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/classLoaderDataGraph.hpp"
  27 #include "code/codeCache.hpp"
  28 #include "gc/g1/g1BarrierSet.hpp"
  29 #include "gc/g1/g1CollectedHeap.inline.hpp"
  30 #include "gc/g1/g1CollectorState.hpp"
  31 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  32 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp"
  33 #include "gc/g1/g1HeapVerifier.hpp"
  34 #include "gc/g1/g1OopClosures.inline.hpp"
  35 #include "gc/g1/g1Policy.hpp"
  36 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
  37 #include "gc/g1/g1StringDedup.hpp"
  38 #include "gc/g1/g1ThreadLocalData.hpp"
  39 #include "gc/g1/heapRegion.inline.hpp"
  40 #include "gc/g1/heapRegionRemSet.hpp"
  41 #include "gc/g1/heapRegionSet.inline.hpp"
  42 #include "gc/shared/adaptiveSizePolicy.hpp"
  43 #include "gc/shared/gcId.hpp"
  44 #include "gc/shared/gcTimer.hpp"
  45 #include "gc/shared/gcTrace.hpp"
  46 #include "gc/shared/gcTraceTime.inline.hpp"
  47 #include "gc/shared/genOopClosures.inline.hpp"
  48 #include "gc/shared/referencePolicy.hpp"
  49 #include "gc/shared/strongRootsScope.hpp"
  50 #include "gc/shared/suspendibleThreadSet.hpp"
  51 #include "gc/shared/taskqueue.inline.hpp"
  52 #include "gc/shared/vmGCOperations.hpp"
  53 #include "gc/shared/weakProcessor.inline.hpp"
  54 #include "include/jvm.h"
  55 #include "logging/log.hpp"
  56 #include "memory/allocation.hpp"
  57 #include "memory/resourceArea.hpp"
  58 #include "oops/access.inline.hpp"
  59 #include "oops/oop.inline.hpp"
  60 #include "runtime/atomic.hpp"
  61 #include "runtime/handles.inline.hpp"
  62 #include "runtime/java.hpp"
  63 #include "runtime/prefetch.inline.hpp"
  64 #include "services/memTracker.hpp"
  65 #include "utilities/align.hpp"
  66 #include "utilities/growableArray.hpp"
  67 
  68 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
  69   assert(addr < _cm->finger(), "invariant");
  70   assert(addr >= _task->finger(), "invariant");
  71 
  72   // We move that task's local finger along.
  73   _task->move_finger_to(addr);
  74 
  75   _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
  76   // we only partially drain the local queue and global stack
  77   _task->drain_local_queue(true);
  78   _task->drain_global_stack(true);
  79 
  80   // if the has_aborted flag has been raised, we need to bail out of
  81   // the iteration
  82   return !_task->has_aborted();
  83 }
  84 
  85 G1CMMarkStack::G1CMMarkStack() :
  86   _max_chunk_capacity(0),
  87   _base(NULL),
  88   _chunk_capacity(0) {
  89   set_empty();
  90 }
  91 
  92 bool G1CMMarkStack::resize(size_t new_capacity) {
  93   assert(is_empty(), "Only resize when stack is empty.");
  94   assert(new_capacity <= _max_chunk_capacity,
  95          "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
  96 
  97   TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
  98 
  99   if (new_base == NULL) {
 100     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));
 101     return false;
 102   }
 103   // Release old mapping.
 104   if (_base != NULL) {
 105     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 106   }
 107 
 108   _base = new_base;
 109   _chunk_capacity = new_capacity;
 110   set_empty();
 111 
 112   return true;
 113 }
 114 
 115 size_t G1CMMarkStack::capacity_alignment() {
 116   return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
 117 }
 118 
 119 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
 120   guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
 121 
 122   size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
 123 
 124   _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 125   size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 126 
 127   guarantee(initial_chunk_capacity <= _max_chunk_capacity,
 128             "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
 129             _max_chunk_capacity,
 130             initial_chunk_capacity);
 131 
 132   log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
 133                 initial_chunk_capacity, _max_chunk_capacity);
 134 
 135   return resize(initial_chunk_capacity);
 136 }
 137 
 138 void G1CMMarkStack::expand() {
 139   if (_chunk_capacity == _max_chunk_capacity) {
 140     log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
 141     return;
 142   }
 143   size_t old_capacity = _chunk_capacity;
 144   // Double capacity if possible
 145   size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
 146 
 147   if (resize(new_capacity)) {
 148     log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 149                   old_capacity, new_capacity);
 150   } else {
 151     log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 152                     old_capacity, new_capacity);
 153   }
 154 }
 155 
 156 G1CMMarkStack::~G1CMMarkStack() {
 157   if (_base != NULL) {
 158     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 159   }
 160 }
 161 
 162 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
 163   elem->next = *list;
 164   *list = elem;
 165 }
 166 
 167 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
 168   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 169   add_chunk_to_list(&_chunk_list, elem);
 170   _chunks_in_chunk_list++;
 171 }
 172 
 173 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
 174   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 175   add_chunk_to_list(&_free_list, elem);
 176 }
 177 
 178 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
 179   TaskQueueEntryChunk* result = *list;
 180   if (result != NULL) {
 181     *list = (*list)->next;
 182   }
 183   return result;
 184 }
 185 
 186 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
 187   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 188   TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
 189   if (result != NULL) {
 190     _chunks_in_chunk_list--;
 191   }
 192   return result;
 193 }
 194 
 195 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
 196   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 197   return remove_chunk_from_list(&_free_list);
 198 }
 199 
 200 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
 201   // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
 202   // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
 203   // wraparound of _hwm.
 204   if (_hwm >= _chunk_capacity) {
 205     return NULL;
 206   }
 207 
 208   size_t cur_idx = Atomic::add(1u, &_hwm) - 1;
 209   if (cur_idx >= _chunk_capacity) {
 210     return NULL;
 211   }
 212 
 213   TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
 214   result->next = NULL;
 215   return result;
 216 }
 217 
 218 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
 219   // Get a new chunk.
 220   TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
 221 
 222   if (new_chunk == NULL) {
 223     // Did not get a chunk from the free list. Allocate from backing memory.
 224     new_chunk = allocate_new_chunk();
 225 
 226     if (new_chunk == NULL) {
 227       return false;
 228     }
 229   }
 230 
 231   Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 232 
 233   add_chunk_to_chunk_list(new_chunk);
 234 
 235   return true;
 236 }
 237 
 238 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
 239   TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
 240 
 241   if (cur == NULL) {
 242     return false;
 243   }
 244 
 245   Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 246 
 247   add_chunk_to_free_list(cur);
 248   return true;
 249 }
 250 
 251 void G1CMMarkStack::set_empty() {
 252   _chunks_in_chunk_list = 0;
 253   _hwm = 0;
 254   _chunk_list = NULL;
 255   _free_list = NULL;
 256 }
 257 
 258 G1CMRootRegions::G1CMRootRegions() :
 259   _survivors(NULL), _cm(NULL), _scan_in_progress(false),
 260   _should_abort(false), _claimed_survivor_index(0) { }
 261 
 262 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
 263   _survivors = survivors;
 264   _cm = cm;
 265 }
 266 
 267 void G1CMRootRegions::prepare_for_scan() {
 268   assert(!scan_in_progress(), "pre-condition");
 269 
 270   // Currently, only survivors can be root regions.
 271   _claimed_survivor_index = 0;
 272   _scan_in_progress = _survivors->regions()->is_nonempty();
 273   _should_abort = false;
 274 }
 275 
 276 HeapRegion* G1CMRootRegions::claim_next() {
 277   if (_should_abort) {
 278     // If someone has set the should_abort flag, we return NULL to
 279     // force the caller to bail out of their loop.
 280     return NULL;
 281   }
 282 
 283   // Currently, only survivors can be root regions.
 284   const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
 285 
 286   int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
 287   if (claimed_index < survivor_regions->length()) {
 288     return survivor_regions->at(claimed_index);
 289   }
 290   return NULL;
 291 }
 292 
 293 uint G1CMRootRegions::num_root_regions() const {
 294   return (uint)_survivors->regions()->length();
 295 }
 296 
 297 void G1CMRootRegions::notify_scan_done() {
 298   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 299   _scan_in_progress = false;
 300   RootRegionScan_lock->notify_all();
 301 }
 302 
 303 void G1CMRootRegions::cancel_scan() {
 304   notify_scan_done();
 305 }
 306 
 307 void G1CMRootRegions::scan_finished() {
 308   assert(scan_in_progress(), "pre-condition");
 309 
 310   // Currently, only survivors can be root regions.
 311   if (!_should_abort) {
 312     assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
 313     assert((uint)_claimed_survivor_index >= _survivors->length(),
 314            "we should have claimed all survivors, claimed index = %u, length = %u",
 315            (uint)_claimed_survivor_index, _survivors->length());
 316   }
 317 
 318   notify_scan_done();
 319 }
 320 
 321 bool G1CMRootRegions::wait_until_scan_finished() {
 322   if (!scan_in_progress()) {
 323     return false;
 324   }
 325 
 326   {
 327     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 328     while (scan_in_progress()) {
 329       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 330     }
 331   }
 332   return true;
 333 }
 334 
 335 // Returns the maximum number of workers to be used in a concurrent
 336 // phase based on the number of GC workers being used in a STW
 337 // phase.
 338 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
 339   return MAX2((num_gc_workers + 2) / 4, 1U);
 340 }
 341 
 342 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
 343                                    G1RegionToSpaceMapper* prev_bitmap_storage,
 344                                    G1RegionToSpaceMapper* next_bitmap_storage) :
 345   // _cm_thread set inside the constructor
 346   _g1h(g1h),
 347   _completed_initialization(false),
 348 
 349   _mark_bitmap_1(),
 350   _mark_bitmap_2(),
 351   _prev_mark_bitmap(&_mark_bitmap_1),
 352   _next_mark_bitmap(&_mark_bitmap_2),
 353 
 354   _heap(_g1h->reserved_region()),
 355 
 356   _root_regions(),
 357 
 358   _global_mark_stack(),
 359 
 360   // _finger set in set_non_marking_state
 361 
 362   _worker_id_offset(DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
 363   _max_num_tasks(ParallelGCThreads),
 364   // _num_active_tasks set in set_non_marking_state()
 365   // _tasks set inside the constructor
 366 
 367   _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
 368   _terminator(ParallelTaskTerminator((int) _max_num_tasks, _task_queues)),
 369 
 370   _first_overflow_barrier_sync(),
 371   _second_overflow_barrier_sync(),
 372 
 373   _has_overflown(false),
 374   _concurrent(false),
 375   _has_aborted(false),
 376   _restart_for_overflow(false),
 377   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 378   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 379 
 380   // _verbose_level set below
 381 
 382   _init_times(),
 383   _remark_times(),
 384   _remark_mark_times(),
 385   _remark_weak_ref_times(),
 386   _cleanup_times(),
 387   _total_cleanup_time(0.0),
 388 
 389   _accum_task_vtime(NULL),
 390 
 391   _concurrent_workers(NULL),
 392   _num_concurrent_workers(0),
 393   _max_concurrent_workers(0),
 394 
 395   _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)),
 396   _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC))
 397 {
 398   _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 399   _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
 400 
 401   // Create & start ConcurrentMark thread.
 402   _cm_thread = new G1ConcurrentMarkThread(this);
 403   if (_cm_thread->osthread() == NULL) {
 404     vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 405   }
 406 
 407   assert(CGC_lock != NULL, "CGC_lock must be initialized");
 408 
 409   _root_regions.init(_g1h->survivor(), this);
 410 
 411   if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
 412     // Calculate the number of concurrent worker threads by scaling
 413     // the number of parallel GC threads.
 414     uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
 415     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 416   }
 417 
 418   assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
 419   if (ConcGCThreads > ParallelGCThreads) {
 420     log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
 421                     ConcGCThreads, ParallelGCThreads);
 422     return;
 423   }
 424 
 425   log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset);
 426   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 427 
 428   _num_concurrent_workers = ConcGCThreads;
 429   _max_concurrent_workers = _num_concurrent_workers;
 430 
 431   _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true);
 432   _concurrent_workers->initialize_workers();
 433 
 434   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 435     size_t mark_stack_size =
 436       MIN2(MarkStackSizeMax,
 437           MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
 438     // Verify that the calculated value for MarkStackSize is in range.
 439     // It would be nice to use the private utility routine from Arguments.
 440     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 441       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 442                       "must be between 1 and " SIZE_FORMAT,
 443                       mark_stack_size, MarkStackSizeMax);
 444       return;
 445     }
 446     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 447   } else {
 448     // Verify MarkStackSize is in range.
