1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/metadataOnStackMark.hpp"
  27 #include "classfile/symbolTable.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc/g1/g1BarrierSet.hpp"
  30 #include "gc/g1/g1CollectedHeap.inline.hpp"
  31 #include "gc/g1/g1CollectorState.hpp"
  32 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  33 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp"
  34 #include "gc/g1/g1HeapVerifier.hpp"
  35 #include "gc/g1/g1OopClosures.inline.hpp"
  36 #include "gc/g1/g1Policy.hpp"
  37 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
  38 #include "gc/g1/g1StringDedup.hpp"
  39 #include "gc/g1/g1ThreadLocalData.hpp"
  40 #include "gc/g1/heapRegion.inline.hpp"
  41 #include "gc/g1/heapRegionRemSet.hpp"
  42 #include "gc/g1/heapRegionSet.inline.hpp"
  43 #include "gc/shared/adaptiveSizePolicy.hpp"
  44 #include "gc/shared/gcId.hpp"
  45 #include "gc/shared/gcTimer.hpp"
  46 #include "gc/shared/gcTrace.hpp"
  47 #include "gc/shared/gcTraceTime.inline.hpp"
  48 #include "gc/shared/genOopClosures.inline.hpp"
  49 #include "gc/shared/referencePolicy.hpp"
  50 #include "gc/shared/strongRootsScope.hpp"
  51 #include "gc/shared/suspendibleThreadSet.hpp"
  52 #include "gc/shared/taskqueue.inline.hpp"
  53 #include "gc/shared/vmGCOperations.hpp"
  54 #include "gc/shared/weakProcessor.hpp"
  55 #include "include/jvm.h"
  56 #include "logging/log.hpp"
  57 #include "memory/allocation.hpp"
  58 #include "memory/resourceArea.hpp"
  59 #include "oops/access.inline.hpp"
  60 #include "oops/oop.inline.hpp"
  61 #include "runtime/atomic.hpp"
  62 #include "runtime/handles.inline.hpp"
  63 #include "runtime/java.hpp"
  64 #include "runtime/prefetch.inline.hpp"
  65 #include "services/memTracker.hpp"
  66 #include "utilities/align.hpp"
  67 #include "utilities/growableArray.hpp"
  68 
  69 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
  70   assert(addr < _cm->finger(), "invariant");
  71   assert(addr >= _task->finger(), "invariant");
  72 
  73   // We move that task's local finger along.
  74   _task->move_finger_to(addr);
  75 
  76   _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
  77   // we only partially drain the local queue and global stack
  78   _task->drain_local_queue(true);
  79   _task->drain_global_stack(true);
  80 
  81   // if the has_aborted flag has been raised, we need to bail out of
  82   // the iteration
  83   return !_task->has_aborted();
  84 }
  85 
  86 G1CMMarkStack::G1CMMarkStack() :
  87   _max_chunk_capacity(0),
  88   _base(NULL),
  89   _chunk_capacity(0) {
  90   set_empty();
  91 }
  92 
  93 bool G1CMMarkStack::resize(size_t new_capacity) {
  94   assert(is_empty(), "Only resize when stack is empty.");
  95   assert(new_capacity <= _max_chunk_capacity,
  96          "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
  97 
  98   TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
  99 
 100   if (new_base == NULL) {
 101     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));
 102     return false;
 103   }
 104   // Release old mapping.
 105   if (_base != NULL) {
 106     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 107   }
 108 
 109   _base = new_base;
 110   _chunk_capacity = new_capacity;
 111   set_empty();
 112 
 113   return true;
 114 }
 115 
 116 size_t G1CMMarkStack::capacity_alignment() {
 117   return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
 118 }
 119 
 120 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
 121   guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
 122 
 123   size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
 124 
 125   _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 126   size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 127 
 128   guarantee(initial_chunk_capacity <= _max_chunk_capacity,
 129             "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
 130             _max_chunk_capacity,
 131             initial_chunk_capacity);
 132 
 133   log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
 134                 initial_chunk_capacity, _max_chunk_capacity);
 135 
 136   return resize(initial_chunk_capacity);
 137 }
 138 
 139 void G1CMMarkStack::expand() {
 140   if (_chunk_capacity == _max_chunk_capacity) {
 141     log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
 142     return;
 143   }
 144   size_t old_capacity = _chunk_capacity;
 145   // Double capacity if possible
 146   size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
 147 
 148   if (resize(new_capacity)) {
 149     log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 150                   old_capacity, new_capacity);
 151   } else {
 152     log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 153                     old_capacity, new_capacity);
 154   }
 155 }
 156 
 157 G1CMMarkStack::~G1CMMarkStack() {
 158   if (_base != NULL) {
 159     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 160   }
 161 }
 162 
 163 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
 164   elem->next = *list;
 165   *list = elem;
 166 }
 167 
 168 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
 169   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 170   add_chunk_to_list(&_chunk_list, elem);
 171   _chunks_in_chunk_list++;
 172 }
 173 
 174 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
 175   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 176   add_chunk_to_list(&_free_list, elem);
 177 }
 178 
 179 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
 180   TaskQueueEntryChunk* result = *list;
 181   if (result != NULL) {
 182     *list = (*list)->next;
 183   }
 184   return result;
 185 }
 186 
 187 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
 188   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 189   TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
 190   if (result != NULL) {
 191     _chunks_in_chunk_list--;
 192   }
 193   return result;
 194 }
 195 
 196 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
 197   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 198   return remove_chunk_from_list(&_free_list);
 199 }
 200 
 201 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
 202   // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
 203   // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
 204   // wraparound of _hwm.
 205   if (_hwm >= _chunk_capacity) {
 206     return NULL;
 207   }
 208 
 209   size_t cur_idx = Atomic::add(1u, &_hwm) - 1;
 210   if (cur_idx >= _chunk_capacity) {
 211     return NULL;
 212   }
 213 
 214   TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
 215   result->next = NULL;
 216   return result;
 217 }
 218 
 219 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
 220   // Get a new chunk.
 221   TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
 222 
 223   if (new_chunk == NULL) {
 224     // Did not get a chunk from the free list. Allocate from backing memory.
 225     new_chunk = allocate_new_chunk();
 226 
 227     if (new_chunk == NULL) {
 228       return false;
 229     }
 230   }
 231 
 232   Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 233 
 234   add_chunk_to_chunk_list(new_chunk);
 235 
 236   return true;
 237 }
 238 
 239 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
 240   TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
 241 
 242   if (cur == NULL) {
 243     return false;
 244   }
 245 
 246   Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 247 
 248   add_chunk_to_free_list(cur);
 249   return true;
 250 }
 251 
 252 void G1CMMarkStack::set_empty() {
 253   _chunks_in_chunk_list = 0;
 254   _hwm = 0;
 255   _chunk_list = NULL;
 256   _free_list = NULL;
 257 }
 258 
 259 G1CMRootRegions::G1CMRootRegions() :
 260   _survivors(NULL), _cm(NULL), _scan_in_progress(false),
 261   _should_abort(false), _claimed_survivor_index(0) { }
 262 
 263 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
 264   _survivors = survivors;
 265   _cm = cm;
 266 }
 267 
 268 void G1CMRootRegions::prepare_for_scan() {
 269   assert(!scan_in_progress(), "pre-condition");
 270 
 271   // Currently, only survivors can be root regions.
 272   _claimed_survivor_index = 0;
 273   _scan_in_progress = _survivors->regions()->is_nonempty();
 274   _should_abort = false;
 275 }
 276 
 277 HeapRegion* G1CMRootRegions::claim_next() {
 278   if (_should_abort) {
 279     // If someone has set the should_abort flag, we return NULL to
 280     // force the caller to bail out of their loop.
 281     return NULL;
 282   }
 283 
 284   // Currently, only survivors can be root regions.
 285   const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
 286 
 287   int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
 288   if (claimed_index < survivor_regions->length()) {
 289     return survivor_regions->at(claimed_index);
 290   }
 291   return NULL;
 292 }
 293 
 294 uint G1CMRootRegions::num_root_regions() const {
 295   return (uint)_survivors->regions()->length();
 296 }
 297 
 298 void G1CMRootRegions::notify_scan_done() {
 299   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 300   _scan_in_progress = false;
 301   RootRegionScan_lock->notify_all();
 302 }
 303 
 304 void G1CMRootRegions::cancel_scan() {
 305   notify_scan_done();
 306 }
 307 
 308 void G1CMRootRegions::scan_finished() {
 309   assert(scan_in_progress(), "pre-condition");
 310 
 311   // Currently, only survivors can be root regions.
 312   if (!_should_abort) {
 313     assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
 314     assert((uint)_claimed_survivor_index >= _survivors->length(),
 315            "we should have claimed all survivors, claimed index = %u, length = %u",
 316            (uint)_claimed_survivor_index, _survivors->length());
 317   }
 318 
 319   notify_scan_done();
 320 }
 321 
 322 bool G1CMRootRegions::wait_until_scan_finished() {
 323   if (!scan_in_progress()) {
 324     return false;
 325   }
 326 
 327   {
 328     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 329     while (scan_in_progress()) {
 330       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 331     }
 332   }
 333   return true;
 334 }
 335 
 336 // Returns the maximum number of workers to be used in a concurrent
 337 // phase based on the number of GC workers being used in a STW
 338 // phase.
 339 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
 340   return MAX2((num_gc_workers + 2) / 4, 1U);
 341 }
 342 
 343 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
 344                                    G1RegionToSpaceMapper* prev_bitmap_storage,
 345                                    G1RegionToSpaceMapper* next_bitmap_storage) :
 346   // _cm_thread set inside the constructor
 347   _g1h(g1h),
 348   _completed_initialization(false),
 349 
 350   _mark_bitmap_1(),
 351   _mark_bitmap_2(),
 352   _prev_mark_bitmap(&_mark_bitmap_1),
 353   _next_mark_bitmap(&_mark_bitmap_2),
 354 
 355   _heap(_g1h->reserved_region()),
 356 
 357   _root_regions(),
 358 
 359   _global_mark_stack(),
 360 
 361   // _finger set in set_non_marking_state
 362 
 363   _worker_id_offset(DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
 364   _max_num_tasks(ParallelGCThreads),
 365   // _num_active_tasks set in set_non_marking_state()
 366   // _tasks set inside the constructor
 367 
 368   _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
 369   _terminator(ParallelTaskTerminator((int) _max_num_tasks, _task_queues)),
 370 
 371   _first_overflow_barrier_sync(),
 372   _second_overflow_barrier_sync(),
 373 
 374   _has_overflown(false),
 375   _concurrent(false),
 376   _has_aborted(false),
 377   _restart_for_overflow(false),
 378   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 379   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 380 
 381   // _verbose_level set below
 382 
 383   _init_times(),
 384   _remark_times(),
 385   _remark_mark_times(),
 386   _remark_weak_ref_times(),
 387   _cleanup_times(),
 388   _total_cleanup_time(0.0),
 389 
 390   _accum_task_vtime(NULL),
 391 
 392   _concurrent_workers(NULL),
 393   _num_concurrent_workers(0),
 394   _max_concurrent_workers(0),
 395 
 396   _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)),
 397   _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC))
 398 {
 399   _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 400   _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
 401 
 402   // Create & start ConcurrentMark thread.
 403   _cm_thread = new G1ConcurrentMarkThread(this);
 404   if (_cm_thread->osthread() == NULL) {
 405     vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 406   }
 407 
 408   assert(CGC_lock != NULL, "CGC_lock must be initialized");
 409 
 410   SATBMarkQueueSet& satb_qs = G1BarrierSet::satb_mark_queue_set();
 411   satb_qs.set_buffer_size(G1SATBBufferSize);
 412 
 413   _root_regions.init(_g1h->survivor(), this);
 414 
 415   if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
 416     // Calculate the number of concurrent worker threads by scaling
 417     // the number of parallel GC threads.
 418     uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
 419     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 420   }
 421 
 422   assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
 423   if (ConcGCThreads > ParallelGCThreads) {
 424     log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
 425                     ConcGCThreads, ParallelGCThreads);
 426     return;
 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 WorkGang("G1 Conc", _max_concurrent_workers, false, true);
 436   _concurrent_workers->initialize_workers();
 437 
 438   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 439     size_t mark_stack_size =
 440       MIN2(MarkStackSizeMax,
 441           MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
 442     // Verify that the calculated value for MarkStackSize is in range.
 443     // It would be nice to use the private utility routine from Arguments.
 444     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 445       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 446                       "must be between 1 and " SIZE_FORMAT,
 447                       mark_stack_size, MarkStackSizeMax);
 448       return;
 449     }
 450     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 451   } else {
 452     // Verify MarkStackSize is in range.
