1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/classLoaderDataGraph.hpp"
  27 #include "code/codeCache.hpp"
  28 #include "gc/g1/g1BarrierSet.hpp"
  29 #include "gc/g1/g1CollectedHeap.inline.hpp"
  30 #include "gc/g1/g1CollectorState.hpp"
  31 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  32 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp"
  33 #include "gc/g1/g1HeapVerifier.hpp"
  34 #include "gc/g1/g1OopClosures.inline.hpp"
  35 #include "gc/g1/g1Policy.hpp"
  36 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp"
  37 #include "gc/g1/g1StringDedup.hpp"
  38 #include "gc/g1/g1ThreadLocalData.hpp"
  39 #include "gc/g1/heapRegion.inline.hpp"
  40 #include "gc/g1/heapRegionRemSet.hpp"
  41 #include "gc/g1/heapRegionSet.inline.hpp"
  42 #include "gc/shared/adaptiveSizePolicy.hpp"
  43 #include "gc/shared/gcId.hpp"
  44 #include "gc/shared/gcTimer.hpp"
  45 #include "gc/shared/gcTrace.hpp"
  46 #include "gc/shared/gcTraceTime.inline.hpp"
  47 #include "gc/shared/genOopClosures.inline.hpp"
  48 #include "gc/shared/referencePolicy.hpp"
  49 #include "gc/shared/strongRootsScope.hpp"
  50 #include "gc/shared/suspendibleThreadSet.hpp"
  51 #include "gc/shared/taskqueue.inline.hpp"
  52 #include "gc/shared/vmGCOperations.hpp"
  53 #include "gc/shared/weakProcessor.inline.hpp"
  54 #include "include/jvm.h"
  55 #include "logging/log.hpp"
  56 #include "memory/allocation.hpp"
  57 #include "memory/resourceArea.hpp"
  58 #include "oops/access.inline.hpp"
  59 #include "oops/oop.inline.hpp"
  60 #include "runtime/atomic.hpp"
  61 #include "runtime/handles.inline.hpp"
  62 #include "runtime/java.hpp"
  63 #include "runtime/prefetch.inline.hpp"
  64 #include "services/memTracker.hpp"
  65 #include "utilities/align.hpp"
  66 #include "utilities/growableArray.hpp"
  67 
  68 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
  69   assert(addr < _cm->finger(), "invariant");
  70   assert(addr >= _task->finger(), "invariant");
  71 
  72   // We move that task's local finger along.
  73   _task->move_finger_to(addr);
  74 
  75   _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
  76   // we only partially drain the local queue and global stack
  77   _task->drain_local_queue(true);
  78   _task->drain_global_stack(true);
  79 
  80   // if the has_aborted flag has been raised, we need to bail out of
  81   // the iteration
  82   return !_task->has_aborted();
  83 }
  84 
  85 G1CMMarkStack::G1CMMarkStack() :
  86   _max_chunk_capacity(0),
  87   _base(NULL),
  88   _chunk_capacity(0) {
  89   set_empty();
  90 }
  91 
  92 bool G1CMMarkStack::resize(size_t new_capacity) {
  93   assert(is_empty(), "Only resize when stack is empty.");
  94   assert(new_capacity <= _max_chunk_capacity,
  95          "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
  96 
  97   TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
  98 
  99   if (new_base == NULL) {
 100     log_warning(gc)("Failed to reserve memory for new overflow mark stack with " SIZE_FORMAT " chunks and size " SIZE_FORMAT "B.", new_capacity, new_capacity * sizeof(TaskQueueEntryChunk));
 101     return false;
 102   }
 103   // Release old mapping.
 104   if (_base != NULL) {
 105     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 106   }
 107 
 108   _base = new_base;
 109   _chunk_capacity = new_capacity;
 110   set_empty();
 111 
 112   return true;
 113 }
 114 
 115 size_t G1CMMarkStack::capacity_alignment() {
 116   return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
 117 }
 118 
 119 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
 120   guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
 121 
 122   size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
 123 
 124   _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 125   size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 126 
 127   guarantee(initial_chunk_capacity <= _max_chunk_capacity,
 128             "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
 129             _max_chunk_capacity,
 130             initial_chunk_capacity);
 131 
 132   log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
 133                 initial_chunk_capacity, _max_chunk_capacity);
 134 
 135   return resize(initial_chunk_capacity);
 136 }
 137 
 138 void G1CMMarkStack::expand() {
 139   if (_chunk_capacity == _max_chunk_capacity) {
 140     log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
 141     return;
 142   }
 143   size_t old_capacity = _chunk_capacity;
 144   // Double capacity if possible
 145   size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
 146 
 147   if (resize(new_capacity)) {
 148     log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 149                   old_capacity, new_capacity);
 150   } else {
 151     log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 152                     old_capacity, new_capacity);
 153   }
 154 }
 155 
 156 G1CMMarkStack::~G1CMMarkStack() {
 157   if (_base != NULL) {
 158     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 159   }
 160 }
 161 
 162 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
 163   elem->next = *list;
 164   *list = elem;
 165 }
 166 
 167 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
 168   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 169   add_chunk_to_list(&_chunk_list, elem);
 170   _chunks_in_chunk_list++;
 171 }
 172 
 173 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
 174   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 175   add_chunk_to_list(&_free_list, elem);
 176 }
 177 
 178 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
 179   TaskQueueEntryChunk* result = *list;
 180   if (result != NULL) {
 181     *list = (*list)->next;
 182   }
 183   return result;
 184 }
 185 
 186 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
 187   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 188   TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
 189   if (result != NULL) {
 190     _chunks_in_chunk_list--;
 191   }
 192   return result;
 193 }
 194 
 195 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
 196   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 197   return remove_chunk_from_list(&_free_list);
 198 }
 199 
 200 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
 201   // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
 202   // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
 203   // wraparound of _hwm.
 204   if (_hwm >= _chunk_capacity) {
 205     return NULL;
 206   }
 207 
 208   size_t cur_idx = Atomic::add(1u, &_hwm) - 1;
 209   if (cur_idx >= _chunk_capacity) {
 210     return NULL;
 211   }
 212 
 213   TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
 214   result->next = NULL;
 215   return result;
 216 }
 217 
 218 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
 219   // Get a new chunk.
 220   TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
 221 
 222   if (new_chunk == NULL) {
 223     // Did not get a chunk from the free list. Allocate from backing memory.
 224     new_chunk = allocate_new_chunk();
 225 
 226     if (new_chunk == NULL) {
 227       return false;
 228     }
 229   }
 230 
 231   Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 232 
 233   add_chunk_to_chunk_list(new_chunk);
 234 
 235   return true;
 236 }
 237 
 238 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
 239   TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
 240 
 241   if (cur == NULL) {
 242     return false;
 243   }
 244 
 245   Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 246 
 247   add_chunk_to_free_list(cur);
 248   return true;
 249 }
 250 
 251 void G1CMMarkStack::set_empty() {
 252   _chunks_in_chunk_list = 0;
 253   _hwm = 0;
 254   _chunk_list = NULL;
 255   _free_list = NULL;
 256 }
 257 
 258 G1CMRootRegions::G1CMRootRegions() :
 259   _survivors(NULL), _cm(NULL), _scan_in_progress(false),
 260   _should_abort(false), _claimed_survivor_index(0) { }
 261 
 262 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
 263   _survivors = survivors;
 264   _cm = cm;
 265 }
 266 
 267 void G1CMRootRegions::prepare_for_scan() {
 268   assert(!scan_in_progress(), "pre-condition");
 269 
 270   // Currently, only survivors can be root regions.
 271   _claimed_survivor_index = 0;
 272   _scan_in_progress = _survivors->regions()->is_nonempty();
 273   _should_abort = false;
 274 }
 275 
 276 HeapRegion* G1CMRootRegions::claim_next() {
 277   if (_should_abort) {
 278     // If someone has set the should_abort flag, we return NULL to
 279     // force the caller to bail out of their loop.