 449     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 450       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 451         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 452           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 453                           "must be between 1 and " SIZE_FORMAT,
 454                           MarkStackSize, MarkStackSizeMax);
 455           return;
 456         }
 457       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 458         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 459           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 460                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 461                           MarkStackSize, MarkStackSizeMax);
 462           return;
 463         }
 464       }
 465     }
 466   }
 467 
 468   if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
 469     vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
 470   }
 471 
 472   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
 473   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
 474 
 475   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 476   _num_active_tasks = _max_num_tasks;
 477 
 478   for (uint i = 0; i < _max_num_tasks; ++i) {
 479     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 480     task_queue->initialize();
 481     _task_queues->register_queue(i, task_queue);
 482 
 483     _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions());
 484 
 485     _accum_task_vtime[i] = 0.0;
 486   }
 487 
 488   reset_at_marking_complete();
 489   _completed_initialization = true;
 490 }
 491 
 492 void G1ConcurrentMark::reset() {
 493   _has_aborted = false;
 494 
 495   reset_marking_for_restart();
 496 
 497   // Reset all tasks, since different phases will use different number of active
 498   // threads. So, it's easiest to have all of them ready.
 499   for (uint i = 0; i < _max_num_tasks; ++i) {
 500     _tasks[i]->reset(_next_mark_bitmap);
 501   }
 502 
 503   uint max_regions = _g1h->max_regions();
 504   for (uint i = 0; i < max_regions; i++) {
 505     _top_at_rebuild_starts[i] = NULL;
 506     _region_mark_stats[i].clear();
 507   }
 508 }
 509 
 510 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
 511   for (uint j = 0; j < _max_num_tasks; ++j) {
 512     _tasks[j]->clear_mark_stats_cache(region_idx);
 513   }
 514   _top_at_rebuild_starts[region_idx] = NULL;
 515   _region_mark_stats[region_idx].clear();
 516 }
 517 
 518 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
 519   uint const region_idx = r->hrm_index();
 520   if (r->is_humongous()) {
 521     assert(r->is_starts_humongous(), "Got humongous continues region here");
 522     uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size());
 523     for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
 524       clear_statistics_in_region(j);
 525     }
 526   } else {
 527     clear_statistics_in_region(region_idx);
 528   }
 529 }
 530 
 531 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
 532   if (bitmap->is_marked(addr)) {
 533     bitmap->clear(addr);
 534   }
 535 }
 536 
 537 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
 538   assert_at_safepoint_on_vm_thread();
 539 
 540   // Need to clear all mark bits of the humongous object.
 541   clear_mark_if_set(_prev_mark_bitmap, r->bottom());
 542   clear_mark_if_set(_next_mark_bitmap, r->bottom());
 543 
 544   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
 545     return;
 546   }
 547 
 548   // Clear any statistics about the region gathered so far.
 549   clear_statistics(r);
 550 }
 551 
 552 void G1ConcurrentMark::reset_marking_for_restart() {
 553   _global_mark_stack.set_empty();
 554 
 555   // Expand the marking stack, if we have to and if we can.
 556   if (has_overflown()) {
 557     _global_mark_stack.expand();
 558 
 559     uint max_regions = _g1h->max_regions();
 560     for (uint i = 0; i < max_regions; i++) {
 561       _region_mark_stats[i].clear_during_overflow();
 562     }
 563   }
 564 
 565   clear_has_overflown();
 566   _finger = _heap.start();
 567 
 568   for (uint i = 0; i < _max_num_tasks; ++i) {
 569     G1CMTaskQueue* queue = _task_queues->queue(i);
 570     queue->set_empty();
 571   }
 572 }
 573 
 574 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 575   assert(active_tasks <= _max_num_tasks, "we should not have more");
 576 
 577   _num_active_tasks = active_tasks;
 578   // Need to update the three data structures below according to the
 579   // number of active threads for this phase.
 580   _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
 581   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 582   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 583 }
 584 
 585 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 586   set_concurrency(active_tasks);
 587 
 588   _concurrent = concurrent;
 589 
 590   if (!concurrent) {
 591     // At this point we should be in a STW phase, and completed marking.
 592     assert_at_safepoint_on_vm_thread();
 593     assert(out_of_regions(),
 594            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 595            p2i(_finger), p2i(_heap.end()));
 596   }
 597 }
 598 
 599 void G1ConcurrentMark::reset_at_marking_complete() {
 600   // We set the global marking state to some default values when we're
 601   // not doing marking.
 602   reset_marking_for_restart();
 603   _num_active_tasks = 0;
 604 }
 605 
 606 G1ConcurrentMark::~G1ConcurrentMark() {
 607   FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
 608   FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
 609   // The G1ConcurrentMark instance is never freed.
 610   ShouldNotReachHere();
 611 }
 612 
 613 class G1ClearBitMapTask : public AbstractGangTask {
 614 public:
 615   static size_t chunk_size() { return M; }
 616 
 617 private:
 618   // Heap region closure used for clearing the given mark bitmap.
 619   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 620   private:
 621     G1CMBitMap* _bitmap;
 622     G1ConcurrentMark* _cm;
 623   public:
 624     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _bitmap(bitmap), _cm(cm) {
 625     }
 626 
 627     virtual bool do_heap_region(HeapRegion* r) {
 628       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 629 
 630       HeapWord* cur = r->bottom();
 631       HeapWord* const end = r->end();
 632 
 633       while (cur < end) {
 634         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 635         _bitmap->clear_range(mr);
 636 
 637         cur += chunk_size_in_words;
 638 
 639         // Abort iteration if after yielding the marking has been aborted.
 640         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 641           return true;
 642         }
 643         // Repeat the asserts from before the start of the closure. We will do them
 644         // as asserts here to minimize their overhead on the product. However, we
 645         // will have them as guarantees at the beginning / end of the bitmap
 646         // clearing to get some checking in the product.
 647         assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
 648         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 649       }
 650       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 651 
 652       return false;
 653     }
 654   };
 655 
 656   G1ClearBitmapHRClosure _cl;
 657   HeapRegionClaimer _hr_claimer;
 658   bool _suspendible; // If the task is suspendible, workers must join the STS.
 659 
 660 public:
 661   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 662     AbstractGangTask("G1 Clear Bitmap"),
 663     _cl(bitmap, suspendible ? cm : NULL),
 664     _hr_claimer(n_workers),
 665     _suspendible(suspendible)
 666   { }
 667 
 668   void work(uint worker_id) {
 669     SuspendibleThreadSetJoiner sts_join(_suspendible);
 670     G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
 671   }
 672 
 673   bool is_complete() {
 674     return _cl.is_complete();
 675   }
 676 };
 677 
 678 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 679   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 680 
 681   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 682   size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 683 
 684   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 685 
 686   G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
 687 
 688   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 689   workers->run_task(&cl, num_workers);
 690   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 691 }
 692 
 693 void G1ConcurrentMark::cleanup_for_next_mark() {
 694   // Make sure that the concurrent mark thread looks to still be in
 695   // the current cycle.
 696   guarantee(cm_thread()->during_cycle(), "invariant");
 697 
 698   // We are finishing up the current cycle by clearing the next
 699   // marking bitmap and getting it ready for the next cycle. During
 700   // this time no other cycle can start. So, let's make sure that this
 701   // is the case.
 702   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 703 
 704   clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
 705 
 706   // Repeat the asserts from above.
 707   guarantee(cm_thread()->during_cycle(), "invariant");
 708   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 709 }
 710 
 711 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 712   assert_at_safepoint_on_vm_thread();
 713   clear_bitmap(_prev_mark_bitmap, workers, false);
 714 }
 715 
 716 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
 717 public:
 718   bool do_heap_region(HeapRegion* r) {
 719     r->note_start_of_marking();
 720     return false;
 721   }
 722 };
 723 
 724 void G1ConcurrentMark::pre_initial_mark() {
 725   // Initialize marking structures. This has to be done in a STW phase.
 726   reset();
 727 
 728   // For each region note start of marking.
 729   NoteStartOfMarkHRClosure startcl;
 730   _g1h->heap_region_iterate(&startcl);
 731 }
 732 
 733 
 734 void G1ConcurrentMark::post_initial_mark() {
 735   // Start Concurrent Marking weak-reference discovery.
 736   ReferenceProcessor* rp = _g1h->ref_processor_cm();
 737   // enable ("weak") refs discovery
 738   rp->enable_discovery();
 739   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 740 
 741   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
 742   // This is the start of  the marking cycle, we're expected all
 743   // threads to have SATB queues with active set to false.
 744   satb_mq_set.set_active_all_threads(true, /* new active value */
 745                                      false /* expected_active */);
 746 
 747   _root_regions.prepare_for_scan();
 748 
 749   // update_g1_committed() will be called at the end of an evac pause
 750   // when marking is on. So, it's also called at the end of the
 751   // initial-mark pause to update the heap end, if the heap expands
 752   // during it. No need to call it here.
 753 }
 754 
 755 /*
 756  * Notice that in the next two methods, we actually leave the STS
 757  * during the barrier sync and join it immediately afterwards. If we
 758  * do not do this, the following deadlock can occur: one thread could
 759  * be in the barrier sync code, waiting for the other thread to also
 760  * sync up, whereas another one could be trying to yield, while also
 761  * waiting for the other threads to sync up too.
 762  *
 763  * Note, however, that this code is also used during remark and in
 764  * this case we should not attempt to leave / enter the STS, otherwise
 765  * we'll either hit an assert (debug / fastdebug) or deadlock
 766  * (product). So we should only leave / enter the STS if we are
 767  * operating concurrently.
 768  *
 769  * Because the thread that does the sync barrier has left the STS, it
 770  * is possible to be suspended for a Full GC or an evacuation pause
 771  * could occur. This is actually safe, since the entering the sync
 772  * barrier is one of the last things do_marking_step() does, and it
 773  * doesn't manipulate any data structures afterwards.
 774  */
 775 
 776 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 777   bool barrier_aborted;
 778   {
 779     SuspendibleThreadSetLeaver sts_leave(concurrent());
 780     barrier_aborted = !_first_overflow_barrier_sync.enter();
 781   }
 782 
 783   // at this point everyone should have synced up and not be doing any
 784   // more work
 785 
 786   if (barrier_aborted) {
 787     // If the barrier aborted we ignore the overflow condition and
 788     // just abort the whole marking phase as quickly as possible.
 789     return;
 790   }
 791 }
 792 
 793 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 794   SuspendibleThreadSetLeaver sts_leave(concurrent());
 795   _second_overflow_barrier_sync.enter();
 796 
 797   // at this point everything should be re-initialized and ready to go
 798 }
 799 
 800 class G1CMConcurrentMarkingTask : public AbstractGangTask {
 801   G1ConcurrentMark*     _cm;
 802 
 803 public:
 804   void work(uint worker_id) {
 805     assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
 806     ResourceMark rm;
 807 
 808     double start_vtime = os::elapsedVTime();
 809 
 810     {
 811       SuspendibleThreadSetJoiner sts_join;
 812 
 813       assert(worker_id < _cm->active_tasks(), "invariant");
 814 
 815       G1CMTask* task = _cm->task(worker_id);
 816       task->record_start_time();
 817       if (!_cm->has_aborted()) {
 818         do {
 819           task->do_marking_step(G1ConcMarkStepDurationMillis,
 820                                 true  /* do_termination */,
 821                                 false /* is_serial*/);
 822 
 823           _cm->do_yield_check();
 824         } while (!_cm->has_aborted() && task->has_aborted());
 825       }
 826       task->record_end_time();
 827       guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
 828     }
 829 
 830     double end_vtime = os::elapsedVTime();
 831     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 832   }
 833 
 834   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
 835       AbstractGangTask("Concurrent Mark"), _cm(cm) { }
 836 
 837   ~G1CMConcurrentMarkingTask() { }
 838 };
 839 
 840 uint G1ConcurrentMark::calc_active_marking_workers() {
 841   uint result = 0;
 842   if (!UseDynamicNumberOfGCThreads ||
 843       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 844        !ForceDynamicNumberOfGCThreads)) {
 845     result = _max_concurrent_workers;
 846   } else {
 847     result =
 848       AdaptiveSizePolicy::calc_default_active_workers(_max_concurrent_workers,
 849                                                       1, /* Minimum workers */
 850                                                       _num_concurrent_workers,
 851                                                       Threads::number_of_non_daemon_threads());
 852     // Don't scale the result down by scale_concurrent_workers() because
 853     // that scaling has already gone into "_max_concurrent_workers".
 854   }
 855   assert(result > 0 && result <= _max_concurrent_workers,
 856          "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
 857          _max_concurrent_workers, result);
 858   return result;
 859 }
 860 
 861 void G1ConcurrentMark::scan_root_region(HeapRegion* hr, uint worker_id) {
 862   // Currently, only survivors can be root regions.