 453     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 454       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 455         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 456           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 457                           "must be between 1 and " SIZE_FORMAT,
 458                           MarkStackSize, MarkStackSizeMax);
 459           return;
 460         }
 461       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 462         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 463           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 464                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 465                           MarkStackSize, MarkStackSizeMax);
 466           return;
 467         }
 468       }
 469     }
 470   }
 471 
 472   if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
 473     vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
 474   }
 475 
 476   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
 477   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
 478 
 479   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 480   _num_active_tasks = _max_num_tasks;
 481 
 482   for (uint i = 0; i < _max_num_tasks; ++i) {
 483     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 484     task_queue->initialize();
 485     _task_queues->register_queue(i, task_queue);
 486 
 487     _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions());
 488 
 489     _accum_task_vtime[i] = 0.0;
 490   }
 491 
 492   reset_at_marking_complete();
 493   _completed_initialization = true;
 494 }
 495 
 496 void G1ConcurrentMark::reset() {
 497   _has_aborted = false;
 498 
 499   reset_marking_for_restart();
 500 
 501   // Reset all tasks, since different phases will use different number of active
 502   // threads. So, it's easiest to have all of them ready.
 503   for (uint i = 0; i < _max_num_tasks; ++i) {
 504     _tasks[i]->reset(_next_mark_bitmap);
 505   }
 506 
 507   uint max_regions = _g1h->max_regions();
 508   for (uint i = 0; i < max_regions; i++) {
 509     _top_at_rebuild_starts[i] = NULL;
 510     _region_mark_stats[i].clear();
 511   }
 512 }
 513 
 514 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
 515   for (uint j = 0; j < _max_num_tasks; ++j) {
 516     _tasks[j]->clear_mark_stats_cache(region_idx);
 517   }
 518   _top_at_rebuild_starts[region_idx] = NULL;
 519   _region_mark_stats[region_idx].clear();
 520 }
 521 
 522 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
 523   uint const region_idx = r->hrm_index();
 524   if (r->is_humongous()) {
 525     assert(r->is_starts_humongous(), "Got humongous continues region here");
 526     uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size());
 527     for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
 528       clear_statistics_in_region(j);
 529     }
 530   } else {
 531     clear_statistics_in_region(region_idx);
 532   }
 533 }
 534 
 535 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
 536   if (bitmap->is_marked(addr)) {
 537     bitmap->clear(addr);
 538   }
 539 }
 540 
 541 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
 542   assert_at_safepoint_on_vm_thread();
 543 
 544   // Need to clear all mark bits of the humongous object.
 545   clear_mark_if_set(_prev_mark_bitmap, r->bottom());
 546   clear_mark_if_set(_next_mark_bitmap, r->bottom());
 547 
 548   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
 549     return;
 550   }
 551 
 552   // Clear any statistics about the region gathered so far.
 553   clear_statistics(r);
 554 }
 555 
 556 void G1ConcurrentMark::reset_marking_for_restart() {
 557   _global_mark_stack.set_empty();
 558 
 559   // Expand the marking stack, if we have to and if we can.
 560   if (has_overflown()) {
 561     _global_mark_stack.expand();
 562 
 563     uint max_regions = _g1h->max_regions();
 564     for (uint i = 0; i < max_regions; i++) {
 565       _region_mark_stats[i].clear_during_overflow();
 566     }
 567   }
 568 
 569   clear_has_overflown();
 570   _finger = _heap.start();
 571 
 572   for (uint i = 0; i < _max_num_tasks; ++i) {
 573     G1CMTaskQueue* queue = _task_queues->queue(i);
 574     queue->set_empty();
 575   }
 576 }
 577 
 578 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 579   assert(active_tasks <= _max_num_tasks, "we should not have more");
 580 
 581   _num_active_tasks = active_tasks;
 582   // Need to update the three data structures below according to the
 583   // number of active threads for this phase.
 584   _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
 585   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 586   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 587 }
 588 
 589 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 590   set_concurrency(active_tasks);
 591 
 592   _concurrent = concurrent;
 593 
 594   if (!concurrent) {
 595     // At this point we should be in a STW phase, and completed marking.
 596     assert_at_safepoint_on_vm_thread();
 597     assert(out_of_regions(),
 598            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 599            p2i(_finger), p2i(_heap.end()));
 600   }
 601 }
 602 
 603 void G1ConcurrentMark::reset_at_marking_complete() {
 604   // We set the global marking state to some default values when we're
 605   // not doing marking.
 606   reset_marking_for_restart();
 607   _num_active_tasks = 0;
 608 }
 609 
 610 G1ConcurrentMark::~G1ConcurrentMark() {
 611   FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
 612   FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
 613   // The G1ConcurrentMark instance is never freed.
 614   ShouldNotReachHere();
 615 }
 616 
 617 class G1ClearBitMapTask : public AbstractGangTask {
 618 public:
 619   static size_t chunk_size() { return M; }
 620 
 621 private:
 622   // Heap region closure used for clearing the given mark bitmap.
 623   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 624   private:
 625     G1CMBitMap* _bitmap;
 626     G1ConcurrentMark* _cm;
 627   public:
 628     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
 629     }
 630 
 631     virtual bool do_heap_region(HeapRegion* r) {
 632       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 633 
 634       HeapWord* cur = r->bottom();
 635       HeapWord* const end = r->end();
 636 
 637       while (cur < end) {
 638         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 639         _bitmap->clear_range(mr);
 640 
 641         cur += chunk_size_in_words;
 642 
 643         // Abort iteration if after yielding the marking has been aborted.
 644         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 645           return true;
 646         }
 647         // Repeat the asserts from before the start of the closure. We will do them
 648         // as asserts here to minimize their overhead on the product. However, we
 649         // will have them as guarantees at the beginning / end of the bitmap
 650         // clearing to get some checking in the product.
 651         assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
 652         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 653       }
 654       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 655 
 656       return false;
 657     }
 658   };
 659 
 660   G1ClearBitmapHRClosure _cl;
 661   HeapRegionClaimer _hr_claimer;
 662   bool _suspendible; // If the task is suspendible, workers must join the STS.
 663 
 664 public:
 665   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 666     AbstractGangTask("G1 Clear Bitmap"),
 667     _cl(bitmap, suspendible ? cm : NULL),
 668     _hr_claimer(n_workers),
 669     _suspendible(suspendible)
 670   { }
 671 
 672   void work(uint worker_id) {
 673     SuspendibleThreadSetJoiner sts_join(_suspendible);
 674     G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
 675   }
 676 
 677   bool is_complete() {
 678     return _cl.is_complete();
 679   }
 680 };
 681 
 682 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 683   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 684 
 685   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 686   size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 687 
 688   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 689 
 690   G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
 691 
 692   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 693   workers->run_task(&cl, num_workers);
 694   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 695 }
 696 
 697 void G1ConcurrentMark::cleanup_for_next_mark() {
 698   // Make sure that the concurrent mark thread looks to still be in
 699   // the current cycle.
 700   guarantee(cm_thread()->during_cycle(), "invariant");
 701 
 702   // We are finishing up the current cycle by clearing the next
 703   // marking bitmap and getting it ready for the next cycle. During
 704   // this time no other cycle can start. So, let's make sure that this
 705   // is the case.
 706   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 707 
 708   clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
 709 
 710   // Repeat the asserts from above.
 711   guarantee(cm_thread()->during_cycle(), "invariant");
 712   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 713 }
 714 
 715 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 716   assert_at_safepoint_on_vm_thread();
 717   clear_bitmap(_prev_mark_bitmap, workers, false);
 718 }
 719 
 720 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 721   G1CMBitMap* _bitmap;
 722  public:
 723   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 724   }
 725 
 726   virtual bool do_heap_region(HeapRegion* r) {
 727     // This closure can be called concurrently to the mutator, so we must make sure
 728     // that the result of the getNextMarkedWordAddress() call is compared to the
 729     // value passed to it as limit to detect any found bits.
 730     // end never changes in G1.
 731     HeapWord* end = r->end();
 732     return _bitmap->get_next_marked_addr(r->bottom(), end) != end;
 733   }
 734 };
 735 
 736 bool G1ConcurrentMark::next_mark_bitmap_is_clear() {
 737   CheckBitmapClearHRClosure cl(_next_mark_bitmap);
 738   _g1h->heap_region_iterate(&cl);
 739   return cl.is_complete();
 740 }
 741 
 742 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
 743 public:
 744   bool do_heap_region(HeapRegion* r) {
 745     r->note_start_of_marking();
 746     return false;
 747   }
 748 };
 749 
 750 void G1ConcurrentMark::pre_initial_mark() {
 751   // Initialize marking structures. This has to be done in a STW phase.
 752   reset();
 753 
 754   // For each region note start of marking.
 755   NoteStartOfMarkHRClosure startcl;
 756   _g1h->heap_region_iterate(&startcl);
 757 }
 758 
 759 
 760 void G1ConcurrentMark::post_initial_mark() {
 761   // Start Concurrent Marking weak-reference discovery.
 762   ReferenceProcessor* rp = _g1h->ref_processor_cm();
 763   // enable ("weak") refs discovery
 764   rp->enable_discovery();
 765   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 766 
 767   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
 768   // This is the start of  the marking cycle, we're expected all
 769   // threads to have SATB queues with active set to false.
 770   satb_mq_set.set_active_all_threads(true, /* new active value */
 771                                      false /* expected_active */);
 772 
 773   _root_regions.prepare_for_scan();
 774 
 775   // update_g1_committed() will be called at the end of an evac pause
 776   // when marking is on. So, it's also called at the end of the
 777   // initial-mark pause to update the heap end, if the heap expands
 778   // during it. No need to call it here.
 779 }
 780 
 781 /*
 782  * Notice that in the next two methods, we actually leave the STS
 783  * during the barrier sync and join it immediately afterwards. If we
 784  * do not do this, the following deadlock can occur: one thread could
 785  * be in the barrier sync code, waiting for the other thread to also
 786  * sync up, whereas another one could be trying to yield, while also
 787  * waiting for the other threads to sync up too.
 788  *
 789  * Note, however, that this code is also used during remark and in
 790  * this case we should not attempt to leave / enter the STS, otherwise
 791  * we'll either hit an assert (debug / fastdebug) or deadlock
 792  * (product). So we should only leave / enter the STS if we are
 793  * operating concurrently.
 794  *
 795  * Because the thread that does the sync barrier has left the STS, it
 796  * is possible to be suspended for a Full GC or an evacuation pause
 797  * could occur. This is actually safe, since the entering the sync
 798  * barrier is one of the last things do_marking_step() does, and it
 799  * doesn't manipulate any data structures afterwards.
 800  */
 801 
 802 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 803   bool barrier_aborted;
 804   {
 805     SuspendibleThreadSetLeaver sts_leave(concurrent());
 806     barrier_aborted = !_first_overflow_barrier_sync.enter();
 807   }
 808 
 809   // at this point everyone should have synced up and not be doing any
 810   // more work
 811 
 812   if (barrier_aborted) {
 813     // If the barrier aborted we ignore the overflow condition and
 814     // just abort the whole marking phase as quickly as possible.
 815     return;
 816   }
 817 }
 818 
 819 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 820   SuspendibleThreadSetLeaver sts_leave(concurrent());
 821   _second_overflow_barrier_sync.enter();
 822 
 823   // at this point everything should be re-initialized and ready to go
 824 }
 825 
 826 class G1CMConcurrentMarkingTask : public AbstractGangTask {
 827   G1ConcurrentMark*     _cm;
 828 
 829 public:
 830   void work(uint worker_id) {
 831     assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
 832     ResourceMark rm;
 833 
 834     double start_vtime = os::elapsedVTime();
 835 
 836     {
 837       SuspendibleThreadSetJoiner sts_join;
 838 
 839       assert(worker_id < _cm->active_tasks(), "invariant");
 840 
 841       G1CMTask* task = _cm->task(worker_id);
 842       task->record_start_time();
 843       if (!_cm->has_aborted()) {
 844         do {
 845           task->do_marking_step(G1ConcMarkStepDurationMillis,
 846                                 true  /* do_termination */,
 847                                 false /* is_serial*/);
 848 
 849           _cm->do_yield_check();
 850         } while (!_cm->has_aborted() && task->has_aborted());
 851       }
 852       task->record_end_time();
 853       guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
 854     }
 855 
 856     double end_vtime = os::elapsedVTime();
 857     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 858   }
 859 
 860   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
 861       AbstractGangTask("Concurrent Mark"), _cm(cm) { }
 862 
 863   ~G1CMConcurrentMarkingTask() { }
 864 };
 865 
 866 uint G1ConcurrentMark::calc_active_marking_workers() {
 867   uint result = 0;
 868   if (!UseDynamicNumberOfGCThreads ||
 869       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 870        !ForceDynamicNumberOfGCThreads)) {
 871     result = _max_concurrent_workers;
 872   } else {
 873     result =
 874       AdaptiveSizePolicy::calc_default_active_workers(_max_concurrent_workers,
 875                                                       1, /* Minimum workers */
 876                                                       _num_concurrent_workers,
 877                                                       Threads::number_of_non_daemon_threads());
 878     // Don't scale the result down by scale_concurrent_workers() because
 879     // that scaling has already gone into "_max_concurrent_workers".
