimed_root_regions >= num_root_regions(),
 341            "we should have claimed all root regions, claimed " SIZE_FORMAT ", length = %u",
 342            _claimed_root_regions, num_root_regions());
 343   }
 344 
 345   notify_scan_done();
 346 }
 347 
 348 bool G1CMRootMemRegions::wait_until_scan_finished() {
 349   if (!scan_in_progress()) {
 350     return false;
 351   }
 352 
 353   {
 354     MonitorLocker ml(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 355     while (scan_in_progress()) {
 356       ml.wait();
 357     }
 358   }
 359   return true;
 360 }
 361 
 362 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
 363                                    G1RegionToSpaceMapper* prev_bitmap_storage,
 364                                    G1RegionToSpaceMapper* next_bitmap_storage) :
 365   // _cm_thread set inside the constructor
 366   _g1h(g1h),
 367 
 368   _mark_bitmap_1(),
 369   _mark_bitmap_2(),
 370   _prev_mark_bitmap(&_mark_bitmap_1),
 371   _next_mark_bitmap(&_mark_bitmap_2),
 372 
 373   _heap(_g1h->reserved()),
 374 
 375   _root_regions(_g1h->max_regions()),
 376 
 377   _global_mark_stack(),
 378 
 379   // _finger set in set_non_marking_state
 380 
 381   _worker_id_offset(G1DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
 382   _max_num_tasks(MAX2(ConcGCThreads, ParallelGCThreads)),
 383   // _num_active_tasks set in set_non_marking_state()
 384   // _tasks set inside the constructor
 385 
 386   _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
 387   _terminator((int) _max_num_tasks, _task_queues),
 388 
 389   _first_overflow_barrier_sync(),
 390   _second_overflow_barrier_sync(),
 391 
 392   _has_overflown(false),
 393   _concurrent(false),
 394   _has_aborted(false),
 395   _restart_for_overflow(false),
 396   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 397   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 398 
 399   // _verbose_level set below
 400 
 401   _init_times(),
 402   _remark_times(),
 403   _remark_mark_times(),
 404   _remark_weak_ref_times(),
 405   _cleanup_times(),
 406   _total_cleanup_time(0.0),
 407 
 408   _accum_task_vtime(NULL),
 409 
 410   _concurrent_workers(NULL),
 411   _num_concurrent_workers(0),
 412   _max_concurrent_workers(0),
 413 
 414   _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_reserved_regions(), mtGC)),
 415   _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_reserved_regions(), mtGC)),
 416   _needs_remembered_set_rebuild(false)
 417 {
 418   assert(CGC_lock != NULL, "CGC_lock must be initialized");
 419 
 420   _mark_bitmap_1.initialize(g1h->reserved(), prev_bitmap_storage);
 421   _mark_bitmap_2.initialize(g1h->reserved(), next_bitmap_storage);
 422 
 423   // Create & start ConcurrentMark thread.
 424   _cm_thread = new G1ConcurrentMarkThread(this);
 425   if (_cm_thread->osthread() == NULL) {
 426     vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 427   }
 428 
 429   log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset);
 430   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 431 
 432   _num_concurrent_workers = ConcGCThreads;
 433   _max_concurrent_workers = _num_concurrent_workers;
 434 
 435   _concurrent_workers = new WorkerThreads("G1 Conc", _max_concurrent_workers);
 436   _concurrent_workers->initialize_workers();
 437 
 438   if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
 439     vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
 440   }
 441 
 442   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
 443   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
 444 
 445   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 446   _num_active_tasks = _max_num_tasks;
 447 
 448   for (uint i = 0; i < _max_num_tasks; ++i) {
 449     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 450     _task_queues->register_queue(i, task_queue);
 451 
 452     _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats);
 453 
 454     _accum_task_vtime[i] = 0.0;
 455   }
 456 
 457   reset_at_marking_complete();
 458 }
 459 
 460 void G1ConcurrentMark::reset() {
 461   _has_aborted = false;
 462 
 463   reset_marking_for_restart();
 464 
 465   // Reset all tasks, since different phases will use different number of active
 466   // threads. So, it's easiest to have all of them ready.
 467   for (uint i = 0; i < _max_num_tasks; ++i) {
 468     _tasks[i]->reset(_next_mark_bitmap);
 469   }
 470 
 471   uint max_reserved_regions = _g1h->max_reserved_regions();
 472   for (uint i = 0; i < max_reserved_regions; i++) {
 473     _top_at_rebuild_starts[i] = NULL;
 474     _region_mark_stats[i].clear();
 475   }
 476 
 477   _root_regions.reset();
 478 }
 479 
 480 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
 481   for (uint j = 0; j < _max_num_tasks; ++j) {
 482     _tasks[j]->clear_mark_stats_cache(region_idx);
 483   }
 484   _top_at_rebuild_starts[region_idx] = NULL;
 485   _region_mark_stats[region_idx].clear();
 486 }
 487 
 488 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
 489   uint const region_idx = r->hrm_index();
 490   if (r->is_humongous()) {
 491     assert(r->is_starts_humongous(), "Got humongous continues region here");
 492     uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(cast_to_oop(r->humongous_start_region()->bottom())->size());
 493     for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
 494       clear_statistics_in_region(j);
 495     }
 496   } else {
 497     clear_statistics_in_region(region_idx);
 498   }
 499 }
 500 
 501 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
 502   if (bitmap->is_marked(addr)) {
 503     bitmap->clear(addr);
 504   }
 505 }
 506 
 507 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
 508   assert_at_safepoint();
 509 
 510   // Need to clear all mark bits of the humongous object.
 511   clear_mark_if_set(_prev_mark_bitmap, r->bottom());
 512   clear_mark_if_set(_next_mark_bitmap, r->bottom());
 513 
 514   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
 515     return;
 516   }
 517 
 518   // Clear any statistics about the region gathered so far.
 519   clear_statistics(r);
 520 }
 521 
 522 void G1ConcurrentMark::reset_marking_for_restart() {
 523   _global_mark_stack.set_empty();
 524 
 525   // Expand the marking stack, if we have to and if we can.
 526   if (has_overflown()) {
 527     _global_mark_stack.expand();
 528 
 529     uint max_reserved_regions = _g1h->max_reserved_regions();
 530     for (uint i = 0; i < max_reserved_regions; i++) {
 531       _region_mark_stats[i].clear_during_overflow();
 532     }
 533   }
 534 
 535   clear_has_overflown();
 536   _finger = _heap.start();
 537 
 538   for (uint i = 0; i < _max_num_tasks; ++i) {
 539     G1CMTaskQueue* queue = _task_queues->queue(i);
 540     queue->set_empty();
 541   }
 542 }
 543 
 544 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 545   assert(active_tasks <= _max_num_tasks, "we should not have more");
 546 
 547   _num_active_tasks = active_tasks;
 548   // Need to update the three data structures below according to the
 549   // number of active threads for this phase.
 550   _terminator.reset_for_reuse(active_tasks);
 551   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 552   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 553 }
 554 
 555 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 556   set_concurrency(active_tasks);
 557 
 558   _concurrent = concurrent;
 559 
 560   if (!concurrent) {
 561     // At this point we should be in a STW phase, and completed marking.
 562     assert_at_safepoint_on_vm_thread();
 563     assert(out_of_regions(),
 564            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 565            p2i(_finger), p2i(_heap.end()));
 566   }
 567 }
 568 
 569 void G1ConcurrentMark::reset_at_marking_complete() {
 570   // We set the global marking state to some default values when we're
 571   // not doing marking.
 572   reset_marking_for_restart();
 573   _num_active_tasks = 0;
 574 }
 575 
 576 G1ConcurrentMark::~G1ConcurrentMark() {
 577   FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
 578   FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
 579   // The G1ConcurrentMark instance is never freed.
 580   ShouldNotReachHere();
 581 }
 582 
 583 class G1ClearBitMapTask : public WorkerTask {
 584 public:
 585   static size_t chunk_size() { return M; }
 586 
 587 private:
 588   // Heap region closure used for clearing the _next_mark_bitmap.
 589   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 590   private:
 591     G1ConcurrentMark* _cm;
 592     G1CMBitMap* _bitmap;
 593     bool _suspendible; // If suspendible, do yield checks.
 594 
 595     bool suspendible() {
 596       return _suspendible;
 597     }
 598 
 599     bool is_clear_concurrent_undo() {
 600       return suspendible() && _cm->cm_thread()->in_undo_mark();
 601     }
 602 
 603     bool has_aborted() {
 604       if (suspendible()) {
 605         _cm->do_yield_check();
 606         return _cm->has_aborted();
 607       }
 608       return false;
 609     }
 610 
 611     HeapWord* region_clear_limit(HeapRegion* r) {
 612       // During a Concurrent Undo Mark cycle, the _next_mark_bitmap is  cleared
 613       // without swapping with the _prev_mark_bitmap. Therefore, the per region
 614       // next_top_at_mark_start and live_words data are current wrt
 615       // _next_mark_bitmap. We use this information to only clear ranges of the
 616       // bitmap that require clearing.
 617       if (is_clear_concurrent_undo()) {
 618         // No need to clear bitmaps for empty regions.
 619         if (_cm->live_words(r->hrm_index()) == 0) {
 620           assert(_bitmap->get_next_marked_addr(r->bottom(), r->end()) == r->end(), "Should not have marked bits");
 621           return r->bottom();
 622         }
 623         assert(_bitmap->get_next_marked_addr(r->next_top_at_mark_start(), r->end()) == r->end(), "Should not have marked bits above ntams");
 624       }
 625       return r->end();
 626     }
 627 
 628   public:
 629     G1ClearBitmapHRClosure(G1ConcurrentMark* cm, bool suspendible) :
 630       HeapRegionClosure(),
 631       _cm(cm),
 632       _bitmap(cm->next_mark_bitmap()),
 633       _suspendible(suspendible)
 634     { }
 635 
 636     virtual bool do_heap_region(HeapRegion* r) {
 637       if (has_aborted()) {
 638         return true;
 639       }
 640 
 641       HeapWord* cur = r->bottom();
 642       HeapWord* const end = region_clear_limit(r);
 643 
 644       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 645 
 646       while (cur < end) {
 647 
 648         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 649         _bitmap->clear_range(mr);
 650 
 651         cur += chunk_size_in_words;
 652 
 653         // Repeat the asserts from before the start of the closure. We will do them
 654         // as asserts here to minimize their overhead on the product. However, we
 655         // will have them as guarantees at the beginning / end of the bitmap
 656         // clearing to get some checking in the product.
 657         assert(!suspendible() || _cm->cm_thread()->in_progress(), "invariant");
 658         assert(!suspendible() || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 659 
 660         // Abort iteration if necessary.
 661         if (has_aborted()) {
 662           return true;
 663         }
 664       }
 665       assert(cur >= end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 666 
 667       return false;
 668     }
 669   };
 670 
 671   G1ClearBitmapHRClosure _cl;
 672   HeapRegionClaimer _hr_claimer;
 673   bool _suspendible; // If the task is suspendible, workers must join the STS.
 674 
 675 public:
 676   G1ClearBitMapTask(G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 677     WorkerTask("G1 Clear Bitmap"),
 678     _cl(cm, suspendible),
 679     _hr_claimer(n_workers),
 680     _suspendible(suspendible)
 681   { }
 682 
 683   void work(uint worker_id) {
 684     SuspendibleThreadSetJoiner sts_join(_suspendible);
 685     G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
 686   }
 687 
 688   bool is_complete() {
 689     return _cl.is_complete();
 690   }
 691 };
 692 
 693 void G1ConcurrentMark::clear_next_bitmap(WorkerThreads* workers, bool may_yield) {
 694   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 695 
 696   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 697   size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 698 
 699   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 700 
 701   G1ClearBitMapTask cl(this, num_workers, may_yield);
 702 
 703   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 704   workers->run_task(&cl, num_workers);
 705   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 706 }
 707 
 708 void G1ConcurrentMark::cleanup_for_next_mark() {
 709   // Make sure that the concurrent mark thread looks to still be in
 710   // the current cycle.
 711   guarantee(cm_thread()->in_progress(), "invariant");
 712 
 713   // We are finishing up the current cycle by clearing the next
 714   // marking bitmap and getting it ready for the next cycle. During
 715   // this time no other cycle can start. So, let's make sure that this
 716   // is the case.
 717   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 718 
 719   clear_next_bitmap(_concurrent_workers, true);
 720 
 721   // Repeat the asserts from above.