 863   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 864   G1RootRegionScanClosure cl(_g1h, this, worker_id);
 865 
 866   const uintx interval = PrefetchScanIntervalInBytes;
 867   HeapWord* curr = hr->bottom();
 868   const HeapWord* end = hr->top();
 869   while (curr < end) {
 870     Prefetch::read(curr, interval);
 871     oop obj = oop(curr);
 872     int size = obj->oop_iterate_size(&cl);
 873     assert(size == obj->size(), "sanity");
 874     curr += size;
 875   }
 876 }
 877 
 878 class G1CMRootRegionScanTask : public AbstractGangTask {
 879   G1ConcurrentMark* _cm;
 880 public:
 881   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 882     AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
 883 
 884   void work(uint worker_id) {
 885     assert(Thread::current()->is_ConcurrentGC_thread(),
 886            "this should only be done by a conc GC thread");
 887 
 888     G1CMRootRegions* root_regions = _cm->root_regions();
 889     HeapRegion* hr = root_regions->claim_next();
 890     while (hr != NULL) {
 891       _cm->scan_root_region(hr, worker_id);
 892       hr = root_regions->claim_next();
 893     }
 894   }
 895 };
 896 
 897 void G1ConcurrentMark::scan_root_regions() {
 898   // scan_in_progress() will have been set to true only if there was
 899   // at least one root region to scan. So, if it's false, we
 900   // should not attempt to do any further work.
 901   if (root_regions()->scan_in_progress()) {
 902     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 903 
 904     _num_concurrent_workers = MIN2(calc_active_marking_workers(),
 905                                    // We distribute work on a per-region basis, so starting
 906                                    // more threads than that is useless.
 907                                    root_regions()->num_root_regions());
 908     assert(_num_concurrent_workers <= _max_concurrent_workers,
 909            "Maximum number of marking threads exceeded");
 910 
 911     G1CMRootRegionScanTask task(this);
 912     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
 913                         task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
 914     _concurrent_workers->run_task(&task, _num_concurrent_workers);
 915 
 916     // It's possible that has_aborted() is true here without actually
 917     // aborting the survivor scan earlier. This is OK as it's
 918     // mainly used for sanity checking.
 919     root_regions()->scan_finished();
 920   }
 921 }
 922 
 923 void G1ConcurrentMark::concurrent_cycle_start() {
 924   _gc_timer_cm->register_gc_start();
 925 
 926   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
 927 
 928   _g1h->trace_heap_before_gc(_gc_tracer_cm);
 929 }
 930 
 931 void G1ConcurrentMark::concurrent_cycle_end() {
 932   _g1h->collector_state()->set_clearing_next_bitmap(false);
 933 
 934   _g1h->trace_heap_after_gc(_gc_tracer_cm);
 935 
 936   if (has_aborted()) {
 937     log_info(gc, marking)("Concurrent Mark Abort");
 938     _gc_tracer_cm->report_concurrent_mode_failure();
 939   }
 940 
 941   _gc_timer_cm->register_gc_end();
 942 
 943   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
 944 }
 945 
 946 void G1ConcurrentMark::mark_from_roots() {
 947   _restart_for_overflow = false;
 948 
 949   _num_concurrent_workers = calc_active_marking_workers();
 950 
 951   uint active_workers = MAX2(1U, _num_concurrent_workers);
 952 
 953   // Setting active workers is not guaranteed since fewer
 954   // worker threads may currently exist and more may not be
 955   // available.
 956   active_workers = _concurrent_workers->update_active_workers(active_workers);
 957   log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
 958 
 959   // Parallel task terminator is set in "set_concurrency_and_phase()"
 960   set_concurrency_and_phase(active_workers, true /* concurrent */);
 961 
 962   G1CMConcurrentMarkingTask marking_task(this);
 963   _concurrent_workers->run_task(&marking_task);
 964   print_stats();
 965 }
 966 
 967 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
 968   G1HeapVerifier* verifier = _g1h->verifier();
 969 
 970   verifier->verify_region_sets_optional();
 971 
 972   if (VerifyDuringGC) {
 973     GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm);
 974 
 975     size_t const BufLen = 512;
 976     char buffer[BufLen];
 977 
 978     jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
 979     verifier->verify(type, vo, buffer);
 980   }
 981 
 982   verifier->check_bitmaps(caller);
 983 }
 984 
 985 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask {
 986   G1CollectedHeap* _g1h;
 987   G1ConcurrentMark* _cm;
 988   HeapRegionClaimer _hrclaimer;
 989   uint volatile _total_selected_for_rebuild;
 990 
 991   G1PrintRegionLivenessInfoClosure _cl;
 992 
 993   class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
 994     G1CollectedHeap* _g1h;
 995     G1ConcurrentMark* _cm;
 996 
 997     G1PrintRegionLivenessInfoClosure* _cl;
 998 
 999     uint _num_regions_selected_for_rebuild;  // The number of regions actually selected for rebuild.
1000 
1001     void update_remset_before_rebuild(HeapRegion* hr) {
1002       G1RemSetTrackingPolicy* tracking_policy = _g1h->g1_policy()->remset_tracker();
1003 
1004       bool selected_for_rebuild;
1005       if (hr->is_humongous()) {
1006         bool const is_live = _cm->liveness(hr->humongous_start_region()->hrm_index()) > 0;
1007         selected_for_rebuild = tracking_policy->update_humongous_before_rebuild(hr, is_live);
1008       } else {
1009         size_t const live_bytes = _cm->liveness(hr->hrm_index());
1010         selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1011       }
1012       if (selected_for_rebuild) {
1013         _num_regions_selected_for_rebuild++;
1014       }
1015       _cm->update_top_at_rebuild_start(hr);
1016     }
1017 
1018     // Distribute the given words across the humongous object starting with hr and
1019     // note end of marking.
1020     void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1021       uint const region_idx = hr->hrm_index();
1022       size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size();
1023       uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1024 
1025       // "Distributing" zero words means that we only note end of marking for these
1026       // regions.
1027       assert(marked_words == 0 || obj_size_in_words == marked_words,
1028              "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1029              obj_size_in_words, marked_words);
1030 
1031       for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1032         HeapRegion* const r = _g1h->region_at(i);
1033         size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1034 
1035         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)",
1036                                words_to_add, i, r->get_type_str());
1037         add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1038         marked_words -= words_to_add;
1039       }
1040       assert(marked_words == 0,
1041              SIZE_FORMAT " words left after distributing space across %u regions",
1042              marked_words, num_regions_in_humongous);
1043     }
1044 
1045     void update_marked_bytes(HeapRegion* hr) {
1046       uint const region_idx = hr->hrm_index();
1047       size_t const marked_words = _cm->liveness(region_idx);
1048       // The marking attributes the object's size completely to the humongous starts
1049       // region. We need to distribute this value across the entire set of regions a
1050       // humongous object spans.
1051       if (hr->is_humongous()) {
1052         assert(hr->is_starts_humongous() || marked_words == 0,
1053                "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1054                marked_words, region_idx, hr->get_type_str());
1055         if (hr->is_starts_humongous()) {
1056           distribute_marked_bytes(hr, marked_words);
1057         }
1058       } else {
1059         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1060         add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize);
1061       }
1062     }
1063 
1064     void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1065       hr->add_to_marked_bytes(marked_bytes);
1066       _cl->do_heap_region(hr);
1067       hr->note_end_of_marking();
1068     }
1069 
1070   public:
1071     G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1072       _g1h(g1h), _cm(cm), _cl(cl), _num_regions_selected_for_rebuild(0) { }
1073 
1074     virtual bool do_heap_region(HeapRegion* r) {
1075       update_remset_before_rebuild(r);
1076       update_marked_bytes(r);
1077 
1078       return false;
1079     }
1080 
1081     uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1082   };
1083 
1084 public:
1085   G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1086     AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"),
1087     _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1088 
1089   virtual void work(uint worker_id) {
1090     G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1091     _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1092     Atomic::add(update_cl.num_selected_for_rebuild(), &_total_selected_for_rebuild);
1093   }
1094 
1095   uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1096 
1097   // Number of regions for which roughly one thread should be spawned for this work.
1098   static const uint RegionsPerThread = 384;
1099 };
1100 
1101 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1102   G1CollectedHeap* _g1h;
1103 public:
1104   G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1105 
1106   virtual bool do_heap_region(HeapRegion* r) {
1107     _g1h->g1_policy()->remset_tracker()->update_after_rebuild(r);
1108     return false;
1109   }
1110 };
1111 
1112 void G1ConcurrentMark::remark() {
1113   assert_at_safepoint_on_vm_thread();
1114 
1115   // If a full collection has happened, we should not continue. However we might
1116   // have ended up here as the Remark VM operation has been scheduled already.
1117   if (has_aborted()) {
1118     return;
1119   }
1120 
1121   G1Policy* g1p = _g1h->g1_policy();
1122   g1p->record_concurrent_mark_remark_start();
1123 
1124   double start = os::elapsedTime();
1125 
1126   verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1127 
1128   {
1129     GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm);
1130     finalize_marking();
1131   }
1132 
1133   double mark_work_end = os::elapsedTime();
1134 
1135   bool const mark_finished = !has_overflown();
1136   if (mark_finished) {
1137     weak_refs_work(false /* clear_all_soft_refs */);
1138 
1139     SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1140     // We're done with marking.
1141     // This is the end of the marking cycle, we're expected all
1142     // threads to have SATB queues with active set to true.
1143     satb_mq_set.set_active_all_threads(false, /* new active value */
1144                                        true /* expected_active */);
1145 
1146     {
1147       GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm);
1148       flush_all_task_caches();
1149     }
1150 
1151     // Install newly created mark bitmap as "prev".
1152     swap_mark_bitmaps();
1153     {
1154       GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1155 
1156       uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1157                                        G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1158       uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1159 
1160       G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1161       log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1162       _g1h->workers()->run_task(&cl, num_workers);
1163 
1164       log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1165                                       _g1h->num_regions(), cl.total_selected_for_rebuild());
1166     }
1167     {
1168       GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm);
1169       reclaim_empty_regions();
1170     }
1171 
1172     // Clean out dead classes
1173     if (ClassUnloadingWithConcurrentMark) {
1174       GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm);
1175       ClassLoaderDataGraph::purge();
1176     }
1177 
1178     _g1h->resize_heap_if_necessary();
1179 
1180     compute_new_sizes();
1181 
1182     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1183 
1184     assert(!restart_for_overflow(), "sanity");
1185     // Completely reset the marking state since marking completed
1186     reset_at_marking_complete();
1187   } else {
1188     // We overflowed.  Restart concurrent marking.
1189     _restart_for_overflow = true;
1190 
1191     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1192 
1193     // Clear the marking state because we will be restarting
1194     // marking due to overflowing the global mark stack.
1195     reset_marking_for_restart();
1196   }
1197 
1198   {
1199     GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm);
1200     report_object_count(mark_finished);
1201   }
1202 
1203   // Statistics
1204   double now = os::elapsedTime();
1205   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1206   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1207   _remark_times.add((now - start) * 1000.0);
1208 
1209   g1p->record_concurrent_mark_remark_end();
1210 }
1211 
1212 class G1ReclaimEmptyRegionsTask : public AbstractGangTask {
1213   // Per-region work during the Cleanup pause.
1214   class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1215     G1CollectedHeap* _g1h;
1216     size_t _freed_bytes;
1217     FreeRegionList* _local_cleanup_list;
1218     uint _old_regions_removed;
1219     uint _humongous_regions_removed;
1220 
1221   public:
1222     G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1223                                  FreeRegionList* local_cleanup_list) :
1224       _g1h(g1h),
1225       _freed_bytes(0),
1226       _local_cleanup_list(local_cleanup_list),
1227       _old_regions_removed(0),
1228       _humongous_regions_removed(0) { }
1229 
1230     size_t freed_bytes() { return _freed_bytes; }
1231     const uint old_regions_removed() { return _old_regions_removed; }
1232     const uint humongous_regions_removed() { return _humongous_regions_removed; }
1233 
1234     bool do_heap_region(HeapRegion *hr) {
1235       if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1236         _freed_bytes += hr->used();
1237         hr->set_containing_set(NULL);
1238         if (hr->is_humongous()) {
1239           _humongous_regions_removed++;
1240           _g1h->free_humongous_region(hr, _local_cleanup_list);
1241         } else {
1242           _old_regions_removed++;
1243           _g1h->free_region(hr, _local_cleanup_list, false /* skip_remset */, false /* skip_hcc */, true /* locked */);
1244         }
1245         hr->clear_cardtable();
1246         _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1247         log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1248       }
1249 
1250       return false;
1251     }
1252   };
1253 
1254   G1CollectedHeap* _g1h;
1255   FreeRegionList* _cleanup_list;
1256   HeapRegionClaimer _hrclaimer;
1257 
1258 public:
1259   G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1260     AbstractGangTask("G1 Cleanup"),
1261     _g1h(g1h),
1262     _cleanup_list(cleanup_list),
1263     _hrclaimer(n_workers) {
1264   }
1265 
1266   void work(uint worker_id) {
1267     FreeRegionList local_cleanup_list("Local Cleanup List");
1268     G1ReclaimEmptyRegionsClosure cl(_g1h, &local_cleanup_list);
1269     _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1270     assert(cl.is_complete(), "Shouldn't have aborted!");
1271 
1272     // Now update the old/humongous region sets
1273     _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed());
1274     {
1275       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1276       _g1h->decrement_summary_bytes(cl.freed_bytes());
1277 
1278       _cleanup_list->add_ordered(&local_cleanup_list);
1279       assert(local_cleanup_list.is_empty(), "post-condition");
1280     }
1281   }
1282 };
1283 
1284 void G1ConcurrentMark::reclaim_empty_regions() {
1285   WorkGang* workers = _g1h->workers();
1286   FreeRegionList empty_regions_list("Empty Regions After Mark List");
1287 
1288   G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1289   workers->run_task(&cl);
1290 
1291   if (!empty_regions_list.is_empty()) {
1292     log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1293     // Now print the empty regions list.