 880   }
 881   assert(result > 0 && result <= _max_concurrent_workers,
 882          "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
 883          _max_concurrent_workers, result);
 884   return result;
 885 }
 886 
 887 void G1ConcurrentMark::scan_root_region(HeapRegion* hr, uint worker_id) {
 888   // Currently, only survivors can be root regions.
 889   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 890   G1RootRegionScanClosure cl(_g1h, this, worker_id);
 891 
 892   const uintx interval = PrefetchScanIntervalInBytes;
 893   HeapWord* curr = hr->bottom();
 894   const HeapWord* end = hr->top();
 895   while (curr < end) {
 896     Prefetch::read(curr, interval);
 897     oop obj = oop(curr);
 898     int size = obj->oop_iterate_size(&cl);
 899     assert(size == obj->size(), "sanity");
 900     curr += size;
 901   }
 902 }
 903 
 904 class G1CMRootRegionScanTask : public AbstractGangTask {
 905   G1ConcurrentMark* _cm;
 906 public:
 907   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 908     AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
 909 
 910   void work(uint worker_id) {
 911     assert(Thread::current()->is_ConcurrentGC_thread(),
 912            "this should only be done by a conc GC thread");
 913 
 914     G1CMRootRegions* root_regions = _cm->root_regions();
 915     HeapRegion* hr = root_regions->claim_next();
 916     while (hr != NULL) {
 917       _cm->scan_root_region(hr, worker_id);
 918       hr = root_regions->claim_next();
 919     }
 920   }
 921 };
 922 
 923 void G1ConcurrentMark::scan_root_regions() {
 924   // scan_in_progress() will have been set to true only if there was
 925   // at least one root region to scan. So, if it's false, we
 926   // should not attempt to do any further work.
 927   if (root_regions()->scan_in_progress()) {
 928     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 929 
 930     _num_concurrent_workers = MIN2(calc_active_marking_workers(),
 931                                    // We distribute work on a per-region basis, so starting
 932                                    // more threads than that is useless.
 933                                    root_regions()->num_root_regions());
 934     assert(_num_concurrent_workers <= _max_concurrent_workers,
 935            "Maximum number of marking threads exceeded");
 936 
 937     G1CMRootRegionScanTask task(this);
 938     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
 939                         task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
 940     _concurrent_workers->run_task(&task, _num_concurrent_workers);
 941 
 942     // It's possible that has_aborted() is true here without actually
 943     // aborting the survivor scan earlier. This is OK as it's
 944     // mainly used for sanity checking.
 945     root_regions()->scan_finished();
 946   }
 947 }
 948 
 949 void G1ConcurrentMark::concurrent_cycle_start() {
 950   _gc_timer_cm->register_gc_start();
 951 
 952   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
 953 
 954   _g1h->trace_heap_before_gc(_gc_tracer_cm);
 955 }
 956 
 957 void G1ConcurrentMark::concurrent_cycle_end() {
 958   _g1h->collector_state()->set_clearing_next_bitmap(false);
 959 
 960   _g1h->trace_heap_after_gc(_gc_tracer_cm);
 961 
 962   if (has_aborted()) {
 963     log_info(gc, marking)("Concurrent Mark Abort");
 964     _gc_tracer_cm->report_concurrent_mode_failure();
 965   }
 966 
 967   _gc_timer_cm->register_gc_end();
 968 
 969   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
 970 }
 971 
 972 void G1ConcurrentMark::mark_from_roots() {
 973   _restart_for_overflow = false;
 974 
 975   _num_concurrent_workers = calc_active_marking_workers();
 976 
 977   uint active_workers = MAX2(1U, _num_concurrent_workers);
 978 
 979   // Setting active workers is not guaranteed since fewer
 980   // worker threads may currently exist and more may not be
 981   // available.
 982   active_workers = _concurrent_workers->update_active_workers(active_workers);
 983   log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
 984 
 985   // Parallel task terminator is set in "set_concurrency_and_phase()"
 986   set_concurrency_and_phase(active_workers, true /* concurrent */);
 987 
 988   G1CMConcurrentMarkingTask marking_task(this);
 989   _concurrent_workers->run_task(&marking_task);
 990   print_stats();
 991 }
 992 
 993 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
 994   G1HeapVerifier* verifier = _g1h->verifier();
 995 
 996   verifier->verify_region_sets_optional();
 997 
 998   if (VerifyDuringGC) {
 999     GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm);
1000 
1001     size_t const BufLen = 512;
1002     char buffer[BufLen];
1003 
1004     jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
1005     verifier->verify(type, vo, buffer);
1006   }
1007 
1008   verifier->check_bitmaps(caller);
1009 }
1010 
1011 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask {
1012   G1CollectedHeap* _g1h;
1013   G1ConcurrentMark* _cm;
1014   HeapRegionClaimer _hrclaimer;
1015   uint volatile _total_selected_for_rebuild;
1016 
1017   G1PrintRegionLivenessInfoClosure _cl;
1018 
1019   class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
1020     G1CollectedHeap* _g1h;
1021     G1ConcurrentMark* _cm;
1022 
1023     G1PrintRegionLivenessInfoClosure* _cl;
1024 
1025     uint _num_regions_selected_for_rebuild;  // The number of regions actually selected for rebuild.
1026 
1027     void update_remset_before_rebuild(HeapRegion* hr) {
1028       G1RemSetTrackingPolicy* tracking_policy = _g1h->g1_policy()->remset_tracker();
1029 
1030       bool selected_for_rebuild;
1031       if (hr->is_humongous()) {
1032         bool const is_live = _cm->liveness(hr->humongous_start_region()->hrm_index()) > 0;
1033         selected_for_rebuild = tracking_policy->update_humongous_before_rebuild(hr, is_live);
1034       } else {
1035         size_t const live_bytes = _cm->liveness(hr->hrm_index());
1036         selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1037       }
1038       if (selected_for_rebuild) {
1039         _num_regions_selected_for_rebuild++;
1040       }
1041       _cm->update_top_at_rebuild_start(hr);
1042     }
1043 
1044     // Distribute the given words across the humongous object starting with hr and
1045     // note end of marking.
1046     void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1047       uint const region_idx = hr->hrm_index();
1048       size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size();
1049       uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1050 
1051       // "Distributing" zero words means that we only note end of marking for these
1052       // regions.
1053       assert(marked_words == 0 || obj_size_in_words == marked_words,
1054              "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1055              obj_size_in_words, marked_words);
1056 
1057       for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1058         HeapRegion* const r = _g1h->region_at(i);
1059         size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1060 
1061         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)",
1062                                words_to_add, i, r->get_type_str());
1063         add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1064         marked_words -= words_to_add;
1065       }
1066       assert(marked_words == 0,
1067              SIZE_FORMAT " words left after distributing space across %u regions",
1068              marked_words, num_regions_in_humongous);
1069     }
1070 
1071     void update_marked_bytes(HeapRegion* hr) {
1072       uint const region_idx = hr->hrm_index();
1073       size_t const marked_words = _cm->liveness(region_idx);
1074       // The marking attributes the object's size completely to the humongous starts
1075       // region. We need to distribute this value across the entire set of regions a
1076       // humongous object spans.
1077       if (hr->is_humongous()) {
1078         assert(hr->is_starts_humongous() || marked_words == 0,
1079                "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1080                marked_words, region_idx, hr->get_type_str());
1081         if (hr->is_starts_humongous()) {
1082           distribute_marked_bytes(hr, marked_words);
1083         }
1084       } else {
1085         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1086         add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize);
1087       }
1088     }
1089 
1090     void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1091       hr->add_to_marked_bytes(marked_bytes);
1092       _cl->do_heap_region(hr);
1093       hr->note_end_of_marking();
1094     }
1095 
1096   public:
1097     G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1098       _g1h(g1h), _cm(cm), _num_regions_selected_for_rebuild(0), _cl(cl) { }
1099 
1100     virtual bool do_heap_region(HeapRegion* r) {
1101       update_remset_before_rebuild(r);
1102       update_marked_bytes(r);
1103 
1104       return false;
1105     }
1106 
1107     uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1108   };
1109 
1110 public:
1111   G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1112     AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"),
1113     _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1114 
1115   virtual void work(uint worker_id) {
1116     G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1117     _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1118     Atomic::add(update_cl.num_selected_for_rebuild(), &_total_selected_for_rebuild);
1119   }
1120 
1121   uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1122 
1123   // Number of regions for which roughly one thread should be spawned for this work.
1124   static const uint RegionsPerThread = 384;
1125 };
1126 
1127 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1128   G1CollectedHeap* _g1h;
1129 public:
1130   G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1131 
1132   virtual bool do_heap_region(HeapRegion* r) {
1133     _g1h->g1_policy()->remset_tracker()->update_after_rebuild(r);
1134     return false;
1135   }
1136 };
1137 
1138 void G1ConcurrentMark::remark() {
1139   assert_at_safepoint_on_vm_thread();
1140 
1141   // If a full collection has happened, we should not continue. However we might
1142   // have ended up here as the Remark VM operation has been scheduled already.
1143   if (has_aborted()) {
1144     return;
1145   }
1146 
1147   G1Policy* g1p = _g1h->g1_policy();
1148   g1p->record_concurrent_mark_remark_start();
1149 
1150   double start = os::elapsedTime();
1151 
1152   verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1153 
1154   {
1155     GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm);
1156     finalize_marking();
1157   }
1158 
1159   double mark_work_end = os::elapsedTime();
1160 
1161   bool const mark_finished = !has_overflown();
1162   if (mark_finished) {
1163     weak_refs_work(false /* clear_all_soft_refs */);
1164 
1165     SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1166     // We're done with marking.
1167     // This is the end of the marking cycle, we're expected all
1168     // threads to have SATB queues with active set to true.
1169     satb_mq_set.set_active_all_threads(false, /* new active value */
1170                                        true /* expected_active */);
1171 
1172     {
1173       GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm);
1174       flush_all_task_caches();
1175     }
1176 
1177     // Install newly created mark bitmap as "prev".
1178     swap_mark_bitmaps();
1179     {
1180       GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1181 
1182       uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1183                                        G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1184       uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1185 
1186       G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1187       log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1188       _g1h->workers()->run_task(&cl, num_workers);
1189 
1190       log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1191                                       _g1h->num_regions(), cl.total_selected_for_rebuild());
1192     }
1193     {
1194       GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm);
1195       reclaim_empty_regions();
1196     }
1197 
1198     // Clean out dead classes
1199     if (ClassUnloadingWithConcurrentMark) {
1200       GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm);
1201       ClassLoaderDataGraph::purge();
1202     }
1203 
1204     compute_new_sizes();
1205 
1206     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1207 
1208     assert(!restart_for_overflow(), "sanity");
1209     // Completely reset the marking state since marking completed
1210     reset_at_marking_complete();
1211   } else {
1212     // We overflowed.  Restart concurrent marking.
1213     _restart_for_overflow = true;
1214 
1215     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1216 
1217     // Clear the marking state because we will be restarting
1218     // marking due to overflowing the global mark stack.
1219     reset_marking_for_restart();
1220   }
1221 
1222   {
1223     GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm);
1224     report_object_count(mark_finished);
1225   }
1226 
1227   // Statistics
1228   double now = os::elapsedTime();
1229   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1230   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1231   _remark_times.add((now - start) * 1000.0);
1232 
1233   g1p->record_concurrent_mark_remark_end();
1234 }
1235 
1236 class G1ReclaimEmptyRegionsTask : public AbstractGangTask {
1237   // Per-region work during the Cleanup pause.