 280     return NULL;
 281   }
 282 
 283   // Currently, only survivors can be root regions.
 284   const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
 285 
 286   int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
 287   if (claimed_index < survivor_regions->length()) {
 288     return survivor_regions->at(claimed_index);
 289   }
 290   return NULL;
 291 }
 292 
 293 uint G1CMRootRegions::num_root_regions() const {
 294   return (uint)_survivors->regions()->length();
 295 }
 296 
 297 void G1CMRootRegions::notify_scan_done() {
 298   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 299   _scan_in_progress = false;
 300   RootRegionScan_lock->notify_all();
 301 }
 302 
 303 void G1CMRootRegions::cancel_scan() {
 304   notify_scan_done();
 305 }
 306 
 307 void G1CMRootRegions::scan_finished() {
 308   assert(scan_in_progress(), "pre-condition");
 309 
 310   // Currently, only survivors can be root regions.
 311   if (!_should_abort) {
 312     assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
 313     assert((uint)_claimed_survivor_index >= _survivors->length(),
 314            "we should have claimed all survivors, claimed index = %u, length = %u",
 315            (uint)_claimed_survivor_index, _survivors->length());
 316   }
 317 
 318   notify_scan_done();
 319 }
 320 
 321 bool G1CMRootRegions::wait_until_scan_finished() {
 322   if (!scan_in_progress()) {
 323     return false;
 324   }
 325 
 326   {
 327     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 328     while (scan_in_progress()) {
 329       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 330     }
 331   }
 332   return true;
 333 }
 334 
 335 // Returns the maximum number of workers to be used in a concurrent
 336 // phase based on the number of GC workers being used in a STW
 337 // phase.
 338 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
 339   return MAX2((num_gc_workers + 2) / 4, 1U);
 340 }
 341 
 342 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
 343                                    G1RegionToSpaceMapper* prev_bitmap_storage,
 344                                    G1RegionToSpaceMapper* next_bitmap_storage) :
 345   // _cm_thread set inside the constructor
 346   _g1h(g1h),
 347   _completed_initialization(false),
 348 
 349   _mark_bitmap_1(),
 350   _mark_bitmap_2(),
 351   _prev_mark_bitmap(&_mark_bitmap_1),
 352   _next_mark_bitmap(&_mark_bitmap_2),
 353 
 354   _heap(_g1h->reserved_region()),
 355 
 356   _root_regions(),
 357 
 358   _global_mark_stack(),
 359 
 360   // _finger set in set_non_marking_state
 361 
 362   _worker_id_offset(DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
 363   _max_num_tasks(ParallelGCThreads),
 364   // _num_active_tasks set in set_non_marking_state()
 365   // _tasks set inside the constructor
 366 
 367   _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
 368   _terminator(ParallelTaskTerminator((int) _max_num_tasks, _task_queues)),
 369 
 370   _first_overflow_barrier_sync(),
 371   _second_overflow_barrier_sync(),
 372 
 373   _has_overflown(false),
 374   _concurrent(false),
 375   _has_aborted(false),
 376   _restart_for_overflow(false),
 377   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 378   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 379 
 380   // _verbose_level set below
 381 
 382   _init_times(),
 383   _remark_times(),
 384   _remark_mark_times(),
 385   _remark_weak_ref_times(),
 386   _cleanup_times(),
 387   _total_cleanup_time(0.0),
 388 
 389   _accum_task_vtime(NULL),
 390 
 391   _concurrent_workers(NULL),
 392   _num_concurrent_workers(0),
 393   _max_concurrent_workers(0),
 394 
 395   _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)),
 396   _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC))
 397 {
 398   _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 399   _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
 400 
 401   // Create & start ConcurrentMark thread.
 402   _cm_thread = new G1ConcurrentMarkThread(this);
 403   if (_cm_thread->osthread() == NULL) {
 404     vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 405   }
 406 
 407   assert(CGC_lock != NULL, "CGC_lock must be initialized");
 408 
 409   SATBMarkQueueSet& satb_qs = G1BarrierSet::satb_mark_queue_set();
 410   satb_qs.set_buffer_size(G1SATBBufferSize);
 411 
 412   _root_regions.init(_g1h->survivor(), this);
 413 
 414   if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
 415     // Calculate the number of concurrent worker threads by scaling
 416     // the number of parallel GC threads.
 417     uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
 418     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 419   }
 420 
 421   assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
 422   if (ConcGCThreads > ParallelGCThreads) {
 423     log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
 424                     ConcGCThreads, ParallelGCThreads);
 425     return;
 426   }
 427 
 428   log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset);
 429   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 430 
 431   _num_concurrent_workers = ConcGCThreads;
 432   _max_concurrent_workers = _num_concurrent_workers;
 433 
 434   _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true);
 435   _concurrent_workers->initialize_workers();
 436 
 437   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 438     size_t mark_stack_size =
 439       MIN2(MarkStackSizeMax,
 440           MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
 441     // Verify that the calculated value for MarkStackSize is in range.
 442     // It would be nice to use the private utility routine from Arguments.
 443     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 444       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 445                       "must be between 1 and " SIZE_FORMAT,
 446                       mark_stack_size, MarkStackSizeMax);
 447       return;
 448     }
 449     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 450   } else {
 451     // Verify MarkStackSize is in range.
 452     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 453       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 454         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 455           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 456                           "must be between 1 and " SIZE_FORMAT,
 457                           MarkStackSize, MarkStackSizeMax);
 458           return;
 459         }
 460       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 461         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 462           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 463                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 464                           MarkStackSize, MarkStackSizeMax);
 465           return;
 466         }
 467       }
 468     }
 469   }
 470 
 471   if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
 472     vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
 473   }
 474 
 475   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
 476   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
 477 
 478   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 479   _num_active_tasks = _max_num_tasks;
 480 
 481   for (uint i = 0; i < _max_num_tasks; ++i) {
 482     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 483     task_queue->initialize();
 484     _task_queues->register_queue(i, task_queue);
 485 
 486     _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions());
 487 
 488     _accum_task_vtime[i] = 0.0;
 489   }
 490 
 491   reset_at_marking_complete();
 492   _completed_initialization = true;
 493 }
 494 
 495 void G1ConcurrentMark::reset() {
 496   _has_aborted = false;
 497 
 498   reset_marking_for_restart();
 499 
 500   // Reset all tasks, since different phases will use different number of active
 501   // threads. So, it's easiest to have all of them ready.
 502   for (uint i = 0; i < _max_num_tasks; ++i) {
 503     _tasks[i]->reset(_next_mark_bitmap);
 504   }
 505 
 506   uint max_regions = _g1h->max_regions();
 507   for (uint i = 0; i < max_regions; i++) {
 508     _top_at_rebuild_starts[i] = NULL;
 509     _region_mark_stats[i].clear();
 510   }
 511 }
 512 
 513 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
 514   for (uint j = 0; j < _max_num_tasks; ++j) {
 515     _tasks[j]->clear_mark_stats_cache(region_idx);
 516   }
 517   _top_at_rebuild_starts[region_idx] = NULL;
 518   _region_mark_stats[region_idx].clear();
 519 }
 520 
 521 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
 522   uint const region_idx = r->hrm_index();
 523   if (r->is_humongous()) {
 524     assert(r->is_starts_humongous(), "Got humongous continues region here");
 525     uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size());
 526     for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
 527       clear_statistics_in_region(j);
 528     }
 529   } else {
 530     clear_statistics_in_region(region_idx);
 531   }
 532 }
 533 
 534 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
 535   if (bitmap->is_marked(addr)) {
 536     bitmap->clear(addr);
 537   }
 538 }
 539 
 540 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
 541   assert_at_safepoint_on_vm_thread();
 542 
 543   // Need to clear all mark bits of the humongous object.