1294     G1HRPrinter* hrp = _g1h->hr_printer();
1295     if (hrp->is_active()) {
1296       FreeRegionListIterator iter(&empty_regions_list);
1297       while (iter.more_available()) {
1298         HeapRegion* hr = iter.get_next();
1299         hrp->cleanup(hr);
1300       }
1301     }
1302     // And actually make them available.
1303     _g1h->prepend_to_freelist(&empty_regions_list);
1304   }
1305 }
1306 
1307 void G1ConcurrentMark::compute_new_sizes() {
1308   MetaspaceGC::compute_new_size();
1309 
1310   // Cleanup will have freed any regions completely full of garbage.
1311   // Update the soft reference policy with the new heap occupancy.
1312   Universe::update_heap_info_at_gc();
1313 
1314   // We reclaimed old regions so we should calculate the sizes to make
1315   // sure we update the old gen/space data.
1316   _g1h->g1mm()->update_sizes();
1317 }
1318 
1319 void G1ConcurrentMark::cleanup() {
1320   assert_at_safepoint_on_vm_thread();
1321 
1322   // If a full collection has happened, we shouldn't do this.
1323   if (has_aborted()) {
1324     return;
1325   }
1326 
1327   G1Policy* g1p = _g1h->g1_policy();
1328   g1p->record_concurrent_mark_cleanup_start();
1329 
1330   double start = os::elapsedTime();
1331 
1332   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1333 
1334   {
1335     GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1336     G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1337     _g1h->heap_region_iterate(&cl);
1338   }
1339 
1340   if (log_is_enabled(Trace, gc, liveness)) {
1341     G1PrintRegionLivenessInfoClosure cl("Post-Cleanup");
1342     _g1h->heap_region_iterate(&cl);
1343   }
1344 
1345   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1346 
1347   // We need to make this be a "collection" so any collection pause that
1348   // races with it goes around and waits for Cleanup to finish.
1349   _g1h->increment_total_collections();
1350 
1351   // Local statistics
1352   double recent_cleanup_time = (os::elapsedTime() - start);
1353   _total_cleanup_time += recent_cleanup_time;
1354   _cleanup_times.add(recent_cleanup_time);
1355 
1356   {
1357     GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1358     _g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1359   }
1360 }
1361 
1362 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1363 // Uses the G1CMTask associated with a worker thread (for serial reference
1364 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1365 // trace referent objects.
1366 //
1367 // Using the G1CMTask and embedded local queues avoids having the worker
1368 // threads operating on the global mark stack. This reduces the risk
1369 // of overflowing the stack - which we would rather avoid at this late
1370 // state. Also using the tasks' local queues removes the potential
1371 // of the workers interfering with each other that could occur if
1372 // operating on the global stack.
1373 
1374 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1375   G1ConcurrentMark* _cm;
1376   G1CMTask*         _task;
1377   uint              _ref_counter_limit;
1378   uint              _ref_counter;
1379   bool              _is_serial;
1380 public:
1381   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1382     _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval),
1383     _ref_counter(_ref_counter_limit), _is_serial(is_serial) {
1384     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1385   }
1386 
1387   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1388   virtual void do_oop(      oop* p) { do_oop_work(p); }
1389 
1390   template <class T> void do_oop_work(T* p) {
1391     if (_cm->has_overflown()) {
1392       return;
1393     }
1394     if (!_task->deal_with_reference(p)) {
1395       // We did not add anything to the mark bitmap (or mark stack), so there is
1396       // no point trying to drain it.
1397       return;
1398     }
1399     _ref_counter--;
1400 
1401     if (_ref_counter == 0) {
1402       // We have dealt with _ref_counter_limit references, pushing them
1403       // and objects reachable from them on to the local stack (and
1404       // possibly the global stack). Call G1CMTask::do_marking_step() to
1405       // process these entries.
1406       //
1407       // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1408       // there's nothing more to do (i.e. we're done with the entries that
1409       // were pushed as a result of the G1CMTask::deal_with_reference() calls
1410       // above) or we overflow.
1411       //
1412       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1413       // flag while there may still be some work to do. (See the comment at
1414       // the beginning of G1CMTask::do_marking_step() for those conditions -
1415       // one of which is reaching the specified time target.) It is only
1416       // when G1CMTask::do_marking_step() returns without setting the
1417       // has_aborted() flag that the marking step has completed.
1418       do {
1419         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1420         _task->do_marking_step(mark_step_duration_ms,
1421                                false      /* do_termination */,
1422                                _is_serial);
1423       } while (_task->has_aborted() && !_cm->has_overflown());
1424       _ref_counter = _ref_counter_limit;
1425     }
1426   }
1427 };
1428 
1429 // 'Drain' oop closure used by both serial and parallel reference processing.
1430 // Uses the G1CMTask associated with a given worker thread (for serial
1431 // reference processing the G1CMtask for worker 0 is used). Calls the
1432 // do_marking_step routine, with an unbelievably large timeout value,
1433 // to drain the marking data structures of the remaining entries
1434 // added by the 'keep alive' oop closure above.
1435 
1436 class G1CMDrainMarkingStackClosure : public VoidClosure {
1437   G1ConcurrentMark* _cm;
1438   G1CMTask*         _task;
1439   bool              _is_serial;
1440  public:
1441   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1442     _cm(cm), _task(task), _is_serial(is_serial) {
1443     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1444   }
1445 
1446   void do_void() {
1447     do {
1448       // We call G1CMTask::do_marking_step() to completely drain the local
1449       // and global marking stacks of entries pushed by the 'keep alive'
1450       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1451       //
1452       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1453       // if there's nothing more to do (i.e. we've completely drained the
1454       // entries that were pushed as a a result of applying the 'keep alive'
1455       // closure to the entries on the discovered ref lists) or we overflow
1456       // the global marking stack.
1457       //
1458       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1459       // flag while there may still be some work to do. (See the comment at
1460       // the beginning of G1CMTask::do_marking_step() for those conditions -
1461       // one of which is reaching the specified time target.) It is only
1462       // when G1CMTask::do_marking_step() returns without setting the
1463       // has_aborted() flag that the marking step has completed.
1464 
1465       _task->do_marking_step(1000000000.0 /* something very large */,
1466                              true         /* do_termination */,
1467                              _is_serial);
1468     } while (_task->has_aborted() && !_cm->has_overflown());
1469   }
1470 };
1471 
1472 // Implementation of AbstractRefProcTaskExecutor for parallel
1473 // reference processing at the end of G1 concurrent marking
1474 
1475 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor {
1476 private:
1477   G1CollectedHeap*  _g1h;
1478   G1ConcurrentMark* _cm;
1479   WorkGang*         _workers;
1480   uint              _active_workers;
1481 
1482 public:
1483   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1484                           G1ConcurrentMark* cm,
1485                           WorkGang* workers,
1486                           uint n_workers) :
1487     _g1h(g1h), _cm(cm),
1488     _workers(workers), _active_workers(n_workers) { }
1489 
1490   virtual void execute(ProcessTask& task, uint ergo_workers);
1491 };
1492 
1493 class G1CMRefProcTaskProxy : public AbstractGangTask {
1494   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1495   ProcessTask&      _proc_task;
1496   G1CollectedHeap*  _g1h;
1497   G1ConcurrentMark* _cm;
1498 
1499 public:
1500   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1501                        G1CollectedHeap* g1h,
1502                        G1ConcurrentMark* cm) :
1503     AbstractGangTask("Process reference objects in parallel"),
1504     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1505     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1506     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1507   }
1508 
1509   virtual void work(uint worker_id) {
1510     ResourceMark rm;
1511     HandleMark hm;
1512     G1CMTask* task = _cm->task(worker_id);
1513     G1CMIsAliveClosure g1_is_alive(_g1h);
1514     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1515     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1516 
1517     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1518   }
1519 };
1520 
1521 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task, uint ergo_workers) {
1522   assert(_workers != NULL, "Need parallel worker threads.");
1523   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1524   assert(_workers->active_workers() >= ergo_workers,
1525          "Ergonomically chosen workers(%u) should be less than or equal to active workers(%u)",
1526          ergo_workers, _workers->active_workers());
1527 
1528   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1529 
1530   // We need to reset the concurrency level before each
1531   // proxy task execution, so that the termination protocol
1532   // and overflow handling in G1CMTask::do_marking_step() knows
1533   // how many workers to wait for.
1534   _cm->set_concurrency(ergo_workers);
1535   _workers->run_task(&proc_task_proxy, ergo_workers);
1536 }
1537 
1538 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1539   ResourceMark rm;
1540   HandleMark   hm;
1541 
1542   // Is alive closure.
1543   G1CMIsAliveClosure g1_is_alive(_g1h);
1544 
1545   // Inner scope to exclude the cleaning of the string table
1546   // from the displayed time.
1547   {
1548     GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm);
1549 
1550     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1551 
1552     // See the comment in G1CollectedHeap::ref_processing_init()
1553     // about how reference processing currently works in G1.
1554 
1555     // Set the soft reference policy
1556     rp->setup_policy(clear_all_soft_refs);
1557     assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1558 
1559     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1560     // in serial reference processing. Note these closures are also
1561     // used for serially processing (by the the current thread) the
1562     // JNI references during parallel reference processing.
1563     //
1564     // These closures do not need to synchronize with the worker
1565     // threads involved in parallel reference processing as these
1566     // instances are executed serially by the current thread (e.g.
1567     // reference processing is not multi-threaded and is thus
1568     // performed by the current thread instead of a gang worker).
1569     //
1570     // The gang tasks involved in parallel reference processing create
1571     // their own instances of these closures, which do their own
1572     // synchronization among themselves.
1573     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1574     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1575 
1576     // We need at least one active thread. If reference processing
1577     // is not multi-threaded we use the current (VMThread) thread,
1578     // otherwise we use the work gang from the G1CollectedHeap and
1579     // we utilize all the worker threads we can.
1580     bool processing_is_mt = rp->processing_is_mt();
1581     uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U);
1582     active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U);
1583 
1584     // Parallel processing task executor.
1585     G1CMRefProcTaskExecutor par_task_executor(_g1h, this,
1586                                               _g1h->workers(), active_workers);
1587     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1588 
1589     // Set the concurrency level. The phase was already set prior to
1590     // executing the remark task.
1591     set_concurrency(active_workers);
1592 
1593     // Set the degree of MT processing here.  If the discovery was done MT,
1594     // the number of threads involved during discovery could differ from
1595     // the number of active workers.  This is OK as long as the discovered
1596     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1597     rp->set_active_mt_degree(active_workers);
1598 
1599     ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues());
1600 
1601     // Process the weak references.
1602     const ReferenceProcessorStats& stats =
1603         rp->process_discovered_references(&g1_is_alive,
1604                                           &g1_keep_alive,
1605                                           &g1_drain_mark_stack,
1606                                           executor,
1607                                           &pt);
1608     _gc_tracer_cm->report_gc_reference_stats(stats);
1609     pt.print_all_references();
1610 
1611     // The do_oop work routines of the keep_alive and drain_marking_stack
1612     // oop closures will set the has_overflown flag if we overflow the
1613     // global marking stack.
1614 
1615     assert(has_overflown() || _global_mark_stack.is_empty(),
1616            "Mark stack should be empty (unless it has overflown)");
1617 
1618     assert(rp->num_queues() == active_workers, "why not");
1619 
1620     rp->verify_no_references_recorded();
1621     assert(!rp->discovery_enabled(), "Post condition");
1622   }
1623 
1624   if (has_overflown()) {
1625     // We can not trust g1_is_alive and the contents of the heap if the marking stack
1626     // overflowed while processing references. Exit the VM.