1238   class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1239     G1CollectedHeap* _g1h;
1240     size_t _freed_bytes;
1241     FreeRegionList* _local_cleanup_list;
1242     uint _old_regions_removed;
1243     uint _humongous_regions_removed;
1244     HRRSCleanupTask* _hrrs_cleanup_task;
1245 
1246   public:
1247     G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1248                                  FreeRegionList* local_cleanup_list,
1249                                  HRRSCleanupTask* hrrs_cleanup_task) :
1250       _g1h(g1h),
1251       _freed_bytes(0),
1252       _local_cleanup_list(local_cleanup_list),
1253       _old_regions_removed(0),
1254       _humongous_regions_removed(0),
1255       _hrrs_cleanup_task(hrrs_cleanup_task) { }
1256 
1257     size_t freed_bytes() { return _freed_bytes; }
1258     const uint old_regions_removed() { return _old_regions_removed; }
1259     const uint humongous_regions_removed() { return _humongous_regions_removed; }
1260 
1261     bool do_heap_region(HeapRegion *hr) {
1262       if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1263         _freed_bytes += hr->used();
1264         hr->set_containing_set(NULL);
1265         if (hr->is_humongous()) {
1266           _humongous_regions_removed++;
1267           _g1h->free_humongous_region(hr, _local_cleanup_list);
1268         } else {
1269           _old_regions_removed++;
1270           _g1h->free_region(hr, _local_cleanup_list, false /* skip_remset */, false /* skip_hcc */, true /* locked */);
1271         }
1272         hr->clear_cardtable();
1273         _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1274         log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1275       } else {
1276         hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1277       }
1278 
1279       return false;
1280     }
1281   };
1282 
1283   G1CollectedHeap* _g1h;
1284   FreeRegionList* _cleanup_list;
1285   HeapRegionClaimer _hrclaimer;
1286 
1287 public:
1288   G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1289     AbstractGangTask("G1 Cleanup"),
1290     _g1h(g1h),
1291     _cleanup_list(cleanup_list),
1292     _hrclaimer(n_workers) {
1293 
1294     HeapRegionRemSet::reset_for_cleanup_tasks();
1295   }
1296 
1297   void work(uint worker_id) {
1298     FreeRegionList local_cleanup_list("Local Cleanup List");
1299     HRRSCleanupTask hrrs_cleanup_task;
1300     G1ReclaimEmptyRegionsClosure cl(_g1h,
1301                                     &local_cleanup_list,
1302                                     &hrrs_cleanup_task);
1303     _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1304     assert(cl.is_complete(), "Shouldn't have aborted!");
1305 
1306     // Now update the old/humongous region sets
1307     _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed());
1308     {
1309       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1310       _g1h->decrement_summary_bytes(cl.freed_bytes());
1311 
1312       _cleanup_list->add_ordered(&local_cleanup_list);
1313       assert(local_cleanup_list.is_empty(), "post-condition");
1314 
1315       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1316     }
1317   }
1318 };
1319 
1320 void G1ConcurrentMark::reclaim_empty_regions() {
1321   WorkGang* workers = _g1h->workers();
1322   FreeRegionList empty_regions_list("Empty Regions After Mark List");
1323 
1324   G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1325   workers->run_task(&cl);
1326 
1327   if (!empty_regions_list.is_empty()) {
1328     log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1329     // Now print the empty regions list.
1330     G1HRPrinter* hrp = _g1h->hr_printer();
1331     if (hrp->is_active()) {
1332       FreeRegionListIterator iter(&empty_regions_list);
1333       while (iter.more_available()) {
1334         HeapRegion* hr = iter.get_next();
1335         hrp->cleanup(hr);
1336       }
1337     }
1338     // And actually make them available.
1339     _g1h->prepend_to_freelist(&empty_regions_list);
1340   }
1341 }
1342 
1343 void G1ConcurrentMark::compute_new_sizes() {
1344   MetaspaceGC::compute_new_size();
1345 
1346   // Cleanup will have freed any regions completely full of garbage.
1347   // Update the soft reference policy with the new heap occupancy.
1348   Universe::update_heap_info_at_gc();
1349 
1350   // We reclaimed old regions so we should calculate the sizes to make
1351   // sure we update the old gen/space data.
1352   _g1h->g1mm()->update_sizes();
1353 }
1354 
1355 void G1ConcurrentMark::cleanup() {
1356   assert_at_safepoint_on_vm_thread();
1357 
1358   // If a full collection has happened, we shouldn't do this.
1359   if (has_aborted()) {
1360     return;
1361   }
1362 
1363   G1Policy* g1p = _g1h->g1_policy();
1364   g1p->record_concurrent_mark_cleanup_start();
1365 
1366   double start = os::elapsedTime();
1367 
1368   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1369 
1370   {
1371     GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1372     G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1373     _g1h->heap_region_iterate(&cl);
1374   }
1375 
1376   if (log_is_enabled(Trace, gc, liveness)) {
1377     G1PrintRegionLivenessInfoClosure cl("Post-Cleanup");
1378     _g1h->heap_region_iterate(&cl);
1379   }
1380 
1381   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1382 
1383   // We need to make this be a "collection" so any collection pause that
1384   // races with it goes around and waits for Cleanup to finish.
1385   _g1h->increment_total_collections();
1386 
1387   // Local statistics
1388   double recent_cleanup_time = (os::elapsedTime() - start);
1389   _total_cleanup_time += recent_cleanup_time;
1390   _cleanup_times.add(recent_cleanup_time);
1391 
1392   {
1393     GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1394     _g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1395   }
1396 }
1397 
1398 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1399 // Uses the G1CMTask associated with a worker thread (for serial reference
1400 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1401 // trace referent objects.
1402 //
1403 // Using the G1CMTask and embedded local queues avoids having the worker
1404 // threads operating on the global mark stack. This reduces the risk
1405 // of overflowing the stack - which we would rather avoid at this late
1406 // state. Also using the tasks' local queues removes the potential
1407 // of the workers interfering with each other that could occur if
1408 // operating on the global stack.
1409 
1410 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1411   G1ConcurrentMark* _cm;
1412   G1CMTask*         _task;
1413   uint              _ref_counter_limit;
1414   uint              _ref_counter;
1415   bool              _is_serial;
1416 public:
1417   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1418     _cm(cm), _task(task), _is_serial(is_serial),
1419     _ref_counter_limit(G1RefProcDrainInterval) {
1420     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1421     _ref_counter = _ref_counter_limit;
1422   }
1423 
1424   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1425   virtual void do_oop(      oop* p) { do_oop_work(p); }
1426 
1427   template <class T> void do_oop_work(T* p) {
1428     if (_cm->has_overflown()) {
1429       return;
1430     }
1431     if (!_task->deal_with_reference(p)) {
1432       // We did not add anything to the mark bitmap (or mark stack), so there is
1433       // no point trying to drain it.
1434       return;
1435     }
1436     _ref_counter--;
1437 
1438     if (_ref_counter == 0) {
1439       // We have dealt with _ref_counter_limit references, pushing them
1440       // and objects reachable from them on to the local stack (and
1441       // possibly the global stack). Call G1CMTask::do_marking_step() to
1442       // process these entries.
1443       //
1444       // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1445       // there's nothing more to do (i.e. we're done with the entries that
1446       // were pushed as a result of the G1CMTask::deal_with_reference() calls
1447       // above) or we overflow.
1448       //
1449       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1450       // flag while there may still be some work to do. (See the comment at
1451       // the beginning of G1CMTask::do_marking_step() for those conditions -
1452       // one of which is reaching the specified time target.) It is only
1453       // when G1CMTask::do_marking_step() returns without setting the
1454       // has_aborted() flag that the marking step has completed.
1455       do {
1456         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1457         _task->do_marking_step(mark_step_duration_ms,
1458                                false      /* do_termination */,
1459                                _is_serial);
1460       } while (_task->has_aborted() && !_cm->has_overflown());
1461       _ref_counter = _ref_counter_limit;
1462     }
1463   }
1464 };
1465 
1466 // 'Drain' oop closure used by both serial and parallel reference processing.
1467 // Uses the G1CMTask associated with a given worker thread (for serial
1468 // reference processing the G1CMtask for worker 0 is used). Calls the
1469 // do_marking_step routine, with an unbelievably large timeout value,
1470 // to drain the marking data structures of the remaining entries
1471 // added by the 'keep alive' oop closure above.
1472 
1473 class G1CMDrainMarkingStackClosure : public VoidClosure {
1474   G1ConcurrentMark* _cm;
1475   G1CMTask*         _task;
1476   bool              _is_serial;
1477  public:
1478   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1479     _cm(cm), _task(task), _is_serial(is_serial) {
1480     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1481   }
1482 
1483   void do_void() {
1484     do {
1485       // We call G1CMTask::do_marking_step() to completely drain the local
1486       // and global marking stacks of entries pushed by the 'keep alive'
1487       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1488       //
1489       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1490       // if there's nothing more to do (i.e. we've completely drained the
1491       // entries that were pushed as a a result of applying the 'keep alive'
1492       // closure to the entries on the discovered ref lists) or we overflow
1493       // the global marking stack.
1494       //
1495       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1496       // flag while there may still be some work to do. (See the comment at
1497       // the beginning of G1CMTask::do_marking_step() for those conditions -
1498       // one of which is reaching the specified time target.) It is only
1499       // when G1CMTask::do_marking_step() returns without setting the
1500       // has_aborted() flag that the marking step has completed.
1501 
1502       _task->do_marking_step(1000000000.0 /* something very large */,
1503                              true         /* do_termination */,
1504                              _is_serial);
1505     } while (_task->has_aborted() && !_cm->has_overflown());
1506   }
1507 };
1508 
1509 // Implementation of AbstractRefProcTaskExecutor for parallel
1510 // reference processing at the end of G1 concurrent marking
1511 
1512 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor {
1513 private:
1514   G1CollectedHeap*  _g1h;
1515   G1ConcurrentMark* _cm;
1516   WorkGang*         _workers;
1517   uint              _active_workers;
1518 
1519 public:
1520   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1521                           G1ConcurrentMark* cm,
1522                           WorkGang* workers,
1523                           uint n_workers) :
1524     _g1h(g1h), _cm(cm),
1525     _workers(workers), _active_workers(n_workers) { }
1526 
1527   virtual void execute(ProcessTask& task, uint ergo_workers);
1528 };
1529 
1530 class G1CMRefProcTaskProxy : public AbstractGangTask {
1531   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1532   ProcessTask&      _proc_task;
1533   G1CollectedHeap*  _g1h;
1534   G1ConcurrentMark* _cm;
1535 
1536 public:
1537   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1538                        G1CollectedHeap* g1h,
1539                        G1ConcurrentMark* cm) :
1540     AbstractGangTask("Process reference objects in parallel"),
1541     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1542     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1543     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1544   }
1545 
1546   virtual void work(uint worker_id) {
1547     ResourceMark rm;
1548     HandleMark hm;
1549     G1CMTask* task = _cm->task(worker_id);
1550     G1CMIsAliveClosure g1_is_alive(_g1h);
1551     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1552     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1553 
1554     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1555   }
1556 };
1557 
1558 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task, uint ergo_workers) {
1559   assert(_workers != NULL, "Need parallel worker threads.");
1560   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1561   assert(_workers->active_workers() >= ergo_workers,
1562          "Ergonomically chosen workers(%u) should be less than or equal to active workers(%u)",
1563          ergo_workers, _workers->active_workers());
1564 
1565   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1566 
1567   // We need to reset the concurrency level before each
1568   // proxy task execution, so that the termination protocol
1569   // and overflow handling in G1CMTask::do_marking_step() knows
1570   // how many workers to wait for.
1571   _cm->set_concurrency(ergo_workers);
1572   _workers->run_task(&proc_task_proxy, ergo_workers);
1573 }
1574 
1575 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1576   ResourceMark rm;
1577   HandleMark   hm;
1578 
1579   // Is alive closure.
1580   G1CMIsAliveClosure g1_is_alive(_g1h);
1581 
1582   // Inner scope to exclude the cleaning of the string and symbol
1583   // tables from the displayed time.
1584   {
1585     GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm);
1586 
1587     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1588 
1589     // See the comment in G1CollectedHeap::ref_processing_init()
1590     // about how reference processing currently works in G1.
1591 
1592     // Set the soft reference policy
1593     rp->setup_policy(clear_all_soft_refs);
1594     assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1595 
1596     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1597     // in serial reference processing. Note these closures are also
1598     // used for serially processing (by the the current thread) the
1599     // JNI references during parallel reference processing.
1600     //
1601     // These closures do not need to synchronize with the worker
1602     // threads involved in parallel reference processing as these
1603     // instances are executed serially by the current thread (e.g.
1604     // reference processing is not multi-threaded and is thus
1605     // performed by the current thread instead of a gang worker).
1606     //
1607     // The gang tasks involved in parallel reference processing create
1608     // their own instances of these closures, which do their own
1609     // synchronization among themselves.
1610     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1611     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1612 
1613     // We need at least one active thread. If reference processing
1614     // is not multi-threaded we use the current (VMThread) thread,
1615     // otherwise we use the work gang from the G1CollectedHeap and
1616     // we utilize all the worker threads we can.
1617     bool processing_is_mt = rp->processing_is_mt();
1618     uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U);
1619     active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U);
1620 
1621     // Parallel processing task executor.
1622     G1CMRefProcTaskExecutor par_task_executor(_g1h, this,
1623                                               _g1h->workers(), active_workers);
1624     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1625 
1626     // Set the concurrency level. The phase was already set prior to
1627     // executing the remark task.
1628     set_concurrency(active_workers);
1629 
1630     // Set the degree of MT processing here.  If the discovery was done MT,
1631     // the number of threads involved during discovery could differ from
1632     // the number of active workers.  This is OK as long as the discovered
1633     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1634     rp->set_active_mt_degree(active_workers);
1635 
1636     ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues());
1637 
1638     // Process the weak references.