 722   guarantee(cm_thread()->in_progress(), "invariant");
 723   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 724 }
 725 
 726 void G1ConcurrentMark::clear_next_bitmap(WorkerThreads* workers) {
 727   assert_at_safepoint_on_vm_thread();
 728   // To avoid fragmentation the full collection requesting to clear the bitmap
 729   // might use fewer workers than available. To ensure the bitmap is cleared
 730   // as efficiently as possible the number of active workers are temporarily
 731   // increased to include all currently created workers.
 732   WithActiveWorkers update(workers, workers->created_workers());
 733   clear_next_bitmap(workers, false);
 734 }
 735 
 736 class G1PreConcurrentStartTask : public G1BatchedTask {
 737   // Concurrent start needs claim bits to keep track of the marked-through CLDs.
 738   class CLDClearClaimedMarksTask;
 739   // Reset marking state.
 740   class ResetMarkingStateTask;
 741   // For each region note start of marking.
 742   class NoteStartOfMarkTask;
 743 
 744 public:
 745   G1PreConcurrentStartTask(GCCause::Cause cause, G1ConcurrentMark* cm);
 746 };
 747 
 748 class G1PreConcurrentStartTask::CLDClearClaimedMarksTask : public G1AbstractSubTask {
 749 public:
 750   CLDClearClaimedMarksTask() : G1AbstractSubTask(G1GCPhaseTimes::CLDClearClaimedMarks) { }
 751 
 752   double worker_cost() const override { return 1.0; }
 753   void do_work(uint worker_id) override;
 754 };
 755 
 756 class G1PreConcurrentStartTask::ResetMarkingStateTask : public G1AbstractSubTask {
 757   G1ConcurrentMark* _cm;
 758 public:
 759   ResetMarkingStateTask(G1ConcurrentMark* cm) : G1AbstractSubTask(G1GCPhaseTimes::ResetMarkingState), _cm(cm) { }
 760 
 761   double worker_cost() const override { return 1.0; }
 762   void do_work(uint worker_id) override;
 763 };
 764 
 765 class G1PreConcurrentStartTask::NoteStartOfMarkTask : public G1AbstractSubTask {
 766   HeapRegionClaimer _claimer;
 767 public:
 768   NoteStartOfMarkTask() : G1AbstractSubTask(G1GCPhaseTimes::NoteStartOfMark), _claimer(0) { }
 769 
 770   double worker_cost() const override {
 771     // The work done per region is very small, therefore we choose this magic number to cap the number
 772     // of threads used when there are few regions.
 773     const uint regions_per_thread = 1000;
 774     return _claimer.n_regions() / regions_per_thread;
 775   }
 776 
 777   void set_max_workers(uint max_workers) override;
 778   void do_work(uint worker_id) override;
 779 };
 780 
 781 void G1PreConcurrentStartTask::CLDClearClaimedMarksTask::do_work(uint worker_id) {
 782   ClassLoaderDataGraph::clear_claimed_marks();
 783 }
 784 
 785 void G1PreConcurrentStartTask::ResetMarkingStateTask::do_work(uint worker_id) {
 786   // Reset marking state.
 787   _cm->reset();
 788 }
 789 
 790 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
 791 public:
 792   bool do_heap_region(HeapRegion* r) override {
 793     r->note_start_of_marking();
 794     return false;
 795   }
 796 };
 797 
 798 void G1PreConcurrentStartTask::NoteStartOfMarkTask::do_work(uint worker_id) {
 799   NoteStartOfMarkHRClosure start_cl;
 800   G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&start_cl, &_claimer, worker_id);
 801 }
 802 
 803 void G1PreConcurrentStartTask::NoteStartOfMarkTask::set_max_workers(uint max_workers) {
 804   _claimer.set_n_workers(max_workers);
 805 }
 806 
 807 G1PreConcurrentStartTask::G1PreConcurrentStartTask(GCCause::Cause cause, G1ConcurrentMark* cm) :
 808   G1BatchedTask("Pre Concurrent Start", G1CollectedHeap::heap()->phase_times()) {
 809   add_serial_task(new CLDClearClaimedMarksTask());
 810   add_serial_task(new ResetMarkingStateTask(cm));
 811   add_parallel_task(new NoteStartOfMarkTask());
 812 };
 813 
 814 void G1ConcurrentMark::pre_concurrent_start(GCCause::Cause cause) {
 815   assert_at_safepoint_on_vm_thread();


 816 
 817   G1PreConcurrentStartTask cl(cause, this);
 818   G1CollectedHeap::heap()->run_batch_task(&cl);
 819 
 820   _gc_tracer_cm->set_gc_cause(cause);
 821 }
 822 
 823 
 824 void G1ConcurrentMark::post_concurrent_mark_start() {
 825   // Start Concurrent Marking weak-reference discovery.
 826   ReferenceProcessor* rp = _g1h->ref_processor_cm();
 827   rp->start_discovery(false /* always_clear */);
 828 
 829   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
 830   // This is the start of  the marking cycle, we're expected all
 831   // threads to have SATB queues with active set to false.
 832   satb_mq_set.set_active_all_threads(true, /* new active value */
 833                                      false /* expected_active */);
 834 
 835   _root_regions.prepare_for_scan();
 836 
 837   // update_g1_committed() will be called at the end of an evac pause
 838   // when marking is on. So, it's also called at the end of the
 839   // concurrent start pause to update the heap end, if the heap expands
 840   // during it. No need to call it here.
 841 }
 842 
 843 void G1ConcurrentMark::post_concurrent_undo_start() {
 844   root_regions()->cancel_scan();
 845 }
 846 
 847 /*
 848  * Notice that in the next two methods, we actually leave the STS
 849  * during the barrier sync and join it immediately afterwards. If we
 850  * do not do this, the following deadlock can occur: one thread could
 851  * be in the barrier sync code, waiting for the other thread to also
 852  * sync up, whereas another one could be trying to yield, while also
 853  * waiting for the other threads to sync up too.
 854  *
 855  * Note, however, that this code is also used during remark and in
 856  * this case we should not attempt to leave / enter the STS, otherwise
 857  * we'll either hit an assert (debug / fastdebug) or deadlock
 858  * (product). So we should only leave / enter the STS if we are
 859  * operating concurrently.
 860  *
 861  * Because the thread that does the sync barrier has left the STS, it
 862  * is possible to be suspended for a Full GC or an evacuation pause
 863  * could occur. This is actually safe, since the entering the sync
 864  * barrier is one of the last things do_marking_step() does, and it
 865  * doesn't manipulate any data structures afterwards.
 866  */
 867 
 868 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 869   bool barrier_aborted;
 870   {
 871     SuspendibleThreadSetLeaver sts_leave(concurrent());
 872     barrier_aborted = !_first_overflow_barrier_sync.enter();
 873   }
 874 
 875   // at this point everyone should have synced up and not be doing any
 876   // more work
 877 
 878   if (barrier_aborted) {
 879     // If the barrier aborted we ignore the overflow condition and
 880     // just abort the whole marking phase as quickly as possible.
 881     return;
 882   }
 883 }
 884 
 885 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 886   SuspendibleThreadSetLeaver sts_leave(concurrent());
 887   _second_overflow_barrier_sync.enter();
 888 
 889   // at this point everything should be re-initialized and ready to go
 890 }
 891 
 892 class G1CMConcurrentMarkingTask : public WorkerTask {
 893   G1ConcurrentMark*     _cm;
 894 
 895 public:
 896   void work(uint worker_id) {
 897     ResourceMark rm;
 898 
 899     double start_vtime = os::elapsedVTime();
 900 
 901     {
 902       SuspendibleThreadSetJoiner sts_join;
 903 
 904       assert(worker_id < _cm->active_tasks(), "invariant");
 905 
 906       G1CMTask* task = _cm->task(worker_id);
 907       task->record_start_time();
 908       if (!_cm->has_aborted()) {
 909         do {
 910           task->do_marking_step(G1ConcMarkStepDurationMillis,
 911                                 true  /* do_termination */,
 912                                 false /* is_serial*/);
 913 
 914           _cm->do_yield_check();
 915         } while (!_cm->has_aborted() && task->has_aborted());
 916       }
 917       task->record_end_time();
 918       guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
 919     }
 920 
 921     double end_vtime = os::elapsedVTime();
 922     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 923   }
 924 
 925   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
 926       WorkerTask("Concurrent Mark"), _cm(cm) { }
 927 
 928   ~G1CMConcurrentMarkingTask() { }
 929 };
 930 
 931 uint G1ConcurrentMark::calc_active_marking_workers() {
 932   uint result = 0;
 933   if (!UseDynamicNumberOfGCThreads || !FLAG_IS_DEFAULT(ConcGCThreads)) {
 934     result = _max_concurrent_workers;
 935   } else {
 936     result =
 937       WorkerPolicy::calc_default_active_workers(_max_concurrent_workers,
 938                                                 1, /* Minimum workers */
 939                                                 _num_concurrent_workers,
 940                                                 Threads::number_of_non_daemon_threads());
 941     // Don't scale the result down by scale_concurrent_workers() because
 942     // that scaling has already gone into "_max_concurrent_workers".
 943   }
 944   assert(result > 0 && result <= _max_concurrent_workers,
 945          "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
 946          _max_concurrent_workers, result);
 947   return result;
 948 }
 949 
 950 void G1ConcurrentMark::scan_root_region(const MemRegion* region, uint worker_id) {
 951 #ifdef ASSERT
 952   HeapWord* last = region->last();
 953   HeapRegion* hr = _g1h->heap_region_containing(last);
 954   assert(hr->is_old() || hr->next_top_at_mark_start() == hr->bottom(),
 955          "Root regions must be old or survivor/eden but region %u is %s", hr->hrm_index(), hr->get_type_str());
 956   assert(hr->next_top_at_mark_start() == region->start(),
 957          "MemRegion start should be equal to nTAMS");
 958 #endif
 959 
 960   G1RootRegionScanClosure cl(_g1h, this, worker_id);
 961 
 962   const uintx interval = PrefetchScanIntervalInBytes;
 963   HeapWord* curr = region->start();
 964   const HeapWord* end = region->end();
 965   while (curr < end) {
 966     Prefetch::read(curr, interval);
 967     oop obj = cast_to_oop(curr);
 968     size_t size = obj->oop_iterate_size(&cl);
 969     assert(size == obj->size(), "sanity");
 970     curr += size;
 971   }
 972 }
 973 
 974 class G1CMRootRegionScanTask : public WorkerTask {
 975   G1ConcurrentMark* _cm;
 976 public:
 977   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 978     WorkerTask("G1 Root Region Scan"), _cm(cm) { }
 979 
 980   void work(uint worker_id) {
 981     G1CMRootMemRegions* root_regions = _cm->root_regions();
 982     const MemRegion* region = root_regions->claim_next();
 983     while (region != NULL) {
 984       _cm->scan_root_region(region, worker_id);
 985       region = root_regions->claim_next();
 986     }
 987   }
 988 };
 989 
 990 void G1ConcurrentMark::scan_root_regions() {
 991   // scan_in_progress() will have been set to true only if there was
 992   // at least one root region to scan. So, if it's false, we
 993   // should not attempt to do any further work.
 994   if (root_regions()->scan_in_progress()) {
 995     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 996 
 997     _num_concurrent_workers = MIN2(calc_active_marking_workers(),
 998                                    // We distribute work on a per-region basis, so starting
 999                                    // more threads than that is useless.
1000                                    root_regions()->num_root_regions());
1001     assert(_num_concurrent_workers <= _max_concurrent_workers,
1002            "Maximum number of marking threads exceeded");
1003 
1004     G1CMRootRegionScanTask task(this);
1005     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
1006                         task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
1007     _concurrent_workers->run_task(&task, _num_concurrent_workers);
1008 
1009     // It's possible that has_aborted() is true here without actually
1010     // aborting the survivor scan earlier. This is OK as it's
1011     // mainly used for sanity checking.