1627     fatal("Overflow during reference processing, can not continue. Please "
1628           "increase MarkStackSizeMax (current value: " SIZE_FORMAT ") and "
1629           "restart.", MarkStackSizeMax);
1630     return;
1631   }
1632 
1633   assert(_global_mark_stack.is_empty(), "Marking should have completed");
1634 
1635   {
1636     GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1637     WeakProcessor::weak_oops_do(_g1h->workers(), &g1_is_alive, &do_nothing_cl, 1);
1638   }
1639 
1640   // Unload Klasses, String, Code Cache, etc.
1641   if (ClassUnloadingWithConcurrentMark) {
1642     GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1643     bool purged_classes = SystemDictionary::do_unloading(_gc_timer_cm);
1644     _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1645   } else {
1646     GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1647     // No need to clean string table as it is treated as strong roots when
1648     // class unloading is disabled.
1649     _g1h->partial_cleaning(&g1_is_alive, false, G1StringDedup::is_enabled());
1650   }
1651 }
1652 
1653 class G1PrecleanYieldClosure : public YieldClosure {
1654   G1ConcurrentMark* _cm;
1655 
1656 public:
1657   G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
1658 
1659   virtual bool should_return() {
1660     return _cm->has_aborted();
1661   }
1662 
1663   virtual bool should_return_fine_grain() {
1664     _cm->do_yield_check();
1665     return _cm->has_aborted();
1666   }
1667 };
1668 
1669 void G1ConcurrentMark::preclean() {
1670   assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
1671 
1672   SuspendibleThreadSetJoiner joiner;
1673 
1674   G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */);
1675   G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */);
1676 
1677   set_concurrency_and_phase(1, true);
1678 
1679   G1PrecleanYieldClosure yield_cl(this);
1680 
1681   ReferenceProcessor* rp = _g1h->ref_processor_cm();
1682   // Precleaning is single threaded. Temporarily disable MT discovery.
1683   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
1684   rp->preclean_discovered_references(rp->is_alive_non_header(),
1685                                      &keep_alive,
1686                                      &drain_mark_stack,
1687                                      &yield_cl,
1688                                      _gc_timer_cm);
1689 }
1690 
1691 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1692 // the prev bitmap determining liveness.
1693 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1694   G1CollectedHeap* _g1h;
1695 public:
1696   G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1697 
1698   bool do_object_b(oop obj) {
1699     HeapWord* addr = (HeapWord*)obj;
1700     return addr != NULL &&
1701            (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_dead(obj));
1702   }
1703 };
1704 
1705 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1706   // Depending on the completion of the marking liveness needs to be determined
1707   // using either the next or prev bitmap.
1708   if (mark_completed) {
1709     G1ObjectCountIsAliveClosure is_alive(_g1h);
1710     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1711   } else {
1712     G1CMIsAliveClosure is_alive(_g1h);
1713     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1714   }
1715 }
1716 
1717 
1718 void G1ConcurrentMark::swap_mark_bitmaps() {
1719   G1CMBitMap* temp = _prev_mark_bitmap;
1720   _prev_mark_bitmap = _next_mark_bitmap;
1721   _next_mark_bitmap = temp;
1722   _g1h->collector_state()->set_clearing_next_bitmap(true);
1723 }
1724 
1725 // Closure for marking entries in SATB buffers.
1726 class G1CMSATBBufferClosure : public SATBBufferClosure {
1727 private:
1728   G1CMTask* _task;
1729   G1CollectedHeap* _g1h;
1730 
1731   // This is very similar to G1CMTask::deal_with_reference, but with
1732   // more relaxed requirements for the argument, so this must be more
1733   // circumspect about treating the argument as an object.
1734   void do_entry(void* entry) const {
1735     _task->increment_refs_reached();
1736     oop const obj = static_cast<oop>(entry);
1737     _task->make_reference_grey(obj);
1738   }
1739 
1740 public:
1741   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1742     : _task(task), _g1h(g1h) { }
1743 
1744   virtual void do_buffer(void** buffer, size_t size) {
1745     for (size_t i = 0; i < size; ++i) {
1746       do_entry(buffer[i]);
1747     }
1748   }
1749 };
1750 
1751 class G1RemarkThreadsClosure : public ThreadClosure {
1752   G1CMSATBBufferClosure _cm_satb_cl;
1753   G1CMOopClosure _cm_cl;
1754   MarkingCodeBlobClosure _code_cl;
1755   int _thread_parity;
1756 
1757  public:
1758   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1759     _cm_satb_cl(task, g1h),
1760     _cm_cl(g1h, task),
1761     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1762     _thread_parity(Threads::thread_claim_parity()) {}
1763 
1764   void do_thread(Thread* thread) {
1765     if (thread->is_Java_thread()) {
1766       if (thread->claim_oops_do(true, _thread_parity)) {
1767         JavaThread* jt = (JavaThread*)thread;
1768 
1769         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1770         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1771         // * Alive if on the stack of an executing method
1772         // * Weakly reachable otherwise
1773         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1774         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1775         jt->nmethods_do(&_code_cl);
1776 
1777         G1ThreadLocalData::satb_mark_queue(jt).apply_closure_and_empty(&_cm_satb_cl);
1778       }
1779     } else if (thread->is_VM_thread()) {
1780       if (thread->claim_oops_do(true, _thread_parity)) {
1781         G1BarrierSet::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1782       }
1783     }
1784   }
1785 };
1786 
1787 class G1CMRemarkTask : public AbstractGangTask {
1788   G1ConcurrentMark* _cm;
1789 public:
1790   void work(uint worker_id) {
1791     G1CMTask* task = _cm->task(worker_id);
1792     task->record_start_time();
1793     {
1794       ResourceMark rm;
1795       HandleMark hm;
1796 
1797       G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1798       Threads::threads_do(&threads_f);
1799     }
1800 
1801     do {
1802       task->do_marking_step(1000000000.0 /* something very large */,
1803                             true         /* do_termination       */,
1804                             false        /* is_serial            */);
1805     } while (task->has_aborted() && !_cm->has_overflown());
1806     // If we overflow, then we do not want to restart. We instead
1807     // want to abort remark and do concurrent marking again.
1808     task->record_end_time();
1809   }
1810 
1811   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1812     AbstractGangTask("Par Remark"), _cm(cm) {
1813     _cm->terminator()->reset_for_reuse(active_workers);
1814   }
1815 };
1816 
1817 void G1ConcurrentMark::finalize_marking() {
1818   ResourceMark rm;
1819   HandleMark   hm;
1820 
1821   _g1h->ensure_parsability(false);
1822 
1823   // this is remark, so we'll use up all active threads
1824   uint active_workers = _g1h->workers()->active_workers();
1825   set_concurrency_and_phase(active_workers, false /* concurrent */);
1826   // Leave _parallel_marking_threads at it's
1827   // value originally calculated in the G1ConcurrentMark
1828   // constructor and pass values of the active workers
1829   // through the gang in the task.
1830 
1831   {
1832     StrongRootsScope srs(active_workers);
1833 
1834     G1CMRemarkTask remarkTask(this, active_workers);
1835     // We will start all available threads, even if we decide that the
1836     // active_workers will be fewer. The extra ones will just bail out
1837     // immediately.
1838     _g1h->workers()->run_task(&remarkTask);
1839   }
1840 
1841   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1842   guarantee(has_overflown() ||
1843             satb_mq_set.completed_buffers_num() == 0,
1844             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1845             BOOL_TO_STR(has_overflown()),
1846             satb_mq_set.completed_buffers_num());
1847 
1848   print_stats();
1849 }
1850 
1851 void G1ConcurrentMark::flush_all_task_caches() {
1852   size_t hits = 0;
1853   size_t misses = 0;
1854   for (uint i = 0; i < _max_num_tasks; i++) {
1855     Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1856     hits += stats.first;
1857     misses += stats.second;
1858   }
1859   size_t sum = hits + misses;
1860   log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1861                        hits, misses, percent_of(hits, sum));
1862 }
1863 
1864 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1865   _prev_mark_bitmap->clear_range(mr);
1866 }
1867 
1868 HeapRegion*
1869 G1ConcurrentMark::claim_region(uint worker_id) {
1870   // "checkpoint" the finger
1871   HeapWord* finger = _finger;
1872 
1873   while (finger < _heap.end()) {
1874     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1875 
1876     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1877     // Make sure that the reads below do not float before loading curr_region.
1878     OrderAccess::loadload();
1879     // Above heap_region_containing may return NULL as we always scan claim
1880     // until the end of the heap. In this case, just jump to the next region.
1881     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1882 
1883     // Is the gap between reading the finger and doing the CAS too long?
1884     HeapWord* res = Atomic::cmpxchg(end, &_finger, finger);
1885     if (res == finger && curr_region != NULL) {
1886       // we succeeded
1887       HeapWord*   bottom        = curr_region->bottom();
1888       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1889 
1890       // notice that _finger == end cannot be guaranteed here since,
1891       // someone else might have moved the finger even further
1892       assert(_finger >= end, "the finger should have moved forward");
1893 
1894       if (limit > bottom) {
1895         return curr_region;
1896       } else {
1897         assert(limit == bottom,
1898                "the region limit should be at bottom");
1899         // we return NULL and the caller should try calling
1900         // claim_region() again.
1901         return NULL;
1902       }
1903     } else {
1904       assert(_finger > finger, "the finger should have moved forward");
1905       // read it again
1906       finger = _finger;
1907     }
1908   }
1909 
1910   return NULL;
1911 }
1912 
1913 #ifndef PRODUCT
1914 class VerifyNoCSetOops {
1915   G1CollectedHeap* _g1h;
1916   const char* _phase;
1917   int _info;
1918 
1919 public:
1920   VerifyNoCSetOops(const char* phase, int info = -1) :
1921     _g1h(G1CollectedHeap::heap()),
1922     _phase(phase),
1923     _info(info)
1924   { }
1925 
1926   void operator()(G1TaskQueueEntry task_entry) const {
1927     if (task_entry.is_array_slice()) {
1928       guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1929       return;
1930     }
1931     guarantee(oopDesc::is_oop(task_entry.obj()),
1932               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1933               p2i(task_entry.obj()), _phase, _info);
1934     guarantee(!_g1h->is_in_cset(task_entry.obj()),
1935               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1936               p2i(task_entry.obj()), _phase, _info);
1937   }
1938 };
1939 
1940 void G1ConcurrentMark::verify_no_cset_oops() {
1941   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1942   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1943     return;
1944   }
1945 
1946   // Verify entries on the global mark stack
1947   _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1948 
1949   // Verify entries on the task queues
1950   for (uint i = 0; i < _max_num_tasks; ++i) {
1951     G1CMTaskQueue* queue = _task_queues->queue(i);
1952     queue->iterate(VerifyNoCSetOops("Queue", i));
1953   }
1954 
1955   // Verify the global finger
1956   HeapWord* global_finger = finger();
1957   if (global_finger != NULL && global_finger < _heap.end()) {
1958     // Since we always iterate over all regions, we might get a NULL HeapRegion
1959     // here.
1960     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1961     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1962               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1963               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1964   }
1965 
1966   // Verify the task fingers
1967   assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
1968   for (uint i = 0; i < _num_concurrent_workers; ++i) {
1969     G1CMTask* task = _tasks[i];
1970     HeapWord* task_finger = task->finger();
1971     if (task_finger != NULL && task_finger < _heap.end()) {
1972       // See above note on the global finger verification.
1973       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
1974       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
1975                 !task_hr->in_collection_set(),
1976                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
1977                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
1978     }
1979   }
1980 }
1981 #endif // PRODUCT
1982 
1983 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
1984   _g1h->g1_rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
1985 }
1986 
1987 void G1ConcurrentMark::print_stats() {
1988   if (!log_is_enabled(Debug, gc, stats)) {
1989     return;
1990   }
1991   log_debug(gc, stats)("---------------------------------------------------------------------");
1992   for (size_t i = 0; i < _num_active_tasks; ++i) {
1993     _tasks[i]->print_stats();
1994     log_debug(gc, stats)("---------------------------------------------------------------------");
1995   }
1996 }
1997 
1998 void G1ConcurrentMark::concurrent_cycle_abort() {
1999   if (!cm_thread()->during_cycle() || _has_aborted) {
2000     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2001     return;
2002   }
2003 
2004   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2005   // concurrent bitmap clearing.