1639     const ReferenceProcessorStats& stats =
1640         rp->process_discovered_references(&g1_is_alive,
1641                                           &g1_keep_alive,
1642                                           &g1_drain_mark_stack,
1643                                           executor,
1644                                           &pt);
1645     _gc_tracer_cm->report_gc_reference_stats(stats);
1646     pt.print_all_references();
1647 
1648     // The do_oop work routines of the keep_alive and drain_marking_stack
1649     // oop closures will set the has_overflown flag if we overflow the
1650     // global marking stack.
1651 
1652     assert(has_overflown() || _global_mark_stack.is_empty(),
1653            "Mark stack should be empty (unless it has overflown)");
1654 
1655     assert(rp->num_queues() == active_workers, "why not");
1656 
1657     rp->verify_no_references_recorded();
1658     assert(!rp->discovery_enabled(), "Post condition");
1659   }
1660 
1661   if (has_overflown()) {
1662     // We can not trust g1_is_alive and the contents of the heap if the marking stack
1663     // overflowed while processing references. Exit the VM.
1664     fatal("Overflow during reference processing, can not continue. Please "
1665           "increase MarkStackSizeMax (current value: " SIZE_FORMAT ") and "
1666           "restart.", MarkStackSizeMax);
1667     return;
1668   }
1669 
1670   assert(_global_mark_stack.is_empty(), "Marking should have completed");
1671 
1672   {
1673     GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1674     WeakProcessor::weak_oops_do(&g1_is_alive, &do_nothing_cl);
1675   }
1676 
1677   // Unload Klasses, String, Symbols, Code Cache, etc.
1678   if (ClassUnloadingWithConcurrentMark) {
1679     GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1680     bool purged_classes = SystemDictionary::do_unloading(_gc_timer_cm, false /* Defer cleaning */);
1681     _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1682   } else {
1683     GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1684     // No need to clean string table and symbol table as they are treated as strong roots when
1685     // class unloading is disabled.
1686     _g1h->partial_cleaning(&g1_is_alive, false, false, G1StringDedup::is_enabled());
1687   }
1688 }
1689 
1690 class G1PrecleanYieldClosure : public YieldClosure {
1691   G1ConcurrentMark* _cm;
1692 
1693 public:
1694   G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
1695 
1696   virtual bool should_return() {
1697     return _cm->has_aborted();
1698   }
1699 
1700   virtual bool should_return_fine_grain() {
1701     _cm->do_yield_check();
1702     return _cm->has_aborted();
1703   }
1704 };
1705 
1706 void G1ConcurrentMark::preclean() {
1707   assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
1708 
1709   SuspendibleThreadSetJoiner joiner;
1710 
1711   G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */);
1712   G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */);
1713 
1714   set_concurrency_and_phase(1, true);
1715 
1716   G1PrecleanYieldClosure yield_cl(this);
1717 
1718   ReferenceProcessor* rp = _g1h->ref_processor_cm();
1719   // Precleaning is single threaded. Temporarily disable MT discovery.
1720   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
1721   rp->preclean_discovered_references(rp->is_alive_non_header(),
1722                                      &keep_alive,
1723                                      &drain_mark_stack,
1724                                      &yield_cl,
1725                                      _gc_timer_cm);
1726 }
1727 
1728 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1729 // the prev bitmap determining liveness.
1730 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1731   G1CollectedHeap* _g1h;
1732 public:
1733   G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1734 
1735   bool do_object_b(oop obj) {
1736     HeapWord* addr = (HeapWord*)obj;
1737     return addr != NULL &&
1738            (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_dead(obj));
1739   }
1740 };
1741 
1742 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1743   // Depending on the completion of the marking liveness needs to be determined
1744   // using either the next or prev bitmap.
1745   if (mark_completed) {
1746     G1ObjectCountIsAliveClosure is_alive(_g1h);
1747     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1748   } else {
1749     G1CMIsAliveClosure is_alive(_g1h);
1750     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1751   }
1752 }
1753 
1754 
1755 void G1ConcurrentMark::swap_mark_bitmaps() {
1756   G1CMBitMap* temp = _prev_mark_bitmap;
1757   _prev_mark_bitmap = _next_mark_bitmap;
1758   _next_mark_bitmap = temp;
1759   _g1h->collector_state()->set_clearing_next_bitmap(true);
1760 }
1761 
1762 // Closure for marking entries in SATB buffers.
1763 class G1CMSATBBufferClosure : public SATBBufferClosure {
1764 private:
1765   G1CMTask* _task;
1766   G1CollectedHeap* _g1h;
1767 
1768   // This is very similar to G1CMTask::deal_with_reference, but with
1769   // more relaxed requirements for the argument, so this must be more
1770   // circumspect about treating the argument as an object.
1771   void do_entry(void* entry) const {
1772     _task->increment_refs_reached();
1773     oop const obj = static_cast<oop>(entry);
1774     _task->make_reference_grey(obj);
1775   }
1776 
1777 public:
1778   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1779     : _task(task), _g1h(g1h) { }
1780 
1781   virtual void do_buffer(void** buffer, size_t size) {
1782     for (size_t i = 0; i < size; ++i) {
1783       do_entry(buffer[i]);
1784     }
1785   }
1786 };
1787 
1788 class G1RemarkThreadsClosure : public ThreadClosure {
1789   G1CMSATBBufferClosure _cm_satb_cl;
1790   G1CMOopClosure _cm_cl;
1791   MarkingCodeBlobClosure _code_cl;
1792   int _thread_parity;
1793 
1794  public:
1795   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1796     _cm_satb_cl(task, g1h),
1797     _cm_cl(g1h, task),
1798     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1799     _thread_parity(Threads::thread_claim_parity()) {}
1800 
1801   void do_thread(Thread* thread) {
1802     if (thread->is_Java_thread()) {
1803       if (thread->claim_oops_do(true, _thread_parity)) {
1804         JavaThread* jt = (JavaThread*)thread;
1805 
1806         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1807         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1808         // * Alive if on the stack of an executing method
1809         // * Weakly reachable otherwise
1810         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1811         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1812         jt->nmethods_do(&_code_cl);
1813 
1814         G1ThreadLocalData::satb_mark_queue(jt).apply_closure_and_empty(&_cm_satb_cl);
1815       }
1816     } else if (thread->is_VM_thread()) {
1817       if (thread->claim_oops_do(true, _thread_parity)) {
1818         G1BarrierSet::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1819       }
1820     }
1821   }
1822 };
1823 
1824 class G1CMRemarkTask : public AbstractGangTask {
1825   G1ConcurrentMark* _cm;
1826 public:
1827   void work(uint worker_id) {
1828     G1CMTask* task = _cm->task(worker_id);
1829     task->record_start_time();
1830     {
1831       ResourceMark rm;
1832       HandleMark hm;
1833 
1834       G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1835       Threads::threads_do(&threads_f);
1836     }
1837 
1838     do {
1839       task->do_marking_step(1000000000.0 /* something very large */,
1840                             true         /* do_termination       */,
1841                             false        /* is_serial            */);
1842     } while (task->has_aborted() && !_cm->has_overflown());
1843     // If we overflow, then we do not want to restart. We instead
1844     // want to abort remark and do concurrent marking again.
1845     task->record_end_time();
1846   }
1847 
1848   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1849     AbstractGangTask("Par Remark"), _cm(cm) {
1850     _cm->terminator()->reset_for_reuse(active_workers);
1851   }
1852 };
1853 
1854 void G1ConcurrentMark::finalize_marking() {
1855   ResourceMark rm;
1856   HandleMark   hm;
1857 
1858   _g1h->ensure_parsability(false);
1859 
1860   // this is remark, so we'll use up all active threads
1861   uint active_workers = _g1h->workers()->active_workers();
1862   set_concurrency_and_phase(active_workers, false /* concurrent */);
1863   // Leave _parallel_marking_threads at it's
1864   // value originally calculated in the G1ConcurrentMark
1865   // constructor and pass values of the active workers
1866   // through the gang in the task.
1867 
1868   {
1869     StrongRootsScope srs(active_workers);
1870 
1871     G1CMRemarkTask remarkTask(this, active_workers);
1872     // We will start all available threads, even if we decide that the
1873     // active_workers will be fewer. The extra ones will just bail out
1874     // immediately.
1875     _g1h->workers()->run_task(&remarkTask);
1876   }
1877 
1878   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1879   guarantee(has_overflown() ||
1880             satb_mq_set.completed_buffers_num() == 0,
1881             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1882             BOOL_TO_STR(has_overflown()),
1883             satb_mq_set.completed_buffers_num());
1884 
1885   print_stats();
1886 }
1887 
1888 void G1ConcurrentMark::flush_all_task_caches() {
1889   size_t hits = 0;
1890   size_t misses = 0;
1891   for (uint i = 0; i < _max_num_tasks; i++) {
1892     Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1893     hits += stats.first;
1894     misses += stats.second;
1895   }
1896   size_t sum = hits + misses;
1897   log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1898                        hits, misses, percent_of(hits, sum));
1899 }
1900 
1901 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1902   _prev_mark_bitmap->clear_range(mr);
1903 }
1904 
1905 HeapRegion*
1906 G1ConcurrentMark::claim_region(uint worker_id) {
1907   // "checkpoint" the finger
1908   HeapWord* finger = _finger;
1909 
1910   while (finger < _heap.end()) {
1911     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1912 
1913     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1914     // Make sure that the reads below do not float before loading curr_region.
1915     OrderAccess::loadload();
1916     // Above heap_region_containing may return NULL as we always scan claim
1917     // until the end of the heap. In this case, just jump to the next region.
1918     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1919 
1920     // Is the gap between reading the finger and doing the CAS too long?
1921     HeapWord* res = Atomic::cmpxchg(end, &_finger, finger);
1922     if (res == finger && curr_region != NULL) {
1923       // we succeeded
1924       HeapWord*   bottom        = curr_region->bottom();
1925       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1926 
1927       // notice that _finger == end cannot be guaranteed here since,
1928       // someone else might have moved the finger even further
1929       assert(_finger >= end, "the finger should have moved forward");
1930 
1931       if (limit > bottom) {
1932         return curr_region;
1933       } else {
1934         assert(limit == bottom,
1935                "the region limit should be at bottom");
1936         // we return NULL and the caller should try calling
1937         // claim_region() again.
1938         return NULL;
1939       }
1940     } else {
1941       assert(_finger > finger, "the finger should have moved forward");
1942       // read it again
1943       finger = _finger;
1944     }
1945   }
1946 
1947   return NULL;
1948 }
1949 
1950 #ifndef PRODUCT
1951 class VerifyNoCSetOops {
1952   G1CollectedHeap* _g1h;
1953   const char* _phase;
1954   int _info;
1955 
1956 public:
1957   VerifyNoCSetOops(const char* phase, int info = -1) :
1958     _g1h(G1CollectedHeap::heap()),
1959     _phase(phase),
1960     _info(info)
1961   { }
1962 
1963   void operator()(G1TaskQueueEntry task_entry) const {
1964     if (task_entry.is_array_slice()) {
1965       guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1966       return;
1967     }
1968     guarantee(oopDesc::is_oop(task_entry.obj()),
1969               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1970               p2i(task_entry.obj()), _phase, _info);
1971     guarantee(!_g1h->is_in_cset(task_entry.obj()),
1972               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1973               p2i(task_entry.obj()), _phase, _info);
1974   }
1975 };
1976 
1977 void G1ConcurrentMark::verify_no_cset_oops() {
1978   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1979   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1980     return;
1981   }
1982 
1983   // Verify entries on the global mark stack
1984   _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1985 
1986   // Verify entries on the task queues
1987   for (uint i = 0; i < _max_num_tasks; ++i) {
1988     G1CMTaskQueue* queue = _task_queues->queue(i);
1989     queue->iterate(VerifyNoCSetOops("Queue", i));
1990   }
1991 
1992   // Verify the global finger
1993   HeapWord* global_finger = finger();
1994   if (global_finger != NULL && global_finger < _heap.end()) {
1995     // Since we always iterate over all regions, we might get a NULL HeapRegion
1996     // here.
1997     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1998     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1999               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
2000               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
2001   }
2002 
2003   // Verify the task fingers
2004   assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
2005   for (uint i = 0; i < _num_concurrent_workers; ++i) {
2006     G1CMTask* task = _tasks[i];
2007     HeapWord* task_finger = task->finger();
2008     if (task_finger != NULL && task_finger < _heap.end()) {
2009       // See above note on the global finger verification.
2010       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
2011       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2012                 !task_hr->in_collection_set(),
2013                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2014                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
2015     }
2016   }
2017 }
2018 #endif // PRODUCT
2019 
2020 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
2021   _g1h->g1_rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
2022 }
2023 
2024 void G1ConcurrentMark::print_stats() {
2025   if (!log_is_enabled(Debug, gc, stats)) {
2026     return;
2027   }
2028   log_debug(gc, stats)("---------------------------------------------------------------------");
2029   for (size_t i = 0; i < _num_active_tasks; ++i) {
2030     _tasks[i]->print_stats();
2031     log_debug(gc, stats)("---------------------------------------------------------------------");
2032   }
2033 }
2034 
2035 void G1ConcurrentMark::concurrent_cycle_abort() {
2036   if (!cm_thread()->during_cycle() || _has_aborted) {
2037     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2038     return;
2039   }
2040 
2041   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2042   // concurrent bitmap clearing.