 544   clear_mark_if_set(_prev_mark_bitmap, r->bottom());
 545   clear_mark_if_set(_next_mark_bitmap, r->bottom());
 546 
 547   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
 548     return;
 549   }
 550 
 551   // Clear any statistics about the region gathered so far.
 552   clear_statistics(r);
 553 }
 554 
 555 void G1ConcurrentMark::reset_marking_for_restart() {
 556   _global_mark_stack.set_empty();
 557 
 558   // Expand the marking stack, if we have to and if we can.
 559   if (has_overflown()) {
 560     _global_mark_stack.expand();
 561 
 562     uint max_regions = _g1h->max_regions();
 563     for (uint i = 0; i < max_regions; i++) {
 564       _region_mark_stats[i].clear_during_overflow();
 565     }
 566   }
 567 
 568   clear_has_overflown();
 569   _finger = _heap.start();
 570 
 571   for (uint i = 0; i < _max_num_tasks; ++i) {
 572     G1CMTaskQueue* queue = _task_queues->queue(i);
 573     queue->set_empty();
 574   }
 575 }
 576 
 577 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 578   assert(active_tasks <= _max_num_tasks, "we should not have more");
 579 
 580   _num_active_tasks = active_tasks;
 581   // Need to update the three data structures below according to the
 582   // number of active threads for this phase.
 583   _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
 584   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 585   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 586 }
 587 
 588 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 589   set_concurrency(active_tasks);
 590 
 591   _concurrent = concurrent;
 592 
 593   if (!concurrent) {
 594     // At this point we should be in a STW phase, and completed marking.
 595     assert_at_safepoint_on_vm_thread();
 596     assert(out_of_regions(),
 597            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 598            p2i(_finger), p2i(_heap.end()));
 599   }
 600 }
 601 
 602 void G1ConcurrentMark::reset_at_marking_complete() {
 603   // We set the global marking state to some default values when we're
 604   // not doing marking.
 605   reset_marking_for_restart();
 606   _num_active_tasks = 0;
 607 }
 608 
 609 G1ConcurrentMark::~G1ConcurrentMark() {
 610   FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
 611   FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
 612   // The G1ConcurrentMark instance is never freed.
 613   ShouldNotReachHere();
 614 }
 615 
 616 class G1ClearBitMapTask : public AbstractGangTask {
 617 public:
 618   static size_t chunk_size() { return M; }
 619 
 620 private:
 621   // Heap region closure used for clearing the given mark bitmap.
 622   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 623   private:
 624     G1CMBitMap* _bitmap;
 625     G1ConcurrentMark* _cm;
 626   public:
 627     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _bitmap(bitmap), _cm(cm) {
 628     }
 629 
 630     virtual bool do_heap_region(HeapRegion* r) {
 631       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 632 
 633       HeapWord* cur = r->bottom();
 634       HeapWord* const end = r->end();
 635 
 636       while (cur < end) {
 637         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 638         _bitmap->clear_range(mr);
 639 
 640         cur += chunk_size_in_words;
 641 
 642         // Abort iteration if after yielding the marking has been aborted.
 643         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 644           return true;
 645         }
 646         // Repeat the asserts from before the start of the closure. We will do them
 647         // as asserts here to minimize their overhead on the product. However, we
 648         // will have them as guarantees at the beginning / end of the bitmap
 649         // clearing to get some checking in the product.
 650         assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
 651         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 652       }
 653       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 654 
 655       return false;
 656     }
 657   };
 658 
 659   G1ClearBitmapHRClosure _cl;
 660   HeapRegionClaimer _hr_claimer;
 661   bool _suspendible; // If the task is suspendible, workers must join the STS.
 662 
 663 public:
 664   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 665     AbstractGangTask("G1 Clear Bitmap"),
 666     _cl(bitmap, suspendible ? cm : NULL),
 667     _hr_claimer(n_workers),
 668     _suspendible(suspendible)
 669   { }
 670 
 671   void work(uint worker_id) {
 672     SuspendibleThreadSetJoiner sts_join(_suspendible);
 673     G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
 674   }
 675 
 676   bool is_complete() {
 677     return _cl.is_complete();
 678   }
 679 };
 680 
 681 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 682   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 683 
 684   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 685   size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 686 
 687   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 688 
 689   G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
 690 
 691   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 692   workers->run_task(&cl, num_workers);
 693   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 694 }
 695 
 696 void G1ConcurrentMark::cleanup_for_next_mark() {
 697   // Make sure that the concurrent mark thread looks to still be in
 698   // the current cycle.
 699   guarantee(cm_thread()->during_cycle(), "invariant");
 700 
 701   // We are finishing up the current cycle by clearing the next
 702   // marking bitmap and getting it ready for the next cycle. During
 703   // this time no other cycle can start. So, let's make sure that this
 704   // is the case.
 705   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 706 
 707   clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
 708 
 709   // Repeat the asserts from above.
 710   guarantee(cm_thread()->during_cycle(), "invariant");
 711   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 712 }
 713 
 714 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 715   assert_at_safepoint_on_vm_thread();
 716   clear_bitmap(_prev_mark_bitmap, workers, false);
 717 }
 718 
 719 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
 720 public:
 721   bool do_heap_region(HeapRegion* r) {
 722     r->note_start_of_marking();
 723     return false;
 724   }
 725 };
 726 
 727 void G1ConcurrentMark::pre_initial_mark() {
 728   // Initialize marking structures. This has to be done in a STW phase.
 729   reset();
 730 
 731   // For each region note start of marking.
 732   NoteStartOfMarkHRClosure startcl;
 733   _g1h->heap_region_iterate(&startcl);
 734 }
 735 
 736 
 737 void G1ConcurrentMark::post_initial_mark() {
 738   // Start Concurrent Marking weak-reference discovery.
 739   ReferenceProcessor* rp = _g1h->ref_processor_cm();
 740   // enable ("weak") refs discovery
 741   rp->enable_discovery();
 742   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 743 
 744   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
 745   // This is the start of  the marking cycle, we're expected all
 746   // threads to have SATB queues with active set to false.
 747   satb_mq_set.set_active_all_threads(true, /* new active value */
 748                                      false /* expected_active */);
 749 
 750   _root_regions.prepare_for_scan();
 751 
 752   // update_g1_committed() will be called at the end of an evac pause
 753   // when marking is on. So, it's also called at the end of the
 754   // initial-mark pause to update the heap end, if the heap expands
 755   // during it. No need to call it here.
 756 }
 757 
 758 /*
 759  * Notice that in the next two methods, we actually leave the STS
 760  * during the barrier sync and join it immediately afterwards. If we
 761  * do not do this, the following deadlock can occur: one thread could
 762  * be in the barrier sync code, waiting for the other thread to also
 763  * sync up, whereas another one could be trying to yield, while also
 764  * waiting for the other threads to sync up too.
 765  *
 766  * Note, however, that this code is also used during remark and in
 767  * this case we should not attempt to leave / enter the STS, otherwise
 768  * we'll either hit an assert (debug / fastdebug) or deadlock
 769  * (product). So we should only leave / enter the STS if we are
 770  * operating concurrently.
 771  *
 772  * Because the thread that does the sync barrier has left the STS, it
 773  * is possible to be suspended for a Full GC or an evacuation pause
 774  * could occur. This is actually safe, since the entering the sync
 775  * barrier is one of the last things do_marking_step() does, and it
 776  * doesn't manipulate any data structures afterwards.
 777  */
 778 
 779 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 780   bool barrier_aborted;
 781   {
 782     SuspendibleThreadSetLeaver sts_leave(concurrent());
 783     barrier_aborted = !_first_overflow_barrier_sync.enter();
 784   }
 785 
 786   // at this point everyone should have synced up and not be doing any
 787   // more work
 788 
 789   if (barrier_aborted) {
 790     // If the barrier aborted we ignore the overflow condition and
 791     // just abort the whole marking phase as quickly as possible.
