1 /*
   2  * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/metadataOnStackMark.hpp"
  27 #include "classfile/symbolTable.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc/g1/concurrentMarkThread.inline.hpp"
  30 #include "gc/g1/g1CollectedHeap.inline.hpp"
  31 #include "gc/g1/g1CollectorState.hpp"
  32 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  33 #include "gc/g1/g1HeapVerifier.hpp"
  34 #include "gc/g1/g1OopClosures.inline.hpp"
  35 #include "gc/g1/g1CardLiveData.inline.hpp"
  36 #include "gc/g1/g1Policy.hpp"
  37 #include "gc/g1/g1StringDedup.hpp"
  38 #include "gc/g1/heapRegion.inline.hpp"
  39 #include "gc/g1/heapRegionRemSet.hpp"
  40 #include "gc/g1/heapRegionSet.inline.hpp"
  41 #include "gc/g1/suspendibleThreadSet.hpp"
  42 #include "gc/shared/gcId.hpp"
  43 #include "gc/shared/gcTimer.hpp"
  44 #include "gc/shared/gcTrace.hpp"
  45 #include "gc/shared/gcTraceTime.inline.hpp"
  46 #include "gc/shared/genOopClosures.inline.hpp"
  47 #include "gc/shared/referencePolicy.hpp"
  48 #include "gc/shared/strongRootsScope.hpp"
  49 #include "gc/shared/taskqueue.inline.hpp"
  50 #include "gc/shared/vmGCOperations.hpp"
  51 #include "logging/log.hpp"
  52 #include "memory/allocation.hpp"
  53 #include "memory/resourceArea.hpp"
  54 #include "oops/oop.inline.hpp"
  55 #include "runtime/atomic.hpp"
  56 #include "runtime/handles.inline.hpp"
  57 #include "runtime/java.hpp"
  58 #include "runtime/prefetch.inline.hpp"
  59 #include "services/memTracker.hpp"
  60 #include "utilities/align.hpp"
  61 #include "utilities/growableArray.hpp"
  62 
  63 void G1CMBitMap::print_on_error(outputStream* st, const char* prefix) const {
  64   _bm.print_on_error(st, prefix);
  65 }
  66 
  67 size_t G1CMBitMap::compute_size(size_t heap_size) {
  68   return ReservedSpace::allocation_align_size_up(heap_size / mark_distance());
  69 }
  70 
  71 size_t G1CMBitMap::mark_distance() {
  72   return MinObjAlignmentInBytes * BitsPerByte;
  73 }
  74 
  75 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) {
  76   _covered = heap;
  77 
  78   _bm = BitMapView((BitMap::bm_word_t*) storage->reserved().start(), _covered.word_size() >> _shifter);
  79 
  80   storage->set_mapping_changed_listener(&_listener);
  81 }
  82 
  83 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) {
  84   if (zero_filled) {
  85     return;
  86   }
  87   // We need to clear the bitmap on commit, removing any existing information.
  88   MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords);
  89   _bm->clear_range(mr);
  90 }
  91 
  92 void G1CMBitMap::clear_range(MemRegion mr) {
  93   MemRegion intersection = mr.intersection(_covered);
  94   assert(!intersection.is_empty(), "Given range from " PTR_FORMAT " to " PTR_FORMAT " is completely outside the heap", p2i(mr.start()), p2i(mr.end()));
  95   // convert address range into offset range
  96   _bm.at_put_range(addr_to_offset(intersection.start()),
  97                    addr_to_offset(intersection.end()), false);
  98 }
  99 
 100 G1CMMarkStack::G1CMMarkStack() :
 101   _max_chunk_capacity(0),
 102   _base(NULL),
 103   _chunk_capacity(0) {
 104   set_empty();
 105 }
 106 
 107 bool G1CMMarkStack::resize(size_t new_capacity) {
 108   assert(is_empty(), "Only resize when stack is empty.");
 109   assert(new_capacity <= _max_chunk_capacity,
 110          "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
 111 
 112   TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
 113 
 114   if (new_base == NULL) {
 115     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));
 116     return false;
 117   }
 118   // Release old mapping.
 119   if (_base != NULL) {
 120     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 121   }
 122 
 123   _base = new_base;
 124   _chunk_capacity = new_capacity;
 125   set_empty();
 126 
 127   return true;
 128 }
 129 
 130 size_t G1CMMarkStack::capacity_alignment() {
 131   return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
 132 }
 133 
 134 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
 135   guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
 136 
 137   size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
 138 
 139   _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 140   size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 141 
 142   guarantee(initial_chunk_capacity <= _max_chunk_capacity,
 143             "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
 144             _max_chunk_capacity,
 145             initial_chunk_capacity);
 146 
 147   log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
 148                 initial_chunk_capacity, _max_chunk_capacity);
 149 
 150   return resize(initial_chunk_capacity);
 151 }
 152 
 153 void G1CMMarkStack::expand() {
 154   if (_chunk_capacity == _max_chunk_capacity) {
 155     log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
 156     return;
 157   }
 158   size_t old_capacity = _chunk_capacity;
 159   // Double capacity if possible
 160   size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
 161 
 162   if (resize(new_capacity)) {
 163     log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 164                   old_capacity, new_capacity);
 165   } else {
 166     log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 167                     old_capacity, new_capacity);
 168   }
 169 }
 170 
 171 G1CMMarkStack::~G1CMMarkStack() {
 172   if (_base != NULL) {
 173     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 174   }
 175 }
 176 
 177 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
 178   elem->next = *list;
 179   *list = elem;
 180 }
 181 
 182 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
 183   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 184   add_chunk_to_list(&_chunk_list, elem);
 185   _chunks_in_chunk_list++;
 186 }
 187 
 188 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
 189   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 190   add_chunk_to_list(&_free_list, elem);
 191 }
 192 
 193 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
 194   TaskQueueEntryChunk* result = *list;
 195   if (result != NULL) {
 196     *list = (*list)->next;
 197   }
 198   return result;
 199 }
 200 
 201 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
 202   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 203   TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
 204   if (result != NULL) {
 205     _chunks_in_chunk_list--;
 206   }
 207   return result;
 208 }
 209 
 210 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
 211   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 212   return remove_chunk_from_list(&_free_list);
 213 }
 214 
 215 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
 216   // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
 217   // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
 218   // wraparound of _hwm.
 219   if (_hwm >= _chunk_capacity) {
 220     return NULL;
 221   }
 222 
 223   size_t cur_idx = Atomic::add(1, &_hwm) - 1;
 224   if (cur_idx >= _chunk_capacity) {
 225     return NULL;
 226   }
 227 
 228   TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
 229   result->next = NULL;
 230   return result;
 231 }
 232 
 233 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
 234   // Get a new chunk.
 235   TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
 236 
 237   if (new_chunk == NULL) {
 238     // Did not get a chunk from the free list. Allocate from backing memory.
 239     new_chunk = allocate_new_chunk();
 240 
 241     if (new_chunk == NULL) {
 242       return false;
 243     }
 244   }
 245 
 246   Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 247 
 248   add_chunk_to_chunk_list(new_chunk);
 249 
 250   return true;
 251 }
 252 
 253 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
 254   TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
 255 
 256   if (cur == NULL) {
 257     return false;
 258   }
 259 
 260   Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 261 
 262   add_chunk_to_free_list(cur);
 263   return true;
 264 }
 265 
 266 void G1CMMarkStack::set_empty() {
 267   _chunks_in_chunk_list = 0;
 268   _hwm = 0;
 269   _chunk_list = NULL;
 270   _free_list = NULL;
 271 }
 272 
 273 G1CMRootRegions::G1CMRootRegions() :
 274   _cm(NULL), _scan_in_progress(false),
 275   _should_abort(false), _claimed_survivor_index(0) { }
 276 
 277 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
 278   _survivors = survivors;
 279   _cm = cm;
 280 }
 281 
 282 void G1CMRootRegions::prepare_for_scan() {
 283   assert(!scan_in_progress(), "pre-condition");
 284 
 285   // Currently, only survivors can be root regions.
 286   _claimed_survivor_index = 0;
 287   _scan_in_progress = _survivors->regions()->is_nonempty();
 288   _should_abort = false;
 289 }
 290 
 291 HeapRegion* G1CMRootRegions::claim_next() {
 292   if (_should_abort) {
 293     // If someone has set the should_abort flag, we return NULL to
 294     // force the caller to bail out of their loop.
 295     return NULL;
 296   }
 297 
 298   // Currently, only survivors can be root regions.
 299   const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
 300 
 301   int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
 302   if (claimed_index < survivor_regions->length()) {
 303     return survivor_regions->at(claimed_index);
 304   }
 305   return NULL;
 306 }
 307 
 308 uint G1CMRootRegions::num_root_regions() const {
 309   return (uint)_survivors->regions()->length();
 310 }
 311 
 312 void G1CMRootRegions::notify_scan_done() {
 313   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 314   _scan_in_progress = false;
 315   RootRegionScan_lock->notify_all();
 316 }
 317 
 318 void G1CMRootRegions::cancel_scan() {
 319   notify_scan_done();
 320 }
 321 
 322 void G1CMRootRegions::scan_finished() {
 323   assert(scan_in_progress(), "pre-condition");
 324 
 325   // Currently, only survivors can be root regions.
 326   if (!_should_abort) {
 327     assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
 328     assert((uint)_claimed_survivor_index >= _survivors->length(),
 329            "we should have claimed all survivors, claimed index = %u, length = %u",
 330            (uint)_claimed_survivor_index, _survivors->length());
 331   }
 332 
 333   notify_scan_done();
 334 }
 335 
 336 bool G1CMRootRegions::wait_until_scan_finished() {
 337   if (!scan_in_progress()) return false;
 338 
 339   {
 340     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 341     while (scan_in_progress()) {
 342       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 343     }
 344   }
 345   return true;
 346 }
 347 
 348 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) {
 349   return MAX2((n_par_threads + 2) / 4, 1U);
 350 }
 351 
 352 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) :
 353   _g1h(g1h),
 354   _markBitMap1(),
 355   _markBitMap2(),
 356   _parallel_marking_threads(0),
 357   _max_parallel_marking_threads(0),
 358   _sleep_factor(0.0),
 359   _marking_task_overhead(1.0),
 360   _cleanup_list("Cleanup List"),
 361 
 362   _prevMarkBitMap(&_markBitMap1),
 363   _nextMarkBitMap(&_markBitMap2),
 364 
 365   _global_mark_stack(),
 366   // _finger set in set_non_marking_state
 367 
 368   _max_worker_id(ParallelGCThreads),
 369   // _active_tasks set in set_non_marking_state
 370   // _tasks set inside the constructor
 371   _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)),
 372   _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)),
 373 
 374   _has_overflown(false),
 375   _concurrent(false),
 376   _has_aborted(false),
 377   _restart_for_overflow(false),
 378   _concurrent_marking_in_progress(false),
 379   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 380   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 381 
 382   // _verbose_level set below
 383 
 384   _init_times(),
 385   _remark_times(), _remark_mark_times(), _remark_weak_ref_times(),
 386   _cleanup_times(),
 387   _total_counting_time(0.0),
 388   _total_rs_scrub_time(0.0),
 389 
 390   _parallel_workers(NULL),
 391 
 392   _completed_initialization(false) {
 393 
 394   _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 395   _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage);
 396 
 397   // Create & start a ConcurrentMark thread.
 398   _cmThread = new ConcurrentMarkThread(this);
 399   assert(cmThread() != NULL, "CM Thread should have been created");
 400   assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm");
 401   if (_cmThread->osthread() == NULL) {
 402       vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 403   }
 404 
 405   assert(CGC_lock != NULL, "Where's the CGC_lock?");
 406 
 407   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 408   satb_qs.set_buffer_size(G1SATBBufferSize);
 409 
 410   _root_regions.init(_g1h->survivor(), this);
 411 
 412   if (ConcGCThreads > ParallelGCThreads) {
 413     log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).",
 414                     ConcGCThreads, ParallelGCThreads);
 415     return;
 416   }
 417   if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) {
 418     // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent
 419     // if both are set
 420     _sleep_factor             = 0.0;
 421     _marking_task_overhead    = 1.0;
 422   } else if (G1MarkingOverheadPercent > 0) {
 423     // We will calculate the number of parallel marking threads based
 424     // on a target overhead with respect to the soft real-time goal
 425     double marking_overhead = (double) G1MarkingOverheadPercent / 100.0;
 426     double overall_cm_overhead =
 427       (double) MaxGCPauseMillis * marking_overhead /
 428       (double) GCPauseIntervalMillis;
 429     double cpu_ratio = 1.0 / os::initial_active_processor_count();
 430     double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio);
 431     double marking_task_overhead =
 432       overall_cm_overhead / marking_thread_num * os::initial_active_processor_count();
 433     double sleep_factor =
 434                        (1.0 - marking_task_overhead) / marking_task_overhead;
 435 
 436     FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num);
 437     _sleep_factor             = sleep_factor;
 438     _marking_task_overhead    = marking_task_overhead;
 439   } else {
 440     // Calculate the number of parallel marking threads by scaling
 441     // the number of parallel GC threads.