1012     root_regions()->scan_finished();
1013   }
1014 }
1015 
1016 void G1ConcurrentMark::concurrent_cycle_start() {
1017   _gc_timer_cm->register_gc_start();
1018 
1019   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
1020 
1021   _g1h->trace_heap_before_gc(_gc_tracer_cm);
1022 }
1023 
1024 void G1ConcurrentMark::concurrent_cycle_end() {
1025   _g1h->collector_state()->set_clearing_next_bitmap(false);
1026 
1027   _g1h->trace_heap_after_gc(_gc_tracer_cm);
1028 
1029   if (has_aborted()) {
1030     log_info(gc, marking)("Concurrent Mark Abort");
1031     _gc_tracer_cm->report_concurrent_mode_failure();
1032   }
1033 
1034   _gc_timer_cm->register_gc_end();
1035 
1036   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1037 }
1038 
1039 void G1ConcurrentMark::mark_from_roots() {
1040   _restart_for_overflow = false;
1041 
1042   _num_concurrent_workers = calc_active_marking_workers();
1043 
1044   uint active_workers = MAX2(1U, _num_concurrent_workers);
1045 
1046   // Setting active workers is not guaranteed since fewer
1047   // worker threads may currently exist and more may not be
1048   // available.
1049   active_workers = _concurrent_workers->set_active_workers(active_workers);
1050   log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->max_workers());
1051 
1052   // Parallel task terminator is set in "set_concurrency_and_phase()"
1053   set_concurrency_and_phase(active_workers, true /* concurrent */);
1054 
1055   G1CMConcurrentMarkingTask marking_task(this);
1056   _concurrent_workers->run_task(&marking_task);
1057   print_stats();
1058 }
1059 
1060 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
1061   G1HeapVerifier* verifier = _g1h->verifier();
1062 
1063   verifier->verify_region_sets_optional();
1064 
1065   if (VerifyDuringGC) {
1066     GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm);
1067 
1068     size_t const BufLen = 512;
1069     char buffer[BufLen];
1070 
1071     jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
1072     verifier->verify(type, vo, buffer);
1073   }
1074 
1075   verifier->check_bitmaps(caller);
1076 }
1077 
1078 class G1UpdateRemSetTrackingBeforeRebuildTask : public WorkerTask {
1079   G1CollectedHeap* _g1h;
1080   G1ConcurrentMark* _cm;
1081   HeapRegionClaimer _hrclaimer;
1082   uint volatile _total_selected_for_rebuild;
1083 
1084   G1PrintRegionLivenessInfoClosure _cl;
1085 
1086   class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
1087     G1CollectedHeap* _g1h;
1088     G1ConcurrentMark* _cm;
1089 
1090     G1PrintRegionLivenessInfoClosure* _cl;
1091 
1092     uint _num_regions_selected_for_rebuild;  // The number of regions actually selected for rebuild.
1093 
1094     void update_remset_before_rebuild(HeapRegion* hr) {
1095       G1RemSetTrackingPolicy* tracking_policy = _g1h->policy()->remset_tracker();
1096 
1097       bool selected_for_rebuild;
1098       if (hr->is_humongous()) {
1099         bool const is_live = _cm->live_words(hr->humongous_start_region()->hrm_index()) > 0;
1100         selected_for_rebuild = tracking_policy->update_humongous_before_rebuild(hr, is_live);
1101       } else {
1102         size_t const live_bytes = _cm->live_bytes(hr->hrm_index());
1103         selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1104       }
1105       if (selected_for_rebuild) {
1106         _num_regions_selected_for_rebuild++;
1107       }
1108       _cm->update_top_at_rebuild_start(hr);
1109     }
1110 
1111     // Distribute the given words across the humongous object starting with hr and
1112     // note end of marking.
1113     void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1114       uint const region_idx = hr->hrm_index();
1115       size_t const obj_size_in_words = (size_t)cast_to_oop(hr->bottom())->size();
1116       uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1117 
1118       // "Distributing" zero words means that we only note end of marking for these
1119       // regions.
1120       assert(marked_words == 0 || obj_size_in_words == marked_words,
1121              "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1122              obj_size_in_words, marked_words);
1123 
1124       for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1125         HeapRegion* const r = _g1h->region_at(i);
1126         size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1127 
1128         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)",
1129                                words_to_add, i, r->get_type_str());
1130         add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1131         marked_words -= words_to_add;
1132       }
1133       assert(marked_words == 0,
1134              SIZE_FORMAT " words left after distributing space across %u regions",
1135              marked_words, num_regions_in_humongous);
1136     }
1137 
1138     void update_marked_bytes(HeapRegion* hr) {
1139       uint const region_idx = hr->hrm_index();
1140       size_t const marked_words = _cm->live_words(region_idx);
1141       // The marking attributes the object's size completely to the humongous starts
1142       // region. We need to distribute this value across the entire set of regions a
1143       // humongous object spans.
1144       if (hr->is_humongous()) {
1145         assert(hr->is_starts_humongous() || marked_words == 0,
1146                "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1147                marked_words, region_idx, hr->get_type_str());
1148         if (hr->is_starts_humongous()) {
1149           distribute_marked_bytes(hr, marked_words);
1150         }
1151       } else {
1152         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1153         add_marked_bytes_and_note_end(hr, _cm->live_bytes(region_idx));
1154       }
1155     }
1156 
1157     void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1158       hr->add_to_marked_bytes(marked_bytes);
1159       _cl->do_heap_region(hr);
1160       hr->note_end_of_marking();
1161     }
1162 
1163   public:
1164     G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1165       _g1h(g1h), _cm(cm), _cl(cl), _num_regions_selected_for_rebuild(0) { }
1166 
1167     virtual bool do_heap_region(HeapRegion* r) {
1168       update_remset_before_rebuild(r);
1169       update_marked_bytes(r);
1170 
1171       return false;
1172     }
1173 
1174     uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1175   };
1176 
1177 public:
1178   G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1179     WorkerTask("G1 Update RemSet Tracking Before Rebuild"),
1180     _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1181 
1182   virtual void work(uint worker_id) {
1183     G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1184     _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1185     Atomic::add(&_total_selected_for_rebuild, update_cl.num_selected_for_rebuild());
1186   }
1187 
1188   uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1189 
1190   // Number of regions for which roughly one thread should be spawned for this work.
1191   static const uint RegionsPerThread = 384;
1192 };
1193 
1194 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1195   G1CollectedHeap* _g1h;
1196 public:
1197   G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1198 
1199   virtual bool do_heap_region(HeapRegion* r) {
1200     _g1h->policy()->remset_tracker()->update_after_rebuild(r);
1201     return false;
1202   }
1203 };
1204 
1205 void G1ConcurrentMark::remark() {
1206   assert_at_safepoint_on_vm_thread();
1207 
1208   // If a full collection has happened, we should not continue. However we might
1209   // have ended up here as the Remark VM operation has been scheduled already.
1210   if (has_aborted()) {
1211     return;
1212   }
1213 
1214   G1Policy* policy = _g1h->policy();
1215   policy->record_concurrent_mark_remark_start();
1216 
1217   double start = os::elapsedTime();
1218 
1219   verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1220 
1221   {
1222     GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm);
1223     finalize_marking();
1224   }
1225 
1226   double mark_work_end = os::elapsedTime();
1227 
1228   bool const mark_finished = !has_overflown();
1229   if (mark_finished) {
1230     weak_refs_work();
1231 
1232     SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1233     // We're done with marking.
1234     // This is the end of the marking cycle, we're expected all
1235     // threads to have SATB queues with active set to true.
1236     satb_mq_set.set_active_all_threads(false, /* new active value */
1237                                        true /* expected_active */);
1238 
1239     {
1240       GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm);
1241       flush_all_task_caches();
1242     }
1243 
1244     // Install newly created mark bitmap as "prev".
1245     swap_mark_bitmaps();
1246 
1247     _g1h->collector_state()->set_clearing_next_bitmap(true);
1248     {
1249       GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1250 
1251       uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1252                                        G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1253       uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1254 
1255       G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1256       log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1257       _g1h->workers()->run_task(&cl, num_workers);
1258 
1259       log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1260                                       _g1h->num_regions(), cl.total_selected_for_rebuild());
1261 
1262       _needs_remembered_set_rebuild = (cl.total_selected_for_rebuild() > 0);
1263     }
1264     {
1265       GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm);
1266       reclaim_empty_regions();
1267     }
1268 
1269     // Clean out dead classes
1270     if (ClassUnloadingWithConcurrentMark) {
1271       GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm);
1272       ClassLoaderDataGraph::purge(/*at_safepoint*/true);
1273     }
1274 
1275     _g1h->resize_heap_if_necessary();
1276     _g1h->uncommit_regions_if_necessary();
1277 
1278     compute_new_sizes();
1279 
1280     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1281 
1282     assert(!restart_for_overflow(), "sanity");
1283     // Completely reset the marking state since marking completed
1284     reset_at_marking_complete();
1285   } else {
1286     // We overflowed.  Restart concurrent marking.
1287     _restart_for_overflow = true;
1288 
1289     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1290 
1291     // Clear the marking state because we will be restarting
1292     // marking due to overflowing the global mark stack.
1293     reset_marking_for_restart();
1294   }
1295 
1296   {
1297     GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm);
1298     report_object_count(mark_finished);
1299   }
1300 
1301   // Statistics
1302   double now = os::elapsedTime();
1303   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1304   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1305   _remark_times.add((now - start) * 1000.0);
1306 
1307   policy->record_concurrent_mark_remark_end();

1308 }
1309 
1310 class G1ReclaimEmptyRegionsTask : public WorkerTask {
1311   // Per-region work during the Cleanup pause.
1312   class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1313     G1CollectedHeap* _g1h;
1314     size_t _freed_bytes;
1315     FreeRegionList* _local_cleanup_list;
1316     uint _old_regions_removed;
1317     uint _archive_regions_removed;
1318     uint _humongous_regions_removed;
1319 
1320   public:
1321     G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1322                                  FreeRegionList* local_cleanup_list) :
1323       _g1h(g1h),
1324       _freed_bytes(0),
1325       _local_cleanup_list(local_cleanup_list),
1326       _old_regions_removed(0),
1327       _archive_regions_removed(0),
1328       _humongous_regions_removed(0) { }
1329 
1330     size_t freed_bytes() { return _freed_bytes; }
1331     const uint old_regions_removed() { return _old_regions_removed; }
1332     const uint archive_regions_removed() { return _archive_regions_removed; }
1333     const uint humongous_regions_removed() { return _humongous_regions_removed; }
1334 
1335     bool do_heap_region(HeapRegion *hr) {
1336       if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_closed_archive()) {
1337         log_trace(gc)("Reclaimed empty old gen region %u (%s) bot " PTR_FORMAT,
1338                       hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1339         _freed_bytes += hr->used();
1340         hr->set_containing_set(NULL);
1341         if (hr->is_humongous()) {
1342           _humongous_regions_removed++;
1343           _g1h->free_humongous_region(hr, _local_cleanup_list);
1344         } else if (hr->is_open_archive()) {
1345           _archive_regions_removed++;
1346           _g1h->free_region(hr, _local_cleanup_list);
1347         } else {
1348           _old_regions_removed++;
1349           _g1h->free_region(hr, _local_cleanup_list);
1350         }
1351         hr->clear_cardtable();
1352         _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1353       }
1354 
1355       return false;
1356     }
1357   };
1358 
1359   G1CollectedHeap* _g1h;
1360   FreeRegionList* _cleanup_list;
1361   HeapRegionClaimer _hrclaimer;
1362 
1363 public:
1364   G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1365     WorkerTask("G1 Cleanup"),
1366     _g1h(g1h),
1367     _cleanup_list(cleanup_list),
1368     _hrclaimer(n_workers) {
1369   }
1370 
1371   void work(uint worker_id) {
1372     FreeRegionList local_cleanup_list("Local Cleanup List");
1373     G1ReclaimEmptyRegionsClosure cl(_g1h, &local_cleanup_list);
1374     _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1375     assert(cl.is_complete(), "Shouldn't have aborted!");
1376 
1377     // Now update the old/archive/humongous region sets
1378     _g1h->remove_from_old_gen_sets(cl.old_regions_removed(),
1379                                    cl.archive_regions_removed(),
1380                                    cl.humongous_regions_removed());
1381     {
1382       MutexLocker x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1383       _g1h->decrement_summary_bytes(cl.freed_bytes());
1384 
1385       _cleanup_list->add_ordered(&local_cleanup_list);
1386       assert(local_cleanup_list.is_empty(), "post-condition");
1387     }
1388   }
1389 };
1390 
1391 void G1ConcurrentMark::reclaim_empty_regions() {
1392   WorkerThreads* workers = _g1h->workers();
1393   FreeRegionList empty_regions_list("Empty Regions After Mark List");
1394 
1395   G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1396   workers->run_task(&cl);
1397 
1398   if (!empty_regions_list.is_empty()) {
1399     log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1400     // Now print the empty regions list.