 792     return;
 793   }
 794 }
 795 
 796 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 797   SuspendibleThreadSetLeaver sts_leave(concurrent());
 798   _second_overflow_barrier_sync.enter();
 799 
 800   // at this point everything should be re-initialized and ready to go
 801 }
 802 
 803 class G1CMConcurrentMarkingTask : public AbstractGangTask {
 804   G1ConcurrentMark*     _cm;
 805 
 806 public:
 807   void work(uint worker_id) {
 808     assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
 809     ResourceMark rm;
 810 
 811     double start_vtime = os::elapsedVTime();
 812 
 813     {
 814       SuspendibleThreadSetJoiner sts_join;
 815 
 816       assert(worker_id < _cm->active_tasks(), "invariant");
 817 
 818       G1CMTask* task = _cm->task(worker_id);
 819       task->record_start_time();
 820       if (!_cm->has_aborted()) {
 821         do {
 822           task->do_marking_step(G1ConcMarkStepDurationMillis,
 823                                 true  /* do_termination */,
 824                                 false /* is_serial*/);
 825 
 826           _cm->do_yield_check();
 827         } while (!_cm->has_aborted() && task->has_aborted());
 828       }
 829       task->record_end_time();
 830       guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
 831     }
 832 
 833     double end_vtime = os::elapsedVTime();
 834     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 835   }
 836 
 837   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
 838       AbstractGangTask("Concurrent Mark"), _cm(cm) { }
 839 
 840   ~G1CMConcurrentMarkingTask() { }
 841 };
 842 
 843 uint G1ConcurrentMark::calc_active_marking_workers() {
 844   uint result = 0;
 845   if (!UseDynamicNumberOfGCThreads ||
 846       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 847        !ForceDynamicNumberOfGCThreads)) {
 848     result = _max_concurrent_workers;
 849   } else {
 850     result =
 851       AdaptiveSizePolicy::calc_default_active_workers(_max_concurrent_workers,
 852                                                       1, /* Minimum workers */
 853                                                       _num_concurrent_workers,
 854                                                       Threads::number_of_non_daemon_threads());
 855     // Don't scale the result down by scale_concurrent_workers() because
 856     // that scaling has already gone into "_max_concurrent_workers".
 857   }
 858   assert(result > 0 && result <= _max_concurrent_workers,
 859          "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
 860          _max_concurrent_workers, result);
 861   return result;
 862 }
 863 
 864 void G1ConcurrentMark::scan_root_region(HeapRegion* hr, uint worker_id) {
 865   // Currently, only survivors can be root regions.
 866   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 867   G1RootRegionScanClosure cl(_g1h, this, worker_id);
 868 
 869   const uintx interval = PrefetchScanIntervalInBytes;
 870   HeapWord* curr = hr->bottom();
 871   const HeapWord* end = hr->top();
 872   while (curr < end) {
 873     Prefetch::read(curr, interval);
 874     oop obj = oop(curr);
 875     int size = obj->oop_iterate_size(&cl);
 876     assert(size == obj->size(), "sanity");
 877     curr += size;
 878   }
 879 }
 880 
 881 class G1CMRootRegionScanTask : public AbstractGangTask {
 882   G1ConcurrentMark* _cm;
 883 public:
 884   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 885     AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
 886 
 887   void work(uint worker_id) {
 888     assert(Thread::current()->is_ConcurrentGC_thread(),
 889            "this should only be done by a conc GC thread");
 890 
 891     G1CMRootRegions* root_regions = _cm->root_regions();
 892     HeapRegion* hr = root_regions->claim_next();
 893     while (hr != NULL) {
 894       _cm->scan_root_region(hr, worker_id);
 895       hr = root_regions->claim_next();
 896     }
 897   }
 898 };
 899 
 900 void G1ConcurrentMark::scan_root_regions() {
 901   // scan_in_progress() will have been set to true only if there was
 902   // at least one root region to scan. So, if it's false, we
 903   // should not attempt to do any further work.
 904   if (root_regions()->scan_in_progress()) {
 905     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 906 
 907     _num_concurrent_workers = MIN2(calc_active_marking_workers(),
 908                                    // We distribute work on a per-region basis, so starting
 909                                    // more threads than that is useless.
 910                                    root_regions()->num_root_regions());
 911     assert(_num_concurrent_workers <= _max_concurrent_workers,
 912            "Maximum number of marking threads exceeded");
 913 
 914     G1CMRootRegionScanTask task(this);
 915     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
 916                         task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
 917     _concurrent_workers->run_task(&task, _num_concurrent_workers);
 918 
 919     // It's possible that has_aborted() is true here without actually
 920     // aborting the survivor scan earlier. This is OK as it's
 921     // mainly used for sanity checking.
 922     root_regions()->scan_finished();
 923   }
 924 }
 925 
 926 void G1ConcurrentMark::concurrent_cycle_start() {
 927   _gc_timer_cm->register_gc_start();
 928 
 929   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
 930 
 931   _g1h->trace_heap_before_gc(_gc_tracer_cm);
 932   // Record start, but take no time
 933   TraceConcMemoryManagerStats tms(TraceConcMemoryManagerStats::CycleStart, _g1h->gc_cause());
 934 }
 935 
 936 void G1ConcurrentMark::concurrent_cycle_end() {
 937   _g1h->collector_state()->set_clearing_next_bitmap(false);
 938 
 939   _g1h->trace_heap_after_gc(_gc_tracer_cm);
 940 
 941   if (has_aborted()) {
 942     log_info(gc, marking)("Concurrent Mark Abort");
 943     _gc_tracer_cm->report_concurrent_mode_failure();
 944   }
 945 
 946   _gc_timer_cm->register_gc_end();
 947 
 948   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
 949   // Record end, but take no time
 950   TraceConcMemoryManagerStats tms(TraceConcMemoryManagerStats::CycleEnd, _g1h->gc_cause());
 951 }
 952 
 953 void G1ConcurrentMark::mark_from_roots() {
 954   _restart_for_overflow = false;
 955 
 956   _num_concurrent_workers = calc_active_marking_workers();
 957 
 958   uint active_workers = MAX2(1U, _num_concurrent_workers);
 959 
 960   // Setting active workers is not guaranteed since fewer
 961   // worker threads may currently exist and more may not be
 962   // available.
 963   active_workers = _concurrent_workers->update_active_workers(active_workers);
 964   log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
 965 
 966   // Parallel task terminator is set in "set_concurrency_and_phase()"
 967   set_concurrency_and_phase(active_workers, true /* concurrent */);
 968 
 969   G1CMConcurrentMarkingTask marking_task(this);
 970   _concurrent_workers->run_task(&marking_task);
 971   print_stats();
 972 }
 973 
 974 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
 975   G1HeapVerifier* verifier = _g1h->verifier();
 976 
 977   verifier->verify_region_sets_optional();
 978 
 979   if (VerifyDuringGC) {
 980     GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm);
 981 
 982     size_t const BufLen = 512;
 983     char buffer[BufLen];
 984 
 985     jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
 986     verifier->verify(type, vo, buffer);
 987   }
 988 
 989   verifier->check_bitmaps(caller);
 990 }
 991 
 992 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask {
 993   G1CollectedHeap* _g1h;
 994   G1ConcurrentMark* _cm;
 995   HeapRegionClaimer _hrclaimer;
 996   uint volatile _total_selected_for_rebuild;
 997 
 998   G1PrintRegionLivenessInfoClosure _cl;
 999 
1000   class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
1001     G1CollectedHeap* _g1h;
1002     G1ConcurrentMark* _cm;
1003 
1004     G1PrintRegionLivenessInfoClosure* _cl;
1005 
1006     uint _num_regions_selected_for_rebuild;  // The number of regions actually selected for rebuild.