 442     uint marking_thread_num = scale_parallel_threads(ParallelGCThreads);
 443     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 444     _sleep_factor             = 0.0;
 445     _marking_task_overhead    = 1.0;
 446   }
 447 
 448   assert(ConcGCThreads > 0, "Should have been set");
 449   log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
 450   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 451   _parallel_marking_threads = ConcGCThreads;
 452   _max_parallel_marking_threads = _parallel_marking_threads;
 453 
 454   _parallel_workers = new WorkGang("G1 Marker",
 455        _max_parallel_marking_threads, false, true);
 456   if (_parallel_workers == NULL) {
 457     vm_exit_during_initialization("Failed necessary allocation.");
 458   } else {
 459     _parallel_workers->initialize_workers();
 460   }
 461 
 462   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 463     size_t mark_stack_size =
 464       MIN2(MarkStackSizeMax,
 465           MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE)));
 466     // Verify that the calculated value for MarkStackSize is in range.
 467     // It would be nice to use the private utility routine from Arguments.
 468     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 469       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 470                       "must be between 1 and " SIZE_FORMAT,
 471                       mark_stack_size, MarkStackSizeMax);
 472       return;
 473     }
 474     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 475   } else {
 476     // Verify MarkStackSize is in range.
 477     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 478       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 479         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 480           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 481                           "must be between 1 and " SIZE_FORMAT,
 482                           MarkStackSize, MarkStackSizeMax);
 483           return;
 484         }
 485       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 486         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 487           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 488                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 489                           MarkStackSize, MarkStackSizeMax);
 490           return;
 491         }
 492       }
 493     }
 494   }
 495 
 496   if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
 497     vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
 498   }
 499 
 500   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC);
 501   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC);
 502 
 503   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 504   _active_tasks = _max_worker_id;
 505 
 506   for (uint i = 0; i < _max_worker_id; ++i) {
 507     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 508     task_queue->initialize();
 509     _task_queues->register_queue(i, task_queue);
 510 
 511     _tasks[i] = new G1CMTask(i, this, task_queue, _task_queues);
 512 
 513     _accum_task_vtime[i] = 0.0;
 514   }
 515 
 516   // so that the call below can read a sensible value
 517   _heap_start = g1h->reserved_region().start();
 518   set_non_marking_state();
 519   _completed_initialization = true;
 520 }
 521 
 522 void G1ConcurrentMark::reset() {
 523   // Starting values for these two. This should be called in a STW
 524   // phase.
 525   MemRegion reserved = _g1h->g1_reserved();
 526   _heap_start = reserved.start();
 527   _heap_end   = reserved.end();
 528 
 529   // Separated the asserts so that we know which one fires.
 530   assert(_heap_start != NULL, "heap bounds should look ok");
 531   assert(_heap_end != NULL, "heap bounds should look ok");
 532   assert(_heap_start < _heap_end, "heap bounds should look ok");
 533 
 534   // Reset all the marking data structures and any necessary flags
 535   reset_marking_state();
 536 
 537   // We do reset all of them, since different phases will use
 538   // different number of active threads. So, it's easiest to have all
 539   // of them ready.
 540   for (uint i = 0; i < _max_worker_id; ++i) {
 541     _tasks[i]->reset(_nextMarkBitMap);
 542   }
 543 
 544   // we need this to make sure that the flag is on during the evac
 545   // pause with initial mark piggy-backed
 546   set_concurrent_marking_in_progress();
 547 }
 548 
 549 
 550 void G1ConcurrentMark::reset_marking_state() {
 551   _global_mark_stack.set_empty();
 552 
 553   // Expand the marking stack, if we have to and if we can.
 554   if (has_overflown()) {
 555     _global_mark_stack.expand();
 556   }
 557 
 558   clear_has_overflown();
 559   _finger = _heap_start;
 560 
 561   for (uint i = 0; i < _max_worker_id; ++i) {
 562     G1CMTaskQueue* queue = _task_queues->queue(i);
 563     queue->set_empty();
 564   }
 565 }
 566 
 567 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 568   assert(active_tasks <= _max_worker_id, "we should not have more");
 569 
 570   _active_tasks = active_tasks;
 571   // Need to update the three data structures below according to the
 572   // number of active threads for this phase.
 573   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 574   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 575   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 576 }
 577 
 578 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 579   set_concurrency(active_tasks);
 580 
 581   _concurrent = concurrent;
 582   // We propagate this to all tasks, not just the active ones.
 583   for (uint i = 0; i < _max_worker_id; ++i)
 584     _tasks[i]->set_concurrent(concurrent);
 585 
 586   if (concurrent) {
 587     set_concurrent_marking_in_progress();
 588   } else {
 589     // We currently assume that the concurrent flag has been set to
 590     // false before we start remark. At this point we should also be
 591     // in a STW phase.
 592     assert(!concurrent_marking_in_progress(), "invariant");
 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::set_non_marking_state() {
 600   // We set the global marking state to some default values when we're
 601   // not doing marking.
 602   reset_marking_state();
 603   _active_tasks = 0;
 604   clear_concurrent_marking_in_progress();
 605 }
 606 
 607 G1ConcurrentMark::~G1ConcurrentMark() {
 608   // The G1ConcurrentMark instance is never freed.
 609   ShouldNotReachHere();
 610 }
 611 
 612 class G1ClearBitMapTask : public AbstractGangTask {
 613 public:
 614   static size_t chunk_size() { return M; }
 615 
 616 private:
 617   // Heap region closure used for clearing the given mark bitmap.
 618   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 619   private:
 620     G1CMBitMap* _bitmap;
 621     G1ConcurrentMark* _cm;
 622   public:
 623     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
 624     }
 625 
 626     virtual bool doHeapRegion(HeapRegion* r) {
 627       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 628 
 629       HeapWord* cur = r->bottom();
 630       HeapWord* const end = r->end();
 631 
 632       while (cur < end) {
 633         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 634         _bitmap->clear_range(mr);
 635 
 636         cur += chunk_size_in_words;
 637 
 638         // Abort iteration if after yielding the marking has been aborted.
 639         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 640           return true;
 641         }
 642         // Repeat the asserts from before the start of the closure. We will do them
 643         // as asserts here to minimize their overhead on the product. However, we
 644         // will have them as guarantees at the beginning / end of the bitmap
 645         // clearing to get some checking in the product.
 646         assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant");
 647         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
 648       }
 649       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 650 
 651       return false;
 652     }
 653   };
 654 
 655   G1ClearBitmapHRClosure _cl;
 656   HeapRegionClaimer _hr_claimer;
 657   bool _suspendible; // If the task is suspendible, workers must join the STS.
 658 
 659 public:
 660   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 661     AbstractGangTask("G1 Clear Bitmap"),
 662     _cl(bitmap, suspendible ? cm : NULL),
 663     _hr_claimer(n_workers),
 664     _suspendible(suspendible)
 665   { }
 666 
 667   void work(uint worker_id) {
 668     SuspendibleThreadSetJoiner sts_join(_suspendible);
 669     G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer);
 670   }
 671 
 672   bool is_complete() {
 673     return _cl.complete();
 674   }
 675 };
 676 
 677 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 678   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 679 
 680   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 681   size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 682 
 683   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 684 
 685   G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
 686 
 687   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 688   workers->run_task(&cl, num_workers);
 689   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 690 }
 691 
 692 void G1ConcurrentMark::cleanup_for_next_mark() {
 693   // Make sure that the concurrent mark thread looks to still be in
 694   // the current cycle.
 695   guarantee(cmThread()->during_cycle(), "invariant");
 696 
 697   // We are finishing up the current cycle by clearing the next
 698   // marking bitmap and getting it ready for the next cycle. During
 699   // this time no other cycle can start. So, let's make sure that this
 700   // is the case.
 701   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 702 
 703   clear_bitmap(_nextMarkBitMap, _parallel_workers, true);
 704 
 705   // Clear the live count data. If the marking has been aborted, the abort()
 706   // call already did that.
 707   if (!has_aborted()) {
 708     clear_live_data(_parallel_workers);
 709     DEBUG_ONLY(verify_live_data_clear());
 710   }
 711 
 712   // Repeat the asserts from above.
 713   guarantee(cmThread()->during_cycle(), "invariant");
 714   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 715 }
 716 
 717 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 718   assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint.");
 719   clear_bitmap(_prevMarkBitMap, workers, false);
 720 }
 721 
 722 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 723   G1CMBitMap* _bitmap;
 724   bool _error;
 725  public:
 726   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 727   }
 728 
 729   virtual bool doHeapRegion(HeapRegion* r) {
 730     // This closure can be called concurrently to the mutator, so we must make sure
 731     // that the result of the getNextMarkedWordAddress() call is compared to the
 732     // value passed to it as limit to detect any found bits.
 733     // end never changes in G1.
 734     HeapWord* end = r->end();
 735     return _bitmap->get_next_marked_addr(r->bottom(), end) != end;
 736   }
 737 };
 738 
 739 bool G1ConcurrentMark::nextMarkBitmapIsClear() {
 740   CheckBitmapClearHRClosure cl(_nextMarkBitMap);
 741   _g1h->heap_region_iterate(&cl);
 742   return cl.complete();
 743 }
 744 
 745 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 746 public:
 747   bool doHeapRegion(HeapRegion* r) {
 748     r->note_start_of_marking();
 749     return false;
 750   }
 751 };
 752 
 753 void G1ConcurrentMark::checkpointRootsInitialPre() {
 754   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 755 
 756   _has_aborted = false;
 757 
 758   // Initialize marking structures. This has to be done in a STW phase.
 759   reset();
 760 
 761   // For each region note start of marking.
 762   NoteStartOfMarkHRClosure startcl;
 763   g1h->heap_region_iterate(&startcl);
 764 }
 765 
 766 
 767 void G1ConcurrentMark::checkpointRootsInitialPost() {
 768   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 769 
 770   // Start Concurrent Marking weak-reference discovery.
 771   ReferenceProcessor* rp = g1h->ref_processor_cm();
 772   // enable ("weak") refs discovery
 773   rp->enable_discovery();
 774   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 775 
 776   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 777   // This is the start of  the marking cycle, we're expected all
 778   // threads to have SATB queues with active set to false.
 779   satb_mq_set.set_active_all_threads(true, /* new active value */
 780                                      false /* expected_active */);
 781 
 782   _root_regions.prepare_for_scan();
 783 
 784   // update_g1_committed() will be called at the end of an evac pause
 785   // when marking is on. So, it's also called at the end of the
 786   // initial-mark pause to update the heap end, if the heap expands
 787   // during it. No need to call it here.
 788 }
 789 
 790 /*
 791  * Notice that in the next two methods, we actually leave the STS
 792  * during the barrier sync and join it immediately afterwards. If we
 793  * do not do this, the following deadlock can occur: one thread could
 794  * be in the barrier sync code, waiting for the other thread to also
 795  * sync up, whereas another one could be trying to yield, while also
 796  * waiting for the other threads to sync up too.
 797  *
 798  * Note, however, that this code is also used during remark and in
 799  * this case we should not attempt to leave / enter the STS, otherwise
 800  * we'll either hit an assert (debug / fastdebug) or deadlock
 801  * (product). So we should only leave / enter the STS if we are
 802  * operating concurrently.
 803  *
 804  * Because the thread that does the sync barrier has left the STS, it
 805  * is possible to be suspended for a Full GC or an evacuation pause
 806  * could occur. This is actually safe, since the entering the sync
 807  * barrier is one of the last things do_marking_step() does, and it
 808  * doesn't manipulate any data structures afterwards.
 809  */
 810 
 811 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 812   bool barrier_aborted;
 813   {
 814     SuspendibleThreadSetLeaver sts_leave(concurrent());
 815     barrier_aborted = !_first_overflow_barrier_sync.enter();
 816   }
 817 
 818   // at this point everyone should have synced up and not be doing any
 819   // more work
 820 
 821   if (barrier_aborted) {
 822     // If the barrier aborted we ignore the overflow condition and
 823     // just abort the whole marking phase as quickly as possible.
 824     return;
 825   }
 826 
 827   // If we're executing the concurrent phase of marking, reset the marking
 828   // state; otherwise the marking state is reset after reference processing,
 829   // during the remark pause.
 830   // If we reset here as a result of an overflow during the remark we will
 831   // see assertion failures from any subsequent set_concurrency_and_phase()
 832   // calls.
 833   if (concurrent()) {
 834     // let the task associated with with worker 0 do this
 835     if (worker_id == 0) {
 836       // task 0 is responsible for clearing the global data structures
 837       // We should be here because of an overflow. During STW we should
 838       // not clear the overflow flag since we rely on it being true when
 839       // we exit this method to abort the pause and restart concurrent
 840       // marking.