1401     _g1h->hr_printer()->cleanup(&empty_regions_list);
1402     // And actually make them available.
1403     _g1h->prepend_to_freelist(&empty_regions_list);
1404   }
1405 }
1406 
1407 void G1ConcurrentMark::compute_new_sizes() {
1408   MetaspaceGC::compute_new_size();
1409 
1410   // Cleanup will have freed any regions completely full of garbage.
1411   // Update the soft reference policy with the new heap occupancy.
1412   Universe::heap()->update_capacity_and_used_at_gc();
1413 
1414   // We reclaimed old regions so we should calculate the sizes to make
1415   // sure we update the old gen/space data.
1416   _g1h->monitoring_support()->update_sizes();
1417 }
1418 
1419 void G1ConcurrentMark::cleanup() {
1420   assert_at_safepoint_on_vm_thread();
1421 
1422   // If a full collection has happened, we shouldn't do this.
1423   if (has_aborted()) {
1424     return;
1425   }
1426 
1427   G1Policy* policy = _g1h->policy();
1428   policy->record_concurrent_mark_cleanup_start();
1429 
1430   double start = os::elapsedTime();
1431 
1432   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1433 
1434   if (needs_remembered_set_rebuild()) {
1435     GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1436     G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1437     _g1h->heap_region_iterate(&cl);
1438   } else {
1439     log_debug(gc, phases)("No Remembered Sets to update after rebuild");
1440   }
1441 
1442   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1443 
1444   // We need to make this be a "collection" so any collection pause that
1445   // races with it goes around and waits for Cleanup to finish.
1446   _g1h->increment_total_collections();
1447 
1448   // Local statistics
1449   double recent_cleanup_time = (os::elapsedTime() - start);
1450   _total_cleanup_time += recent_cleanup_time;
1451   _cleanup_times.add(recent_cleanup_time);
1452 
1453   {
1454     GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1455     policy->record_concurrent_mark_cleanup_end(needs_remembered_set_rebuild());
1456   }
1457 }
1458 
1459 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1460 // Uses the G1CMTask associated with a worker thread (for serial reference
1461 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1462 // trace referent objects.
1463 //
1464 // Using the G1CMTask and embedded local queues avoids having the worker
1465 // threads operating on the global mark stack. This reduces the risk
1466 // of overflowing the stack - which we would rather avoid at this late
1467 // state. Also using the tasks' local queues removes the potential
1468 // of the workers interfering with each other that could occur if
1469 // operating on the global stack.
1470 
1471 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1472   G1ConcurrentMark* _cm;
1473   G1CMTask*         _task;
1474   uint              _ref_counter_limit;
1475   uint              _ref_counter;
1476   bool              _is_serial;
1477 public:
1478   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1479     _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval),
1480     _ref_counter(_ref_counter_limit), _is_serial(is_serial) {
1481     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1482   }
1483 
1484   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1485   virtual void do_oop(      oop* p) { do_oop_work(p); }
1486 
1487   template <class T> void do_oop_work(T* p) {
1488     if (_cm->has_overflown()) {
1489       return;
1490     }
1491     if (!_task->deal_with_reference(p)) {
1492       // We did not add anything to the mark bitmap (or mark stack), so there is
1493       // no point trying to drain it.
1494       return;
1495     }
1496     _ref_counter--;
1497 
1498     if (_ref_counter == 0) {
1499       // We have dealt with _ref_counter_limit references, pushing them
1500       // and objects reachable from them on to the local stack (and
1501       // possibly the global stack). Call G1CMTask::do_marking_step() to
1502       // process these entries.
1503       //
1504       // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1505       // there's nothing more to do (i.e. we're done with the entries that
1506       // were pushed as a result of the G1CMTask::deal_with_reference() calls
1507       // above) or we overflow.
1508       //
1509       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1510       // flag while there may still be some work to do. (See the comment at
1511       // the beginning of G1CMTask::do_marking_step() for those conditions -
1512       // one of which is reaching the specified time target.) It is only
1513       // when G1CMTask::do_marking_step() returns without setting the
1514       // has_aborted() flag that the marking step has completed.
1515       do {
1516         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1517         _task->do_marking_step(mark_step_duration_ms,
1518                                false      /* do_termination */,
1519                                _is_serial);
1520       } while (_task->has_aborted() && !_cm->has_overflown());
1521       _ref_counter = _ref_counter_limit;
1522     }
1523   }
1524 };
1525 
1526 // 'Drain' oop closure used by both serial and parallel reference processing.
1527 // Uses the G1CMTask associated with a given worker thread (for serial
1528 // reference processing the G1CMtask for worker 0 is used). Calls the
1529 // do_marking_step routine, with an unbelievably large timeout value,
1530 // to drain the marking data structures of the remaining entries
1531 // added by the 'keep alive' oop closure above.
1532 
1533 class G1CMDrainMarkingStackClosure : public VoidClosure {
1534   G1ConcurrentMark* _cm;
1535   G1CMTask*         _task;
1536   bool              _is_serial;
1537  public:
1538   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1539     _cm(cm), _task(task), _is_serial(is_serial) {
1540     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1541   }
1542 
1543   void do_void() {
1544     do {
1545       // We call G1CMTask::do_marking_step() to completely drain the local
1546       // and global marking stacks of entries pushed by the 'keep alive'
1547       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1548       //
1549       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1550       // if there's nothing more to do (i.e. we've completely drained the
1551       // entries that were pushed as a a result of applying the 'keep alive'
1552       // closure to the entries on the discovered ref lists) or we overflow
1553       // the global marking stack.
1554       //
1555       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1556       // flag while there may still be some work to do. (See the comment at
1557       // the beginning of G1CMTask::do_marking_step() for those conditions -
1558       // one of which is reaching the specified time target.) It is only
1559       // when G1CMTask::do_marking_step() returns without setting the
1560       // has_aborted() flag that the marking step has completed.
1561 
1562       _task->do_marking_step(1000000000.0 /* something very large */,
1563                              true         /* do_termination */,
1564                              _is_serial);
1565     } while (_task->has_aborted() && !_cm->has_overflown());
1566   }
1567 };
1568 
1569 class G1CMRefProcProxyTask : public RefProcProxyTask {
1570   G1CollectedHeap& _g1h;
1571   G1ConcurrentMark& _cm;
1572 
1573 public:
1574   G1CMRefProcProxyTask(uint max_workers, G1CollectedHeap& g1h, G1ConcurrentMark &cm)
1575     : RefProcProxyTask("G1CMRefProcProxyTask", max_workers),
1576       _g1h(g1h),
1577       _cm(cm) {}
1578 
1579   void work(uint worker_id) override {
1580     assert(worker_id < _max_workers, "sanity");
1581     G1CMIsAliveClosure is_alive(&_g1h);
1582     uint index = (_tm == RefProcThreadModel::Single) ? 0 : worker_id;
1583     G1CMKeepAliveAndDrainClosure keep_alive(&_cm, _cm.task(index), _tm == RefProcThreadModel::Single);
1584     BarrierEnqueueDiscoveredFieldClosure enqueue;
1585     G1CMDrainMarkingStackClosure complete_gc(&_cm, _cm.task(index), _tm == RefProcThreadModel::Single);
1586     _rp_task->rp_work(worker_id, &is_alive, &keep_alive, &enqueue, &complete_gc);
1587   }
1588 
1589   void prepare_run_task_hook() override {
1590     // We need to reset the concurrency level before each
1591     // proxy task execution, so that the termination protocol
1592     // and overflow handling in G1CMTask::do_marking_step() knows
1593     // how many workers to wait for.
1594     _cm.set_concurrency(_queue_count);
1595   }
1596 };
1597 
1598 void G1ConcurrentMark::weak_refs_work() {
1599   ResourceMark rm;
1600 
1601   // Is alive closure.
1602   G1CMIsAliveClosure g1_is_alive(_g1h);
1603 
1604   {
1605     GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm);
1606 
1607     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1608 
1609     // See the comment in G1CollectedHeap::ref_processing_init()
1610     // about how reference processing currently works in G1.
1611 
1612     assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1613 
1614     // We need at least one active thread. If reference processing
1615     // is not multi-threaded we use the current (VMThread) thread,
1616     // otherwise we use the workers from the G1CollectedHeap and
1617     // we utilize all the worker threads we can.
1618     uint active_workers = (ParallelRefProcEnabled ? _g1h->workers()->active_workers() : 1U);
1619     active_workers = clamp(active_workers, 1u, _max_num_tasks);
1620 
1621     // Set the degree of MT processing here.  If the discovery was done MT,
1622     // the number of threads involved during discovery could differ from
1623     // the number of active workers.  This is OK as long as the discovered
1624     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1625     rp->set_active_mt_degree(active_workers);
1626 
1627     // Parallel processing task executor.
1628     G1CMRefProcProxyTask task(rp->max_num_queues(), *_g1h, *this);
1629     ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues());
1630 
1631     // Process the weak references.
1632     const ReferenceProcessorStats& stats = rp->process_discovered_references(task, pt);
1633     _gc_tracer_cm->report_gc_reference_stats(stats);
1634     pt.print_all_references();
1635 
1636     // The do_oop work routines of the keep_alive and drain_marking_stack
1637     // oop closures will set the has_overflown flag if we overflow the
1638     // global marking stack.
1639 
1640     assert(has_overflown() || _global_mark_stack.is_empty(),
1641            "Mark stack should be empty (unless it has overflown)");
1642 
1643     assert(rp->num_queues() == active_workers, "why not");
1644   }
1645 
1646   if (has_overflown()) {
1647     // We can not trust g1_is_alive and the contents of the heap if the marking stack
1648     // overflowed while processing references. Exit the VM.
1649     fatal("Overflow during reference processing, can not continue. Please "
1650           "increase MarkStackSizeMax (current value: " SIZE_FORMAT ") and "
1651           "restart.", MarkStackSizeMax);
1652     return;
1653   }
1654 
1655   assert(_global_mark_stack.is_empty(), "Marking should have completed");
1656 
1657   {
1658     GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1659     WeakProcessor::weak_oops_do(_g1h->workers(), &g1_is_alive, &do_nothing_cl, 1);
1660   }
1661 
1662   // Unload Klasses, String, Code Cache, etc.
1663   if (ClassUnloadingWithConcurrentMark) {
1664     GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1665     bool purged_classes = SystemDictionary::do_unloading(_gc_timer_cm);
1666     _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1667   }
1668 }
1669 
1670 class G1PrecleanYieldClosure : public YieldClosure {
1671   G1ConcurrentMark* _cm;
1672 
1673 public:
1674   G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
1675 
1676   virtual bool should_return() {
1677     return _cm->has_aborted();
1678   }
1679 
1680   virtual bool should_return_fine_grain() {
1681     _cm->do_yield_check();
1682     return _cm->has_aborted();
1683   }
1684 };
1685 
1686 void G1ConcurrentMark::preclean() {
1687   assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
1688 
1689   SuspendibleThreadSetJoiner joiner;
1690 
1691   BarrierEnqueueDiscoveredFieldClosure enqueue;
1692 
1693   set_concurrency_and_phase(1, true);
1694 
1695   G1PrecleanYieldClosure yield_cl(this);
1696 
1697   ReferenceProcessor* rp = _g1h->ref_processor_cm();
1698   // Precleaning is single threaded. Temporarily disable MT discovery.
1699   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
1700   rp->preclean_discovered_references(rp->is_alive_non_header(),
1701                                      &enqueue,
1702                                      &yield_cl,
1703                                      _gc_timer_cm);
1704 }
1705 
1706 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1707 // the prev bitmap determining liveness.
1708 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1709   G1CollectedHeap* _g1h;
1710 public:
1711   G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1712 
1713   bool do_object_b(oop obj) {
1714     return obj != NULL &&
1715            (!_g1h->is_in_reserved(obj) || !_g1h->is_obj_dead(obj));
1716   }
1717 };
1718 
1719 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1720   // Depending on the completion of the marking liveness needs to be determined
1721   // using either the next or prev bitmap.