1007 
1008     void update_remset_before_rebuild(HeapRegion* hr) {
1009       G1RemSetTrackingPolicy* tracking_policy = _g1h->g1_policy()->remset_tracker();
1010 
1011       bool selected_for_rebuild;
1012       if (hr->is_humongous()) {
1013         bool const is_live = _cm->liveness(hr->humongous_start_region()->hrm_index()) > 0;
1014         selected_for_rebuild = tracking_policy->update_humongous_before_rebuild(hr, is_live);
1015       } else {
1016         size_t const live_bytes = _cm->liveness(hr->hrm_index());
1017         selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1018       }
1019       if (selected_for_rebuild) {
1020         _num_regions_selected_for_rebuild++;
1021       }
1022       _cm->update_top_at_rebuild_start(hr);
1023     }
1024 
1025     // Distribute the given words across the humongous object starting with hr and
1026     // note end of marking.
1027     void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1028       uint const region_idx = hr->hrm_index();
1029       size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size();
1030       uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1031 
1032       // "Distributing" zero words means that we only note end of marking for these
1033       // regions.
1034       assert(marked_words == 0 || obj_size_in_words == marked_words,
1035              "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1036              obj_size_in_words, marked_words);
1037 
1038       for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1039         HeapRegion* const r = _g1h->region_at(i);
1040         size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1041 
1042         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)",
1043                                words_to_add, i, r->get_type_str());
1044         add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1045         marked_words -= words_to_add;
1046       }
1047       assert(marked_words == 0,
1048              SIZE_FORMAT " words left after distributing space across %u regions",
1049              marked_words, num_regions_in_humongous);
1050     }
1051 
1052     void update_marked_bytes(HeapRegion* hr) {
1053       uint const region_idx = hr->hrm_index();
1054       size_t const marked_words = _cm->liveness(region_idx);
1055       // The marking attributes the object's size completely to the humongous starts
1056       // region. We need to distribute this value across the entire set of regions a
1057       // humongous object spans.
1058       if (hr->is_humongous()) {
1059         assert(hr->is_starts_humongous() || marked_words == 0,
1060                "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1061                marked_words, region_idx, hr->get_type_str());
1062         if (hr->is_starts_humongous()) {
1063           distribute_marked_bytes(hr, marked_words);
1064         }
1065       } else {
1066         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1067         add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize);
1068       }
1069     }
1070 
1071     void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1072       hr->add_to_marked_bytes(marked_bytes);
1073       _cl->do_heap_region(hr);
1074       hr->note_end_of_marking();
1075     }
1076 
1077   public:
1078     G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1079       _g1h(g1h), _cm(cm), _cl(cl), _num_regions_selected_for_rebuild(0) { }
1080 
1081     virtual bool do_heap_region(HeapRegion* r) {
1082       update_remset_before_rebuild(r);
1083       update_marked_bytes(r);
1084 
1085       return false;
1086     }
1087 
1088     uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1089   };
1090 
1091 public:
1092   G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1093     AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"),
1094     _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1095 
1096   virtual void work(uint worker_id) {
1097     G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1098     _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1099     Atomic::add(update_cl.num_selected_for_rebuild(), &_total_selected_for_rebuild);
1100   }
1101 
1102   uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1103 
1104   // Number of regions for which roughly one thread should be spawned for this work.
1105   static const uint RegionsPerThread = 384;
1106 };
1107 
1108 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1109   G1CollectedHeap* _g1h;
1110 public:
1111   G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1112 
1113   virtual bool do_heap_region(HeapRegion* r) {
1114     _g1h->g1_policy()->remset_tracker()->update_after_rebuild(r);
1115     return false;
1116   }
1117 };
1118 
1119 void G1ConcurrentMark::remark() {
1120   assert_at_safepoint_on_vm_thread();
1121 
1122   // If a full collection has happened, we should not continue. However we might
1123   // have ended up here as the Remark VM operation has been scheduled already.
1124   if (has_aborted()) {
1125     return;
1126   }
1127 
1128   TraceConcMemoryManagerStats tms(TraceConcMemoryManagerStats::Remark, _g1h->gc_cause());
1129 
1130   G1Policy* g1p = _g1h->g1_policy();
1131   g1p->record_concurrent_mark_remark_start();
1132 
1133   double start = os::elapsedTime();
1134 
1135   verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1136 
1137   {
1138     GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm);
1139     finalize_marking();
1140   }
1141 
1142   double mark_work_end = os::elapsedTime();
1143 
1144   bool const mark_finished = !has_overflown();
1145   if (mark_finished) {
1146     weak_refs_work(false /* clear_all_soft_refs */);
1147 
1148     SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1149     // We're done with marking.
1150     // This is the end of the marking cycle, we're expected all
1151     // threads to have SATB queues with active set to true.
1152     satb_mq_set.set_active_all_threads(false, /* new active value */
1153                                        true /* expected_active */);
1154 
1155     {
1156       GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm);
1157       flush_all_task_caches();
1158     }
1159 
1160     // Install newly created mark bitmap as "prev".
1161     swap_mark_bitmaps();
1162     {
1163       GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1164 
1165       uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1166                                        G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1167       uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1168 
1169       G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1170       log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1171       _g1h->workers()->run_task(&cl, num_workers);
1172 
1173       log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1174                                       _g1h->num_regions(), cl.total_selected_for_rebuild());
1175     }
1176     {
1177       GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm);
1178       reclaim_empty_regions();
1179     }
1180 
1181     // Clean out dead classes
1182     if (ClassUnloadingWithConcurrentMark) {
1183       GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm);
1184       ClassLoaderDataGraph::purge();
1185     }
1186 
1187     compute_new_sizes();
1188 
1189     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1190 
1191     assert(!restart_for_overflow(), "sanity");
1192     // Completely reset the marking state since marking completed
1193     reset_at_marking_complete();
1194   } else {
1195     // We overflowed.  Restart concurrent marking.
1196     _restart_for_overflow = true;
1197 
1198     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1199 
1200     // Clear the marking state because we will be restarting
1201     // marking due to overflowing the global mark stack.
1202     reset_marking_for_restart();
1203   }
1204 
1205   {
1206     GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm);
1207     report_object_count(mark_finished);
1208   }
1209 
1210   // Statistics
1211   double now = os::elapsedTime();
1212   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1213   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1214   _remark_times.add((now - start) * 1000.0);
1215 
1216   g1p->record_concurrent_mark_remark_end();
1217 }
1218 
1219 class G1ReclaimEmptyRegionsTask : public AbstractGangTask {
1220   // Per-region work during the Cleanup pause.