 841       reset_marking_state();
 842 
 843       log_info(gc, marking)("Concurrent Mark reset for overflow");
 844     }
 845   }
 846 
 847   // after this, each task should reset its own data structures then
 848   // then go into the second barrier
 849 }
 850 
 851 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 852   SuspendibleThreadSetLeaver sts_leave(concurrent());
 853   _second_overflow_barrier_sync.enter();
 854 
 855   // at this point everything should be re-initialized and ready to go
 856 }
 857 
 858 class G1CMConcurrentMarkingTask: public AbstractGangTask {
 859 private:
 860   G1ConcurrentMark*     _cm;
 861   ConcurrentMarkThread* _cmt;
 862 
 863 public:
 864   void work(uint worker_id) {
 865     assert(Thread::current()->is_ConcurrentGC_thread(),
 866            "this should only be done by a conc GC thread");
 867     ResourceMark rm;
 868 
 869     double start_vtime = os::elapsedVTime();
 870 
 871     {
 872       SuspendibleThreadSetJoiner sts_join;
 873 
 874       assert(worker_id < _cm->active_tasks(), "invariant");
 875       G1CMTask* the_task = _cm->task(worker_id);
 876       the_task->record_start_time();
 877       if (!_cm->has_aborted()) {
 878         do {
 879           double start_vtime_sec = os::elapsedVTime();
 880           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
 881 
 882           the_task->do_marking_step(mark_step_duration_ms,
 883                                     true  /* do_termination */,
 884                                     false /* is_serial*/);
 885 
 886           double end_vtime_sec = os::elapsedVTime();
 887           double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec;
 888           _cm->do_yield_check();
 889 
 890           jlong sleep_time_ms;
 891           if (!_cm->has_aborted() && the_task->has_aborted()) {
 892             sleep_time_ms =
 893               (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0);
 894             {
 895               SuspendibleThreadSetLeaver sts_leave;
 896               os::sleep(Thread::current(), sleep_time_ms, false);
 897             }
 898           }
 899         } while (!_cm->has_aborted() && the_task->has_aborted());
 900       }
 901       the_task->record_end_time();
 902       guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant");
 903     }
 904 
 905     double end_vtime = os::elapsedVTime();
 906     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 907   }
 908 
 909   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm,
 910                             ConcurrentMarkThread* cmt) :
 911       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
 912 
 913   ~G1CMConcurrentMarkingTask() { }
 914 };
 915 
 916 // Calculates the number of active workers for a concurrent
 917 // phase.
 918 uint G1ConcurrentMark::calc_parallel_marking_threads() {
 919   uint n_conc_workers = 0;
 920   if (!UseDynamicNumberOfGCThreads ||
 921       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 922        !ForceDynamicNumberOfGCThreads)) {
 923     n_conc_workers = max_parallel_marking_threads();
 924   } else {
 925     n_conc_workers =
 926       AdaptiveSizePolicy::calc_default_active_workers(max_parallel_marking_threads(),
 927                                                       1, /* Minimum workers */
 928                                                       parallel_marking_threads(),
 929                                                       Threads::number_of_non_daemon_threads());
 930     // Don't scale down "n_conc_workers" by scale_parallel_threads() because
 931     // that scaling has already gone into "_max_parallel_marking_threads".
 932   }
 933   assert(n_conc_workers > 0 && n_conc_workers <= max_parallel_marking_threads(),
 934          "Calculated number of workers must be larger than zero and at most the maximum %u, but is %u",
 935          max_parallel_marking_threads(), n_conc_workers);
 936   return n_conc_workers;
 937 }
 938 
 939 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr) {
 940   // Currently, only survivors can be root regions.
 941   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 942   G1RootRegionScanClosure cl(_g1h, this);
 943 
 944   const uintx interval = PrefetchScanIntervalInBytes;
 945   HeapWord* curr = hr->bottom();
 946   const HeapWord* end = hr->top();
 947   while (curr < end) {
 948     Prefetch::read(curr, interval);
 949     oop obj = oop(curr);
 950     int size = obj->oop_iterate_size(&cl);
 951     assert(size == obj->size(), "sanity");
 952     curr += size;
 953   }
 954 }
 955 
 956 class G1CMRootRegionScanTask : public AbstractGangTask {
 957 private:
 958   G1ConcurrentMark* _cm;
 959 
 960 public:
 961   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 962     AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
 963 
 964   void work(uint worker_id) {
 965     assert(Thread::current()->is_ConcurrentGC_thread(),
 966            "this should only be done by a conc GC thread");
 967 
 968     G1CMRootRegions* root_regions = _cm->root_regions();
 969     HeapRegion* hr = root_regions->claim_next();
 970     while (hr != NULL) {
 971       _cm->scanRootRegion(hr);
 972       hr = root_regions->claim_next();
 973     }
 974   }
 975 };
 976 
 977 void G1ConcurrentMark::scan_root_regions() {
 978   // scan_in_progress() will have been set to true only if there was
 979   // at least one root region to scan. So, if it's false, we
 980   // should not attempt to do any further work.
 981   if (root_regions()->scan_in_progress()) {
 982     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 983 
 984     _parallel_marking_threads = MIN2(calc_parallel_marking_threads(),
 985                                      // We distribute work on a per-region basis, so starting
 986                                      // more threads than that is useless.
 987                                      root_regions()->num_root_regions());
 988     assert(parallel_marking_threads() <= max_parallel_marking_threads(),
 989            "Maximum number of marking threads exceeded");
 990 
 991     G1CMRootRegionScanTask task(this);
 992     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
 993                         task.name(), _parallel_marking_threads, root_regions()->num_root_regions());
 994     _parallel_workers->run_task(&task, _parallel_marking_threads);
 995 
 996     // It's possible that has_aborted() is true here without actually
 997     // aborting the survivor scan earlier. This is OK as it's
 998     // mainly used for sanity checking.
 999     root_regions()->scan_finished();
1000   }
1001 }
1002 
1003 void G1ConcurrentMark::concurrent_cycle_start() {
1004   _gc_timer_cm->register_gc_start();
1005 
1006   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
1007 
1008   _g1h->trace_heap_before_gc(_gc_tracer_cm);
1009 }
1010 
1011 void G1ConcurrentMark::concurrent_cycle_end() {
1012   _g1h->trace_heap_after_gc(_gc_tracer_cm);
1013 
1014   if (has_aborted()) {
1015     _gc_tracer_cm->report_concurrent_mode_failure();
1016   }
1017 
1018   _gc_timer_cm->register_gc_end();
1019 
1020   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
1021 }
1022 
1023 void G1ConcurrentMark::mark_from_roots() {
1024   // we might be tempted to assert that:
1025   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
1026   //        "inconsistent argument?");
1027   // However that wouldn't be right, because it's possible that
1028   // a safepoint is indeed in progress as a younger generation
1029   // stop-the-world GC happens even as we mark in this generation.
1030 
1031   _restart_for_overflow = false;
1032 
1033   // _g1h has _n_par_threads
1034   _parallel_marking_threads = calc_parallel_marking_threads();
1035   assert(parallel_marking_threads() <= max_parallel_marking_threads(),
1036     "Maximum number of marking threads exceeded");
1037 
1038   uint active_workers = MAX2(1U, parallel_marking_threads());
1039   assert(active_workers > 0, "Should have been set");
1040 
1041   // Setting active workers is not guaranteed since fewer
1042   // worker threads may currently exist and more may not be
1043   // available.
1044   active_workers = _parallel_workers->update_active_workers(active_workers);
1045   log_info(gc, task)("Using %u workers of %u for marking", active_workers, _parallel_workers->total_workers());
1046 
1047   // Parallel task terminator is set in "set_concurrency_and_phase()"
1048   set_concurrency_and_phase(active_workers, true /* concurrent */);
1049 
1050   G1CMConcurrentMarkingTask markingTask(this, cmThread());
1051   _parallel_workers->run_task(&markingTask);
1052   print_stats();
1053 }
1054 
1055 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) {
1056   // world is stopped at this checkpoint
1057   assert(SafepointSynchronize::is_at_safepoint(),
1058          "world should be stopped");
1059 
1060   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1061 
1062   // If a full collection has happened, we shouldn't do this.
1063   if (has_aborted()) {
1064     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1065     return;
1066   }
1067 
1068   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1069 
1070   if (VerifyDuringGC) {
1071     HandleMark hm;  // handle scope
1072     g1h->prepare_for_verify();
1073     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1074   }
1075   g1h->verifier()->check_bitmaps("Remark Start");
1076 
1077   G1Policy* g1p = g1h->g1_policy();
1078   g1p->record_concurrent_mark_remark_start();
1079 
1080   double start = os::elapsedTime();
1081 
1082   checkpointRootsFinalWork();
1083 
1084   double mark_work_end = os::elapsedTime();
1085 
1086   weakRefsWork(clear_all_soft_refs);
1087 
1088   if (has_overflown()) {
1089     // We overflowed.  Restart concurrent marking.
1090     _restart_for_overflow = true;
1091 
1092     // Verify the heap w.r.t. the previous marking bitmap.
1093     if (VerifyDuringGC) {
1094       HandleMark hm;  // handle scope
1095       g1h->prepare_for_verify();
1096       Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)");
1097     }
1098 
1099     // Clear the marking state because we will be restarting
1100     // marking due to overflowing the global mark stack.
1101     reset_marking_state();
1102   } else {
1103     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1104     // We're done with marking.
1105     // This is the end of  the marking cycle, we're expected all
1106     // threads to have SATB queues with active set to true.
1107     satb_mq_set.set_active_all_threads(false, /* new active value */
1108                                        true /* expected_active */);
1109 
1110     if (VerifyDuringGC) {
1111       HandleMark hm;  // handle scope
1112       g1h->prepare_for_verify();
1113       Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)");
1114     }
1115     g1h->verifier()->check_bitmaps("Remark End");
1116     assert(!restart_for_overflow(), "sanity");
1117     // Completely reset the marking state since marking completed
1118     set_non_marking_state();
1119   }
1120 
1121   // Statistics
1122   double now = os::elapsedTime();
1123   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1124   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1125   _remark_times.add((now - start) * 1000.0);
1126 
1127   g1p->record_concurrent_mark_remark_end();
1128 
1129   G1CMIsAliveClosure is_alive(g1h);
1130   _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1131 }
1132 
1133 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1134   G1CollectedHeap* _g1;
1135   size_t _freed_bytes;
1136   FreeRegionList* _local_cleanup_list;
1137   uint _old_regions_removed;
1138   uint _humongous_regions_removed;
1139   HRRSCleanupTask* _hrrs_cleanup_task;
1140 
1141 public:
1142   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1143                              FreeRegionList* local_cleanup_list,
1144                              HRRSCleanupTask* hrrs_cleanup_task) :
1145     _g1(g1),
1146     _freed_bytes(0),
1147     _local_cleanup_list(local_cleanup_list),
1148     _old_regions_removed(0),
1149     _humongous_regions_removed(0),
1150     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1151 
1152   size_t freed_bytes() { return _freed_bytes; }
1153   const uint old_regions_removed() { return _old_regions_removed; }
1154   const uint humongous_regions_removed() { return _humongous_regions_removed; }
1155 
1156   bool doHeapRegion(HeapRegion *hr) {
1157     if (hr->is_archive()) {
1158       return false;
1159     }
1160     _g1->reset_gc_time_stamps(hr);
1161     hr->note_end_of_marking();
1162 
1163     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) {
1164       _freed_bytes += hr->used();
1165       hr->set_containing_set(NULL);
1166       if (hr->is_humongous()) {
1167         _humongous_regions_removed++;
1168         _g1->free_humongous_region(hr, _local_cleanup_list, true /* skip_remset */);
1169       } else {
1170         _old_regions_removed++;
1171         _g1->free_region(hr, _local_cleanup_list, true /* skip_remset */);
1172       }
1173     } else {
1174       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1175     }
1176 
1177     return false;
1178   }
1179 };
1180 
1181 class G1ParNoteEndTask: public AbstractGangTask {
1182   friend class G1NoteEndOfConcMarkClosure;
1183 
1184 protected:
1185   G1CollectedHeap* _g1h;
1186   FreeRegionList* _cleanup_list;
1187   HeapRegionClaimer _hrclaimer;
1188 
1189 public:
1190   G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1191       AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1192   }
1193 
1194   void work(uint worker_id) {
1195     FreeRegionList local_cleanup_list("Local Cleanup List");
1196     HRRSCleanupTask hrrs_cleanup_task;
1197     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1198                                            &hrrs_cleanup_task);
1199     _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer);
1200     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1201 
1202     // Now update the lists
1203     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1204     {
1205       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1206       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1207 
1208       // If we iterate over the global cleanup list at the end of
1209       // cleanup to do this printing we will not guarantee to only
1210       // generate output for the newly-reclaimed regions (the list
1211       // might not be empty at the beginning of cleanup; we might
1212       // still be working on its previous contents). So we do the
1213       // printing here, before we append the new regions to the global
1214       // cleanup list.
1215 
1216       G1HRPrinter* hr_printer = _g1h->hr_printer();
1217       if (hr_printer->is_active()) {
1218         FreeRegionListIterator iter(&local_cleanup_list);
1219         while (iter.more_available()) {
1220           HeapRegion* hr = iter.get_next();
1221           hr_printer->cleanup(hr);
1222         }
1223       }
1224 
1225       _cleanup_list->add_ordered(&local_cleanup_list);
1226       assert(local_cleanup_list.is_empty(), "post-condition");
1227 
1228       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1229     }
1230   }
1231 };
1232 
1233 void G1ConcurrentMark::cleanup() {
1234   // world is stopped at this checkpoint
1235   assert(SafepointSynchronize::is_at_safepoint(),
1236          "world should be stopped");
1237   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1238 
1239   // If a full collection has happened, we shouldn't do this.