1722   if (mark_completed) {
1723     G1ObjectCountIsAliveClosure is_alive(_g1h);
1724     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1725   } else {
1726     G1CMIsAliveClosure is_alive(_g1h);
1727     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1728   }
1729 }
1730 
1731 
1732 void G1ConcurrentMark::swap_mark_bitmaps() {
1733   G1CMBitMap* temp = _prev_mark_bitmap;
1734   _prev_mark_bitmap = _next_mark_bitmap;
1735   _next_mark_bitmap = temp;
1736 }
1737 
1738 // Closure for marking entries in SATB buffers.
1739 class G1CMSATBBufferClosure : public SATBBufferClosure {
1740 private:
1741   G1CMTask* _task;
1742   G1CollectedHeap* _g1h;
1743 
1744   // This is very similar to G1CMTask::deal_with_reference, but with
1745   // more relaxed requirements for the argument, so this must be more
1746   // circumspect about treating the argument as an object.
1747   void do_entry(void* entry) const {
1748     _task->increment_refs_reached();
1749     oop const obj = cast_to_oop(entry);
1750     _task->make_reference_grey(obj);
1751   }
1752 
1753 public:
1754   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1755     : _task(task), _g1h(g1h) { }
1756 
1757   virtual void do_buffer(void** buffer, size_t size) {
1758     for (size_t i = 0; i < size; ++i) {
1759       do_entry(buffer[i]);
1760     }
1761   }
1762 };
1763 
1764 class G1RemarkThreadsClosure : public ThreadClosure {
1765   G1SATBMarkQueueSet& _qset;
1766   G1CMOopClosure _cm_cl;
1767   MarkingCodeBlobClosure _code_cl;
1768   uintx _claim_token;
1769 
1770  public:
1771   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1772     _qset(G1BarrierSet::satb_mark_queue_set()),
1773     _cm_cl(g1h, task),
1774     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1775     _claim_token(Threads::thread_claim_token()) {}
1776 
1777   void do_thread(Thread* thread) {
1778     if (thread->claim_threads_do(true, _claim_token)) {
1779       // Transfer any partial buffer to the qset for completed buffer processing.
1780       _qset.flush_queue(G1ThreadLocalData::satb_mark_queue(thread));
1781       if (thread->is_Java_thread()) {
1782         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1783         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1784         // * Alive if on the stack of an executing method
1785         // * Weakly reachable otherwise
1786         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1787         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1788         JavaThread::cast(thread)->nmethods_do(&_code_cl);
1789       }
1790     }
1791   }
1792 };
1793 
1794 class G1CMRemarkTask : public WorkerTask {
1795   G1ConcurrentMark* _cm;
1796 public:
1797   void work(uint worker_id) {
1798     G1CMTask* task = _cm->task(worker_id);
1799     task->record_start_time();
1800     {
1801       ResourceMark rm;
1802 
1803       G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1804       Threads::threads_do(&threads_f);
1805     }
1806 
1807     do {
1808       task->do_marking_step(1000000000.0 /* something very large */,
1809                             true         /* do_termination       */,
1810                             false        /* is_serial            */);
1811     } while (task->has_aborted() && !_cm->has_overflown());
1812     // If we overflow, then we do not want to restart. We instead
1813     // want to abort remark and do concurrent marking again.
1814     task->record_end_time();
1815   }
1816 
1817   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1818     WorkerTask("Par Remark"), _cm(cm) {
1819     _cm->terminator()->reset_for_reuse(active_workers);
1820   }
1821 };
1822 
1823 void G1ConcurrentMark::finalize_marking() {
1824   ResourceMark rm;
1825 
1826   _g1h->ensure_parsability(false);
1827 
1828   // this is remark, so we'll use up all active threads
1829   uint active_workers = _g1h->workers()->active_workers();
1830   set_concurrency_and_phase(active_workers, false /* concurrent */);
1831   // Leave _parallel_marking_threads at it's
1832   // value originally calculated in the G1ConcurrentMark
1833   // constructor and pass values of the active workers
1834   // through the task.
1835 
1836   {
1837     StrongRootsScope srs(active_workers);
1838 
1839     G1CMRemarkTask remarkTask(this, active_workers);
1840     // We will start all available threads, even if we decide that the
1841     // active_workers will be fewer. The extra ones will just bail out
1842     // immediately.
1843     _g1h->workers()->run_task(&remarkTask);
1844   }
1845 
1846   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1847   guarantee(has_overflown() ||
1848             satb_mq_set.completed_buffers_num() == 0,
1849             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1850             BOOL_TO_STR(has_overflown()),
1851             satb_mq_set.completed_buffers_num());
1852 
1853   print_stats();
1854 }
1855 
1856 void G1ConcurrentMark::flush_all_task_caches() {
1857   size_t hits = 0;
1858   size_t misses = 0;
1859   for (uint i = 0; i < _max_num_tasks; i++) {
1860     Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1861     hits += stats.first;
1862     misses += stats.second;
1863   }
1864   size_t sum = hits + misses;
1865   log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1866                        hits, misses, percent_of(hits, sum));
1867 }
1868 
1869 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1870   _prev_mark_bitmap->clear_range(mr);
1871 }
1872 
1873 HeapRegion*
1874 G1ConcurrentMark::claim_region(uint worker_id) {
1875   // "checkpoint" the finger
1876   HeapWord* finger = _finger;
1877 
1878   while (finger < _heap.end()) {
1879     assert(_g1h->is_in_reserved(finger), "invariant");
1880 
1881     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1882     // Make sure that the reads below do not float before loading curr_region.
1883     OrderAccess::loadload();
1884     // Above heap_region_containing may return NULL as we always scan claim
1885     // until the end of the heap. In this case, just jump to the next region.
1886     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1887 
1888     // Is the gap between reading the finger and doing the CAS too long?
1889     HeapWord* res = Atomic::cmpxchg(&_finger, finger, end);
1890     if (res == finger && curr_region != NULL) {
1891       // we succeeded
1892       HeapWord*   bottom        = curr_region->bottom();
1893       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1894 
1895       // notice that _finger == end cannot be guaranteed here since,
1896       // someone else might have moved the finger even further
1897       assert(_finger >= end, "the finger should have moved forward");
1898 
1899       if (limit > bottom) {
1900         return curr_region;
1901       } else {
1902         assert(limit == bottom,
1903                "the region limit should be at bottom");
1904         // we return NULL and the caller should try calling
1905         // claim_region() again.
1906         return NULL;
1907       }
1908     } else {
1909       assert(_finger > finger, "the finger should have moved forward");
1910       // read it again
1911       finger = _finger;
1912     }
1913   }
1914 
1915   return NULL;
1916 }
1917 
1918 #ifndef PRODUCT
1919 class VerifyNoCSetOops {
1920   G1CollectedHeap* _g1h;
1921   const char* _phase;
1922   int _info;
1923 
1924 public:
1925   VerifyNoCSetOops(const char* phase, int info = -1) :
1926     _g1h(G1CollectedHeap::heap()),
1927     _phase(phase),
1928     _info(info)
1929   { }
1930 
1931   void operator()(G1TaskQueueEntry task_entry) const {
1932     if (task_entry.is_array_slice()) {
1933       guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1934       return;
1935     }
1936     guarantee(oopDesc::is_oop(task_entry.obj()),
1937               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1938               p2i(task_entry.obj()), _phase, _info);
1939     HeapRegion* r = _g1h->heap_region_containing(task_entry.obj());
1940     guarantee(!(r->in_collection_set() || r->has_index_in_opt_cset()),
1941               "obj " PTR_FORMAT " from %s (%d) in region %u in (optional) collection set",
1942               p2i(task_entry.obj()), _phase, _info, r->hrm_index());
1943   }
1944 };
1945 
1946 void G1ConcurrentMark::verify_no_collection_set_oops() {
1947   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1948   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1949     return;
1950   }
1951 
1952   // Verify entries on the global mark stack
1953   _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1954 
1955   // Verify entries on the task queues
1956   for (uint i = 0; i < _max_num_tasks; ++i) {
1957     G1CMTaskQueue* queue = _task_queues->queue(i);
1958     queue->iterate(VerifyNoCSetOops("Queue", i));
1959   }
1960 
1961   // Verify the global finger
1962   HeapWord* global_finger = finger();
1963   if (global_finger != NULL && global_finger < _heap.end()) {
1964     // Since we always iterate over all regions, we might get a NULL HeapRegion
1965     // here.
1966     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1967     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1968               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1969               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1970   }
1971 
1972   // Verify the task fingers
1973   assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
1974   for (uint i = 0; i < _num_concurrent_workers; ++i) {
1975     G1CMTask* task = _tasks[i];
1976     HeapWord* task_finger = task->finger();
1977     if (task_finger != NULL && task_finger < _heap.end()) {
1978       // See above note on the global finger verification.
1979       HeapRegion* r = _g1h->heap_region_containing(task_finger);
1980       guarantee(r == NULL || task_finger == r->bottom() ||
1981                 !r->in_collection_set() || !r->has_index_in_opt_cset(),
1982                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
1983                 p2i(task_finger), HR_FORMAT_PARAMS(r));
1984     }
1985   }
1986 }
1987 #endif // PRODUCT
1988 
1989 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
1990   // If Remark did not select any regions for RemSet rebuild,
1991   // skip the rebuild remembered set phase
1992   if (!needs_remembered_set_rebuild()) {
1993     log_debug(gc, marking)("Skipping Remembered Set Rebuild. No regions selected for rebuild");
1994     return;
1995   }
1996   _g1h->rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
1997 }
1998 
1999 void G1ConcurrentMark::print_stats() {
2000   if (!log_is_enabled(Debug, gc, stats)) {
2001     return;
2002   }
2003   log_debug(gc, stats)("---------------------------------------------------------------------");
2004   for (size_t i = 0; i < _num_active_tasks; ++i) {
2005     _tasks[i]->print_stats();
2006     log_debug(gc, stats)("---------------------------------------------------------------------");
2007   }
2008 }
2009 
2010 void G1ConcurrentMark::concurrent_cycle_abort() {
2011   if (!cm_thread()->in_progress() || _has_aborted) {
2012     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2013     return;
2014   }
2015 
2016   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2017   // concurrent bitmap clearing.
2018   {
2019     GCTraceTime(Debug, gc) debug("Clear Next Bitmap");
2020     clear_next_bitmap(_g1h->workers());
2021   }
2022   // Note we cannot clear the previous marking bitmap here
2023   // since VerifyDuringGC verifies the objects marked during
2024   // a full GC against the previous bitmap.
2025 
2026   // Empty mark stack
2027   reset_marking_for_restart();
2028   for (uint i = 0; i < _max_num_tasks; ++i) {
2029     _tasks[i]->clear_region_fields();
2030   }
2031   _first_overflow_barrier_sync.abort();
2032   _second_overflow_barrier_sync.abort();
2033   _has_aborted = true;
2034 
2035   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2036   satb_mq_set.abandon_partial_marking();
2037   // This can be called either during or outside marking, we'll read
2038   // the expected_active value from the SATB queue set.
2039   satb_mq_set.set_active_all_threads(
2040                                  false, /* new active value */
2041                                  satb_mq_set.is_active() /* expected_active */);
2042 }
2043 
2044 static void print_ms_time_info(const char* prefix, const char* name,
2045                                NumberSeq& ns) {
2046   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2047                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2048   if (ns.num() > 0) {
2049     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2050                            prefix, ns.sd(), ns.maximum());
2051   }
2052 }
2053 
2054 void G1ConcurrentMark::print_summary_info() {
2055   Log(gc, marking) log;
2056   if (!log.is_trace()) {
2057     return;
2058   }
2059 
2060   log.trace(" Concurrent marking:");
2061   print_ms_time_info("  ", "init marks", _init_times);
2062   print_ms_time_info("  ", "remarks", _remark_times);
2063   {
2064     print_ms_time_info("     ", "final marks", _remark_mark_times);
2065     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2066 
2067   }
2068   print_ms_time_info("  ", "cleanups", _cleanup_times);
2069   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2070             _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2071   log.trace("  Total stop_world time = %8.2f s.",
2072             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2073   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2074             cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2075 }
2076 
2077 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2078   _concurrent_workers->threads_do(tc);
2079 }
2080 
2081 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2082   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2083                p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2084   _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2085   _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2086 }
2087 
2088 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2089   ReferenceProcessor* result = g1h->ref_processor_cm();
2090   assert(result != NULL, "CM reference processor should not be NULL");
2091   return result;
2092 }
2093 
2094 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2095                                G1CMTask* task)
2096   : MetadataVisitingOopIterateClosure(get_cm_oop_closure_ref_processor(g1h)),
2097     _g1h(g1h), _task(task)
2098 { }
2099 
2100 void G1CMTask::setup_for_region(HeapRegion* hr) {
2101   assert(hr != NULL,
2102         "claim_region() should have filtered out NULL regions");
2103   _curr_region  = hr;
2104   _finger       = hr->bottom();
2105   update_region_limit();
2106 }
2107 
2108 void G1CMTask::update_region_limit() {
2109   HeapRegion* hr            = _curr_region;
2110   HeapWord* bottom          = hr->bottom();
2111   HeapWord* limit           = hr->next_top_at_mark_start();
2112 
2113   if (limit == bottom) {
2114     // The region was collected underneath our feet.