2043   {
2044     GCTraceTime(Debug, gc) debug("Clear Next Bitmap");
2045     clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
2046   }
2047   // Note we cannot clear the previous marking bitmap here
2048   // since VerifyDuringGC verifies the objects marked during
2049   // a full GC against the previous bitmap.
2050 
2051   // Empty mark stack
2052   reset_marking_for_restart();
2053   for (uint i = 0; i < _max_num_tasks; ++i) {
2054     _tasks[i]->clear_region_fields();
2055   }
2056   _first_overflow_barrier_sync.abort();
2057   _second_overflow_barrier_sync.abort();
2058   _has_aborted = true;
2059 
2060   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2061   satb_mq_set.abandon_partial_marking();
2062   // This can be called either during or outside marking, we'll read
2063   // the expected_active value from the SATB queue set.
2064   satb_mq_set.set_active_all_threads(
2065                                  false, /* new active value */
2066                                  satb_mq_set.is_active() /* expected_active */);
2067 }
2068 
2069 static void print_ms_time_info(const char* prefix, const char* name,
2070                                NumberSeq& ns) {
2071   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2072                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2073   if (ns.num() > 0) {
2074     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2075                            prefix, ns.sd(), ns.maximum());
2076   }
2077 }
2078 
2079 void G1ConcurrentMark::print_summary_info() {
2080   Log(gc, marking) log;
2081   if (!log.is_trace()) {
2082     return;
2083   }
2084 
2085   log.trace(" Concurrent marking:");
2086   print_ms_time_info("  ", "init marks", _init_times);
2087   print_ms_time_info("  ", "remarks", _remark_times);
2088   {
2089     print_ms_time_info("     ", "final marks", _remark_mark_times);
2090     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2091 
2092   }
2093   print_ms_time_info("  ", "cleanups", _cleanup_times);
2094   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2095             _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2096   log.trace("  Total stop_world time = %8.2f s.",
2097             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2098   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2099             cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2100 }
2101 
2102 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2103   _concurrent_workers->print_worker_threads_on(st);
2104 }
2105 
2106 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2107   _concurrent_workers->threads_do(tc);
2108 }
2109 
2110 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2111   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2112                p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2113   _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2114   _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2115 }
2116 
2117 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2118   ReferenceProcessor* result = g1h->ref_processor_cm();
2119   assert(result != NULL, "CM reference processor should not be NULL");
2120   return result;
2121 }
2122 
2123 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2124                                G1CMTask* task)
2125   : MetadataVisitingOopIterateClosure(get_cm_oop_closure_ref_processor(g1h)),
2126     _g1h(g1h), _task(task)
2127 { }
2128 
2129 void G1CMTask::setup_for_region(HeapRegion* hr) {
2130   assert(hr != NULL,
2131         "claim_region() should have filtered out NULL regions");
2132   _curr_region  = hr;
2133   _finger       = hr->bottom();
2134   update_region_limit();
2135 }
2136 
2137 void G1CMTask::update_region_limit() {
2138   HeapRegion* hr            = _curr_region;
2139   HeapWord* bottom          = hr->bottom();
2140   HeapWord* limit           = hr->next_top_at_mark_start();
2141 
2142   if (limit == bottom) {
2143     // The region was collected underneath our feet.
2144     // We set the finger to bottom to ensure that the bitmap
2145     // iteration that will follow this will not do anything.
2146     // (this is not a condition that holds when we set the region up,
2147     // as the region is not supposed to be empty in the first place)
2148     _finger = bottom;
2149   } else if (limit >= _region_limit) {
2150     assert(limit >= _finger, "peace of mind");
2151   } else {
2152     assert(limit < _region_limit, "only way to get here");
2153     // This can happen under some pretty unusual circumstances.  An
2154     // evacuation pause empties the region underneath our feet (NTAMS
2155     // at bottom). We then do some allocation in the region (NTAMS
2156     // stays at bottom), followed by the region being used as a GC
2157     // alloc region (NTAMS will move to top() and the objects
2158     // originally below it will be grayed). All objects now marked in
2159     // the region are explicitly grayed, if below the global finger,
2160     // and we do not need in fact to scan anything else. So, we simply
2161     // set _finger to be limit to ensure that the bitmap iteration
2162     // doesn't do anything.
2163     _finger = limit;
2164   }
2165 
2166   _region_limit = limit;
2167 }
2168 
2169 void G1CMTask::giveup_current_region() {
2170   assert(_curr_region != NULL, "invariant");
2171   clear_region_fields();
2172 }
2173 
2174 void G1CMTask::clear_region_fields() {
2175   // Values for these three fields that indicate that we're not
2176   // holding on to a region.
2177   _curr_region   = NULL;
2178   _finger        = NULL;
2179   _region_limit  = NULL;
2180 }
2181 
2182 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2183   if (cm_oop_closure == NULL) {
2184     assert(_cm_oop_closure != NULL, "invariant");
2185   } else {
2186     assert(_cm_oop_closure == NULL, "invariant");
2187   }
2188   _cm_oop_closure = cm_oop_closure;
2189 }
2190 
2191 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2192   guarantee(next_mark_bitmap != NULL, "invariant");
2193   _next_mark_bitmap              = next_mark_bitmap;
2194   clear_region_fields();
2195 
2196   _calls                         = 0;
2197   _elapsed_time_ms               = 0.0;
2198   _termination_time_ms           = 0.0;
2199   _termination_start_time_ms     = 0.0;
2200 
2201   _mark_stats_cache.reset();
2202 }
2203 
2204 bool G1CMTask::should_exit_termination() {
2205   regular_clock_call();
2206   // This is called when we are in the termination protocol. We should
2207   // quit if, for some reason, this task wants to abort or the global
2208   // stack is not empty (this means that we can get work from it).
2209   return !_cm->mark_stack_empty() || has_aborted();
2210 }
2211 
2212 void G1CMTask::reached_limit() {
2213   assert(_words_scanned >= _words_scanned_limit ||
2214          _refs_reached >= _refs_reached_limit ,
2215          "shouldn't have been called otherwise");
2216   regular_clock_call();
2217 }
2218 
2219 void G1CMTask::regular_clock_call() {
2220   if (has_aborted()) {
2221     return;
2222   }
2223 
2224   // First, we need to recalculate the words scanned and refs reached
2225   // limits for the next clock call.
2226   recalculate_limits();
2227 
2228   // During the regular clock call we do the following
2229 
2230   // (1) If an overflow has been flagged, then we abort.
2231   if (_cm->has_overflown()) {
2232     set_has_aborted();
2233     return;
2234   }
2235 
2236   // If we are not concurrent (i.e. we're doing remark) we don't need
2237   // to check anything else. The other steps are only needed during
2238   // the concurrent marking phase.
2239   if (!_cm->concurrent()) {
2240     return;
2241   }
2242 
2243   // (2) If marking has been aborted for Full GC, then we also abort.
2244   if (_cm->has_aborted()) {
2245     set_has_aborted();
2246     return;
2247   }
2248 
2249   double curr_time_ms = os::elapsedVTime() * 1000.0;
2250 
2251   // (4) We check whether we should yield. If we have to, then we abort.
2252   if (SuspendibleThreadSet::should_yield()) {
2253     // We should yield. To do this we abort the task. The caller is
2254     // responsible for yielding.
2255     set_has_aborted();
2256     return;
2257   }
2258 
2259   // (5) We check whether we've reached our time quota. If we have,
2260   // then we abort.
2261   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2262   if (elapsed_time_ms > _time_target_ms) {
2263     set_has_aborted();
2264     _has_timed_out = true;
2265     return;
2266   }
2267 
2268   // (6) Finally, we check whether there are enough completed STAB
2269   // buffers available for processing. If there are, we abort.
2270   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2271   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2272     // we do need to process SATB buffers, we'll abort and restart
2273     // the marking task to do so
2274     set_has_aborted();
2275     return;
2276   }
2277 }
2278 
2279 void G1CMTask::recalculate_limits() {
2280   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2281   _words_scanned_limit      = _real_words_scanned_limit;
2282 
2283   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2284   _refs_reached_limit       = _real_refs_reached_limit;
2285 }
2286 
2287 void G1CMTask::decrease_limits() {
2288   // This is called when we believe that we're going to do an infrequent
2289   // operation which will increase the per byte scanned cost (i.e. move
2290   // entries to/from the global stack). It basically tries to decrease the
2291   // scanning limit so that the clock is called earlier.
2292 
2293   _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2294   _refs_reached_limit  = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2295 }
2296 
2297 void G1CMTask::move_entries_to_global_stack() {
2298   // Local array where we'll store the entries that will be popped
2299   // from the local queue.
2300   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2301 
2302   size_t n = 0;
2303   G1TaskQueueEntry task_entry;
2304   while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2305     buffer[n] = task_entry;
2306     ++n;
2307   }
2308   if (n < G1CMMarkStack::EntriesPerChunk) {
2309     buffer[n] = G1TaskQueueEntry();
2310   }
2311 
2312   if (n > 0) {
2313     if (!_cm->mark_stack_push(buffer)) {
2314       set_has_aborted();
2315     }
2316   }
2317 
2318   // This operation was quite expensive, so decrease the limits.
2319   decrease_limits();
2320 }
2321 
2322 bool G1CMTask::get_entries_from_global_stack() {
2323   // Local array where we'll store the entries that will be popped
2324   // from the global stack.
2325   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2326 
2327   if (!_cm->mark_stack_pop(buffer)) {
2328     return false;
2329   }
2330 
2331   // We did actually pop at least one entry.
2332   for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2333     G1TaskQueueEntry task_entry = buffer[i];
2334     if (task_entry.is_null()) {
2335       break;
2336     }
2337     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()));
2338     bool success = _task_queue->push(task_entry);
2339     // We only call this when the local queue is empty or under a
2340     // given target limit. So, we do not expect this push to fail.
2341     assert(success, "invariant");
2342   }
2343 
2344   // This operation was quite expensive, so decrease the limits
2345   decrease_limits();
2346   return true;
2347 }
2348 
2349 void G1CMTask::drain_local_queue(bool partially) {
2350   if (has_aborted()) {
2351     return;
2352   }
2353 
2354   // Decide what the target size is, depending whether we're going to
2355   // drain it partially (so that other tasks can steal if they run out
2356   // of things to do) or totally (at the very end).
2357   size_t target_size;
2358   if (partially) {
2359     target_size = MIN2((size_t)_task_queue->max_elems()/3, (size_t)GCDrainStackTargetSize);
2360   } else {
2361     target_size = 0;
2362   }
2363 
2364   if (_task_queue->size() > target_size) {
2365     G1TaskQueueEntry entry;
2366     bool ret = _task_queue->pop_local(entry);
2367     while (ret) {
2368       scan_task_entry(entry);
2369       if (_task_queue->size() <= target_size || has_aborted()) {
2370         ret = false;
2371       } else {
2372         ret = _task_queue->pop_local(entry);
2373       }
2374     }
2375   }
2376 }
2377 
2378 void G1CMTask::drain_global_stack(bool partially) {
2379   if (has_aborted()) {
2380     return;
2381   }
2382 
2383   // We have a policy to drain the local queue before we attempt to
2384   // drain the global stack.
2385   assert(partially || _task_queue->size() == 0, "invariant");
2386 
2387   // Decide what the target size is, depending whether we're going to
2388   // drain it partially (so that other tasks can steal if they run out
2389   // of things to do) or totally (at the very end).
2390   // Notice that when draining the global mark stack partially, due to the racyness
2391   // of the mark stack size update we might in fact drop below the target. But,
2392   // this is not a problem.
2393   // In case of total draining, we simply process until the global mark stack is
2394   // totally empty, disregarding the size counter.
2395   if (partially) {
2396     size_t const target_size = _cm->partial_mark_stack_size_target();
2397     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2398       if (get_entries_from_global_stack()) {
2399         drain_local_queue(partially);
2400       }
2401     }
2402   } else {
2403     while (!has_aborted() && get_entries_from_global_stack()) {
2404       drain_local_queue(partially);
2405     }
2406   }
2407 }
2408 
2409 // SATB Queue has several assumptions on whether to call the par or
2410 // non-par versions of the methods. this is why some of the code is
2411 // replicated. We should really get rid of the single-threaded version
2412 // of the code to simplify things.