1240   if (has_aborted()) {
1241     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1242     return;
1243   }
1244 
1245   g1h->verifier()->verify_region_sets_optional();
1246 
1247   if (VerifyDuringGC) {
1248     HandleMark hm;  // handle scope
1249     g1h->prepare_for_verify();
1250     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)");
1251   }
1252   g1h->verifier()->check_bitmaps("Cleanup Start");
1253 
1254   G1Policy* g1p = g1h->g1_policy();
1255   g1p->record_concurrent_mark_cleanup_start();
1256 
1257   double start = os::elapsedTime();
1258 
1259   HeapRegionRemSet::reset_for_cleanup_tasks();
1260 
1261   {
1262     GCTraceTime(Debug, gc)("Finalize Live Data");
1263     finalize_live_data();
1264   }
1265 
1266   if (VerifyDuringGC) {
1267     GCTraceTime(Debug, gc)("Verify Live Data");
1268     verify_live_data();
1269   }
1270 
1271   g1h->collector_state()->set_mark_in_progress(false);
1272 
1273   double count_end = os::elapsedTime();
1274   double this_final_counting_time = (count_end - start);
1275   _total_counting_time += this_final_counting_time;
1276 
1277   if (log_is_enabled(Trace, gc, liveness)) {
1278     G1PrintRegionLivenessInfoClosure cl("Post-Marking");
1279     _g1h->heap_region_iterate(&cl);
1280   }
1281 
1282   // Install newly created mark bitMap as "prev".
1283   swapMarkBitMaps();
1284 
1285   g1h->reset_gc_time_stamp();
1286 
1287   uint n_workers = _g1h->workers()->active_workers();
1288 
1289   // Note end of marking in all heap regions.
1290   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1291   g1h->workers()->run_task(&g1_par_note_end_task);
1292   g1h->check_gc_time_stamps();
1293 
1294   if (!cleanup_list_is_empty()) {
1295     // The cleanup list is not empty, so we'll have to process it
1296     // concurrently. Notify anyone else that might be wanting free
1297     // regions that there will be more free regions coming soon.
1298     g1h->set_free_regions_coming();
1299   }
1300 
1301   // call below, since it affects the metric by which we sort the heap
1302   // regions.
1303   if (G1ScrubRemSets) {
1304     double rs_scrub_start = os::elapsedTime();
1305     g1h->scrub_rem_set();
1306     _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start);
1307   }
1308 
1309   // this will also free any regions totally full of garbage objects,
1310   // and sort the regions.
1311   g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1312 
1313   // Statistics.
1314   double end = os::elapsedTime();
1315   _cleanup_times.add((end - start) * 1000.0);
1316 
1317   // Clean up will have freed any regions completely full of garbage.
1318   // Update the soft reference policy with the new heap occupancy.
1319   Universe::update_heap_info_at_gc();
1320 
1321   if (VerifyDuringGC) {
1322     HandleMark hm;  // handle scope
1323     g1h->prepare_for_verify();
1324     Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)");
1325   }
1326 
1327   g1h->verifier()->check_bitmaps("Cleanup End");
1328 
1329   g1h->verifier()->verify_region_sets_optional();
1330 
1331   // We need to make this be a "collection" so any collection pause that
1332   // races with it goes around and waits for completeCleanup to finish.
1333   g1h->increment_total_collections();
1334 
1335   // Clean out dead classes and update Metaspace sizes.
1336   if (ClassUnloadingWithConcurrentMark) {
1337     ClassLoaderDataGraph::purge();
1338   }
1339   MetaspaceGC::compute_new_size();
1340 
1341   // We reclaimed old regions so we should calculate the sizes to make
1342   // sure we update the old gen/space data.
1343   g1h->g1mm()->update_sizes();
1344   g1h->allocation_context_stats().update_after_mark();
1345 }
1346 
1347 void G1ConcurrentMark::complete_cleanup() {
1348   if (has_aborted()) return;
1349 
1350   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1351 
1352   _cleanup_list.verify_optional();
1353   FreeRegionList tmp_free_list("Tmp Free List");
1354 
1355   log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1356                                   "cleanup list has %u entries",
1357                                   _cleanup_list.length());
1358 
1359   // No one else should be accessing the _cleanup_list at this point,
1360   // so it is not necessary to take any locks
1361   while (!_cleanup_list.is_empty()) {
1362     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
1363     assert(hr != NULL, "Got NULL from a non-empty list");
1364     hr->par_clear();
1365     tmp_free_list.add_ordered(hr);
1366 
1367     // Instead of adding one region at a time to the secondary_free_list,
1368     // we accumulate them in the local list and move them a few at a
1369     // time. This also cuts down on the number of notify_all() calls
1370     // we do during this process. We'll also append the local list when
1371     // _cleanup_list is empty (which means we just removed the last
1372     // region from the _cleanup_list).
1373     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
1374         _cleanup_list.is_empty()) {
1375       log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1376                                       "appending %u entries to the secondary_free_list, "
1377                                       "cleanup list still has %u entries",
1378                                       tmp_free_list.length(),
1379                                       _cleanup_list.length());
1380 
1381       {
1382         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
1383         g1h->secondary_free_list_add(&tmp_free_list);
1384         SecondaryFreeList_lock->notify_all();
1385       }
1386 #ifndef PRODUCT
1387       if (G1StressConcRegionFreeing) {
1388         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
1389           os::sleep(Thread::current(), (jlong) 1, false);
1390         }
1391       }
1392 #endif
1393     }
1394   }
1395   assert(tmp_free_list.is_empty(), "post-condition");
1396 }
1397 
1398 // Supporting Object and Oop closures for reference discovery
1399 // and processing in during marking
1400 
1401 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1402   HeapWord* addr = (HeapWord*)obj;
1403   return addr != NULL &&
1404          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
1405 }
1406 
1407 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1408 // Uses the G1CMTask associated with a worker thread (for serial reference
1409 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1410 // trace referent objects.
1411 //
1412 // Using the G1CMTask and embedded local queues avoids having the worker
1413 // threads operating on the global mark stack. This reduces the risk
1414 // of overflowing the stack - which we would rather avoid at this late
1415 // state. Also using the tasks' local queues removes the potential
1416 // of the workers interfering with each other that could occur if
1417 // operating on the global stack.
1418 
1419 class G1CMKeepAliveAndDrainClosure: public OopClosure {
1420   G1ConcurrentMark* _cm;
1421   G1CMTask*         _task;
1422   int               _ref_counter_limit;
1423   int               _ref_counter;
1424   bool              _is_serial;
1425  public:
1426   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1427     _cm(cm), _task(task), _is_serial(is_serial),
1428     _ref_counter_limit(G1RefProcDrainInterval) {
1429     assert(_ref_counter_limit > 0, "sanity");
1430     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1431     _ref_counter = _ref_counter_limit;
1432   }
1433 
1434   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1435   virtual void do_oop(      oop* p) { do_oop_work(p); }
1436 
1437   template <class T> void do_oop_work(T* p) {
1438     if (!_cm->has_overflown()) {
1439       oop obj = oopDesc::load_decode_heap_oop(p);
1440       _task->deal_with_reference(obj);
1441       _ref_counter--;
1442 
1443       if (_ref_counter == 0) {
1444         // We have dealt with _ref_counter_limit references, pushing them
1445         // and objects reachable from them on to the local stack (and
1446         // possibly the global stack). Call G1CMTask::do_marking_step() to
1447         // process these entries.
1448         //
1449         // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1450         // there's nothing more to do (i.e. we're done with the entries that
1451         // were pushed as a result of the G1CMTask::deal_with_reference() calls
1452         // above) or we overflow.
1453         //
1454         // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1455         // flag while there may still be some work to do. (See the comment at
1456         // the beginning of G1CMTask::do_marking_step() for those conditions -
1457         // one of which is reaching the specified time target.) It is only
1458         // when G1CMTask::do_marking_step() returns without setting the
1459         // has_aborted() flag that the marking step has completed.
1460         do {
1461           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1462           _task->do_marking_step(mark_step_duration_ms,
1463                                  false      /* do_termination */,
1464                                  _is_serial);
1465         } while (_task->has_aborted() && !_cm->has_overflown());
1466         _ref_counter = _ref_counter_limit;
1467       }
1468     }
1469   }
1470 };
1471 
1472 // 'Drain' oop closure used by both serial and parallel reference processing.
1473 // Uses the G1CMTask associated with a given worker thread (for serial
1474 // reference processing the G1CMtask for worker 0 is used). Calls the
1475 // do_marking_step routine, with an unbelievably large timeout value,
1476 // to drain the marking data structures of the remaining entries
1477 // added by the 'keep alive' oop closure above.
1478 
1479 class G1CMDrainMarkingStackClosure: public VoidClosure {
1480   G1ConcurrentMark* _cm;
1481   G1CMTask*         _task;
1482   bool              _is_serial;
1483  public:
1484   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1485     _cm(cm), _task(task), _is_serial(is_serial) {
1486     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1487   }
1488 
1489   void do_void() {
1490     do {
1491       // We call G1CMTask::do_marking_step() to completely drain the local
1492       // and global marking stacks of entries pushed by the 'keep alive'
1493       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1494       //
1495       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1496       // if there's nothing more to do (i.e. we've completely drained the
1497       // entries that were pushed as a a result of applying the 'keep alive'
1498       // closure to the entries on the discovered ref lists) or we overflow
1499       // the global marking stack.
1500       //
1501       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1502       // flag while there may still be some work to do. (See the comment at
1503       // the beginning of G1CMTask::do_marking_step() for those conditions -
1504       // one of which is reaching the specified time target.) It is only
1505       // when G1CMTask::do_marking_step() returns without setting the
1506       // has_aborted() flag that the marking step has completed.
1507 
1508       _task->do_marking_step(1000000000.0 /* something very large */,
1509                              true         /* do_termination */,
1510                              _is_serial);
1511     } while (_task->has_aborted() && !_cm->has_overflown());
1512   }
1513 };
1514 
1515 // Implementation of AbstractRefProcTaskExecutor for parallel
1516 // reference processing at the end of G1 concurrent marking
1517 
1518 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
1519 private:
1520   G1CollectedHeap*  _g1h;
1521   G1ConcurrentMark* _cm;
1522   WorkGang*         _workers;
1523   uint              _active_workers;
1524 
1525 public:
1526   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1527                           G1ConcurrentMark* cm,
1528                           WorkGang* workers,
1529                           uint n_workers) :
1530     _g1h(g1h), _cm(cm),
1531     _workers(workers), _active_workers(n_workers) { }
1532 
1533   // Executes the given task using concurrent marking worker threads.
1534   virtual void execute(ProcessTask& task);
1535   virtual void execute(EnqueueTask& task);
1536 };
1537 
1538 class G1CMRefProcTaskProxy: public AbstractGangTask {
1539   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1540   ProcessTask&      _proc_task;
1541   G1CollectedHeap*  _g1h;
1542   G1ConcurrentMark* _cm;
1543 
1544 public:
1545   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1546                        G1CollectedHeap* g1h,
1547                        G1ConcurrentMark* cm) :
1548     AbstractGangTask("Process reference objects in parallel"),
1549     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1550     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1551     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1552   }
1553 
1554   virtual void work(uint worker_id) {
1555     ResourceMark rm;
1556     HandleMark hm;
1557     G1CMTask* task = _cm->task(worker_id);
1558     G1CMIsAliveClosure g1_is_alive(_g1h);
1559     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1560     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1561 
1562     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1563   }
1564 };
1565 
1566 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1567   assert(_workers != NULL, "Need parallel worker threads.");
1568   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1569 
1570   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1571 
1572   // We need to reset the concurrency level before each
1573   // proxy task execution, so that the termination protocol
1574   // and overflow handling in G1CMTask::do_marking_step() knows
1575   // how many workers to wait for.
1576   _cm->set_concurrency(_active_workers);
1577   _workers->run_task(&proc_task_proxy);
1578 }
1579 
1580 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
1581   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1582   EnqueueTask& _enq_task;
1583 
1584 public:
1585   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1586     AbstractGangTask("Enqueue reference objects in parallel"),
1587     _enq_task(enq_task) { }
1588 
1589   virtual void work(uint worker_id) {
1590     _enq_task.work(worker_id);
1591   }
1592 };
1593 
1594 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1595   assert(_workers != NULL, "Need parallel worker threads.");
1596   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1597 
1598   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1599 
1600   // Not strictly necessary but...
1601   //
1602   // We need to reset the concurrency level before each
1603   // proxy task execution, so that the termination protocol
1604   // and overflow handling in G1CMTask::do_marking_step() knows
1605   // how many workers to wait for.
1606   _cm->set_concurrency(_active_workers);
1607   _workers->run_task(&enq_task_proxy);
1608 }
1609 
1610 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) {
1611   if (has_overflown()) {
1612     // Skip processing the discovered references if we have
1613     // overflown the global marking stack. Reference objects
1614     // only get discovered once so it is OK to not
1615     // de-populate the discovered reference lists. We could have,
1616     // but the only benefit would be that, when marking restarts,
1617     // less reference objects are discovered.
1618     return;
1619   }
1620 
1621   ResourceMark rm;
1622   HandleMark   hm;
1623 
1624   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1625 
1626   // Is alive closure.