1221   class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1222     G1CollectedHeap* _g1h;
1223     size_t _freed_bytes;
1224     FreeRegionList* _local_cleanup_list;
1225     uint _old_regions_removed;
1226     uint _humongous_regions_removed;
1227     HRRSCleanupTask* _hrrs_cleanup_task;
1228 
1229   public:
1230     G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1231                                  FreeRegionList* local_cleanup_list,
1232                                  HRRSCleanupTask* hrrs_cleanup_task) :
1233       _g1h(g1h),
1234       _freed_bytes(0),
1235       _local_cleanup_list(local_cleanup_list),
1236       _old_regions_removed(0),
1237       _humongous_regions_removed(0),
1238       _hrrs_cleanup_task(hrrs_cleanup_task) { }
1239 
1240     size_t freed_bytes() { return _freed_bytes; }
1241     const uint old_regions_removed() { return _old_regions_removed; }
1242     const uint humongous_regions_removed() { return _humongous_regions_removed; }
1243 
1244     bool do_heap_region(HeapRegion *hr) {
1245       if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1246         _freed_bytes += hr->used();
1247         hr->set_containing_set(NULL);
1248         if (hr->is_humongous()) {
1249           _humongous_regions_removed++;
1250           _g1h->free_humongous_region(hr, _local_cleanup_list);
1251         } else {
1252           _old_regions_removed++;
1253           _g1h->free_region(hr, _local_cleanup_list, false /* skip_remset */, false /* skip_hcc */, true /* locked */);
1254         }
1255         hr->clear_cardtable();
1256         _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1257         log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1258       } else {
1259         hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1260       }
1261 
1262       return false;
1263     }
1264   };
1265 
1266   G1CollectedHeap* _g1h;
1267   FreeRegionList* _cleanup_list;
1268   HeapRegionClaimer _hrclaimer;
1269 
1270 public:
1271   G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1272     AbstractGangTask("G1 Cleanup"),
1273     _g1h(g1h),
1274     _cleanup_list(cleanup_list),
1275     _hrclaimer(n_workers) {
1276 
1277     HeapRegionRemSet::reset_for_cleanup_tasks();
1278   }
1279 
1280   void work(uint worker_id) {
1281     FreeRegionList local_cleanup_list("Local Cleanup List");
1282     HRRSCleanupTask hrrs_cleanup_task;
1283     G1ReclaimEmptyRegionsClosure cl(_g1h,
1284                                     &local_cleanup_list,
1285                                     &hrrs_cleanup_task);
1286     _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1287     assert(cl.is_complete(), "Shouldn't have aborted!");
1288 
1289     // Now update the old/humongous region sets
1290     _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed());
1291     {
1292       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1293       _g1h->decrement_summary_bytes(cl.freed_bytes());
1294 
1295       _cleanup_list->add_ordered(&local_cleanup_list);
1296       assert(local_cleanup_list.is_empty(), "post-condition");
1297 
1298       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1299     }
1300   }
1301 };
1302 
1303 void G1ConcurrentMark::reclaim_empty_regions() {
1304   WorkGang* workers = _g1h->workers();
1305   FreeRegionList empty_regions_list("Empty Regions After Mark List");
1306 
1307   G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1308   workers->run_task(&cl);
1309 
1310   if (!empty_regions_list.is_empty()) {
1311     log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1312     // Now print the empty regions list.
1313     G1HRPrinter* hrp = _g1h->hr_printer();
1314     if (hrp->is_active()) {
1315       FreeRegionListIterator iter(&empty_regions_list);
1316       while (iter.more_available()) {
1317         HeapRegion* hr = iter.get_next();
1318         hrp->cleanup(hr);
1319       }
1320     }
1321     // And actually make them available.
1322     _g1h->prepend_to_freelist(&empty_regions_list);
1323   }
1324 }
1325 
1326 void G1ConcurrentMark::compute_new_sizes() {
1327   MetaspaceGC::compute_new_size();
1328 
1329   // Cleanup will have freed any regions completely full of garbage.
1330   // Update the soft reference policy with the new heap occupancy.
1331   Universe::update_heap_info_at_gc();
1332 
1333   // We reclaimed old regions so we should calculate the sizes to make
1334   // sure we update the old gen/space data.
1335   _g1h->g1mm()->update_sizes();
1336 }
1337 
1338 void G1ConcurrentMark::cleanup() {
1339   assert_at_safepoint_on_vm_thread();
1340 
1341   // If a full collection has happened, we shouldn't do this.
1342   if (has_aborted()) {
1343     return;
1344   }
1345 
1346   TraceConcMemoryManagerStats tms(TraceConcMemoryManagerStats::Cleanup, _g1h->gc_cause());
1347 
1348   G1Policy* g1p = _g1h->g1_policy();
1349   g1p->record_concurrent_mark_cleanup_start();
1350 
1351   double start = os::elapsedTime();
1352 
1353   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1354 
1355   {
1356     GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1357     G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1358     _g1h->heap_region_iterate(&cl);
1359   }
1360 
1361   if (log_is_enabled(Trace, gc, liveness)) {
1362     G1PrintRegionLivenessInfoClosure cl("Post-Cleanup");
1363     _g1h->heap_region_iterate(&cl);
1364   }
1365 
1366   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1367 
1368   // We need to make this be a "collection" so any collection pause that
1369   // races with it goes around and waits for Cleanup to finish.
1370   _g1h->increment_total_collections();
1371 
1372   // Local statistics
1373   double recent_cleanup_time = (os::elapsedTime() - start);
1374   _total_cleanup_time += recent_cleanup_time;
1375   _cleanup_times.add(recent_cleanup_time);
1376 
1377   {
1378     GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1379     _g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1380   }
1381 }
1382 
1383 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1384 // Uses the G1CMTask associated with a worker thread (for serial reference
1385 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1386 // trace referent objects.
1387 //
1388 // Using the G1CMTask and embedded local queues avoids having the worker
1389 // threads operating on the global mark stack. This reduces the risk
1390 // of overflowing the stack - which we would rather avoid at this late
1391 // state. Also using the tasks' local queues removes the potential
1392 // of the workers interfering with each other that could occur if
1393 // operating on the global stack.
1394 
1395 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1396   G1ConcurrentMark* _cm;
1397   G1CMTask*         _task;
1398   uint              _ref_counter_limit;
1399   uint              _ref_counter;
1400   bool              _is_serial;
1401 public:
1402   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1403     _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval),
1404     _ref_counter(_ref_counter_limit), _is_serial(is_serial) {
1405     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1406   }
1407 
1408   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1409   virtual void do_oop(      oop* p) { do_oop_work(p); }
1410 
1411   template <class T> void do_oop_work(T* p) {
1412     if (_cm->has_overflown()) {
1413       return;
1414     }
1415     if (!_task->deal_with_reference(p)) {
1416       // We did not add anything to the mark bitmap (or mark stack), so there is
1417       // no point trying to drain it.
1418       return;
1419     }
1420     _ref_counter--;
1421 
1422     if (_ref_counter == 0) {
1423       // We have dealt with _ref_counter_limit references, pushing them
1424       // and objects reachable from them on to the local stack (and
1425       // possibly the global stack). Call G1CMTask::do_marking_step() to
1426       // process these entries.
1427       //
1428       // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1429       // there's nothing more to do (i.e. we're done with the entries that
1430       // were pushed as a result of the G1CMTask::deal_with_reference() calls
1431       // above) or we overflow.
1432       //
1433       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1434       // flag while there may still be some work to do. (See the comment at
1435       // the beginning of G1CMTask::do_marking_step() for those conditions -
1436       // one of which is reaching the specified time target.) It is only
1437       // when G1CMTask::do_marking_step() returns without setting the
1438       // has_aborted() flag that the marking step has completed.
1439       do {
1440         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1441         _task->do_marking_step(mark_step_duration_ms,
1442                                false      /* do_termination */,
1443                                _is_serial);
1444       } while (_task->has_aborted() && !_cm->has_overflown());
1445       _ref_counter = _ref_counter_limit;
1446     }
1447   }
1448 };
1449 
1450 // 'Drain' oop closure used by both serial and parallel reference processing.
1451 // Uses the G1CMTask associated with a given worker thread (for serial
1452 // reference processing the G1CMtask for worker 0 is used). Calls the
1453 // do_marking_step routine, with an unbelievably large timeout value,
1454 // to drain the marking data structures of the remaining entries
1455 // added by the 'keep alive' oop closure above.
1456 
1457 class G1CMDrainMarkingStackClosure : public VoidClosure {
1458   G1ConcurrentMark* _cm;
1459   G1CMTask*         _task;
1460   bool              _is_serial;
1461  public:
1462   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1463     _cm(cm), _task(task), _is_serial(is_serial) {
1464     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1465   }
1466 
1467   void do_void() {
1468     do {
1469       // We call G1CMTask::do_marking_step() to completely drain the local
1470       // and global marking stacks of entries pushed by the 'keep alive'
1471       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1472       //
1473       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1474       // if there's nothing more to do (i.e. we've completely drained the
1475       // entries that were pushed as a a result of applying the 'keep alive'
1476       // closure to the entries on the discovered ref lists) or we overflow
1477       // the global marking stack.