2115     // We set the finger to bottom to ensure that the bitmap
2116     // iteration that will follow this will not do anything.
2117     // (this is not a condition that holds when we set the region up,
2118     // as the region is not supposed to be empty in the first place)
2119     _finger = bottom;
2120   } else if (limit >= _region_limit) {
2121     assert(limit >= _finger, "peace of mind");
2122   } else {
2123     assert(limit < _region_limit, "only way to get here");
2124     // This can happen under some pretty unusual circumstances.  An
2125     // evacuation pause empties the region underneath our feet (NTAMS
2126     // at bottom). We then do some allocation in the region (NTAMS
2127     // stays at bottom), followed by the region being used as a GC
2128     // alloc region (NTAMS will move to top() and the objects
2129     // originally below it will be grayed). All objects now marked in
2130     // the region are explicitly grayed, if below the global finger,
2131     // and we do not need in fact to scan anything else. So, we simply
2132     // set _finger to be limit to ensure that the bitmap iteration
2133     // doesn't do anything.
2134     _finger = limit;
2135   }
2136 
2137   _region_limit = limit;
2138 }
2139 
2140 void G1CMTask::giveup_current_region() {
2141   assert(_curr_region != NULL, "invariant");
2142   clear_region_fields();
2143 }
2144 
2145 void G1CMTask::clear_region_fields() {
2146   // Values for these three fields that indicate that we're not
2147   // holding on to a region.
2148   _curr_region   = NULL;
2149   _finger        = NULL;
2150   _region_limit  = NULL;
2151 }
2152 
2153 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2154   if (cm_oop_closure == NULL) {
2155     assert(_cm_oop_closure != NULL, "invariant");
2156   } else {
2157     assert(_cm_oop_closure == NULL, "invariant");
2158   }
2159   _cm_oop_closure = cm_oop_closure;
2160 }
2161 
2162 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2163   guarantee(next_mark_bitmap != NULL, "invariant");
2164   _next_mark_bitmap              = next_mark_bitmap;
2165   clear_region_fields();
2166 
2167   _calls                         = 0;
2168   _elapsed_time_ms               = 0.0;
2169   _termination_time_ms           = 0.0;
2170   _termination_start_time_ms     = 0.0;
2171 
2172   _mark_stats_cache.reset();
2173 }
2174 
2175 bool G1CMTask::should_exit_termination() {
2176   if (!regular_clock_call()) {
2177     return true;
2178   }
2179 
2180   // This is called when we are in the termination protocol. We should
2181   // quit if, for some reason, this task wants to abort or the global
2182   // stack is not empty (this means that we can get work from it).
2183   return !_cm->mark_stack_empty() || has_aborted();
2184 }
2185 
2186 void G1CMTask::reached_limit() {
2187   assert(_words_scanned >= _words_scanned_limit ||
2188          _refs_reached >= _refs_reached_limit ,
2189          "shouldn't have been called otherwise");
2190   abort_marking_if_regular_check_fail();
2191 }
2192 
2193 bool G1CMTask::regular_clock_call() {
2194   if (has_aborted()) {
2195     return false;
2196   }
2197 
2198   // First, we need to recalculate the words scanned and refs reached
2199   // limits for the next clock call.
2200   recalculate_limits();
2201 
2202   // During the regular clock call we do the following
2203 
2204   // (1) If an overflow has been flagged, then we abort.
2205   if (_cm->has_overflown()) {
2206     return false;
2207   }
2208 
2209   // If we are not concurrent (i.e. we're doing remark) we don't need
2210   // to check anything else. The other steps are only needed during
2211   // the concurrent marking phase.
2212   if (!_cm->concurrent()) {
2213     return true;
2214   }
2215 
2216   // (2) If marking has been aborted for Full GC, then we also abort.
2217   if (_cm->has_aborted()) {
2218     return false;
2219   }
2220 
2221   double curr_time_ms = os::elapsedVTime() * 1000.0;
2222 
2223   // (4) We check whether we should yield. If we have to, then we abort.
2224   if (SuspendibleThreadSet::should_yield()) {
2225     // We should yield. To do this we abort the task. The caller is
2226     // responsible for yielding.
2227     return false;
2228   }
2229 
2230   // (5) We check whether we've reached our time quota. If we have,
2231   // then we abort.
2232   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2233   if (elapsed_time_ms > _time_target_ms) {
2234     _has_timed_out = true;
2235     return false;
2236   }
2237 
2238   // (6) Finally, we check whether there are enough completed STAB
2239   // buffers available for processing. If there are, we abort.
2240   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2241   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2242     // we do need to process SATB buffers, we'll abort and restart
2243     // the marking task to do so
2244     return false;
2245   }
2246   return true;
2247 }
2248 
2249 void G1CMTask::recalculate_limits() {
2250   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2251   _words_scanned_limit      = _real_words_scanned_limit;
2252 
2253   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2254   _refs_reached_limit       = _real_refs_reached_limit;
2255 }
2256 
2257 void G1CMTask::decrease_limits() {
2258   // This is called when we believe that we're going to do an infrequent
2259   // operation which will increase the per byte scanned cost (i.e. move
2260   // entries to/from the global stack). It basically tries to decrease the
2261   // scanning limit so that the clock is called earlier.
2262 
2263   _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2264   _refs_reached_limit  = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2265 }
2266 
2267 void G1CMTask::move_entries_to_global_stack() {
2268   // Local array where we'll store the entries that will be popped
2269   // from the local queue.
2270   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2271 
2272   size_t n = 0;
2273   G1TaskQueueEntry task_entry;
2274   while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2275     buffer[n] = task_entry;
2276     ++n;
2277   }
2278   if (n < G1CMMarkStack::EntriesPerChunk) {
2279     buffer[n] = G1TaskQueueEntry();
2280   }
2281 
2282   if (n > 0) {
2283     if (!_cm->mark_stack_push(buffer)) {
2284       set_has_aborted();
2285     }
2286   }
2287 
2288   // This operation was quite expensive, so decrease the limits.
2289   decrease_limits();
2290 }
2291 
2292 bool G1CMTask::get_entries_from_global_stack() {
2293   // Local array where we'll store the entries that will be popped
2294   // from the global stack.
2295   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2296 
2297   if (!_cm->mark_stack_pop(buffer)) {
2298     return false;
2299   }
2300 
2301   // We did actually pop at least one entry.
2302   for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2303     G1TaskQueueEntry task_entry = buffer[i];
2304     if (task_entry.is_null()) {
2305       break;
2306     }
2307     assert(task_entry.is_array_slice() || oopDesc::is_oop(task_entry.obj()), "Element " PTR_FORMAT " must be an array slice or oop", p2i(task_entry.obj()));
2308     bool success = _task_queue->push(task_entry);
2309     // We only call this when the local queue is empty or under a
2310     // given target limit. So, we do not expect this push to fail.
2311     assert(success, "invariant");
2312   }
2313 
2314   // This operation was quite expensive, so decrease the limits
2315   decrease_limits();
2316   return true;
2317 }
2318 
2319 void G1CMTask::drain_local_queue(bool partially) {
2320   if (has_aborted()) {
2321     return;
2322   }
2323 
2324   // Decide what the target size is, depending whether we're going to
2325   // drain it partially (so that other tasks can steal if they run out
2326   // of things to do) or totally (at the very end).
2327   size_t target_size;
2328   if (partially) {
2329     target_size = MIN2((size_t)_task_queue->max_elems()/3, (size_t)GCDrainStackTargetSize);
2330   } else {
2331     target_size = 0;
2332   }
2333 
2334   if (_task_queue->size() > target_size) {
2335     G1TaskQueueEntry entry;
2336     bool ret = _task_queue->pop_local(entry);
2337     while (ret) {
2338       scan_task_entry(entry);
2339       if (_task_queue->size() <= target_size || has_aborted()) {
2340         ret = false;
2341       } else {
2342         ret = _task_queue->pop_local(entry);
2343       }
2344     }
2345   }
2346 }
2347 
2348 void G1CMTask::drain_global_stack(bool partially) {
2349   if (has_aborted()) {
2350     return;
2351   }
2352 
2353   // We have a policy to drain the local queue before we attempt to
2354   // drain the global stack.
2355   assert(partially || _task_queue->size() == 0, "invariant");
2356 
2357   // Decide what the target size is, depending whether we're going to
2358   // drain it partially (so that other tasks can steal if they run out
2359   // of things to do) or totally (at the very end).
2360   // Notice that when draining the global mark stack partially, due to the racyness
2361   // of the mark stack size update we might in fact drop below the target. But,
2362   // this is not a problem.
2363   // In case of total draining, we simply process until the global mark stack is
2364   // totally empty, disregarding the size counter.
2365   if (partially) {
2366     size_t const target_size = _cm->partial_mark_stack_size_target();
2367     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2368       if (get_entries_from_global_stack()) {
2369         drain_local_queue(partially);
2370       }
2371     }
2372   } else {
2373     while (!has_aborted() && get_entries_from_global_stack()) {
2374       drain_local_queue(partially);
2375     }
2376   }
2377 }
2378 
2379 // SATB Queue has several assumptions on whether to call the par or
2380 // non-par versions of the methods. this is why some of the code is
2381 // replicated. We should really get rid of the single-threaded version
2382 // of the code to simplify things.
2383 void G1CMTask::drain_satb_buffers() {
2384   if (has_aborted()) {
2385     return;
2386   }
2387 
2388   // We set this so that the regular clock knows that we're in the
2389   // middle of draining buffers and doesn't set the abort flag when it
2390   // notices that SATB buffers are available for draining. It'd be
2391   // very counter productive if it did that. :-)
2392   _draining_satb_buffers = true;
2393 
2394   G1CMSATBBufferClosure satb_cl(this, _g1h);
2395   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2396 
2397   // This keeps claiming and applying the closure to completed buffers
2398   // until we run out of buffers or we need to abort.
2399   while (!has_aborted() &&
2400          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2401     abort_marking_if_regular_check_fail();
2402   }
2403 
2404   // Can't assert qset is empty here, even if not aborted.  If concurrent,
2405   // some other thread might be adding to the queue.  If not concurrent,
2406   // some other thread might have won the race for the last buffer, but
2407   // has not yet decremented the count.
2408 
2409   _draining_satb_buffers = false;
2410 
2411   // again, this was a potentially expensive operation, decrease the
2412   // limits to get the regular clock call early
2413   decrease_limits();
2414 }
2415 
2416 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2417   _mark_stats_cache.reset(region_idx);
2418 }
2419 
2420 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2421   return _mark_stats_cache.evict_all();
2422 }
2423 
2424 void G1CMTask::print_stats() {
2425   log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2426   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2427                        _elapsed_time_ms, _termination_time_ms);
2428   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2429                        _step_times_ms.num(),
2430                        _step_times_ms.avg(),
2431                        _step_times_ms.sd(),
2432                        _step_times_ms.maximum(),
2433                        _step_times_ms.sum());
2434   size_t const hits = _mark_stats_cache.hits();
2435   size_t const misses = _mark_stats_cache.misses();
2436   log_debug(gc, stats)("  Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2437                        hits, misses, percent_of(hits, hits + misses));
2438 }
2439 
2440 bool G1ConcurrentMark::try_stealing(uint worker_id, G1TaskQueueEntry& task_entry) {
2441   return _task_queues->steal(worker_id, task_entry);
2442 }
2443 
2444 /*****************************************************************************
2445 
2446     The do_marking_step(time_target_ms, ...) method is the building
2447     block of the parallel marking framework. It can be called in parallel
2448     with other invocations of do_marking_step() on different tasks
2449     (but only one per task, obviously) and concurrently with the
2450     mutator threads, or during remark, hence it eliminates the need
2451     for two versions of the code. When called during remark, it will
2452     pick up from where the task left off during the concurrent marking
2453     phase. Interestingly, tasks are also claimable during evacuation
2454     pauses too, since do_marking_step() ensures that it aborts before
2455     it needs to yield.