1627   G1CMIsAliveClosure g1_is_alive(g1h);
1628 
1629   // Inner scope to exclude the cleaning of the string and symbol
1630   // tables from the displayed time.
1631   {
1632     GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm);
1633 
1634     ReferenceProcessor* rp = g1h->ref_processor_cm();
1635 
1636     // See the comment in G1CollectedHeap::ref_processing_init()
1637     // about how reference processing currently works in G1.
1638 
1639     // Set the soft reference policy
1640     rp->setup_policy(clear_all_soft_refs);
1641     assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1642 
1643     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1644     // in serial reference processing. Note these closures are also
1645     // used for serially processing (by the the current thread) the
1646     // JNI references during parallel reference processing.
1647     //
1648     // These closures do not need to synchronize with the worker
1649     // threads involved in parallel reference processing as these
1650     // instances are executed serially by the current thread (e.g.
1651     // reference processing is not multi-threaded and is thus
1652     // performed by the current thread instead of a gang worker).
1653     //
1654     // The gang tasks involved in parallel reference processing create
1655     // their own instances of these closures, which do their own
1656     // synchronization among themselves.
1657     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1658     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1659 
1660     // We need at least one active thread. If reference processing
1661     // is not multi-threaded we use the current (VMThread) thread,
1662     // otherwise we use the work gang from the G1CollectedHeap and
1663     // we utilize all the worker threads we can.
1664     bool processing_is_mt = rp->processing_is_mt();
1665     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
1666     active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U);
1667 
1668     // Parallel processing task executor.
1669     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
1670                                               g1h->workers(), active_workers);
1671     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1672 
1673     // Set the concurrency level. The phase was already set prior to
1674     // executing the remark task.
1675     set_concurrency(active_workers);
1676 
1677     // Set the degree of MT processing here.  If the discovery was done MT,
1678     // the number of threads involved during discovery could differ from
1679     // the number of active workers.  This is OK as long as the discovered
1680     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1681     rp->set_active_mt_degree(active_workers);
1682 
1683     // Process the weak references.
1684     const ReferenceProcessorStats& stats =
1685         rp->process_discovered_references(&g1_is_alive,
1686                                           &g1_keep_alive,
1687                                           &g1_drain_mark_stack,
1688                                           executor,
1689                                           _gc_timer_cm);
1690     _gc_tracer_cm->report_gc_reference_stats(stats);
1691 
1692     // The do_oop work routines of the keep_alive and drain_marking_stack
1693     // oop closures will set the has_overflown flag if we overflow the
1694     // global marking stack.
1695 
1696     assert(has_overflown() || _global_mark_stack.is_empty(),
1697             "Mark stack should be empty (unless it has overflown)");
1698 
1699     assert(rp->num_q() == active_workers, "why not");
1700 
1701     rp->enqueue_discovered_references(executor);
1702 
1703     rp->verify_no_references_recorded();
1704     assert(!rp->discovery_enabled(), "Post condition");
1705   }
1706 
1707   if (has_overflown()) {
1708     // We can not trust g1_is_alive if the marking stack overflowed
1709     return;
1710   }
1711 
1712   assert(_global_mark_stack.is_empty(), "Marking should have completed");
1713 
1714   // Unload Klasses, String, Symbols, Code Cache, etc.
1715   if (ClassUnloadingWithConcurrentMark) {
1716     GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1717     bool purged_classes = SystemDictionary::do_unloading(&g1_is_alive, _gc_timer_cm, false /* Defer cleaning */);
1718     g1h->complete_cleaning(&g1_is_alive, purged_classes);
1719   } else {
1720     GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1721     // No need to clean string table and symbol table as they are treated as strong roots when
1722     // class unloading is disabled.
1723     g1h->partial_cleaning(&g1_is_alive, false, false, G1StringDedup::is_enabled());
1724 
1725   }
1726 }
1727 
1728 void G1ConcurrentMark::swapMarkBitMaps() {
1729   G1CMBitMap* temp = _prevMarkBitMap;
1730   _prevMarkBitMap  = _nextMarkBitMap;
1731   _nextMarkBitMap  = temp;
1732 }
1733 
1734 // Closure for marking entries in SATB buffers.
1735 class G1CMSATBBufferClosure : public SATBBufferClosure {
1736 private:
1737   G1CMTask* _task;
1738   G1CollectedHeap* _g1h;
1739 
1740   // This is very similar to G1CMTask::deal_with_reference, but with
1741   // more relaxed requirements for the argument, so this must be more
1742   // circumspect about treating the argument as an object.
1743   void do_entry(void* entry) const {
1744     _task->increment_refs_reached();
1745     HeapRegion* hr = _g1h->heap_region_containing(entry);
1746     if (entry < hr->next_top_at_mark_start()) {
1747       // Until we get here, we don't know whether entry refers to a valid
1748       // object; it could instead have been a stale reference.
1749       oop obj = static_cast<oop>(entry);
1750       assert(obj->is_oop(true /* ignore mark word */),
1751              "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj));
1752       _task->make_reference_grey(obj);
1753     }
1754   }
1755 
1756 public:
1757   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1758     : _task(task), _g1h(g1h) { }
1759 
1760   virtual void do_buffer(void** buffer, size_t size) {
1761     for (size_t i = 0; i < size; ++i) {
1762       do_entry(buffer[i]);
1763     }
1764   }
1765 };
1766 
1767 class G1RemarkThreadsClosure : public ThreadClosure {
1768   G1CMSATBBufferClosure _cm_satb_cl;
1769   G1CMOopClosure _cm_cl;
1770   MarkingCodeBlobClosure _code_cl;
1771   int _thread_parity;
1772 
1773  public:
1774   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1775     _cm_satb_cl(task, g1h),
1776     _cm_cl(g1h, g1h->concurrent_mark(), task),
1777     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1778     _thread_parity(Threads::thread_claim_parity()) {}
1779 
1780   void do_thread(Thread* thread) {
1781     if (thread->is_Java_thread()) {
1782       if (thread->claim_oops_do(true, _thread_parity)) {
1783         JavaThread* jt = (JavaThread*)thread;
1784 
1785         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1786         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1787         // * Alive if on the stack of an executing method
1788         // * Weakly reachable otherwise
1789         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1790         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1791         jt->nmethods_do(&_code_cl);
1792 
1793         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
1794       }
1795     } else if (thread->is_VM_thread()) {
1796       if (thread->claim_oops_do(true, _thread_parity)) {
1797         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1798       }
1799     }
1800   }
1801 };
1802 
1803 class G1CMRemarkTask: public AbstractGangTask {
1804 private:
1805   G1ConcurrentMark* _cm;
1806 public:
1807   void work(uint worker_id) {
1808     // Since all available tasks are actually started, we should
1809     // only proceed if we're supposed to be active.
1810     if (worker_id < _cm->active_tasks()) {
1811       G1CMTask* task = _cm->task(worker_id);
1812       task->record_start_time();
1813       {
1814         ResourceMark rm;
1815         HandleMark hm;
1816 
1817         G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1818         Threads::threads_do(&threads_f);
1819       }
1820 
1821       do {
1822         task->do_marking_step(1000000000.0 /* something very large */,
1823                               true         /* do_termination       */,
1824                               false        /* is_serial            */);
1825       } while (task->has_aborted() && !_cm->has_overflown());
1826       // If we overflow, then we do not want to restart. We instead
1827       // want to abort remark and do concurrent marking again.
1828       task->record_end_time();
1829     }
1830   }
1831 
1832   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1833     AbstractGangTask("Par Remark"), _cm(cm) {
1834     _cm->terminator()->reset_for_reuse(active_workers);
1835   }
1836 };
1837 
1838 void G1ConcurrentMark::checkpointRootsFinalWork() {
1839   ResourceMark rm;
1840   HandleMark   hm;
1841   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1842 
1843   GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm);
1844 
1845   g1h->ensure_parsability(false);
1846 
1847   // this is remark, so we'll use up all active threads
1848   uint active_workers = g1h->workers()->active_workers();
1849   set_concurrency_and_phase(active_workers, false /* concurrent */);
1850   // Leave _parallel_marking_threads at it's
1851   // value originally calculated in the G1ConcurrentMark
1852   // constructor and pass values of the active workers
1853   // through the gang in the task.
1854 
1855   {
1856     StrongRootsScope srs(active_workers);
1857 
1858     G1CMRemarkTask remarkTask(this, active_workers);
1859     // We will start all available threads, even if we decide that the
1860     // active_workers will be fewer. The extra ones will just bail out
1861     // immediately.
1862     g1h->workers()->run_task(&remarkTask);
1863   }
1864 
1865   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1866   guarantee(has_overflown() ||
1867             satb_mq_set.completed_buffers_num() == 0,
1868             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1869             BOOL_TO_STR(has_overflown()),
1870             satb_mq_set.completed_buffers_num());
1871 
1872   print_stats();
1873 }
1874 
1875 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) {
1876   _prevMarkBitMap->clear_range(mr);
1877 }
1878 
1879 HeapRegion*
1880 G1ConcurrentMark::claim_region(uint worker_id) {
1881   // "checkpoint" the finger
1882   HeapWord* finger = _finger;
1883 
1884   // _heap_end will not change underneath our feet; it only changes at
1885   // yield points.
1886   while (finger < _heap_end) {
1887     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1888 
1889     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1890     // Make sure that the reads below do not float before loading curr_region.
1891     OrderAccess::loadload();
1892     // Above heap_region_containing may return NULL as we always scan claim
1893     // until the end of the heap. In this case, just jump to the next region.
1894     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1895 
1896     // Is the gap between reading the finger and doing the CAS too long?
1897     HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger);
1898     if (res == finger && curr_region != NULL) {
1899       // we succeeded
1900       HeapWord*   bottom        = curr_region->bottom();
1901       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1902 
1903       // notice that _finger == end cannot be guaranteed here since,
1904       // someone else might have moved the finger even further
1905       assert(_finger >= end, "the finger should have moved forward");
1906 
1907       if (limit > bottom) {
1908         return curr_region;
1909       } else {
1910         assert(limit == bottom,
1911                "the region limit should be at bottom");
1912         // we return NULL and the caller should try calling
1913         // claim_region() again.
1914         return NULL;
1915       }
1916     } else {
1917       assert(_finger > finger, "the finger should have moved forward");
1918       // read it again
1919       finger = _finger;
1920     }
1921   }
1922 
1923   return NULL;
1924 }
1925 
1926 #ifndef PRODUCT
1927 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
1928 private:
1929   G1CollectedHeap* _g1h;
1930   const char* _phase;
1931   int _info;
1932 
1933 public:
1934   VerifyNoCSetOops(const char* phase, int info = -1) :
1935     _g1h(G1CollectedHeap::heap()),
1936     _phase(phase),
1937     _info(info)
1938   { }
1939 
1940   void operator()(G1TaskQueueEntry task_entry) const {
1941     if (task_entry.is_array_slice()) {
1942       guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1943       return;
1944     }
1945     guarantee(task_entry.obj()->is_oop(),
1946               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1947               p2i(task_entry.obj()), _phase, _info);
1948     guarantee(!_g1h->is_in_cset(task_entry.obj()),
1949               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1950               p2i(task_entry.obj()), _phase, _info);
1951   }
1952 };
1953 
1954 void G1ConcurrentMark::verify_no_cset_oops() {
1955   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1956   if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
1957     return;
1958   }
1959 
1960   // Verify entries on the global mark stack
1961   _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1962 
1963   // Verify entries on the task queues
1964   for (uint i = 0; i < _max_worker_id; ++i) {
1965     G1CMTaskQueue* queue = _task_queues->queue(i);
1966     queue->iterate(VerifyNoCSetOops("Queue", i));
1967   }
1968 
1969   // Verify the global finger
1970   HeapWord* global_finger = finger();
1971   if (global_finger != NULL && global_finger < _heap_end) {
1972     // Since we always iterate over all regions, we might get a NULL HeapRegion
1973     // here.
1974     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1975     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1976               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1977               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1978   }
1979 
1980   // Verify the task fingers
1981   assert(parallel_marking_threads() <= _max_worker_id, "sanity");
1982   for (uint i = 0; i < parallel_marking_threads(); ++i) {
1983     G1CMTask* task = _tasks[i];
1984     HeapWord* task_finger = task->finger();
1985     if (task_finger != NULL && task_finger < _heap_end) {
1986       // See above note on the global finger verification.