2006   {
2007     GCTraceTime(Debug, gc) debug("Clear Next Bitmap");
2008     clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
2009   }
2010   // Note we cannot clear the previous marking bitmap here
2011   // since VerifyDuringGC verifies the objects marked during
2012   // a full GC against the previous bitmap.
2013 
2014   // Empty mark stack
2015   reset_marking_for_restart();
2016   for (uint i = 0; i < _max_num_tasks; ++i) {
2017     _tasks[i]->clear_region_fields();
2018   }
2019   _first_overflow_barrier_sync.abort();
2020   _second_overflow_barrier_sync.abort();
2021   _has_aborted = true;
2022 
2023   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2024   satb_mq_set.abandon_partial_marking();
2025   // This can be called either during or outside marking, we'll read
2026   // the expected_active value from the SATB queue set.
2027   satb_mq_set.set_active_all_threads(
2028                                  false, /* new active value */
2029                                  satb_mq_set.is_active() /* expected_active */);
2030 }
2031 
2032 static void print_ms_time_info(const char* prefix, const char* name,
2033                                NumberSeq& ns) {
2034   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2035                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2036   if (ns.num() > 0) {
2037     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2038                            prefix, ns.sd(), ns.maximum());
2039   }
2040 }
2041 
2042 void G1ConcurrentMark::print_summary_info() {
2043   Log(gc, marking) log;
2044   if (!log.is_trace()) {
2045     return;
2046   }
2047 
2048   log.trace(" Concurrent marking:");
2049   print_ms_time_info("  ", "init marks", _init_times);
2050   print_ms_time_info("  ", "remarks", _remark_times);
2051   {
2052     print_ms_time_info("     ", "final marks", _remark_mark_times);
2053     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2054 
2055   }
2056   print_ms_time_info("  ", "cleanups", _cleanup_times);
2057   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2058             _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2059   log.trace("  Total stop_world time = %8.2f s.",
2060             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2061   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2062             cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2063 }
2064 
2065 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2066   _concurrent_workers->print_worker_threads_on(st);
2067 }
2068 
2069 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2070   _concurrent_workers->threads_do(tc);
2071 }
2072 
2073 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2074   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2075                p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2076   _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2077   _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2078 }
2079 
2080 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2081   ReferenceProcessor* result = g1h->ref_processor_cm();
2082   assert(result != NULL, "CM reference processor should not be NULL");
2083   return result;
2084 }
2085 
2086 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2087                                G1CMTask* task)
2088   : MetadataVisitingOopIterateClosure(get_cm_oop_closure_ref_processor(g1h)),
2089     _g1h(g1h), _task(task)
2090 { }
2091 
2092 void G1CMTask::setup_for_region(HeapRegion* hr) {
2093   assert(hr != NULL,
2094         "claim_region() should have filtered out NULL regions");
2095   _curr_region  = hr;
2096   _finger       = hr->bottom();
2097   update_region_limit();
2098 }
2099 
2100 void G1CMTask::update_region_limit() {
2101   HeapRegion* hr            = _curr_region;
2102   HeapWord* bottom          = hr->bottom();
2103   HeapWord* limit           = hr->next_top_at_mark_start();
2104 
2105   if (limit == bottom) {
2106     // The region was collected underneath our feet.
2107     // We set the finger to bottom to ensure that the bitmap
2108     // iteration that will follow this will not do anything.
2109     // (this is not a condition that holds when we set the region up,
2110     // as the region is not supposed to be empty in the first place)
2111     _finger = bottom;
2112   } else if (limit >= _region_limit) {
2113     assert(limit >= _finger, "peace of mind");
2114   } else {
2115     assert(limit < _region_limit, "only way to get here");
2116     // This can happen under some pretty unusual circumstances.  An
2117     // evacuation pause empties the region underneath our feet (NTAMS
2118     // at bottom). We then do some allocation in the region (NTAMS
2119     // stays at bottom), followed by the region being used as a GC
2120     // alloc region (NTAMS will move to top() and the objects
2121     // originally below it will be grayed). All objects now marked in
2122     // the region are explicitly grayed, if below the global finger,
2123     // and we do not need in fact to scan anything else. So, we simply
2124     // set _finger to be limit to ensure that the bitmap iteration
2125     // doesn't do anything.
2126     _finger = limit;
2127   }
2128 
2129   _region_limit = limit;
2130 }
2131 
2132 void G1CMTask::giveup_current_region() {
2133   assert(_curr_region != NULL, "invariant");
2134   clear_region_fields();
2135 }
2136 
2137 void G1CMTask::clear_region_fields() {
2138   // Values for these three fields that indicate that we're not
2139   // holding on to a region.
2140   _curr_region   = NULL;
2141   _finger        = NULL;
2142   _region_limit  = NULL;
2143 }
2144 
2145 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2146   if (cm_oop_closure == NULL) {
2147     assert(_cm_oop_closure != NULL, "invariant");
2148   } else {
2149     assert(_cm_oop_closure == NULL, "invariant");
2150   }
2151   _cm_oop_closure = cm_oop_closure;
2152 }
2153 
2154 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2155   guarantee(next_mark_bitmap != NULL, "invariant");
2156   _next_mark_bitmap              = next_mark_bitmap;
2157   clear_region_fields();
2158 
2159   _calls                         = 0;
2160   _elapsed_time_ms               = 0.0;
2161   _termination_time_ms           = 0.0;
2162   _termination_start_time_ms     = 0.0;
2163 
2164   _mark_stats_cache.reset();
2165 }
2166 
2167 bool G1CMTask::should_exit_termination() {
2168   regular_clock_call();
2169   // This is called when we are in the termination protocol. We should
2170   // quit if, for some reason, this task wants to abort or the global
2171   // stack is not empty (this means that we can get work from it).
2172   return !_cm->mark_stack_empty() || has_aborted();
2173 }
2174 
2175 void G1CMTask::reached_limit() {
2176   assert(_words_scanned >= _words_scanned_limit ||
2177          _refs_reached >= _refs_reached_limit ,
2178          "shouldn't have been called otherwise");
2179   regular_clock_call();
2180 }
2181 
2182 void G1CMTask::regular_clock_call() {
2183   if (has_aborted()) {
2184     return;
2185   }
2186 
2187   // First, we need to recalculate the words scanned and refs reached
2188   // limits for the next clock call.
2189   recalculate_limits();
2190 
2191   // During the regular clock call we do the following
2192 
2193   // (1) If an overflow has been flagged, then we abort.
2194   if (_cm->has_overflown()) {
2195     set_has_aborted();
2196     return;
2197   }
2198 
2199   // If we are not concurrent (i.e. we're doing remark) we don't need
2200   // to check anything else. The other steps are only needed during
2201   // the concurrent marking phase.
2202   if (!_cm->concurrent()) {
2203     return;
2204   }
2205 
2206   // (2) If marking has been aborted for Full GC, then we also abort.
2207   if (_cm->has_aborted()) {
2208     set_has_aborted();
2209     return;
2210   }
2211 
2212   double curr_time_ms = os::elapsedVTime() * 1000.0;
2213 
2214   // (4) We check whether we should yield. If we have to, then we abort.
2215   if (SuspendibleThreadSet::should_yield()) {
2216     // We should yield. To do this we abort the task. The caller is
2217     // responsible for yielding.
2218     set_has_aborted();
2219     return;
2220   }
2221 
2222   // (5) We check whether we've reached our time quota. If we have,
2223   // then we abort.
2224   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2225   if (elapsed_time_ms > _time_target_ms) {
2226     set_has_aborted();
2227     _has_timed_out = true;
2228     return;
2229   }
2230 
2231   // (6) Finally, we check whether there are enough completed STAB
2232   // buffers available for processing. If there are, we abort.
2233   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2234   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2235     // we do need to process SATB buffers, we'll abort and restart
2236     // the marking task to do so
2237     set_has_aborted();
2238     return;
2239   }
2240 }
2241 
2242 void G1CMTask::recalculate_limits() {
2243   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2244   _words_scanned_limit      = _real_words_scanned_limit;
2245 
2246   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2247   _refs_reached_limit       = _real_refs_reached_limit;
2248 }
2249 
2250 void G1CMTask::decrease_limits() {
2251   // This is called when we believe that we're going to do an infrequent
2252   // operation which will increase the per byte scanned cost (i.e. move
2253   // entries to/from the global stack). It basically tries to decrease the
2254   // scanning limit so that the clock is called earlier.
2255 
2256   _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2257   _refs_reached_limit  = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2258 }
2259 
2260 void G1CMTask::move_entries_to_global_stack() {
2261   // Local array where we'll store the entries that will be popped
2262   // from the local queue.
2263   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2264 
2265   size_t n = 0;
2266   G1TaskQueueEntry task_entry;
2267   while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2268     buffer[n] = task_entry;
2269     ++n;
2270   }
2271   if (n < G1CMMarkStack::EntriesPerChunk) {
2272     buffer[n] = G1TaskQueueEntry();
2273   }
2274 
2275   if (n > 0) {
2276     if (!_cm->mark_stack_push(buffer)) {
2277       set_has_aborted();
2278     }
2279   }
2280 
2281   // This operation was quite expensive, so decrease the limits.
2282   decrease_limits();
2283 }
2284 
2285 bool G1CMTask::get_entries_from_global_stack() {
2286   // Local array where we'll store the entries that will be popped
2287   // from the global stack.
2288   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2289 
2290   if (!_cm->mark_stack_pop(buffer)) {
2291     return false;
2292   }
2293 
2294   // We did actually pop at least one entry.
2295   for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2296     G1TaskQueueEntry task_entry = buffer[i];
2297     if (task_entry.is_null()) {
2298       break;
2299     }
2300     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()));
2301     bool success = _task_queue->push(task_entry);
2302     // We only call this when the local queue is empty or under a
2303     // given target limit. So, we do not expect this push to fail.
2304     assert(success, "invariant");
2305   }
2306 
2307   // This operation was quite expensive, so decrease the limits
2308   decrease_limits();
2309   return true;
2310 }
2311 
2312 void G1CMTask::drain_local_queue(bool partially) {
2313   if (has_aborted()) {
2314     return;
2315   }
2316 
2317   // Decide what the target size is, depending whether we're going to
2318   // drain it partially (so that other tasks can steal if they run out
2319   // of things to do) or totally (at the very end).
2320   size_t target_size;
2321   if (partially) {
2322     target_size = MIN2((size_t)_task_queue->max_elems()/3, (size_t)GCDrainStackTargetSize);
2323   } else {
2324     target_size = 0;
2325   }
2326 
2327   if (_task_queue->size() > target_size) {
2328     G1TaskQueueEntry entry;
2329     bool ret = _task_queue->pop_local(entry);
2330     while (ret) {
2331       scan_task_entry(entry);
2332       if (_task_queue->size() <= target_size || has_aborted()) {
2333         ret = false;
2334       } else {
2335         ret = _task_queue->pop_local(entry);
2336       }
2337     }
2338   }
2339 }
2340 
2341 void G1CMTask::drain_global_stack(bool partially) {
2342   if (has_aborted()) {
2343     return;
2344   }
2345 
2346   // We have a policy to drain the local queue before we attempt to
2347   // drain the global stack.
2348   assert(partially || _task_queue->size() == 0, "invariant");
2349 
2350   // Decide what the target size is, depending whether we're going to
2351   // drain it partially (so that other tasks can steal if they run out
2352   // of things to do) or totally (at the very end).
2353   // Notice that when draining the global mark stack partially, due to the racyness
2354   // of the mark stack size update we might in fact drop below the target. But,
2355   // this is not a problem.
2356   // In case of total draining, we simply process until the global mark stack is
2357   // totally empty, disregarding the size counter.
2358   if (partially) {
2359     size_t const target_size = _cm->partial_mark_stack_size_target();
2360     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2361       if (get_entries_from_global_stack()) {
2362         drain_local_queue(partially);
2363       }
2364     }
2365   } else {
2366     while (!has_aborted() && get_entries_from_global_stack()) {
2367       drain_local_queue(partially);
2368     }
2369   }
2370 }
2371 
2372 // SATB Queue has several assumptions on whether to call the par or
2373 // non-par versions of the methods. this is why some of the code is
2374 // replicated. We should really get rid of the single-threaded version
2375 // of the code to simplify things.
2376 void G1CMTask::drain_satb_buffers() {
2377   if (has_aborted()) {
2378     return;
2379   }
2380 
2381   // We set this so that the regular clock knows that we're in the
2382   // middle of draining buffers and doesn't set the abort flag when it
2383   // notices that SATB buffers are available for draining. It'd be
2384   // very counter productive if it did that. :-)
2385   _draining_satb_buffers = true;
2386 
2387   G1CMSATBBufferClosure satb_cl(this, _g1h);
2388   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2389 
2390   // This keeps claiming and applying the closure to completed buffers
2391   // until we run out of buffers or we need to abort.