2456 
2457     The data structures that it uses to do marking work are the
2458     following:
2459 
2460       (1) Marking Bitmap. If there are gray objects that appear only
2461       on the bitmap (this happens either when dealing with an overflow
2462       or when the concurrent start pause has simply marked the roots
2463       and didn't push them on the stack), then tasks claim heap
2464       regions whose bitmap they then scan to find gray objects. A
2465       global finger indicates where the end of the last claimed region
2466       is. A local finger indicates how far into the region a task has
2467       scanned. The two fingers are used to determine how to gray an
2468       object (i.e. whether simply marking it is OK, as it will be
2469       visited by a task in the future, or whether it needs to be also
2470       pushed on a stack).
2471 
2472       (2) Local Queue. The local queue of the task which is accessed
2473       reasonably efficiently by the task. Other tasks can steal from
2474       it when they run out of work. Throughout the marking phase, a
2475       task attempts to keep its local queue short but not totally
2476       empty, so that entries are available for stealing by other
2477       tasks. Only when there is no more work, a task will totally
2478       drain its local queue.
2479 
2480       (3) Global Mark Stack. This handles local queue overflow. During
2481       marking only sets of entries are moved between it and the local
2482       queues, as access to it requires a mutex and more fine-grain
2483       interaction with it which might cause contention. If it
2484       overflows, then the marking phase should restart and iterate
2485       over the bitmap to identify gray objects. Throughout the marking
2486       phase, tasks attempt to keep the global mark stack at a small
2487       length but not totally empty, so that entries are available for
2488       popping by other tasks. Only when there is no more work, tasks
2489       will totally drain the global mark stack.
2490 
2491       (4) SATB Buffer Queue. This is where completed SATB buffers are
2492       made available. Buffers are regularly removed from this queue
2493       and scanned for roots, so that the queue doesn't get too
2494       long. During remark, all completed buffers are processed, as
2495       well as the filled in parts of any uncompleted buffers.
2496 
2497     The do_marking_step() method tries to abort when the time target
2498     has been reached. There are a few other cases when the
2499     do_marking_step() method also aborts:
2500 
2501       (1) When the marking phase has been aborted (after a Full GC).
2502 
2503       (2) When a global overflow (on the global stack) has been
2504       triggered. Before the task aborts, it will actually sync up with
2505       the other tasks to ensure that all the marking data structures
2506       (local queues, stacks, fingers etc.)  are re-initialized so that
2507       when do_marking_step() completes, the marking phase can
2508       immediately restart.
2509 
2510       (3) When enough completed SATB buffers are available. The
2511       do_marking_step() method only tries to drain SATB buffers right
2512       at the beginning. So, if enough buffers are available, the
2513       marking step aborts and the SATB buffers are processed at
2514       the beginning of the next invocation.
2515 
2516       (4) To yield. when we have to yield then we abort and yield
2517       right at the end of do_marking_step(). This saves us from a lot
2518       of hassle as, by yielding we might allow a Full GC. If this
2519       happens then objects will be compacted underneath our feet, the
2520       heap might shrink, etc. We save checking for this by just
2521       aborting and doing the yield right at the end.
2522 
2523     From the above it follows that the do_marking_step() method should
2524     be called in a loop (or, otherwise, regularly) until it completes.
2525 
2526     If a marking step completes without its has_aborted() flag being
2527     true, it means it has completed the current marking phase (and
2528     also all other marking tasks have done so and have all synced up).
2529 
2530     A method called regular_clock_call() is invoked "regularly" (in
2531     sub ms intervals) throughout marking. It is this clock method that
2532     checks all the abort conditions which were mentioned above and
2533     decides when the task should abort. A work-based scheme is used to
2534     trigger this clock method: when the number of object words the
2535     marking phase has scanned or the number of references the marking
2536     phase has visited reach a given limit. Additional invocations to
2537     the method clock have been planted in a few other strategic places
2538     too. The initial reason for the clock method was to avoid calling
2539     vtime too regularly, as it is quite expensive. So, once it was in
2540     place, it was natural to piggy-back all the other conditions on it
2541     too and not constantly check them throughout the code.
2542 
2543     If do_termination is true then do_marking_step will enter its
2544     termination protocol.
2545 
2546     The value of is_serial must be true when do_marking_step is being
2547     called serially (i.e. by the VMThread) and do_marking_step should
2548     skip any synchronization in the termination and overflow code.
2549     Examples include the serial remark code and the serial reference
2550     processing closures.
2551 
2552     The value of is_serial must be false when do_marking_step is
2553     being called by any of the worker threads.
2554     Examples include the concurrent marking code (CMMarkingTask),
2555     the MT remark code, and the MT reference processing closures.
2556 
2557  *****************************************************************************/
2558 
2559 void G1CMTask::do_marking_step(double time_target_ms,
2560                                bool do_termination,
2561                                bool is_serial) {
2562   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2563 
2564   _start_time_ms = os::elapsedVTime() * 1000.0;
2565 
2566   // If do_stealing is true then do_marking_step will attempt to
2567   // steal work from the other G1CMTasks. It only makes sense to
2568   // enable stealing when the termination protocol is enabled
2569   // and do_marking_step() is not being called serially.
2570   bool do_stealing = do_termination && !is_serial;
2571 
2572   G1Predictions const& predictor = _g1h->policy()->predictor();
2573   double diff_prediction_ms = predictor.predict_zero_bounded(&_marking_step_diff_ms);
2574   _time_target_ms = time_target_ms - diff_prediction_ms;
2575 
2576   // set up the variables that are used in the work-based scheme to
2577   // call the regular clock method
2578   _words_scanned = 0;
2579   _refs_reached  = 0;
2580   recalculate_limits();
2581 
2582   // clear all flags
2583   clear_has_aborted();
2584   _has_timed_out = false;
2585   _draining_satb_buffers = false;
2586 
2587   ++_calls;
2588 
2589   // Set up the bitmap and oop closures. Anything that uses them is
2590   // eventually called from this method, so it is OK to allocate these
2591   // statically.
2592   G1CMBitMapClosure bitmap_closure(this, _cm);
2593   G1CMOopClosure cm_oop_closure(_g1h, this);
2594   set_cm_oop_closure(&cm_oop_closure);
2595 
2596   if (_cm->has_overflown()) {
2597     // This can happen if the mark stack overflows during a GC pause
2598     // and this task, after a yield point, restarts. We have to abort
2599     // as we need to get into the overflow protocol which happens
2600     // right at the end of this task.
2601     set_has_aborted();
2602   }
2603 
2604   // First drain any available SATB buffers. After this, we will not
2605   // look at SATB buffers before the next invocation of this method.
2606   // If enough completed SATB buffers are queued up, the regular clock
2607   // will abort this task so that it restarts.
2608   drain_satb_buffers();
2609   // ...then partially drain the local queue and the global stack
2610   drain_local_queue(true);
2611   drain_global_stack(true);
2612 
2613   do {
2614     if (!has_aborted() && _curr_region != NULL) {
2615       // This means that we're already holding on to a region.
2616       assert(_finger != NULL, "if region is not NULL, then the finger "
2617              "should not be NULL either");
2618 
2619       // We might have restarted this task after an evacuation pause
2620       // which might have evacuated the region we're holding on to
2621       // underneath our feet. Let's read its limit again to make sure
2622       // that we do not iterate over a region of the heap that
2623       // contains garbage (update_region_limit() will also move
2624       // _finger to the start of the region if it is found empty).
2625       update_region_limit();
2626       // We will start from _finger not from the start of the region,
2627       // as we might be restarting this task after aborting half-way
2628       // through scanning this region. In this case, _finger points to
2629       // the address where we last found a marked object. If this is a
2630       // fresh region, _finger points to start().
2631       MemRegion mr = MemRegion(_finger, _region_limit);
2632 
2633       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2634              "humongous regions should go around loop once only");
2635 
2636       // Some special cases:
2637       // If the memory region is empty, we can just give up the region.
2638       // If the current region is humongous then we only need to check
2639       // the bitmap for the bit associated with the start of the object,
2640       // scan the object if it's live, and give up the region.
2641       // Otherwise, let's iterate over the bitmap of the part of the region
2642       // that is left.
2643       // If the iteration is successful, give up the region.
2644       if (mr.is_empty()) {
2645         giveup_current_region();
2646         abort_marking_if_regular_check_fail();
2647       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2648         if (_next_mark_bitmap->is_marked(mr.start())) {
2649           // The object is marked - apply the closure
2650           bitmap_closure.do_addr(mr.start());
2651         }
2652         // Even if this task aborted while scanning the humongous object
2653         // we can (and should) give up the current region.
2654         giveup_current_region();
2655         abort_marking_if_regular_check_fail();
2656       } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2657         giveup_current_region();
2658         abort_marking_if_regular_check_fail();
2659       } else {
2660         assert(has_aborted(), "currently the only way to do so");
2661         // The only way to abort the bitmap iteration is to return
2662         // false from the do_bit() method. However, inside the
2663         // do_bit() method we move the _finger to point to the
2664         // object currently being looked at. So, if we bail out, we
2665         // have definitely set _finger to something non-null.
2666         assert(_finger != NULL, "invariant");
2667 
2668         // Region iteration was actually aborted. So now _finger
2669         // points to the address of the object we last scanned. If we
2670         // leave it there, when we restart this task, we will rescan
2671         // the object. It is easy to avoid this. We move the finger by
2672         // enough to point to the next possible object header.
2673         assert(_finger < _region_limit, "invariant");
2674         HeapWord* const new_finger = _finger + cast_to_oop(_finger)->size();
2675         // Check if bitmap iteration was aborted while scanning the last object
2676         if (new_finger >= _region_limit) {
2677           giveup_current_region();
2678         } else {
2679           move_finger_to(new_finger);
2680         }
2681       }
2682     }
2683     // At this point we have either completed iterating over the
2684     // region we were holding on to, or we have aborted.
2685 
2686     // We then partially drain the local queue and the global stack.
2687     // (Do we really need this?)
2688     drain_local_queue(true);
2689     drain_global_stack(true);
2690 
2691     // Read the note on the claim_region() method on why it might
2692     // return NULL with potentially more regions available for
2693     // claiming and why we have to check out_of_regions() to determine
2694     // whether we're done or not.
2695     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2696       // We are going to try to claim a new region. We should have
2697       // given up on the previous one.
2698       // Separated the asserts so that we know which one fires.
2699       assert(_curr_region  == NULL, "invariant");
2700       assert(_finger       == NULL, "invariant");
2701       assert(_region_limit == NULL, "invariant");
2702       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2703       if (claimed_region != NULL) {
2704         // Yes, we managed to claim one
2705         setup_for_region(claimed_region);
2706         assert(_curr_region == claimed_region, "invariant");
2707       }
2708       // It is important to call the regular clock here. It might take
2709       // a while to claim a region if, for example, we hit a large
2710       // block of empty regions. So we need to call the regular clock
2711       // method once round the loop to make sure it's called
2712       // frequently enough.
2713       abort_marking_if_regular_check_fail();
2714     }
2715 
2716     if (!has_aborted() && _curr_region == NULL) {
2717       assert(_cm->out_of_regions(),
2718              "at this point we should be out of regions");
2719     }
2720   } while ( _curr_region != NULL && !has_aborted());
2721 
2722   if (!has_aborted()) {
2723     // We cannot check whether the global stack is empty, since other
2724     // tasks might be pushing objects to it concurrently.
2725     assert(_cm->out_of_regions(),
2726            "at this point we should be out of regions");
2727     // Try to reduce the number of available SATB buffers so that
2728     // remark has less work to do.
2729     drain_satb_buffers();
2730   }
2731 
2732   // Since we've done everything else, we can now totally drain the
2733   // local queue and global stack.
2734   drain_local_queue(false);
2735   drain_global_stack(false);
2736 
2737   // Attempt at work stealing from other task's queues.
2738   if (do_stealing && !has_aborted()) {
2739     // We have not aborted. This means that we have finished all that
2740     // we could. Let's try to do some stealing...
2741 
2742     // We cannot check whether the global stack is empty, since other
2743     // tasks might be pushing objects to it concurrently.