1987       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
1988       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
1989                 !task_hr->in_collection_set(),
1990                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
1991                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
1992     }
1993   }
1994 }
1995 #endif // PRODUCT
1996 void G1ConcurrentMark::create_live_data() {
1997   _g1h->g1_rem_set()->create_card_live_data(_parallel_workers, _nextMarkBitMap);
1998 }
1999 
2000 void G1ConcurrentMark::finalize_live_data() {
2001   _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _nextMarkBitMap);
2002 }
2003 
2004 void G1ConcurrentMark::verify_live_data() {
2005   _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _nextMarkBitMap);
2006 }
2007 
2008 void G1ConcurrentMark::clear_live_data(WorkGang* workers) {
2009   _g1h->g1_rem_set()->clear_card_live_data(workers);
2010 }
2011 
2012 #ifdef ASSERT
2013 void G1ConcurrentMark::verify_live_data_clear() {
2014   _g1h->g1_rem_set()->verify_card_live_data_is_clear();
2015 }
2016 #endif
2017 
2018 void G1ConcurrentMark::print_stats() {
2019   if (!log_is_enabled(Debug, gc, stats)) {
2020     return;
2021   }
2022   log_debug(gc, stats)("---------------------------------------------------------------------");
2023   for (size_t i = 0; i < _active_tasks; ++i) {
2024     _tasks[i]->print_stats();
2025     log_debug(gc, stats)("---------------------------------------------------------------------");
2026   }
2027 }
2028 
2029 void G1ConcurrentMark::abort() {
2030   if (!cmThread()->during_cycle() || _has_aborted) {
2031     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2032     return;
2033   }
2034 
2035   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2036   // concurrent bitmap clearing.
2037   {
2038     GCTraceTime(Debug, gc)("Clear Next Bitmap");
2039     clear_bitmap(_nextMarkBitMap, _g1h->workers(), false);
2040   }
2041   // Note we cannot clear the previous marking bitmap here
2042   // since VerifyDuringGC verifies the objects marked during
2043   // a full GC against the previous bitmap.
2044 
2045   {
2046     GCTraceTime(Debug, gc)("Clear Live Data");
2047     clear_live_data(_g1h->workers());
2048   }
2049   DEBUG_ONLY({
2050     GCTraceTime(Debug, gc)("Verify Live Data Clear");
2051     verify_live_data_clear();
2052   })
2053   // Empty mark stack
2054   reset_marking_state();
2055   for (uint i = 0; i < _max_worker_id; ++i) {
2056     _tasks[i]->clear_region_fields();
2057   }
2058   _first_overflow_barrier_sync.abort();
2059   _second_overflow_barrier_sync.abort();
2060   _has_aborted = true;
2061 
2062   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2063   satb_mq_set.abandon_partial_marking();
2064   // This can be called either during or outside marking, we'll read
2065   // the expected_active value from the SATB queue set.
2066   satb_mq_set.set_active_all_threads(
2067                                  false, /* new active value */
2068                                  satb_mq_set.is_active() /* expected_active */);
2069 }
2070 
2071 static void print_ms_time_info(const char* prefix, const char* name,
2072                                NumberSeq& ns) {
2073   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2074                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2075   if (ns.num() > 0) {
2076     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2077                            prefix, ns.sd(), ns.maximum());
2078   }
2079 }
2080 
2081 void G1ConcurrentMark::print_summary_info() {
2082   Log(gc, marking) log;
2083   if (!log.is_trace()) {
2084     return;
2085   }
2086 
2087   log.trace(" Concurrent marking:");
2088   print_ms_time_info("  ", "init marks", _init_times);
2089   print_ms_time_info("  ", "remarks", _remark_times);
2090   {
2091     print_ms_time_info("     ", "final marks", _remark_mark_times);
2092     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2093 
2094   }
2095   print_ms_time_info("  ", "cleanups", _cleanup_times);
2096   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2097             _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2098   if (G1ScrubRemSets) {
2099     log.trace("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
2100               _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2101   }
2102   log.trace("  Total stop_world time = %8.2f s.",
2103             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2104   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2105             cmThread()->vtime_accum(), cmThread()->vtime_mark_accum());
2106 }
2107 
2108 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2109   _parallel_workers->print_worker_threads_on(st);
2110 }
2111 
2112 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2113   _parallel_workers->threads_do(tc);
2114 }
2115 
2116 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2117   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2118       p2i(_prevMarkBitMap), p2i(_nextMarkBitMap));
2119   _prevMarkBitMap->print_on_error(st, " Prev Bits: ");
2120   _nextMarkBitMap->print_on_error(st, " Next Bits: ");
2121 }
2122 
2123 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
2124   assert(addr < _cm->finger(), "invariant");
2125   assert(addr >= _task->finger(), "invariant");
2126 
2127   // We move that task's local finger along.
2128   _task->move_finger_to(addr);
2129 
2130   _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
2131   // we only partially drain the local queue and global stack
2132   _task->drain_local_queue(true);
2133   _task->drain_global_stack(true);
2134 
2135   // if the has_aborted flag has been raised, we need to bail out of
2136   // the iteration
2137   return !_task->has_aborted();
2138 }
2139 
2140 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2141   ReferenceProcessor* result = g1h->ref_processor_cm();
2142   assert(result != NULL, "CM reference processor should not be NULL");
2143   return result;
2144 }
2145 
2146 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2147                                G1ConcurrentMark* cm,
2148                                G1CMTask* task)
2149   : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2150     _g1h(g1h), _cm(cm), _task(task)
2151 { }
2152 
2153 void G1CMTask::setup_for_region(HeapRegion* hr) {
2154   assert(hr != NULL,
2155         "claim_region() should have filtered out NULL regions");
2156   _curr_region  = hr;
2157   _finger       = hr->bottom();
2158   update_region_limit();
2159 }
2160 
2161 void G1CMTask::update_region_limit() {
2162   HeapRegion* hr            = _curr_region;
2163   HeapWord* bottom          = hr->bottom();
2164   HeapWord* limit           = hr->next_top_at_mark_start();
2165 
2166   if (limit == bottom) {
2167     // The region was collected underneath our feet.
2168     // We set the finger to bottom to ensure that the bitmap
2169     // iteration that will follow this will not do anything.
2170     // (this is not a condition that holds when we set the region up,
2171     // as the region is not supposed to be empty in the first place)
2172     _finger = bottom;
2173   } else if (limit >= _region_limit) {
2174     assert(limit >= _finger, "peace of mind");
2175   } else {
2176     assert(limit < _region_limit, "only way to get here");
2177     // This can happen under some pretty unusual circumstances.  An
2178     // evacuation pause empties the region underneath our feet (NTAMS
2179     // at bottom). We then do some allocation in the region (NTAMS
2180     // stays at bottom), followed by the region being used as a GC
2181     // alloc region (NTAMS will move to top() and the objects
2182     // originally below it will be grayed). All objects now marked in
2183     // the region are explicitly grayed, if below the global finger,
2184     // and we do not need in fact to scan anything else. So, we simply
2185     // set _finger to be limit to ensure that the bitmap iteration
2186     // doesn't do anything.
2187     _finger = limit;
2188   }
2189 
2190   _region_limit = limit;
2191 }
2192 
2193 void G1CMTask::giveup_current_region() {
2194   assert(_curr_region != NULL, "invariant");
2195   clear_region_fields();
2196 }
2197 
2198 void G1CMTask::clear_region_fields() {
2199   // Values for these three fields that indicate that we're not
2200   // holding on to a region.
2201   _curr_region   = NULL;
2202   _finger        = NULL;
2203   _region_limit  = NULL;
2204 }
2205 
2206 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2207   if (cm_oop_closure == NULL) {
2208     assert(_cm_oop_closure != NULL, "invariant");
2209   } else {
2210     assert(_cm_oop_closure == NULL, "invariant");
2211   }
2212   _cm_oop_closure = cm_oop_closure;
2213 }
2214 
2215 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) {
2216   guarantee(nextMarkBitMap != NULL, "invariant");
2217   _nextMarkBitMap                = nextMarkBitMap;
2218   clear_region_fields();
2219 
2220   _calls                         = 0;
2221   _elapsed_time_ms               = 0.0;
2222   _termination_time_ms           = 0.0;
2223   _termination_start_time_ms     = 0.0;
2224 }
2225 
2226 bool G1CMTask::should_exit_termination() {
2227   regular_clock_call();
2228   // This is called when we are in the termination protocol. We should
2229   // quit if, for some reason, this task wants to abort or the global
2230   // stack is not empty (this means that we can get work from it).
2231   return !_cm->mark_stack_empty() || has_aborted();
2232 }
2233 
2234 void G1CMTask::reached_limit() {
2235   assert(_words_scanned >= _words_scanned_limit ||
2236          _refs_reached >= _refs_reached_limit ,
2237          "shouldn't have been called otherwise");
2238   regular_clock_call();
2239 }
2240 
2241 void G1CMTask::regular_clock_call() {
2242   if (has_aborted()) return;
2243 
2244   // First, we need to recalculate the words scanned and refs reached
2245   // limits for the next clock call.
2246   recalculate_limits();
2247 
2248   // During the regular clock call we do the following
2249 
2250   // (1) If an overflow has been flagged, then we abort.
2251   if (_cm->has_overflown()) {
2252     set_has_aborted();
2253     return;
2254   }
2255 
2256   // If we are not concurrent (i.e. we're doing remark) we don't need
2257   // to check anything else. The other steps are only needed during
2258   // the concurrent marking phase.
2259   if (!concurrent()) return;
2260 
2261   // (2) If marking has been aborted for Full GC, then we also abort.
2262   if (_cm->has_aborted()) {
2263     set_has_aborted();
2264     return;
2265   }
2266 
2267   double curr_time_ms = os::elapsedVTime() * 1000.0;
2268 
2269   // (4) We check whether we should yield. If we have to, then we abort.
2270   if (SuspendibleThreadSet::should_yield()) {
2271     // We should yield. To do this we abort the task. The caller is
2272     // responsible for yielding.
2273     set_has_aborted();
2274     return;
2275   }
2276 
2277   // (5) We check whether we've reached our time quota. If we have,
2278   // then we abort.
2279   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2280   if (elapsed_time_ms > _time_target_ms) {
2281     set_has_aborted();
2282     _has_timed_out = true;
2283     return;
2284   }
2285 
2286   // (6) Finally, we check whether there are enough completed STAB
2287   // buffers available for processing. If there are, we abort.
2288   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2289   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2290     // we do need to process SATB buffers, we'll abort and restart
2291     // the marking task to do so
2292     set_has_aborted();
2293     return;
2294   }
2295 }
2296 
2297 void G1CMTask::recalculate_limits() {
2298   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2299   _words_scanned_limit      = _real_words_scanned_limit;
2300 
2301   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2302   _refs_reached_limit       = _real_refs_reached_limit;
2303 }
2304 
2305 void G1CMTask::decrease_limits() {
2306   // This is called when we believe that we're going to do an infrequent
2307   // operation which will increase the per byte scanned cost (i.e. move
2308   // entries to/from the global stack). It basically tries to decrease the
2309   // scanning limit so that the clock is called earlier.
2310 
2311   _words_scanned_limit = _real_words_scanned_limit -
2312     3 * words_scanned_period / 4;
2313   _refs_reached_limit  = _real_refs_reached_limit -
2314     3 * refs_reached_period / 4;
2315 }
2316 
2317 void G1CMTask::move_entries_to_global_stack() {
2318   // Local array where we'll store the entries that will be popped
2319   // from the local queue.
2320   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2321 
2322   size_t n = 0;
2323   G1TaskQueueEntry task_entry;
2324   while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2325     buffer[n] = task_entry;
2326     ++n;
2327   }
2328   if (n < G1CMMarkStack::EntriesPerChunk) {
2329     buffer[n] = G1TaskQueueEntry();
2330   }
2331 
2332   if (n > 0) {
2333     if (!_cm->mark_stack_push(buffer)) {
2334       set_has_aborted();
2335     }
2336   }
2337 
2338   // This operation was quite expensive, so decrease the limits.
2339   decrease_limits();
2340 }
2341 
2342 bool G1CMTask::get_entries_from_global_stack() {
2343   // Local array where we'll store the entries that will be popped
2344   // from the global stack.
2345   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2346 
2347   if (!_cm->mark_stack_pop(buffer)) {
2348     return false;
2349   }
2350 
2351   // We did actually pop at least one entry.
2352   for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2353     G1TaskQueueEntry task_entry = buffer[i];
2354     if (task_entry.is_null()) {
2355       break;
2356     }
2357     assert(task_entry.is_array_slice() || task_entry.obj()->is_oop(), "Element " PTR_FORMAT " must be an array slice or oop", p2i(task_entry.obj()));
2358     bool success = _task_queue->push(task_entry);
2359     // We only call this when the local queue is empty or under a
2360     // given target limit. So, we do not expect this push to fail.
2361     assert(success, "invariant");
2362   }
2363 
2364   // This operation was quite expensive, so decrease the limits
2365   decrease_limits();
2366   return true;
2367 }
2368 
2369 void G1CMTask::drain_local_queue(bool partially) {
2370   if (has_aborted()) {
2371     return;
2372   }
2373 
2374   // Decide what the target size is, depending whether we're going to
2375   // drain it partially (so that other tasks can steal if they run out
2376   // of things to do) or totally (at the very end).
2377   size_t target_size;
2378   if (partially) {
2379     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2380   } else {
2381     target_size = 0;
2382   }
2383 
2384   if (_task_queue->size() > target_size) {
2385     G1TaskQueueEntry entry;
2386     bool ret = _task_queue->pop_local(entry);
2387     while (ret) {
2388       scan_task_entry(entry);
2389       if (_task_queue->size() <= target_size || has_aborted()) {
2390         ret = false;
2391       } else {
2392         ret = _task_queue->pop_local(entry);
2393       }
2394     }
2395   }
2396 }
2397 
2398 void G1CMTask::drain_global_stack(bool partially) {
2399   if (has_aborted()) return;
2400 
2401   // We have a policy to drain the local queue before we attempt to
2402   // drain the global stack.