2392   while (!has_aborted() &&
2393          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2394     regular_clock_call();
2395   }
2396 
2397   _draining_satb_buffers = false;
2398 
2399   assert(has_aborted() ||
2400          _cm->concurrent() ||
2401          satb_mq_set.completed_buffers_num() == 0, "invariant");
2402 
2403   // again, this was a potentially expensive operation, decrease the
2404   // limits to get the regular clock call early
2405   decrease_limits();
2406 }
2407 
2408 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2409   _mark_stats_cache.reset(region_idx);
2410 }
2411 
2412 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2413   return _mark_stats_cache.evict_all();
2414 }
2415 
2416 void G1CMTask::print_stats() {
2417   log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2418   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2419                        _elapsed_time_ms, _termination_time_ms);
2420   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2421                        _step_times_ms.num(),
2422                        _step_times_ms.avg(),
2423                        _step_times_ms.sd(),
2424                        _step_times_ms.maximum(),
2425                        _step_times_ms.sum());
2426   size_t const hits = _mark_stats_cache.hits();
2427   size_t const misses = _mark_stats_cache.misses();
2428   log_debug(gc, stats)("  Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2429                        hits, misses, percent_of(hits, hits + misses));
2430 }
2431 
2432 bool G1ConcurrentMark::try_stealing(uint worker_id, G1TaskQueueEntry& task_entry) {
2433   return _task_queues->steal(worker_id, task_entry);
2434 }
2435 
2436 /*****************************************************************************
2437 
2438     The do_marking_step(time_target_ms, ...) method is the building
2439     block of the parallel marking framework. It can be called in parallel
2440     with other invocations of do_marking_step() on different tasks
2441     (but only one per task, obviously) and concurrently with the
2442     mutator threads, or during remark, hence it eliminates the need
2443     for two versions of the code. When called during remark, it will
2444     pick up from where the task left off during the concurrent marking
2445     phase. Interestingly, tasks are also claimable during evacuation
2446     pauses too, since do_marking_step() ensures that it aborts before
2447     it needs to yield.
2448 
2449     The data structures that it uses to do marking work are the
2450     following:
2451 
2452       (1) Marking Bitmap. If there are gray objects that appear only
2453       on the bitmap (this happens either when dealing with an overflow
2454       or when the initial marking phase has simply marked the roots
2455       and didn't push them on the stack), then tasks claim heap
2456       regions whose bitmap they then scan to find gray objects. A
2457       global finger indicates where the end of the last claimed region
2458       is. A local finger indicates how far into the region a task has
2459       scanned. The two fingers are used to determine how to gray an
2460       object (i.e. whether simply marking it is OK, as it will be
2461       visited by a task in the future, or whether it needs to be also
2462       pushed on a stack).
2463 
2464       (2) Local Queue. The local queue of the task which is accessed
2465       reasonably efficiently by the task. Other tasks can steal from
2466       it when they run out of work. Throughout the marking phase, a
2467       task attempts to keep its local queue short but not totally
2468       empty, so that entries are available for stealing by other
2469       tasks. Only when there is no more work, a task will totally
2470       drain its local queue.
2471 
2472       (3) Global Mark Stack. This handles local queue overflow. During
2473       marking only sets of entries are moved between it and the local
2474       queues, as access to it requires a mutex and more fine-grain
2475       interaction with it which might cause contention. If it
2476       overflows, then the marking phase should restart and iterate
2477       over the bitmap to identify gray objects. Throughout the marking
2478       phase, tasks attempt to keep the global mark stack at a small
2479       length but not totally empty, so that entries are available for
2480       popping by other tasks. Only when there is no more work, tasks
2481       will totally drain the global mark stack.
2482 
2483       (4) SATB Buffer Queue. This is where completed SATB buffers are
2484       made available. Buffers are regularly removed from this queue
2485       and scanned for roots, so that the queue doesn't get too
2486       long. During remark, all completed buffers are processed, as
2487       well as the filled in parts of any uncompleted buffers.
2488 
2489     The do_marking_step() method tries to abort when the time target
2490     has been reached. There are a few other cases when the
2491     do_marking_step() method also aborts:
2492 
2493       (1) When the marking phase has been aborted (after a Full GC).
2494 
2495       (2) When a global overflow (on the global stack) has been
2496       triggered. Before the task aborts, it will actually sync up with
2497       the other tasks to ensure that all the marking data structures
2498       (local queues, stacks, fingers etc.)  are re-initialized so that
2499       when do_marking_step() completes, the marking phase can
2500       immediately restart.
2501 
2502       (3) When enough completed SATB buffers are available. The
2503       do_marking_step() method only tries to drain SATB buffers right
2504       at the beginning. So, if enough buffers are available, the
2505       marking step aborts and the SATB buffers are processed at
2506       the beginning of the next invocation.
2507 
2508       (4) To yield. when we have to yield then we abort and yield
2509       right at the end of do_marking_step(). This saves us from a lot
2510       of hassle as, by yielding we might allow a Full GC. If this
2511       happens then objects will be compacted underneath our feet, the
2512       heap might shrink, etc. We save checking for this by just
2513       aborting and doing the yield right at the end.
2514 
2515     From the above it follows that the do_marking_step() method should
2516     be called in a loop (or, otherwise, regularly) until it completes.
2517 
2518     If a marking step completes without its has_aborted() flag being
2519     true, it means it has completed the current marking phase (and
2520     also all other marking tasks have done so and have all synced up).
2521 
2522     A method called regular_clock_call() is invoked "regularly" (in
2523     sub ms intervals) throughout marking. It is this clock method that
2524     checks all the abort conditions which were mentioned above and
2525     decides when the task should abort. A work-based scheme is used to
2526     trigger this clock method: when the number of object words the
2527     marking phase has scanned or the number of references the marking
2528     phase has visited reach a given limit. Additional invocations to
2529     the method clock have been planted in a few other strategic places
2530     too. The initial reason for the clock method was to avoid calling
2531     vtime too regularly, as it is quite expensive. So, once it was in
2532     place, it was natural to piggy-back all the other conditions on it
2533     too and not constantly check them throughout the code.
2534 
2535     If do_termination is true then do_marking_step will enter its
2536     termination protocol.
2537 
2538     The value of is_serial must be true when do_marking_step is being
2539     called serially (i.e. by the VMThread) and do_marking_step should
2540     skip any synchronization in the termination and overflow code.
2541     Examples include the serial remark code and the serial reference
2542     processing closures.
2543 
2544     The value of is_serial must be false when do_marking_step is
2545     being called by any of the worker threads in a work gang.
2546     Examples include the concurrent marking code (CMMarkingTask),
2547     the MT remark code, and the MT reference processing closures.
2548 
2549  *****************************************************************************/
2550 
2551 void G1CMTask::do_marking_step(double time_target_ms,
2552                                bool do_termination,
2553                                bool is_serial) {
2554   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2555 
2556   _start_time_ms = os::elapsedVTime() * 1000.0;
2557 
2558   // If do_stealing is true then do_marking_step will attempt to
2559   // steal work from the other G1CMTasks. It only makes sense to
2560   // enable stealing when the termination protocol is enabled
2561   // and do_marking_step() is not being called serially.
2562   bool do_stealing = do_termination && !is_serial;
2563 
2564   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2565   _time_target_ms = time_target_ms - diff_prediction_ms;
2566 
2567   // set up the variables that are used in the work-based scheme to
2568   // call the regular clock method
2569   _words_scanned = 0;
2570   _refs_reached  = 0;
2571   recalculate_limits();
2572 
2573   // clear all flags
2574   clear_has_aborted();
2575   _has_timed_out = false;
2576   _draining_satb_buffers = false;
2577 
2578   ++_calls;
2579 
2580   // Set up the bitmap and oop closures. Anything that uses them is
2581   // eventually called from this method, so it is OK to allocate these
2582   // statically.
2583   G1CMBitMapClosure bitmap_closure(this, _cm);
2584   G1CMOopClosure cm_oop_closure(_g1h, this);
2585   set_cm_oop_closure(&cm_oop_closure);
2586 
2587   if (_cm->has_overflown()) {
2588     // This can happen if the mark stack overflows during a GC pause
2589     // and this task, after a yield point, restarts. We have to abort
2590     // as we need to get into the overflow protocol which happens
2591     // right at the end of this task.
2592     set_has_aborted();
2593   }
2594 
2595   // First drain any available SATB buffers. After this, we will not
2596   // look at SATB buffers before the next invocation of this method.
2597   // If enough completed SATB buffers are queued up, the regular clock
2598   // will abort this task so that it restarts.
2599   drain_satb_buffers();
2600   // ...then partially drain the local queue and the global stack
2601   drain_local_queue(true);
2602   drain_global_stack(true);
2603 
2604   do {
2605     if (!has_aborted() && _curr_region != NULL) {
2606       // This means that we're already holding on to a region.
2607       assert(_finger != NULL, "if region is not NULL, then the finger "
2608              "should not be NULL either");
2609 
2610       // We might have restarted this task after an evacuation pause
2611       // which might have evacuated the region we're holding on to
2612       // underneath our feet. Let's read its limit again to make sure
2613       // that we do not iterate over a region of the heap that
2614       // contains garbage (update_region_limit() will also move
2615       // _finger to the start of the region if it is found empty).
2616       update_region_limit();
2617       // We will start from _finger not from the start of the region,
2618       // as we might be restarting this task after aborting half-way
2619       // through scanning this region. In this case, _finger points to
2620       // the address where we last found a marked object. If this is a
2621       // fresh region, _finger points to start().
2622       MemRegion mr = MemRegion(_finger, _region_limit);
2623 
2624       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2625              "humongous regions should go around loop once only");
2626 
2627       // Some special cases:
2628       // If the memory region is empty, we can just give up the region.
2629       // If the current region is humongous then we only need to check
2630       // the bitmap for the bit associated with the start of the object,
2631       // scan the object if it's live, and give up the region.
2632       // Otherwise, let's iterate over the bitmap of the part of the region
2633       // that is left.
2634       // If the iteration is successful, give up the region.
2635       if (mr.is_empty()) {
2636         giveup_current_region();
2637         regular_clock_call();
2638       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2639         if (_next_mark_bitmap->is_marked(mr.start())) {
2640           // The object is marked - apply the closure
2641           bitmap_closure.do_addr(mr.start());
2642         }
2643         // Even if this task aborted while scanning the humongous object
2644         // we can (and should) give up the current region.
2645         giveup_current_region();
2646         regular_clock_call();
2647       } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2648         giveup_current_region();
2649         regular_clock_call();
2650       } else {
2651         assert(has_aborted(), "currently the only way to do so");
2652         // The only way to abort the bitmap iteration is to return
2653         // false from the do_bit() method. However, inside the
2654         // do_bit() method we move the _finger to point to the
2655         // object currently being looked at. So, if we bail out, we
2656         // have definitely set _finger to something non-null.
2657         assert(_finger != NULL, "invariant");
2658 
2659         // Region iteration was actually aborted. So now _finger
2660         // points to the address of the object we last scanned. If we
2661         // leave it there, when we restart this task, we will rescan
2662         // the object. It is easy to avoid this. We move the finger by
2663         // enough to point to the next possible object header.
2664         assert(_finger < _region_limit, "invariant");
2665         HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2666         // Check if bitmap iteration was aborted while scanning the last object
2667         if (new_finger >= _region_limit) {
2668           giveup_current_region();
2669         } else {
2670           move_finger_to(new_finger);
2671         }
2672       }
2673     }
2674     // At this point we have either completed iterating over the
2675     // region we were holding on to, or we have aborted.
2676 
2677     // We then partially drain the local queue and the global stack.
2678     // (Do we really need this?)
2679     drain_local_queue(true);
2680     drain_global_stack(true);
2681 
2682     // Read the note on the claim_region() method on why it might
2683     // return NULL with potentially more regions available for
2684     // claiming and why we have to check out_of_regions() to determine
2685     // whether we're done or not.
2686     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2687       // We are going to try to claim a new region. We should have
2688       // given up on the previous one.
2689       // Separated the asserts so that we know which one fires.
2690       assert(_curr_region  == NULL, "invariant");
2691       assert(_finger       == NULL, "invariant");
2692       assert(_region_limit == NULL, "invariant");
2693       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2694       if (claimed_region != NULL) {
2695         // Yes, we managed to claim one
2696         setup_for_region(claimed_region);
2697         assert(_curr_region == claimed_region, "invariant");
2698       }
2699       // It is important to call the regular clock here. It might take
2700       // a while to claim a region if, for example, we hit a large
2701       // block of empty regions. So we need to call the regular clock
2702       // method once round the loop to make sure it's called
2703       // frequently enough.