1478       //
1479       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1480       // flag while there may still be some work to do. (See the comment at
1481       // the beginning of G1CMTask::do_marking_step() for those conditions -
1482       // one of which is reaching the specified time target.) It is only
1483       // when G1CMTask::do_marking_step() returns without setting the
1484       // has_aborted() flag that the marking step has completed.
1485 
1486       _task->do_marking_step(1000000000.0 /* something very large */,
1487                              true         /* do_termination */,
1488                              _is_serial);
1489     } while (_task->has_aborted() && !_cm->has_overflown());
1490   }
1491 };
1492 
1493 // Implementation of AbstractRefProcTaskExecutor for parallel
1494 // reference processing at the end of G1 concurrent marking
1495 
1496 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor {
1497 private:
1498   G1CollectedHeap*  _g1h;
1499   G1ConcurrentMark* _cm;
1500   WorkGang*         _workers;
1501   uint              _active_workers;
1502 
1503 public:
1504   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1505                           G1ConcurrentMark* cm,
1506                           WorkGang* workers,
1507                           uint n_workers) :
1508     _g1h(g1h), _cm(cm),
1509     _workers(workers), _active_workers(n_workers) { }
1510 
1511   virtual void execute(ProcessTask& task, uint ergo_workers);
1512 };
1513 
1514 class G1CMRefProcTaskProxy : public AbstractGangTask {
1515   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1516   ProcessTask&      _proc_task;
1517   G1CollectedHeap*  _g1h;
1518   G1ConcurrentMark* _cm;
1519 
1520 public:
1521   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1522                        G1CollectedHeap* g1h,
1523                        G1ConcurrentMark* cm) :
1524     AbstractGangTask("Process reference objects in parallel"),
1525     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1526     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1527     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1528   }
1529 
1530   virtual void work(uint worker_id) {
1531     ResourceMark rm;
1532     HandleMark hm;
1533     G1CMTask* task = _cm->task(worker_id);
1534     G1CMIsAliveClosure g1_is_alive(_g1h);
1535     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1536     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1537 
1538     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1539   }
1540 };
1541 
1542 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task, uint ergo_workers) {
1543   assert(_workers != NULL, "Need parallel worker threads.");
1544   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1545   assert(_workers->active_workers() >= ergo_workers,
1546          "Ergonomically chosen workers(%u) should be less than or equal to active workers(%u)",
1547          ergo_workers, _workers->active_workers());
1548 
1549   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1550 
1551   // We need to reset the concurrency level before each
1552   // proxy task execution, so that the termination protocol
1553   // and overflow handling in G1CMTask::do_marking_step() knows
1554   // how many workers to wait for.
1555   _cm->set_concurrency(ergo_workers);
1556   _workers->run_task(&proc_task_proxy, ergo_workers);
1557 }
1558 
1559 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1560   ResourceMark rm;
1561   HandleMark   hm;
1562 
1563   // Is alive closure.
1564   G1CMIsAliveClosure g1_is_alive(_g1h);
1565 
1566   // Inner scope to exclude the cleaning of the string table
1567   // from the displayed time.
1568   {
1569     GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm);
1570 
1571     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1572 
1573     // See the comment in G1CollectedHeap::ref_processing_init()
1574     // about how reference processing currently works in G1.
1575 
1576     // Set the soft reference policy
1577     rp->setup_policy(clear_all_soft_refs);
1578     assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1579 
1580     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1581     // in serial reference processing. Note these closures are also
1582     // used for serially processing (by the the current thread) the
1583     // JNI references during parallel reference processing.
1584     //
1585     // These closures do not need to synchronize with the worker
1586     // threads involved in parallel reference processing as these
1587     // instances are executed serially by the current thread (e.g.
1588     // reference processing is not multi-threaded and is thus
1589     // performed by the current thread instead of a gang worker).
1590     //
1591     // The gang tasks involved in parallel reference processing create
1592     // their own instances of these closures, which do their own
1593     // synchronization among themselves.
1594     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1595     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1596 
1597     // We need at least one active thread. If reference processing
1598     // is not multi-threaded we use the current (VMThread) thread,
1599     // otherwise we use the work gang from the G1CollectedHeap and
1600     // we utilize all the worker threads we can.
1601     bool processing_is_mt = rp->processing_is_mt();
1602     uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U);
1603     active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U);
1604 
1605     // Parallel processing task executor.
1606     G1CMRefProcTaskExecutor par_task_executor(_g1h, this,
1607                                               _g1h->workers(), active_workers);
1608     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1609 
1610     // Set the concurrency level. The phase was already set prior to
1611     // executing the remark task.
1612     set_concurrency(active_workers);
1613 
1614     // Set the degree of MT processing here.  If the discovery was done MT,
1615     // the number of threads involved during discovery could differ from
1616     // the number of active workers.  This is OK as long as the discovered
1617     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1618     rp->set_active_mt_degree(active_workers);
1619 
1620     ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues());
1621 
1622     // Process the weak references.
1623     const ReferenceProcessorStats& stats =
1624         rp->process_discovered_references(&g1_is_alive,
1625                                           &g1_keep_alive,
1626                                           &g1_drain_mark_stack,
1627                                           executor,
1628                                           &pt);
1629     _gc_tracer_cm->report_gc_reference_stats(stats);
1630     pt.print_all_references();
1631 
1632     // The do_oop work routines of the keep_alive and drain_marking_stack
1633     // oop closures will set the has_overflown flag if we overflow the
1634     // global marking stack.
1635 
1636     assert(has_overflown() || _global_mark_stack.is_empty(),
1637            "Mark stack should be empty (unless it has overflown)");
1638 
1639     assert(rp->num_queues() == active_workers, "why not");
1640 
1641     rp->verify_no_references_recorded();
1642     assert(!rp->discovery_enabled(), "Post condition");
1643   }
1644 
1645   if (has_overflown()) {
1646     // We can not trust g1_is_alive and the contents of the heap if the marking stack
1647     // overflowed while processing references. Exit the VM.
1648     fatal("Overflow during reference processing, can not continue. Please "
1649           "increase MarkStackSizeMax (current value: " SIZE_FORMAT ") and "
1650           "restart.", MarkStackSizeMax);
1651     return;
1652   }
1653 
1654   assert(_global_mark_stack.is_empty(), "Marking should have completed");
1655 
1656   {
1657     GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1658     WeakProcessor::weak_oops_do(_g1h->workers(), &g1_is_alive, &do_nothing_cl, 1);
1659   }
1660 
1661   // Unload Klasses, String, Code Cache, etc.
1662   if (ClassUnloadingWithConcurrentMark) {
1663     GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1664     bool purged_classes = SystemDictionary::do_unloading(_gc_timer_cm, false /* Defer cleaning */);
1665     _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1666   } else {
1667     GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1668     // No need to clean string table as it is treated as strong roots when
1669     // class unloading is disabled.
1670     _g1h->partial_cleaning(&g1_is_alive, false, G1StringDedup::is_enabled());
1671   }
1672 }
1673 
1674 class G1PrecleanYieldClosure : public YieldClosure {
1675   G1ConcurrentMark* _cm;
1676 
1677 public:
1678   G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
1679 
1680   virtual bool should_return() {
1681     return _cm->has_aborted();
1682   }
1683 
1684   virtual bool should_return_fine_grain() {
1685     _cm->do_yield_check();
1686     return _cm->has_aborted();
1687   }
1688 };
1689 
1690 void G1ConcurrentMark::preclean() {
1691   assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
1692 
1693   SuspendibleThreadSetJoiner joiner;
1694 
1695   G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */);
1696   G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */);
1697 
1698   set_concurrency_and_phase(1, true);
1699 
1700   G1PrecleanYieldClosure yield_cl(this);
1701 
1702   ReferenceProcessor* rp = _g1h->ref_processor_cm();
1703   // Precleaning is single threaded. Temporarily disable MT discovery.