2403   assert(partially || _task_queue->size() == 0, "invariant");
2404 
2405   // Decide what the target size is, depending whether we're going to
2406   // drain it partially (so that other tasks can steal if they run out
2407   // of things to do) or totally (at the very end).
2408   // Notice that when draining the global mark stack partially, due to the racyness
2409   // of the mark stack size update we might in fact drop below the target. But,
2410   // this is not a problem.
2411   // In case of total draining, we simply process until the global mark stack is
2412   // totally empty, disregarding the size counter.
2413   if (partially) {
2414     size_t const target_size = _cm->partial_mark_stack_size_target();
2415     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2416       if (get_entries_from_global_stack()) {
2417         drain_local_queue(partially);
2418       }
2419     }
2420   } else {
2421     while (!has_aborted() && get_entries_from_global_stack()) {
2422       drain_local_queue(partially);
2423     }
2424   }
2425 }
2426 
2427 // SATB Queue has several assumptions on whether to call the par or
2428 // non-par versions of the methods. this is why some of the code is
2429 // replicated. We should really get rid of the single-threaded version
2430 // of the code to simplify things.
2431 void G1CMTask::drain_satb_buffers() {
2432   if (has_aborted()) return;
2433 
2434   // We set this so that the regular clock knows that we're in the
2435   // middle of draining buffers and doesn't set the abort flag when it
2436   // notices that SATB buffers are available for draining. It'd be
2437   // very counter productive if it did that. :-)
2438   _draining_satb_buffers = true;
2439 
2440   G1CMSATBBufferClosure satb_cl(this, _g1h);
2441   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2442 
2443   // This keeps claiming and applying the closure to completed buffers
2444   // until we run out of buffers or we need to abort.
2445   while (!has_aborted() &&
2446          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2447     regular_clock_call();
2448   }
2449 
2450   _draining_satb_buffers = false;
2451 
2452   assert(has_aborted() ||
2453          concurrent() ||
2454          satb_mq_set.completed_buffers_num() == 0, "invariant");
2455 
2456   // again, this was a potentially expensive operation, decrease the
2457   // limits to get the regular clock call early
2458   decrease_limits();
2459 }
2460 
2461 void G1CMTask::print_stats() {
2462   log_debug(gc, stats)("Marking Stats, task = %u, calls = %d",
2463                        _worker_id, _calls);
2464   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2465                        _elapsed_time_ms, _termination_time_ms);
2466   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
2467                        _step_times_ms.num(), _step_times_ms.avg(),
2468                        _step_times_ms.sd());
2469   log_debug(gc, stats)("                    max = %1.2lfms, total = %1.2lfms",
2470                        _step_times_ms.maximum(), _step_times_ms.sum());
2471 }
2472 
2473 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry) {
2474   return _task_queues->steal(worker_id, hash_seed, task_entry);
2475 }
2476 
2477 /*****************************************************************************
2478 
2479     The do_marking_step(time_target_ms, ...) method is the building
2480     block of the parallel marking framework. It can be called in parallel
2481     with other invocations of do_marking_step() on different tasks
2482     (but only one per task, obviously) and concurrently with the
2483     mutator threads, or during remark, hence it eliminates the need
2484     for two versions of the code. When called during remark, it will
2485     pick up from where the task left off during the concurrent marking
2486     phase. Interestingly, tasks are also claimable during evacuation
2487     pauses too, since do_marking_step() ensures that it aborts before
2488     it needs to yield.
2489 
2490     The data structures that it uses to do marking work are the
2491     following:
2492 
2493       (1) Marking Bitmap. If there are gray objects that appear only
2494       on the bitmap (this happens either when dealing with an overflow
2495       or when the initial marking phase has simply marked the roots
2496       and didn't push them on the stack), then tasks claim heap
2497       regions whose bitmap they then scan to find gray objects. A
2498       global finger indicates where the end of the last claimed region
2499       is. A local finger indicates how far into the region a task has
2500       scanned. The two fingers are used to determine how to gray an
2501       object (i.e. whether simply marking it is OK, as it will be
2502       visited by a task in the future, or whether it needs to be also
2503       pushed on a stack).
2504 
2505       (2) Local Queue. The local queue of the task which is accessed
2506       reasonably efficiently by the task. Other tasks can steal from
2507       it when they run out of work. Throughout the marking phase, a
2508       task attempts to keep its local queue short but not totally
2509       empty, so that entries are available for stealing by other
2510       tasks. Only when there is no more work, a task will totally
2511       drain its local queue.
2512 
2513       (3) Global Mark Stack. This handles local queue overflow. During
2514       marking only sets of entries are moved between it and the local
2515       queues, as access to it requires a mutex and more fine-grain
2516       interaction with it which might cause contention. If it
2517       overflows, then the marking phase should restart and iterate
2518       over the bitmap to identify gray objects. Throughout the marking
2519       phase, tasks attempt to keep the global mark stack at a small
2520       length but not totally empty, so that entries are available for
2521       popping by other tasks. Only when there is no more work, tasks
2522       will totally drain the global mark stack.
2523 
2524       (4) SATB Buffer Queue. This is where completed SATB buffers are
2525       made available. Buffers are regularly removed from this queue
2526       and scanned for roots, so that the queue doesn't get too
2527       long. During remark, all completed buffers are processed, as
2528       well as the filled in parts of any uncompleted buffers.
2529 
2530     The do_marking_step() method tries to abort when the time target
2531     has been reached. There are a few other cases when the
2532     do_marking_step() method also aborts:
2533 
2534       (1) When the marking phase has been aborted (after a Full GC).
2535 
2536       (2) When a global overflow (on the global stack) has been
2537       triggered. Before the task aborts, it will actually sync up with
2538       the other tasks to ensure that all the marking data structures
2539       (local queues, stacks, fingers etc.)  are re-initialized so that
2540       when do_marking_step() completes, the marking phase can
2541       immediately restart.
2542 
2543       (3) When enough completed SATB buffers are available. The
2544       do_marking_step() method only tries to drain SATB buffers right
2545       at the beginning. So, if enough buffers are available, the
2546       marking step aborts and the SATB buffers are processed at
2547       the beginning of the next invocation.
2548 
2549       (4) To yield. when we have to yield then we abort and yield
2550       right at the end of do_marking_step(). This saves us from a lot
2551       of hassle as, by yielding we might allow a Full GC. If this
2552       happens then objects will be compacted underneath our feet, the
2553       heap might shrink, etc. We save checking for this by just
2554       aborting and doing the yield right at the end.
2555 
2556     From the above it follows that the do_marking_step() method should
2557     be called in a loop (or, otherwise, regularly) until it completes.
2558 
2559     If a marking step completes without its has_aborted() flag being
2560     true, it means it has completed the current marking phase (and
2561     also all other marking tasks have done so and have all synced up).
2562 
2563     A method called regular_clock_call() is invoked "regularly" (in
2564     sub ms intervals) throughout marking. It is this clock method that
2565     checks all the abort conditions which were mentioned above and
2566     decides when the task should abort. A work-based scheme is used to
2567     trigger this clock method: when the number of object words the
2568     marking phase has scanned or the number of references the marking
2569     phase has visited reach a given limit. Additional invocations to
2570     the method clock have been planted in a few other strategic places
2571     too. The initial reason for the clock method was to avoid calling
2572     vtime too regularly, as it is quite expensive. So, once it was in
2573     place, it was natural to piggy-back all the other conditions on it
2574     too and not constantly check them throughout the code.
2575 
2576     If do_termination is true then do_marking_step will enter its
2577     termination protocol.
2578 
2579     The value of is_serial must be true when do_marking_step is being
2580     called serially (i.e. by the VMThread) and do_marking_step should
2581     skip any synchronization in the termination and overflow code.
2582     Examples include the serial remark code and the serial reference
2583     processing closures.
2584 
2585     The value of is_serial must be false when do_marking_step is
2586     being called by any of the worker threads in a work gang.
2587     Examples include the concurrent marking code (CMMarkingTask),
2588     the MT remark code, and the MT reference processing closures.
2589 
2590  *****************************************************************************/
2591 
2592 void G1CMTask::do_marking_step(double time_target_ms,
2593                                bool do_termination,
2594                                bool is_serial) {
2595   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2596   assert(concurrent() == _cm->concurrent(), "they should be the same");
2597 
2598   G1Policy* g1_policy = _g1h->g1_policy();
2599   assert(_task_queues != NULL, "invariant");
2600   assert(_task_queue != NULL, "invariant");
2601   assert(_task_queues->queue(_worker_id) == _task_queue, "invariant");
2602 
2603   assert(!_claimed,
2604          "only one thread should claim this task at any one time");
2605 
2606   // OK, this doesn't safeguard again all possible scenarios, as it is
2607   // possible for two threads to set the _claimed flag at the same
2608   // time. But it is only for debugging purposes anyway and it will
2609   // catch most problems.
2610   _claimed = true;
2611 
2612   _start_time_ms = os::elapsedVTime() * 1000.0;
2613 
2614   // If do_stealing is true then do_marking_step will attempt to
2615   // steal work from the other G1CMTasks. It only makes sense to
2616   // enable stealing when the termination protocol is enabled
2617   // and do_marking_step() is not being called serially.
2618   bool do_stealing = do_termination && !is_serial;
2619 
2620   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2621   _time_target_ms = time_target_ms - diff_prediction_ms;
2622 
2623   // set up the variables that are used in the work-based scheme to
2624   // call the regular clock method
2625   _words_scanned = 0;
2626   _refs_reached  = 0;
2627   recalculate_limits();
2628 
2629   // clear all flags
2630   clear_has_aborted();
2631   _has_timed_out = false;
2632   _draining_satb_buffers = false;
2633 
2634   ++_calls;
2635 
2636   // Set up the bitmap and oop closures. Anything that uses them is
2637   // eventually called from this method, so it is OK to allocate these
2638   // statically.
2639   G1CMBitMapClosure bitmap_closure(this, _cm);
2640   G1CMOopClosure    cm_oop_closure(_g1h, _cm, this);
2641   set_cm_oop_closure(&cm_oop_closure);
2642 
2643   if (_cm->has_overflown()) {
2644     // This can happen if the mark stack overflows during a GC pause
2645     // and this task, after a yield point, restarts. We have to abort
2646     // as we need to get into the overflow protocol which happens
2647     // right at the end of this task.
2648     set_has_aborted();
2649   }
2650 
2651   // First drain any available SATB buffers. After this, we will not
2652   // look at SATB buffers before the next invocation of this method.
2653   // If enough completed SATB buffers are queued up, the regular clock
2654   // will abort this task so that it restarts.
2655   drain_satb_buffers();
2656   // ...then partially drain the local queue and the global stack
2657   drain_local_queue(true);
2658   drain_global_stack(true);
2659 
2660   do {
2661     if (!has_aborted() && _curr_region != NULL) {
2662       // This means that we're already holding on to a region.
2663       assert(_finger != NULL, "if region is not NULL, then the finger "
2664              "should not be NULL either");
2665 
2666       // We might have restarted this task after an evacuation pause
2667       // which might have evacuated the region we're holding on to
2668       // underneath our feet. Let's read its limit again to make sure
2669       // that we do not iterate over a region of the heap that
2670       // contains garbage (update_region_limit() will also move
2671       // _finger to the start of the region if it is found empty).
2672       update_region_limit();
2673       // We will start from _finger not from the start of the region,
2674       // as we might be restarting this task after aborting half-way
2675       // through scanning this region. In this case, _finger points to
2676       // the address where we last found a marked object. If this is a
2677       // fresh region, _finger points to start().
2678       MemRegion mr = MemRegion(_finger, _region_limit);
2679 
2680       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2681              "humongous regions should go around loop once only");
2682 
2683       // Some special cases:
2684       // If the memory region is empty, we can just give up the region.
2685       // If the current region is humongous then we only need to check
2686       // the bitmap for the bit associated with the start of the object,
2687       // scan the object if it's live, and give up the region.
2688       // Otherwise, let's iterate over the bitmap of the part of the region
2689       // that is left.
2690       // If the iteration is successful, give up the region.
2691       if (mr.is_empty()) {
2692         giveup_current_region();
2693         regular_clock_call();
2694       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2695         if (_nextMarkBitMap->is_marked(mr.start())) {
2696           // The object is marked - apply the closure
2697           bitmap_closure.do_addr(mr.start());
2698         }
2699         // Even if this task aborted while scanning the humongous object
2700         // we can (and should) give up the current region.
2701         giveup_current_region();
2702         regular_clock_call();
2703       } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) {
2704         giveup_current_region();
2705         regular_clock_call();
2706       } else {
2707         assert(has_aborted(), "currently the only way to do so");
2708         // The only way to abort the bitmap iteration is to return
2709         // false from the do_bit() method. However, inside the
2710         // do_bit() method we move the _finger to point to the
2711         // object currently being looked at. So, if we bail out, we
2712         // have definitely set _finger to something non-null.