2413 void G1CMTask::drain_satb_buffers() {
2414   if (has_aborted()) {
2415     return;
2416   }
2417 
2418   // We set this so that the regular clock knows that we're in the
2419   // middle of draining buffers and doesn't set the abort flag when it
2420   // notices that SATB buffers are available for draining. It'd be
2421   // very counter productive if it did that. :-)
2422   _draining_satb_buffers = true;
2423 
2424   G1CMSATBBufferClosure satb_cl(this, _g1h);
2425   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2426 
2427   // This keeps claiming and applying the closure to completed buffers
2428   // until we run out of buffers or we need to abort.
2429   while (!has_aborted() &&
2430          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2431     regular_clock_call();
2432   }
2433 
2434   _draining_satb_buffers = false;
2435 
2436   assert(has_aborted() ||
2437          _cm->concurrent() ||
2438          satb_mq_set.completed_buffers_num() == 0, "invariant");
2439 
2440   // again, this was a potentially expensive operation, decrease the
2441   // limits to get the regular clock call early
2442   decrease_limits();
2443 }
2444 
2445 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2446   _mark_stats_cache.reset(region_idx);
2447 }
2448 
2449 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2450   return _mark_stats_cache.evict_all();
2451 }
2452 
2453 void G1CMTask::print_stats() {
2454   log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2455   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2456                        _elapsed_time_ms, _termination_time_ms);
2457   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2458                        _step_times_ms.num(),
2459                        _step_times_ms.avg(),
2460                        _step_times_ms.sd(),
2461                        _step_times_ms.maximum(),
2462                        _step_times_ms.sum());
2463   size_t const hits = _mark_stats_cache.hits();
2464   size_t const misses = _mark_stats_cache.misses();
2465   log_debug(gc, stats)("  Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2466                        hits, misses, percent_of(hits, hits + misses));
2467 }
2468 
2469 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry) {
2470   return _task_queues->steal(worker_id, hash_seed, task_entry);
2471 }
2472 
2473 /*****************************************************************************
2474 
2475     The do_marking_step(time_target_ms, ...) method is the building
2476     block of the parallel marking framework. It can be called in parallel
2477     with other invocations of do_marking_step() on different tasks
2478     (but only one per task, obviously) and concurrently with the
2479     mutator threads, or during remark, hence it eliminates the need
2480     for two versions of the code. When called during remark, it will
2481     pick up from where the task left off during the concurrent marking
2482     phase. Interestingly, tasks are also claimable during evacuation
2483     pauses too, since do_marking_step() ensures that it aborts before
2484     it needs to yield.
2485 
2486     The data structures that it uses to do marking work are the
2487     following:
2488 
2489       (1) Marking Bitmap. If there are gray objects that appear only
2490       on the bitmap (this happens either when dealing with an overflow
2491       or when the initial marking phase has simply marked the roots
2492       and didn't push them on the stack), then tasks claim heap
2493       regions whose bitmap they then scan to find gray objects. A
2494       global finger indicates where the end of the last claimed region
2495       is. A local finger indicates how far into the region a task has
2496       scanned. The two fingers are used to determine how to gray an
2497       object (i.e. whether simply marking it is OK, as it will be
2498       visited by a task in the future, or whether it needs to be also
2499       pushed on a stack).
2500 
2501       (2) Local Queue. The local queue of the task which is accessed
2502       reasonably efficiently by the task. Other tasks can steal from
2503       it when they run out of work. Throughout the marking phase, a
2504       task attempts to keep its local queue short but not totally
2505       empty, so that entries are available for stealing by other
2506       tasks. Only when there is no more work, a task will totally
2507       drain its local queue.
2508 
2509       (3) Global Mark Stack. This handles local queue overflow. During
2510       marking only sets of entries are moved between it and the local
2511       queues, as access to it requires a mutex and more fine-grain
2512       interaction with it which might cause contention. If it
2513       overflows, then the marking phase should restart and iterate
2514       over the bitmap to identify gray objects. Throughout the marking
2515       phase, tasks attempt to keep the global mark stack at a small
2516       length but not totally empty, so that entries are available for
2517       popping by other tasks. Only when there is no more work, tasks
2518       will totally drain the global mark stack.
2519 
2520       (4) SATB Buffer Queue. This is where completed SATB buffers are
2521       made available. Buffers are regularly removed from this queue
2522       and scanned for roots, so that the queue doesn't get too
2523       long. During remark, all completed buffers are processed, as
2524       well as the filled in parts of any uncompleted buffers.
2525 
2526     The do_marking_step() method tries to abort when the time target
2527     has been reached. There are a few other cases when the
2528     do_marking_step() method also aborts:
2529 
2530       (1) When the marking phase has been aborted (after a Full GC).
2531 
2532       (2) When a global overflow (on the global stack) has been
2533       triggered. Before the task aborts, it will actually sync up with
2534       the other tasks to ensure that all the marking data structures
2535       (local queues, stacks, fingers etc.)  are re-initialized so that
2536       when do_marking_step() completes, the marking phase can
2537       immediately restart.
2538 
2539       (3) When enough completed SATB buffers are available. The
2540       do_marking_step() method only tries to drain SATB buffers right
2541       at the beginning. So, if enough buffers are available, the
2542       marking step aborts and the SATB buffers are processed at
2543       the beginning of the next invocation.
2544 
2545       (4) To yield. when we have to yield then we abort and yield
2546       right at the end of do_marking_step(). This saves us from a lot
2547       of hassle as, by yielding we might allow a Full GC. If this
2548       happens then objects will be compacted underneath our feet, the
2549       heap might shrink, etc. We save checking for this by just
2550       aborting and doing the yield right at the end.
2551 
2552     From the above it follows that the do_marking_step() method should
2553     be called in a loop (or, otherwise, regularly) until it completes.
2554 
2555     If a marking step completes without its has_aborted() flag being
2556     true, it means it has completed the current marking phase (and
2557     also all other marking tasks have done so and have all synced up).
2558 
2559     A method called regular_clock_call() is invoked "regularly" (in
2560     sub ms intervals) throughout marking. It is this clock method that
2561     checks all the abort conditions which were mentioned above and
2562     decides when the task should abort. A work-based scheme is used to
2563     trigger this clock method: when the number of object words the
2564     marking phase has scanned or the number of references the marking
2565     phase has visited reach a given limit. Additional invocations to
2566     the method clock have been planted in a few other strategic places
2567     too. The initial reason for the clock method was to avoid calling
2568     vtime too regularly, as it is quite expensive. So, once it was in
2569     place, it was natural to piggy-back all the other conditions on it
2570     too and not constantly check them throughout the code.
2571 
2572     If do_termination is true then do_marking_step will enter its
2573     termination protocol.
2574 
2575     The value of is_serial must be true when do_marking_step is being
2576     called serially (i.e. by the VMThread) and do_marking_step should
2577     skip any synchronization in the termination and overflow code.
2578     Examples include the serial remark code and the serial reference
2579     processing closures.
2580 
2581     The value of is_serial must be false when do_marking_step is
2582     being called by any of the worker threads in a work gang.
2583     Examples include the concurrent marking code (CMMarkingTask),
2584     the MT remark code, and the MT reference processing closures.
2585 
2586  *****************************************************************************/
2587 
2588 void G1CMTask::do_marking_step(double time_target_ms,
2589                                bool do_termination,
2590                                bool is_serial) {
2591   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2592 
2593   _start_time_ms = os::elapsedVTime() * 1000.0;
2594 
2595   // If do_stealing is true then do_marking_step will attempt to
2596   // steal work from the other G1CMTasks. It only makes sense to
2597   // enable stealing when the termination protocol is enabled
2598   // and do_marking_step() is not being called serially.
2599   bool do_stealing = do_termination && !is_serial;
2600 
2601   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2602   _time_target_ms = time_target_ms - diff_prediction_ms;
2603 
2604   // set up the variables that are used in the work-based scheme to
2605   // call the regular clock method
2606   _words_scanned = 0;
2607   _refs_reached  = 0;
2608   recalculate_limits();
2609 
2610   // clear all flags
2611   clear_has_aborted();
2612   _has_timed_out = false;
2613   _draining_satb_buffers = false;
2614 
2615   ++_calls;
2616 
2617   // Set up the bitmap and oop closures. Anything that uses them is
2618   // eventually called from this method, so it is OK to allocate these
2619   // statically.
2620   G1CMBitMapClosure bitmap_closure(this, _cm);
2621   G1CMOopClosure cm_oop_closure(_g1h, this);
2622   set_cm_oop_closure(&cm_oop_closure);
2623 
2624   if (_cm->has_overflown()) {
2625     // This can happen if the mark stack overflows during a GC pause
2626     // and this task, after a yield point, restarts. We have to abort
2627     // as we need to get into the overflow protocol which happens
2628     // right at the end of this task.
2629     set_has_aborted();
2630   }
2631 
2632   // First drain any available SATB buffers. After this, we will not
2633   // look at SATB buffers before the next invocation of this method.
2634   // If enough completed SATB buffers are queued up, the regular clock
2635   // will abort this task so that it restarts.
2636   drain_satb_buffers();
2637   // ...then partially drain the local queue and the global stack
2638   drain_local_queue(true);
2639   drain_global_stack(true);
2640 
2641   do {
2642     if (!has_aborted() && _curr_region != NULL) {
2643       // This means that we're already holding on to a region.
2644       assert(_finger != NULL, "if region is not NULL, then the finger "
2645              "should not be NULL either");
2646 
2647       // We might have restarted this task after an evacuation pause
2648       // which might have evacuated the region we're holding on to
2649       // underneath our feet. Let's read its limit again to make sure
2650       // that we do not iterate over a region of the heap that
2651       // contains garbage (update_region_limit() will also move
2652       // _finger to the start of the region if it is found empty).
2653       update_region_limit();
2654       // We will start from _finger not from the start of the region,
2655       // as we might be restarting this task after aborting half-way
2656       // through scanning this region. In this case, _finger points to
2657       // the address where we last found a marked object. If this is a
2658       // fresh region, _finger points to start().
2659       MemRegion mr = MemRegion(_finger, _region_limit);
2660 
2661       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2662              "humongous regions should go around loop once only");
2663 
2664       // Some special cases:
2665       // If the memory region is empty, we can just give up the region.
2666       // If the current region is humongous then we only need to check
2667       // the bitmap for the bit associated with the start of the object,
2668       // scan the object if it's live, and give up the region.
2669       // Otherwise, let's iterate over the bitmap of the part of the region
2670       // that is left.
2671       // If the iteration is successful, give up the region.
2672       if (mr.is_empty()) {
2673         giveup_current_region();
2674         regular_clock_call();
2675       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2676         if (_next_mark_bitmap->is_marked(mr.start())) {
2677           // The object is marked - apply the closure
2678           bitmap_closure.do_addr(mr.start());
2679         }
2680         // Even if this task aborted while scanning the humongous object
2681         // we can (and should) give up the current region.
2682         giveup_current_region();
2683         regular_clock_call();
2684       } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2685         giveup_current_region();
2686         regular_clock_call();
2687       } else {
2688         assert(has_aborted(), "currently the only way to do so");
2689         // The only way to abort the bitmap iteration is to return
2690         // false from the do_bit() method. However, inside the
2691         // do_bit() method we move the _finger to point to the
2692         // object currently being looked at. So, if we bail out, we
2693         // have definitely set _finger to something non-null.
2694         assert(_finger != NULL, "invariant");
2695 
2696         // Region iteration was actually aborted. So now _finger
2697         // points to the address of the object we last scanned. If we
2698         // leave it there, when we restart this task, we will rescan
2699         // the object. It is easy to avoid this. We move the finger by
2700         // enough to point to the next possible object header.
2701         assert(_finger < _region_limit, "invariant");
2702         HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2703         // Check if bitmap iteration was aborted while scanning the last object
2704         if (new_finger >= _region_limit) {
2705           giveup_current_region();
2706         } else {
2707           move_finger_to(new_finger);
2708         }
2709       }
2710     }
2711     // At this point we have either completed iterating over the
2712     // region we were holding on to, or we have aborted.
2713 
2714     // We then partially drain the local queue and the global stack.
2715     // (Do we really need this?)
2716     drain_local_queue(true);
2717     drain_global_stack(true);
2718 
2719     // Read the note on the claim_region() method on why it might
2720     // return NULL with potentially more regions available for
2721     // claiming and why we have to check out_of_regions() to determine
2722     // whether we're done or not.
2723     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2724       // We are going to try to claim a new region. We should have
2725       // given up on the previous one.
2726       // Separated the asserts so that we know which one fires.
2727       assert(_curr_region  == NULL, "invariant");
2728       assert(_finger       == NULL, "invariant");
2729       assert(_region_limit == NULL, "invariant");
2730       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2731       if (claimed_region != NULL) {
2732         // Yes, we managed to claim one
2733         setup_for_region(claimed_region);
2734         assert(_curr_region == claimed_region, "invariant");
2735       }
2736       // It is important to call the regular clock here. It might take
2737       // a while to claim a region if, for example, we hit a large
2738       // block of empty regions. So we need to call the regular clock
2739       // method once round the loop to make sure it's called
2740       // frequently enough.