2744     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2745            "only way to reach here");
2746     while (!has_aborted()) {
2747       G1TaskQueueEntry entry;
2748       if (_cm->try_stealing(_worker_id, entry)) {
2749         scan_task_entry(entry);
2750 
2751         // And since we're towards the end, let's totally drain the
2752         // local queue and global stack.
2753         drain_local_queue(false);
2754         drain_global_stack(false);
2755       } else {
2756         break;
2757       }
2758     }
2759   }
2760 
2761   // We still haven't aborted. Now, let's try to get into the
2762   // termination protocol.
2763   if (do_termination && !has_aborted()) {
2764     // We cannot check whether the global stack is empty, since other
2765     // tasks might be concurrently pushing objects on it.
2766     // Separated the asserts so that we know which one fires.
2767     assert(_cm->out_of_regions(), "only way to reach here");
2768     assert(_task_queue->size() == 0, "only way to reach here");
2769     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2770 
2771     // The G1CMTask class also extends the TerminatorTerminator class,
2772     // hence its should_exit_termination() method will also decide
2773     // whether to exit the termination protocol or not.
2774     bool finished = (is_serial ||
2775                      _cm->terminator()->offer_termination(this));
2776     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2777     _termination_time_ms +=
2778       termination_end_time_ms - _termination_start_time_ms;
2779 
2780     if (finished) {
2781       // We're all done.
2782 
2783       // We can now guarantee that the global stack is empty, since
2784       // all other tasks have finished. We separated the guarantees so
2785       // that, if a condition is false, we can immediately find out
2786       // which one.
2787       guarantee(_cm->out_of_regions(), "only way to reach here");
2788       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2789       guarantee(_task_queue->size() == 0, "only way to reach here");
2790       guarantee(!_cm->has_overflown(), "only way to reach here");
2791       guarantee(!has_aborted(), "should never happen if termination has completed");
2792     } else {
2793       // Apparently there's more work to do. Let's abort this task. It
2794       // will restart it and we can hopefully find more things to do.
2795       set_has_aborted();
2796     }
2797   }
2798 
2799   // Mainly for debugging purposes to make sure that a pointer to the
2800   // closure which was statically allocated in this frame doesn't
2801   // escape it by accident.
2802   set_cm_oop_closure(NULL);
2803   double end_time_ms = os::elapsedVTime() * 1000.0;
2804   double elapsed_time_ms = end_time_ms - _start_time_ms;
2805   // Update the step history.
2806   _step_times_ms.add(elapsed_time_ms);
2807 
2808   if (has_aborted()) {
2809     // The task was aborted for some reason.
2810     if (_has_timed_out) {
2811       double diff_ms = elapsed_time_ms - _time_target_ms;
2812       // Keep statistics of how well we did with respect to hitting
2813       // our target only if we actually timed out (if we aborted for
2814       // other reasons, then the results might get skewed).
2815       _marking_step_diff_ms.add(diff_ms);
2816     }
2817 
2818     if (_cm->has_overflown()) {
2819       // This is the interesting one. We aborted because a global
2820       // overflow was raised. This means we have to restart the
2821       // marking phase and start iterating over regions. However, in
2822       // order to do this we have to make sure that all tasks stop
2823       // what they are doing and re-initialize in a safe manner. We
2824       // will achieve this with the use of two barrier sync points.
2825 
2826       if (!is_serial) {
2827         // We only need to enter the sync barrier if being called
2828         // from a parallel context
2829         _cm->enter_first_sync_barrier(_worker_id);
2830 
2831         // When we exit this sync barrier we know that all tasks have
2832         // stopped doing marking work. So, it's now safe to
2833         // re-initialize our data structures.
2834       }
2835 
2836       clear_region_fields();
2837       flush_mark_stats_cache();
2838 
2839       if (!is_serial) {
2840         // If we're executing the concurrent phase of marking, reset the marking
2841         // state; otherwise the marking state is reset after reference processing,
2842         // during the remark pause.
2843         // If we reset here as a result of an overflow during the remark we will
2844         // see assertion failures from any subsequent set_concurrency_and_phase()
2845         // calls.
2846         if (_cm->concurrent() && _worker_id == 0) {
2847           // Worker 0 is responsible for clearing the global data structures because
2848           // of an overflow. During STW we should not clear the overflow flag (in
2849           // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2850           // method to abort the pause and restart concurrent marking.
2851           _cm->reset_marking_for_restart();
2852 
2853           log_info(gc, marking)("Concurrent Mark reset for overflow");
2854         }
2855 
2856         // ...and enter the second barrier.
2857         _cm->enter_second_sync_barrier(_worker_id);
2858       }
2859       // At this point, if we're during the concurrent phase of
2860       // marking, everything has been re-initialized and we're
2861       // ready to restart.
2862     }
2863   }
2864 }
2865 
2866 G1CMTask::G1CMTask(uint worker_id,
2867                    G1ConcurrentMark* cm,
2868                    G1CMTaskQueue* task_queue,
2869                    G1RegionMarkStats* mark_stats) :
2870   _objArray_processor(this),
2871   _worker_id(worker_id),
2872   _g1h(G1CollectedHeap::heap()),
2873   _cm(cm),
2874   _next_mark_bitmap(NULL),
2875   _task_queue(task_queue),
2876   _mark_stats_cache(mark_stats, G1RegionMarkStatsCache::RegionMarkStatsCacheSize),
2877   _calls(0),
2878   _time_target_ms(0.0),
2879   _start_time_ms(0.0),
2880   _cm_oop_closure(NULL),
2881   _curr_region(NULL),
2882   _finger(NULL),
2883   _region_limit(NULL),
2884   _words_scanned(0),
2885   _words_scanned_limit(0),
2886   _real_words_scanned_limit(0),
2887   _refs_reached(0),
2888   _refs_reached_limit(0),
2889   _real_refs_reached_limit(0),
2890   _has_aborted(false),
2891   _has_timed_out(false),
2892   _draining_satb_buffers(false),
2893   _step_times_ms(),
2894   _elapsed_time_ms(0.0),
2895   _termination_time_ms(0.0),
2896   _termination_start_time_ms(0.0),
2897   _marking_step_diff_ms()
2898 {
2899   guarantee(task_queue != NULL, "invariant");
2900 
2901   _marking_step_diff_ms.add(0.5);
2902 }
2903 
2904 // These are formatting macros that are used below to ensure
2905 // consistent formatting. The *_H_* versions are used to format the
2906 // header for a particular value and they should be kept consistent
2907 // with the corresponding macro. Also note that most of the macros add
2908 // the necessary white space (as a prefix) which makes them a bit
2909 // easier to compose.
2910 
2911 // All the output lines are prefixed with this string to be able to
2912 // identify them easily in a large log file.
2913 #define G1PPRL_LINE_PREFIX            "###"
2914 
2915 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2916 #ifdef _LP64
2917 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2918 #else // _LP64
2919 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2920 #endif // _LP64
2921 
2922 // For per-region info
2923 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2924 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2925 #define G1PPRL_STATE_FORMAT           "   %-5s"
2926 #define G1PPRL_STATE_H_FORMAT         "   %5s"
2927 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2928 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2929 #define G1PPRL_DOUBLE_FORMAT          "%14.1f"
2930 #define G1PPRL_GCEFF_FORMAT           "  %14s"
2931 #define G1PPRL_GCEFF_H_FORMAT         "  %14s"
2932 
2933 // For summary info
2934 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2935 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2936 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2937 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2938 
2939 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2940   _total_used_bytes(0), _total_capacity_bytes(0),
2941   _total_prev_live_bytes(0), _total_next_live_bytes(0),
2942   _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2943 {
2944   if (!log_is_enabled(Trace, gc, liveness)) {
2945     return;
2946   }
2947 
2948   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2949   MemRegion reserved = g1h->reserved();
2950   double now = os::elapsedTime();
2951 
2952   // Print the header of the output.
2953   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2954   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2955                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2956                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2957                           p2i(reserved.start()), p2i(reserved.end()),
2958                           HeapRegion::GrainBytes);
2959   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2960   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2961                           G1PPRL_TYPE_H_FORMAT
2962                           G1PPRL_ADDR_BASE_H_FORMAT
2963                           G1PPRL_BYTE_H_FORMAT
2964                           G1PPRL_BYTE_H_FORMAT
2965                           G1PPRL_BYTE_H_FORMAT
2966                           G1PPRL_GCEFF_H_FORMAT
2967                           G1PPRL_BYTE_H_FORMAT
2968                           G1PPRL_STATE_H_FORMAT
2969                           G1PPRL_BYTE_H_FORMAT,
2970                           "type", "address-range",
2971                           "used", "prev-live", "next-live", "gc-eff",
2972                           "remset", "state", "code-roots");
2973   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2974                           G1PPRL_TYPE_H_FORMAT
2975                           G1PPRL_ADDR_BASE_H_FORMAT
2976                           G1PPRL_BYTE_H_FORMAT
2977                           G1PPRL_BYTE_H_FORMAT
2978                           G1PPRL_BYTE_H_FORMAT
2979                           G1PPRL_GCEFF_H_FORMAT
2980                           G1PPRL_BYTE_H_FORMAT
2981                           G1PPRL_STATE_H_FORMAT
2982                           G1PPRL_BYTE_H_FORMAT,
2983                           "", "",
2984                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
2985                           "(bytes)", "", "(bytes)");
2986 }
2987 
2988 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
2989   if (!log_is_enabled(Trace, gc, liveness)) {
2990     return false;
2991   }
2992 
2993   const char* type       = r->get_type_str();
2994   HeapWord* bottom       = r->bottom();
2995   HeapWord* end          = r->end();
2996   size_t capacity_bytes  = r->capacity();
2997   size_t used_bytes      = r->used();
2998   size_t prev_live_bytes = r->live_bytes();
2999   size_t next_live_bytes = r->next_live_bytes();
3000   double gc_eff          = r->gc_efficiency();
3001   size_t remset_bytes    = r->rem_set()->mem_size();
3002   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3003   const char* remset_type = r->rem_set()->get_short_state_str();
3004   FormatBuffer<16> gc_efficiency("");
3005 
3006   _total_used_bytes      += used_bytes;
3007   _total_capacity_bytes  += capacity_bytes;
3008   _total_prev_live_bytes += prev_live_bytes;
3009   _total_next_live_bytes += next_live_bytes;
3010   _total_remset_bytes    += remset_bytes;
3011   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3012 
3013   if(gc_eff < 0) {
3014     gc_efficiency.append("-");
3015   } else {
3016     gc_efficiency.append(G1PPRL_DOUBLE_FORMAT, gc_eff);
3017   }
3018 
3019   // Print a line for this particular region.
3020   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3021                         G1PPRL_TYPE_FORMAT
3022                         G1PPRL_ADDR_BASE_FORMAT
3023                         G1PPRL_BYTE_FORMAT
3024                         G1PPRL_BYTE_FORMAT
3025                         G1PPRL_BYTE_FORMAT
3026                         G1PPRL_GCEFF_FORMAT
3027                         G1PPRL_BYTE_FORMAT
3028                         G1PPRL_STATE_FORMAT
3029                         G1PPRL_BYTE_FORMAT,
3030                         type, p2i(bottom), p2i(end),
3031                         used_bytes, prev_live_bytes, next_live_bytes, gc_efficiency.buffer(),
3032                         remset_bytes, remset_type, strong_code_roots_bytes);
3033 
3034   return false;
3035 }
3036 
3037 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3038   if (!log_is_enabled(Trace, gc, liveness)) {
3039     return;
3040   }
3041 
3042   // add static memory usages to remembered set sizes
3043   _total_remset_bytes += G1CardSetFreePool::free_list_pool()->mem_size() + HeapRegionRemSet::static_mem_size();
3044   // Print the footer of the output.
3045   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3046   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3047                          " SUMMARY"
3048                          G1PPRL_SUM_MB_FORMAT("capacity")
3049                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3050                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3051                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3052                          G1PPRL_SUM_MB_FORMAT("remset")
3053                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3054                          bytes_to_mb(_total_capacity_bytes),
3055                          bytes_to_mb(_total_used_bytes),
3056                          percent_of(_total_used_bytes, _total_capacity_bytes),
3057                          bytes_to_mb(_total_prev_live_bytes),
3058                          percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3059                          bytes_to_mb(_total_next_live_bytes),
3060                          percent_of(_total_next_live_bytes, _total_capacity_bytes),
3061                          bytes_to_mb(_total_remset_bytes),
3062                          bytes_to_mb(_total_strong_code_roots_bytes));
3063 }
--- EOF ---