2713         assert(_finger != NULL, "invariant");
2714 
2715         // Region iteration was actually aborted. So now _finger
2716         // points to the address of the object we last scanned. If we
2717         // leave it there, when we restart this task, we will rescan
2718         // the object. It is easy to avoid this. We move the finger by
2719         // enough to point to the next possible object header (the
2720         // bitmap knows by how much we need to move it as it knows its
2721         // granularity).
2722         assert(_finger < _region_limit, "invariant");
2723         HeapWord* new_finger = _nextMarkBitMap->addr_after_obj(_finger);
2724         // Check if bitmap iteration was aborted while scanning the last object
2725         if (new_finger >= _region_limit) {
2726           giveup_current_region();
2727         } else {
2728           move_finger_to(new_finger);
2729         }
2730       }
2731     }
2732     // At this point we have either completed iterating over the
2733     // region we were holding on to, or we have aborted.
2734 
2735     // We then partially drain the local queue and the global stack.
2736     // (Do we really need this?)
2737     drain_local_queue(true);
2738     drain_global_stack(true);
2739 
2740     // Read the note on the claim_region() method on why it might
2741     // return NULL with potentially more regions available for
2742     // claiming and why we have to check out_of_regions() to determine
2743     // whether we're done or not.
2744     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2745       // We are going to try to claim a new region. We should have
2746       // given up on the previous one.
2747       // Separated the asserts so that we know which one fires.
2748       assert(_curr_region  == NULL, "invariant");
2749       assert(_finger       == NULL, "invariant");
2750       assert(_region_limit == NULL, "invariant");
2751       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2752       if (claimed_region != NULL) {
2753         // Yes, we managed to claim one
2754         setup_for_region(claimed_region);
2755         assert(_curr_region == claimed_region, "invariant");
2756       }
2757       // It is important to call the regular clock here. It might take
2758       // a while to claim a region if, for example, we hit a large
2759       // block of empty regions. So we need to call the regular clock
2760       // method once round the loop to make sure it's called
2761       // frequently enough.
2762       regular_clock_call();
2763     }
2764 
2765     if (!has_aborted() && _curr_region == NULL) {
2766       assert(_cm->out_of_regions(),
2767              "at this point we should be out of regions");
2768     }
2769   } while ( _curr_region != NULL && !has_aborted());
2770 
2771   if (!has_aborted()) {
2772     // We cannot check whether the global stack is empty, since other
2773     // tasks might be pushing objects to it concurrently.
2774     assert(_cm->out_of_regions(),
2775            "at this point we should be out of regions");
2776     // Try to reduce the number of available SATB buffers so that
2777     // remark has less work to do.
2778     drain_satb_buffers();
2779   }
2780 
2781   // Since we've done everything else, we can now totally drain the
2782   // local queue and global stack.
2783   drain_local_queue(false);
2784   drain_global_stack(false);
2785 
2786   // Attempt at work stealing from other task's queues.
2787   if (do_stealing && !has_aborted()) {
2788     // We have not aborted. This means that we have finished all that
2789     // we could. Let's try to do some stealing...
2790 
2791     // We cannot check whether the global stack is empty, since other
2792     // tasks might be pushing objects to it concurrently.
2793     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2794            "only way to reach here");
2795     while (!has_aborted()) {
2796       G1TaskQueueEntry entry;
2797       if (_cm->try_stealing(_worker_id, &_hash_seed, entry)) {
2798         scan_task_entry(entry);
2799 
2800         // And since we're towards the end, let's totally drain the
2801         // local queue and global stack.
2802         drain_local_queue(false);
2803         drain_global_stack(false);
2804       } else {
2805         break;
2806       }
2807     }
2808   }
2809 
2810   // We still haven't aborted. Now, let's try to get into the
2811   // termination protocol.
2812   if (do_termination && !has_aborted()) {
2813     // We cannot check whether the global stack is empty, since other
2814     // tasks might be concurrently pushing objects on it.
2815     // Separated the asserts so that we know which one fires.
2816     assert(_cm->out_of_regions(), "only way to reach here");
2817     assert(_task_queue->size() == 0, "only way to reach here");
2818     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2819 
2820     // The G1CMTask class also extends the TerminatorTerminator class,
2821     // hence its should_exit_termination() method will also decide
2822     // whether to exit the termination protocol or not.
2823     bool finished = (is_serial ||
2824                      _cm->terminator()->offer_termination(this));
2825     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2826     _termination_time_ms +=
2827       termination_end_time_ms - _termination_start_time_ms;
2828 
2829     if (finished) {
2830       // We're all done.
2831 
2832       if (_worker_id == 0) {
2833         // let's allow task 0 to do this
2834         if (concurrent()) {
2835           assert(_cm->concurrent_marking_in_progress(), "invariant");
2836           // we need to set this to false before the next
2837           // safepoint. This way we ensure that the marking phase
2838           // doesn't observe any more heap expansions.
2839           _cm->clear_concurrent_marking_in_progress();
2840         }
2841       }
2842 
2843       // We can now guarantee that the global stack is empty, since
2844       // all other tasks have finished. We separated the guarantees so
2845       // that, if a condition is false, we can immediately find out
2846       // which one.
2847       guarantee(_cm->out_of_regions(), "only way to reach here");
2848       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2849       guarantee(_task_queue->size() == 0, "only way to reach here");
2850       guarantee(!_cm->has_overflown(), "only way to reach here");
2851     } else {
2852       // Apparently there's more work to do. Let's abort this task. It
2853       // will restart it and we can hopefully find more things to do.
2854       set_has_aborted();
2855     }
2856   }
2857 
2858   // Mainly for debugging purposes to make sure that a pointer to the
2859   // closure which was statically allocated in this frame doesn't
2860   // escape it by accident.
2861   set_cm_oop_closure(NULL);
2862   double end_time_ms = os::elapsedVTime() * 1000.0;
2863   double elapsed_time_ms = end_time_ms - _start_time_ms;
2864   // Update the step history.
2865   _step_times_ms.add(elapsed_time_ms);
2866 
2867   if (has_aborted()) {
2868     // The task was aborted for some reason.
2869     if (_has_timed_out) {
2870       double diff_ms = elapsed_time_ms - _time_target_ms;
2871       // Keep statistics of how well we did with respect to hitting
2872       // our target only if we actually timed out (if we aborted for
2873       // other reasons, then the results might get skewed).
2874       _marking_step_diffs_ms.add(diff_ms);
2875     }
2876 
2877     if (_cm->has_overflown()) {
2878       // This is the interesting one. We aborted because a global
2879       // overflow was raised. This means we have to restart the
2880       // marking phase and start iterating over regions. However, in
2881       // order to do this we have to make sure that all tasks stop
2882       // what they are doing and re-initialize in a safe manner. We
2883       // will achieve this with the use of two barrier sync points.
2884 
2885       if (!is_serial) {
2886         // We only need to enter the sync barrier if being called
2887         // from a parallel context
2888         _cm->enter_first_sync_barrier(_worker_id);
2889 
2890         // When we exit this sync barrier we know that all tasks have
2891         // stopped doing marking work. So, it's now safe to
2892         // re-initialize our data structures. At the end of this method,
2893         // task 0 will clear the global data structures.
2894       }
2895 
2896       // We clear the local state of this task...
2897       clear_region_fields();
2898 
2899       if (!is_serial) {
2900         // ...and enter the second barrier.
2901         _cm->enter_second_sync_barrier(_worker_id);
2902       }
2903       // At this point, if we're during the concurrent phase of
2904       // marking, everything has been re-initialized and we're
2905       // ready to restart.
2906     }
2907   }
2908 
2909   _claimed = false;
2910 }
2911 
2912 G1CMTask::G1CMTask(uint worker_id,
2913                    G1ConcurrentMark* cm,
2914                    G1CMTaskQueue* task_queue,
2915                    G1CMTaskQueueSet* task_queues)
2916   : _g1h(G1CollectedHeap::heap()),
2917     _worker_id(worker_id), _cm(cm),
2918     _objArray_processor(this),
2919     _claimed(false),
2920     _nextMarkBitMap(NULL), _hash_seed(17),
2921     _task_queue(task_queue),
2922     _task_queues(task_queues),
2923     _cm_oop_closure(NULL) {
2924   guarantee(task_queue != NULL, "invariant");
2925   guarantee(task_queues != NULL, "invariant");
2926 
2927   _marking_step_diffs_ms.add(0.5);
2928 }
2929 
2930 // These are formatting macros that are used below to ensure
2931 // consistent formatting. The *_H_* versions are used to format the
2932 // header for a particular value and they should be kept consistent
2933 // with the corresponding macro. Also note that most of the macros add
2934 // the necessary white space (as a prefix) which makes them a bit
2935 // easier to compose.
2936 
2937 // All the output lines are prefixed with this string to be able to
2938 // identify them easily in a large log file.
2939 #define G1PPRL_LINE_PREFIX            "###"
2940 
2941 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2942 #ifdef _LP64
2943 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2944 #else // _LP64
2945 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2946 #endif // _LP64
2947 
2948 // For per-region info
2949 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2950 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2951 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2952 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2953 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2954 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2955 
2956 // For summary info
2957 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2958 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2959 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2960 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2961 
2962 G1PrintRegionLivenessInfoClosure::
2963 G1PrintRegionLivenessInfoClosure(const char* phase_name)
2964   : _total_used_bytes(0), _total_capacity_bytes(0),
2965     _total_prev_live_bytes(0), _total_next_live_bytes(0),
2966     _total_remset_bytes(0), _total_strong_code_roots_bytes(0) {
2967   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2968   MemRegion g1_reserved = g1h->g1_reserved();
2969   double now = os::elapsedTime();
2970 
2971   // Print the header of the output.
2972   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2973   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2974                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2975                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2976                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2977                           HeapRegion::GrainBytes);
2978   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2979   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2980                           G1PPRL_TYPE_H_FORMAT
2981                           G1PPRL_ADDR_BASE_H_FORMAT
2982                           G1PPRL_BYTE_H_FORMAT
2983                           G1PPRL_BYTE_H_FORMAT
2984                           G1PPRL_BYTE_H_FORMAT
2985                           G1PPRL_DOUBLE_H_FORMAT
2986                           G1PPRL_BYTE_H_FORMAT
2987                           G1PPRL_BYTE_H_FORMAT,
2988                           "type", "address-range",
2989                           "used", "prev-live", "next-live", "gc-eff",
2990                           "remset", "code-roots");
2991   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2992                           G1PPRL_TYPE_H_FORMAT
2993                           G1PPRL_ADDR_BASE_H_FORMAT
2994                           G1PPRL_BYTE_H_FORMAT
2995                           G1PPRL_BYTE_H_FORMAT
2996                           G1PPRL_BYTE_H_FORMAT
2997                           G1PPRL_DOUBLE_H_FORMAT
2998                           G1PPRL_BYTE_H_FORMAT
2999                           G1PPRL_BYTE_H_FORMAT,
3000                           "", "",
3001                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3002                           "(bytes)", "(bytes)");
3003 }
3004 
3005 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
3006   const char* type       = r->get_type_str();
3007   HeapWord* bottom       = r->bottom();
3008   HeapWord* end          = r->end();
3009   size_t capacity_bytes  = r->capacity();
3010   size_t used_bytes      = r->used();
3011   size_t prev_live_bytes = r->live_bytes();
3012   size_t next_live_bytes = r->next_live_bytes();
3013   double gc_eff          = r->gc_efficiency();
3014   size_t remset_bytes    = r->rem_set()->mem_size();
3015   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3016 
3017   _total_used_bytes      += used_bytes;
3018   _total_capacity_bytes  += capacity_bytes;
3019   _total_prev_live_bytes += prev_live_bytes;
3020   _total_next_live_bytes += next_live_bytes;
3021   _total_remset_bytes    += remset_bytes;
3022   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3023 
3024   // Print a line for this particular region.
3025   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3026                           G1PPRL_TYPE_FORMAT
3027                           G1PPRL_ADDR_BASE_FORMAT
3028                           G1PPRL_BYTE_FORMAT
3029                           G1PPRL_BYTE_FORMAT
3030                           G1PPRL_BYTE_FORMAT
3031                           G1PPRL_DOUBLE_FORMAT
3032                           G1PPRL_BYTE_FORMAT
3033                           G1PPRL_BYTE_FORMAT,
3034                           type, p2i(bottom), p2i(end),
3035                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3036                           remset_bytes, strong_code_roots_bytes);
3037 
3038   return false;
3039 }
3040 
3041 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3042   // add static memory usages to remembered set sizes
3043   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3044   // Print the footer of the output.
3045   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3046   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3047                          " SUMMARY"
3048                          G1PPRL_SUM_MB_FORMAT("capacity")
3049                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3050                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3051                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3052                          G1PPRL_SUM_MB_FORMAT("remset")
3053                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3054                          bytes_to_mb(_total_capacity_bytes),
3055                          bytes_to_mb(_total_used_bytes),
3056                          perc(_total_used_bytes, _total_capacity_bytes),
3057                          bytes_to_mb(_total_prev_live_bytes),
3058                          perc(_total_prev_live_bytes, _total_capacity_bytes),
3059                          bytes_to_mb(_total_next_live_bytes),
3060                          perc(_total_next_live_bytes, _total_capacity_bytes),
3061                          bytes_to_mb(_total_remset_bytes),
3062                          bytes_to_mb(_total_strong_code_roots_bytes));
3063 }