2741       regular_clock_call();
2742     }
2743 
2744     if (!has_aborted() && _curr_region == NULL) {
2745       assert(_cm->out_of_regions(),
2746              "at this point we should be out of regions");
2747     }
2748   } while ( _curr_region != NULL && !has_aborted());
2749 
2750   if (!has_aborted()) {
2751     // We cannot check whether the global stack is empty, since other
2752     // tasks might be pushing objects to it concurrently.
2753     assert(_cm->out_of_regions(),
2754            "at this point we should be out of regions");
2755     // Try to reduce the number of available SATB buffers so that
2756     // remark has less work to do.
2757     drain_satb_buffers();
2758   }
2759 
2760   // Since we've done everything else, we can now totally drain the
2761   // local queue and global stack.
2762   drain_local_queue(false);
2763   drain_global_stack(false);
2764 
2765   // Attempt at work stealing from other task's queues.
2766   if (do_stealing && !has_aborted()) {
2767     // We have not aborted. This means that we have finished all that
2768     // we could. Let's try to do some stealing...
2769 
2770     // We cannot check whether the global stack is empty, since other
2771     // tasks might be pushing objects to it concurrently.
2772     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2773            "only way to reach here");
2774     while (!has_aborted()) {
2775       G1TaskQueueEntry entry;
2776       if (_cm->try_stealing(_worker_id, &_hash_seed, entry)) {
2777         scan_task_entry(entry);
2778 
2779         // And since we're towards the end, let's totally drain the
2780         // local queue and global stack.
2781         drain_local_queue(false);
2782         drain_global_stack(false);
2783       } else {
2784         break;
2785       }
2786     }
2787   }
2788 
2789   // We still haven't aborted. Now, let's try to get into the
2790   // termination protocol.
2791   if (do_termination && !has_aborted()) {
2792     // We cannot check whether the global stack is empty, since other
2793     // tasks might be concurrently pushing objects on it.
2794     // Separated the asserts so that we know which one fires.
2795     assert(_cm->out_of_regions(), "only way to reach here");
2796     assert(_task_queue->size() == 0, "only way to reach here");
2797     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2798 
2799     // The G1CMTask class also extends the TerminatorTerminator class,
2800     // hence its should_exit_termination() method will also decide
2801     // whether to exit the termination protocol or not.
2802     bool finished = (is_serial ||
2803                      _cm->terminator()->offer_termination(this));
2804     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2805     _termination_time_ms +=
2806       termination_end_time_ms - _termination_start_time_ms;
2807 
2808     if (finished) {
2809       // We're all done.
2810 
2811       // We can now guarantee that the global stack is empty, since
2812       // all other tasks have finished. We separated the guarantees so
2813       // that, if a condition is false, we can immediately find out
2814       // which one.
2815       guarantee(_cm->out_of_regions(), "only way to reach here");
2816       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2817       guarantee(_task_queue->size() == 0, "only way to reach here");
2818       guarantee(!_cm->has_overflown(), "only way to reach here");
2819     } else {
2820       // Apparently there's more work to do. Let's abort this task. It
2821       // will restart it and we can hopefully find more things to do.
2822       set_has_aborted();
2823     }
2824   }
2825 
2826   // Mainly for debugging purposes to make sure that a pointer to the
2827   // closure which was statically allocated in this frame doesn't
2828   // escape it by accident.
2829   set_cm_oop_closure(NULL);
2830   double end_time_ms = os::elapsedVTime() * 1000.0;
2831   double elapsed_time_ms = end_time_ms - _start_time_ms;
2832   // Update the step history.
2833   _step_times_ms.add(elapsed_time_ms);
2834 
2835   if (has_aborted()) {
2836     // The task was aborted for some reason.
2837     if (_has_timed_out) {
2838       double diff_ms = elapsed_time_ms - _time_target_ms;
2839       // Keep statistics of how well we did with respect to hitting
2840       // our target only if we actually timed out (if we aborted for
2841       // other reasons, then the results might get skewed).
2842       _marking_step_diffs_ms.add(diff_ms);
2843     }
2844 
2845     if (_cm->has_overflown()) {
2846       // This is the interesting one. We aborted because a global
2847       // overflow was raised. This means we have to restart the
2848       // marking phase and start iterating over regions. However, in
2849       // order to do this we have to make sure that all tasks stop
2850       // what they are doing and re-initialize in a safe manner. We
2851       // will achieve this with the use of two barrier sync points.
2852 
2853       if (!is_serial) {
2854         // We only need to enter the sync barrier if being called
2855         // from a parallel context
2856         _cm->enter_first_sync_barrier(_worker_id);
2857 
2858         // When we exit this sync barrier we know that all tasks have
2859         // stopped doing marking work. So, it's now safe to
2860         // re-initialize our data structures.
2861       }
2862 
2863       clear_region_fields();
2864       flush_mark_stats_cache();
2865 
2866       if (!is_serial) {
2867         // If we're executing the concurrent phase of marking, reset the marking
2868         // state; otherwise the marking state is reset after reference processing,
2869         // during the remark pause.
2870         // If we reset here as a result of an overflow during the remark we will
2871         // see assertion failures from any subsequent set_concurrency_and_phase()
2872         // calls.
2873         if (_cm->concurrent() && _worker_id == 0) {
2874           // Worker 0 is responsible for clearing the global data structures because
2875           // of an overflow. During STW we should not clear the overflow flag (in
2876           // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2877           // method to abort the pause and restart concurrent marking.
2878           _cm->reset_marking_for_restart();
2879 
2880           log_info(gc, marking)("Concurrent Mark reset for overflow");
2881         }
2882 
2883         // ...and enter the second barrier.
2884         _cm->enter_second_sync_barrier(_worker_id);
2885       }
2886       // At this point, if we're during the concurrent phase of
2887       // marking, everything has been re-initialized and we're
2888       // ready to restart.
2889     }
2890   }
2891 }
2892 
2893 G1CMTask::G1CMTask(uint worker_id,
2894                    G1ConcurrentMark* cm,
2895                    G1CMTaskQueue* task_queue,
2896                    G1RegionMarkStats* mark_stats,
2897                    uint max_regions) :
2898   _objArray_processor(this),
2899   _worker_id(worker_id),
2900   _g1h(G1CollectedHeap::heap()),
2901   _cm(cm),
2902   _next_mark_bitmap(NULL),
2903   _task_queue(task_queue),
2904   _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize),
2905   _calls(0),
2906   _time_target_ms(0.0),
2907   _start_time_ms(0.0),
2908   _cm_oop_closure(NULL),
2909   _curr_region(NULL),
2910   _finger(NULL),
2911   _region_limit(NULL),
2912   _words_scanned(0),
2913   _words_scanned_limit(0),
2914   _real_words_scanned_limit(0),
2915   _refs_reached(0),
2916   _refs_reached_limit(0),
2917   _real_refs_reached_limit(0),
2918   _hash_seed(17),
2919   _has_aborted(false),
2920   _has_timed_out(false),
2921   _draining_satb_buffers(false),
2922   _step_times_ms(),
2923   _elapsed_time_ms(0.0),
2924   _termination_time_ms(0.0),
2925   _termination_start_time_ms(0.0),
2926   _marking_step_diffs_ms()
2927 {
2928   guarantee(task_queue != NULL, "invariant");
2929 
2930   _marking_step_diffs_ms.add(0.5);
2931 }
2932 
2933 // These are formatting macros that are used below to ensure
2934 // consistent formatting. The *_H_* versions are used to format the
2935 // header for a particular value and they should be kept consistent
2936 // with the corresponding macro. Also note that most of the macros add
2937 // the necessary white space (as a prefix) which makes them a bit
2938 // easier to compose.
2939 
2940 // All the output lines are prefixed with this string to be able to
2941 // identify them easily in a large log file.
2942 #define G1PPRL_LINE_PREFIX            "###"
2943 
2944 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2945 #ifdef _LP64
2946 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2947 #else // _LP64
2948 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2949 #endif // _LP64
2950 
2951 // For per-region info
2952 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2953 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2954 #define G1PPRL_STATE_FORMAT           "   %-5s"
2955 #define G1PPRL_STATE_H_FORMAT         "   %5s"
2956 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2957 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2958 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2959 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2960 
2961 // For summary info
2962 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2963 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2964 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2965 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2966 
2967 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2968   _total_used_bytes(0), _total_capacity_bytes(0),
2969   _total_prev_live_bytes(0), _total_next_live_bytes(0),
2970   _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2971 {
2972   if (!log_is_enabled(Trace, gc, liveness)) {
2973     return;
2974   }
2975 
2976   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2977   MemRegion g1_reserved = g1h->g1_reserved();
2978   double now = os::elapsedTime();
2979 
2980   // Print the header of the output.
2981   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2982   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2983                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2984                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2985                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2986                           HeapRegion::GrainBytes);
2987   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2988   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2989                           G1PPRL_TYPE_H_FORMAT
2990                           G1PPRL_ADDR_BASE_H_FORMAT
2991                           G1PPRL_BYTE_H_FORMAT
2992                           G1PPRL_BYTE_H_FORMAT
2993                           G1PPRL_BYTE_H_FORMAT
2994                           G1PPRL_DOUBLE_H_FORMAT
2995                           G1PPRL_BYTE_H_FORMAT
2996                           G1PPRL_STATE_H_FORMAT
2997                           G1PPRL_BYTE_H_FORMAT,
2998                           "type", "address-range",
2999                           "used", "prev-live", "next-live", "gc-eff",
3000                           "remset", "state", "code-roots");
3001   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3002                           G1PPRL_TYPE_H_FORMAT
3003                           G1PPRL_ADDR_BASE_H_FORMAT
3004                           G1PPRL_BYTE_H_FORMAT
3005                           G1PPRL_BYTE_H_FORMAT
3006                           G1PPRL_BYTE_H_FORMAT
3007                           G1PPRL_DOUBLE_H_FORMAT
3008                           G1PPRL_BYTE_H_FORMAT
3009                           G1PPRL_STATE_H_FORMAT
3010                           G1PPRL_BYTE_H_FORMAT,
3011                           "", "",
3012                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3013                           "(bytes)", "", "(bytes)");
3014 }
3015 
3016 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
3017   if (!log_is_enabled(Trace, gc, liveness)) {
3018     return false;
3019   }
3020 
3021   const char* type       = r->get_type_str();
3022   HeapWord* bottom       = r->bottom();
3023   HeapWord* end          = r->end();
3024   size_t capacity_bytes  = r->capacity();
3025   size_t used_bytes      = r->used();
3026   size_t prev_live_bytes = r->live_bytes();
3027   size_t next_live_bytes = r->next_live_bytes();
3028   double gc_eff          = r->gc_efficiency();
3029   size_t remset_bytes    = r->rem_set()->mem_size();
3030   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3031   const char* remset_type = r->rem_set()->get_short_state_str();
3032 
3033   _total_used_bytes      += used_bytes;
3034   _total_capacity_bytes  += capacity_bytes;
3035   _total_prev_live_bytes += prev_live_bytes;
3036   _total_next_live_bytes += next_live_bytes;
3037   _total_remset_bytes    += remset_bytes;
3038   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3039 
3040   // Print a line for this particular region.
3041   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3042                           G1PPRL_TYPE_FORMAT
3043                           G1PPRL_ADDR_BASE_FORMAT
3044                           G1PPRL_BYTE_FORMAT
3045                           G1PPRL_BYTE_FORMAT
3046                           G1PPRL_BYTE_FORMAT
3047                           G1PPRL_DOUBLE_FORMAT
3048                           G1PPRL_BYTE_FORMAT
3049                           G1PPRL_STATE_FORMAT
3050                           G1PPRL_BYTE_FORMAT,
3051                           type, p2i(bottom), p2i(end),
3052                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3053                           remset_bytes, remset_type, strong_code_roots_bytes);
3054 
3055   return false;
3056 }
3057 
3058 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3059   if (!log_is_enabled(Trace, gc, liveness)) {
3060     return;
3061   }
3062 
3063   // add static memory usages to remembered set sizes
3064   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3065   // Print the footer of the output.
3066   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3067   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3068                          " SUMMARY"
3069                          G1PPRL_SUM_MB_FORMAT("capacity")
3070                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3071                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3072                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3073                          G1PPRL_SUM_MB_FORMAT("remset")
3074                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3075                          bytes_to_mb(_total_capacity_bytes),
3076                          bytes_to_mb(_total_used_bytes),
3077                          percent_of(_total_used_bytes, _total_capacity_bytes),
3078                          bytes_to_mb(_total_prev_live_bytes),
3079                          percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3080                          bytes_to_mb(_total_next_live_bytes),
3081                          percent_of(_total_next_live_bytes, _total_capacity_bytes),
3082                          bytes_to_mb(_total_remset_bytes),
3083                          bytes_to_mb(_total_strong_code_roots_bytes));
3084 }