2704       regular_clock_call();
2705     }
2706 
2707     if (!has_aborted() && _curr_region == NULL) {
2708       assert(_cm->out_of_regions(),
2709              "at this point we should be out of regions");
2710     }
2711   } while ( _curr_region != NULL && !has_aborted());
2712 
2713   if (!has_aborted()) {
2714     // We cannot check whether the global stack is empty, since other
2715     // tasks might be pushing objects to it concurrently.
2716     assert(_cm->out_of_regions(),
2717            "at this point we should be out of regions");
2718     // Try to reduce the number of available SATB buffers so that
2719     // remark has less work to do.
2720     drain_satb_buffers();
2721   }
2722 
2723   // Since we've done everything else, we can now totally drain the
2724   // local queue and global stack.
2725   drain_local_queue(false);
2726   drain_global_stack(false);
2727 
2728   // Attempt at work stealing from other task's queues.
2729   if (do_stealing && !has_aborted()) {
2730     // We have not aborted. This means that we have finished all that
2731     // we could. Let's try to do some stealing...
2732 
2733     // We cannot check whether the global stack is empty, since other
2734     // tasks might be pushing objects to it concurrently.
2735     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2736            "only way to reach here");
2737     while (!has_aborted()) {
2738       G1TaskQueueEntry entry;
2739       if (_cm->try_stealing(_worker_id, entry)) {
2740         scan_task_entry(entry);
2741 
2742         // And since we're towards the end, let's totally drain the
2743         // local queue and global stack.
2744         drain_local_queue(false);
2745         drain_global_stack(false);
2746       } else {
2747         break;
2748       }
2749     }
2750   }
2751 
2752   // We still haven't aborted. Now, let's try to get into the
2753   // termination protocol.
2754   if (do_termination && !has_aborted()) {
2755     // We cannot check whether the global stack is empty, since other
2756     // tasks might be concurrently pushing objects on it.
2757     // Separated the asserts so that we know which one fires.
2758     assert(_cm->out_of_regions(), "only way to reach here");
2759     assert(_task_queue->size() == 0, "only way to reach here");
2760     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2761 
2762     // The G1CMTask class also extends the TerminatorTerminator class,
2763     // hence its should_exit_termination() method will also decide
2764     // whether to exit the termination protocol or not.
2765     bool finished = (is_serial ||
2766                      _cm->terminator()->offer_termination(this));
2767     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2768     _termination_time_ms +=
2769       termination_end_time_ms - _termination_start_time_ms;
2770 
2771     if (finished) {
2772       // We're all done.
2773 
2774       // We can now guarantee that the global stack is empty, since
2775       // all other tasks have finished. We separated the guarantees so
2776       // that, if a condition is false, we can immediately find out
2777       // which one.
2778       guarantee(_cm->out_of_regions(), "only way to reach here");
2779       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2780       guarantee(_task_queue->size() == 0, "only way to reach here");
2781       guarantee(!_cm->has_overflown(), "only way to reach here");
2782     } else {
2783       // Apparently there's more work to do. Let's abort this task. It
2784       // will restart it and we can hopefully find more things to do.
2785       set_has_aborted();
2786     }
2787   }
2788 
2789   // Mainly for debugging purposes to make sure that a pointer to the
2790   // closure which was statically allocated in this frame doesn't
2791   // escape it by accident.
2792   set_cm_oop_closure(NULL);
2793   double end_time_ms = os::elapsedVTime() * 1000.0;
2794   double elapsed_time_ms = end_time_ms - _start_time_ms;
2795   // Update the step history.
2796   _step_times_ms.add(elapsed_time_ms);
2797 
2798   if (has_aborted()) {
2799     // The task was aborted for some reason.
2800     if (_has_timed_out) {
2801       double diff_ms = elapsed_time_ms - _time_target_ms;
2802       // Keep statistics of how well we did with respect to hitting
2803       // our target only if we actually timed out (if we aborted for
2804       // other reasons, then the results might get skewed).
2805       _marking_step_diffs_ms.add(diff_ms);
2806     }
2807 
2808     if (_cm->has_overflown()) {
2809       // This is the interesting one. We aborted because a global
2810       // overflow was raised. This means we have to restart the
2811       // marking phase and start iterating over regions. However, in
2812       // order to do this we have to make sure that all tasks stop
2813       // what they are doing and re-initialize in a safe manner. We
2814       // will achieve this with the use of two barrier sync points.
2815 
2816       if (!is_serial) {
2817         // We only need to enter the sync barrier if being called
2818         // from a parallel context
2819         _cm->enter_first_sync_barrier(_worker_id);
2820 
2821         // When we exit this sync barrier we know that all tasks have
2822         // stopped doing marking work. So, it's now safe to
2823         // re-initialize our data structures.
2824       }
2825 
2826       clear_region_fields();
2827       flush_mark_stats_cache();
2828 
2829       if (!is_serial) {
2830         // If we're executing the concurrent phase of marking, reset the marking
2831         // state; otherwise the marking state is reset after reference processing,
2832         // during the remark pause.
2833         // If we reset here as a result of an overflow during the remark we will
2834         // see assertion failures from any subsequent set_concurrency_and_phase()
2835         // calls.
2836         if (_cm->concurrent() && _worker_id == 0) {
2837           // Worker 0 is responsible for clearing the global data structures because
2838           // of an overflow. During STW we should not clear the overflow flag (in
2839           // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2840           // method to abort the pause and restart concurrent marking.
2841           _cm->reset_marking_for_restart();
2842 
2843           log_info(gc, marking)("Concurrent Mark reset for overflow");
2844         }
2845 
2846         // ...and enter the second barrier.
2847         _cm->enter_second_sync_barrier(_worker_id);
2848       }
2849       // At this point, if we're during the concurrent phase of
2850       // marking, everything has been re-initialized and we're
2851       // ready to restart.
2852     }
2853   }
2854 }
2855 
2856 G1CMTask::G1CMTask(uint worker_id,
2857                    G1ConcurrentMark* cm,
2858                    G1CMTaskQueue* task_queue,
2859                    G1RegionMarkStats* mark_stats,
2860                    uint max_regions) :
2861   _objArray_processor(this),
2862   _worker_id(worker_id),
2863   _g1h(G1CollectedHeap::heap()),
2864   _cm(cm),
2865   _next_mark_bitmap(NULL),
2866   _task_queue(task_queue),
2867   _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize),
2868   _calls(0),
2869   _time_target_ms(0.0),
2870   _start_time_ms(0.0),
2871   _cm_oop_closure(NULL),
2872   _curr_region(NULL),
2873   _finger(NULL),
2874   _region_limit(NULL),
2875   _words_scanned(0),
2876   _words_scanned_limit(0),
2877   _real_words_scanned_limit(0),
2878   _refs_reached(0),
2879   _refs_reached_limit(0),
2880   _real_refs_reached_limit(0),
2881   _has_aborted(false),
2882   _has_timed_out(false),
2883   _draining_satb_buffers(false),
2884   _step_times_ms(),
2885   _elapsed_time_ms(0.0),
2886   _termination_time_ms(0.0),
2887   _termination_start_time_ms(0.0),
2888   _marking_step_diffs_ms()
2889 {
2890   guarantee(task_queue != NULL, "invariant");
2891 
2892   _marking_step_diffs_ms.add(0.5);
2893 }
2894 
2895 // These are formatting macros that are used below to ensure
2896 // consistent formatting. The *_H_* versions are used to format the
2897 // header for a particular value and they should be kept consistent
2898 // with the corresponding macro. Also note that most of the macros add
2899 // the necessary white space (as a prefix) which makes them a bit
2900 // easier to compose.
2901 
2902 // All the output lines are prefixed with this string to be able to
2903 // identify them easily in a large log file.
2904 #define G1PPRL_LINE_PREFIX            "###"
2905 
2906 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2907 #ifdef _LP64
2908 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2909 #else // _LP64
2910 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2911 #endif // _LP64
2912 
2913 // For per-region info
2914 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2915 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2916 #define G1PPRL_STATE_FORMAT           "   %-5s"
2917 #define G1PPRL_STATE_H_FORMAT         "   %5s"
2918 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2919 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2920 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2921 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2922 
2923 // For summary info
2924 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2925 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2926 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2927 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2928 
2929 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2930   _total_used_bytes(0), _total_capacity_bytes(0),
2931   _total_prev_live_bytes(0), _total_next_live_bytes(0),
2932   _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2933 {
2934   if (!log_is_enabled(Trace, gc, liveness)) {
2935     return;
2936   }
2937 
2938   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2939   MemRegion g1_reserved = g1h->g1_reserved();
2940   double now = os::elapsedTime();
2941 
2942   // Print the header of the output.
2943   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2944   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2945                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2946                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2947                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2948                           HeapRegion::GrainBytes);
2949   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2950   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2951                           G1PPRL_TYPE_H_FORMAT
2952                           G1PPRL_ADDR_BASE_H_FORMAT
2953                           G1PPRL_BYTE_H_FORMAT
2954                           G1PPRL_BYTE_H_FORMAT
2955                           G1PPRL_BYTE_H_FORMAT
2956                           G1PPRL_DOUBLE_H_FORMAT
2957                           G1PPRL_BYTE_H_FORMAT
2958                           G1PPRL_STATE_H_FORMAT
2959                           G1PPRL_BYTE_H_FORMAT,
2960                           "type", "address-range",
2961                           "used", "prev-live", "next-live", "gc-eff",
2962                           "remset", "state", "code-roots");
2963   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2964                           G1PPRL_TYPE_H_FORMAT
2965                           G1PPRL_ADDR_BASE_H_FORMAT
2966                           G1PPRL_BYTE_H_FORMAT
2967                           G1PPRL_BYTE_H_FORMAT
2968                           G1PPRL_BYTE_H_FORMAT
2969                           G1PPRL_DOUBLE_H_FORMAT
2970                           G1PPRL_BYTE_H_FORMAT
2971                           G1PPRL_STATE_H_FORMAT
2972                           G1PPRL_BYTE_H_FORMAT,
2973                           "", "",
2974                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
2975                           "(bytes)", "", "(bytes)");
2976 }
2977 
2978 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
2979   if (!log_is_enabled(Trace, gc, liveness)) {
2980     return false;
2981   }
2982 
2983   const char* type       = r->get_type_str();
2984   HeapWord* bottom       = r->bottom();
2985   HeapWord* end          = r->end();
2986   size_t capacity_bytes  = r->capacity();
2987   size_t used_bytes      = r->used();
2988   size_t prev_live_bytes = r->live_bytes();
2989   size_t next_live_bytes = r->next_live_bytes();
2990   double gc_eff          = r->gc_efficiency();
2991   size_t remset_bytes    = r->rem_set()->mem_size();
2992   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
2993   const char* remset_type = r->rem_set()->get_short_state_str();
2994 
2995   _total_used_bytes      += used_bytes;
2996   _total_capacity_bytes  += capacity_bytes;
2997   _total_prev_live_bytes += prev_live_bytes;
2998   _total_next_live_bytes += next_live_bytes;
2999   _total_remset_bytes    += remset_bytes;
3000   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3001 
3002   // Print a line for this particular region.
3003   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3004                           G1PPRL_TYPE_FORMAT
3005                           G1PPRL_ADDR_BASE_FORMAT
3006                           G1PPRL_BYTE_FORMAT
3007                           G1PPRL_BYTE_FORMAT
3008                           G1PPRL_BYTE_FORMAT
3009                           G1PPRL_DOUBLE_FORMAT
3010                           G1PPRL_BYTE_FORMAT
3011                           G1PPRL_STATE_FORMAT
3012                           G1PPRL_BYTE_FORMAT,
3013                           type, p2i(bottom), p2i(end),
3014                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3015                           remset_bytes, remset_type, strong_code_roots_bytes);
3016 
3017   return false;
3018 }
3019 
3020 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3021   if (!log_is_enabled(Trace, gc, liveness)) {
3022     return;
3023   }
3024 
3025   // add static memory usages to remembered set sizes
3026   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3027   // Print the footer of the output.
3028   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3029   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3030                          " SUMMARY"
3031                          G1PPRL_SUM_MB_FORMAT("capacity")
3032                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3033                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3034                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3035                          G1PPRL_SUM_MB_FORMAT("remset")
3036                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3037                          bytes_to_mb(_total_capacity_bytes),
3038                          bytes_to_mb(_total_used_bytes),
3039                          percent_of(_total_used_bytes, _total_capacity_bytes),
3040                          bytes_to_mb(_total_prev_live_bytes),
3041                          percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3042                          bytes_to_mb(_total_next_live_bytes),
3043                          percent_of(_total_next_live_bytes, _total_capacity_bytes),
3044                          bytes_to_mb(_total_remset_bytes),
3045                          bytes_to_mb(_total_strong_code_roots_bytes));
3046 }