1704   ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
1705   rp->preclean_discovered_references(rp->is_alive_non_header(),
1706                                      &keep_alive,
1707                                      &drain_mark_stack,
1708                                      &yield_cl,
1709                                      _gc_timer_cm);
1710 }
1711 
1712 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1713 // the prev bitmap determining liveness.
1714 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1715   G1CollectedHeap* _g1h;
1716 public:
1717   G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1718 
1719   bool do_object_b(oop obj) {
1720     HeapWord* addr = (HeapWord*)obj;
1721     return addr != NULL &&
1722            (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_dead(obj));
1723   }
1724 };
1725 
1726 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1727   // Depending on the completion of the marking liveness needs to be determined
1728   // using either the next or prev bitmap.
1729   if (mark_completed) {
1730     G1ObjectCountIsAliveClosure is_alive(_g1h);
1731     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1732   } else {
1733     G1CMIsAliveClosure is_alive(_g1h);
1734     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1735   }
1736 }
1737 
1738 
1739 void G1ConcurrentMark::swap_mark_bitmaps() {
1740   G1CMBitMap* temp = _prev_mark_bitmap;
1741   _prev_mark_bitmap = _next_mark_bitmap;
1742   _next_mark_bitmap = temp;
1743   _g1h->collector_state()->set_clearing_next_bitmap(true);
1744 }
1745 
1746 // Closure for marking entries in SATB buffers.
1747 class G1CMSATBBufferClosure : public SATBBufferClosure {
1748 private:
1749   G1CMTask* _task;
1750   G1CollectedHeap* _g1h;
1751 
1752   // This is very similar to G1CMTask::deal_with_reference, but with
1753   // more relaxed requirements for the argument, so this must be more
1754   // circumspect about treating the argument as an object.
1755   void do_entry(void* entry) const {
1756     _task->increment_refs_reached();
1757     oop const obj = static_cast<oop>(entry);
1758     _task->make_reference_grey(obj);
1759   }
1760 
1761 public:
1762   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1763     : _task(task), _g1h(g1h) { }
1764 
1765   virtual void do_buffer(void** buffer, size_t size) {
1766     for (size_t i = 0; i < size; ++i) {
1767       do_entry(buffer[i]);
1768     }
1769   }
1770 };
1771 
1772 class G1RemarkThreadsClosure : public ThreadClosure {
1773   G1CMSATBBufferClosure _cm_satb_cl;
1774   G1CMOopClosure _cm_cl;
1775   MarkingCodeBlobClosure _code_cl;
1776   int _thread_parity;
1777 
1778  public:
1779   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1780     _cm_satb_cl(task, g1h),
1781     _cm_cl(g1h, task),
1782     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1783     _thread_parity(Threads::thread_claim_parity()) {}
1784 
1785   void do_thread(Thread* thread) {
1786     if (thread->is_Java_thread()) {
1787       if (thread->claim_oops_do(true, _thread_parity)) {
1788         JavaThread* jt = (JavaThread*)thread;
1789 
1790         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1791         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1792         // * Alive if on the stack of an executing method
1793         // * Weakly reachable otherwise
1794         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1795         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1796         jt->nmethods_do(&_code_cl);
1797 
1798         G1ThreadLocalData::satb_mark_queue(jt).apply_closure_and_empty(&_cm_satb_cl);
1799       }
1800     } else if (thread->is_VM_thread()) {
1801       if (thread->claim_oops_do(true, _thread_parity)) {
1802         G1BarrierSet::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1803       }
1804     }
1805   }
1806 };
1807 
1808 class G1CMRemarkTask : public AbstractGangTask {
1809   G1ConcurrentMark* _cm;
1810 public:
1811   void work(uint worker_id) {
1812     G1CMTask* task = _cm->task(worker_id);
1813     task->record_start_time();
1814     {
1815       ResourceMark rm;
1816       HandleMark hm;
1817 
1818       G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1819       Threads::threads_do(&threads_f);
1820     }
1821 
1822     do {
1823       task->do_marking_step(1000000000.0 /* something very large */,
1824                             true         /* do_termination       */,
1825                             false        /* is_serial            */);
1826     } while (task->has_aborted() && !_cm->has_overflown());
1827     // If we overflow, then we do not want to restart. We instead
1828     // want to abort remark and do concurrent marking again.
1829     task->record_end_time();
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::finalize_marking() {
1839   ResourceMark rm;
1840   HandleMark   hm;
1841 
1842   _g1h->ensure_parsability(false);
1843 
1844   // this is remark, so we'll use up all active threads
1845   uint active_workers = _g1h->workers()->active_workers();
1846   set_concurrency_and_phase(active_workers, false /* concurrent */);
1847   // Leave _parallel_marking_threads at it's
1848   // value originally calculated in the G1ConcurrentMark
1849   // constructor and pass values of the active workers
1850   // through the gang in the task.
1851 
1852   {
1853     StrongRootsScope srs(active_workers);
1854 
1855     G1CMRemarkTask remarkTask(this, active_workers);
1856     // We will start all available threads, even if we decide that the
1857     // active_workers will be fewer. The extra ones will just bail out
1858     // immediately.
1859     _g1h->workers()->run_task(&remarkTask);
1860   }
1861 
1862   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
1863   guarantee(has_overflown() ||
1864             satb_mq_set.completed_buffers_num() == 0,
1865             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1866             BOOL_TO_STR(has_overflown()),
1867             satb_mq_set.completed_buffers_num());
1868 
1869   print_stats();
1870 }
1871 
1872 void G1ConcurrentMark::flush_all_task_caches() {
1873   size_t hits = 0;
1874   size_t misses = 0;
1875   for (uint i = 0; i < _max_num_tasks; i++) {
1876     Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1877     hits += stats.first;
1878     misses += stats.second;
1879   }
1880   size_t sum = hits + misses;
1881   log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1882                        hits, misses, percent_of(hits, sum));
1883 }
1884 
1885 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1886   _prev_mark_bitmap->clear_range(mr);
1887 }
1888 
1889 HeapRegion*
1890 G1ConcurrentMark::claim_region(uint worker_id) {
1891   // "checkpoint" the finger
1892   HeapWord* finger = _finger;
1893 
1894   while (finger < _heap.end()) {
1895     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1896 
1897     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1898     // Make sure that the reads below do not float before loading curr_region.
1899     OrderAccess::loadload();
1900     // Above heap_region_containing may return NULL as we always scan claim
1901     // until the end of the heap. In this case, just jump to the next region.
1902     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1903 
1904     // Is the gap between reading the finger and doing the CAS too long?
1905     HeapWord* res = Atomic::cmpxchg(end, &_finger, finger);
1906     if (res == finger && curr_region != NULL) {
1907       // we succeeded
1908       HeapWord*   bottom        = curr_region->bottom();
1909       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1910 
1911       // notice that _finger == end cannot be guaranteed here since,
1912       // someone else might have moved the finger even further
1913       assert(_finger >= end, "the finger should have moved forward");
1914 
1915       if (limit > bottom) {
1916         return curr_region;
1917       } else {
1918         assert(limit == bottom,
1919                "the region limit should be at bottom");
1920         // we return NULL and the caller should try calling
1921         // claim_region() again.
1922         return NULL;
1923       }
1924     } else {
1925       assert(_finger > finger, "the finger should have moved forward");
1926       // read it again
1927       finger = _finger;
1928     }
1929   }
1930 
1931   return NULL;
1932 }
1933 
1934 #ifndef PRODUCT
1935 class VerifyNoCSetOops {
1936   G1CollectedHeap* _g1h;
1937   const char* _phase;
1938   int _info;
1939 
1940 public:
1941   VerifyNoCSetOops(const char* phase, int info = -1) :
1942     _g1h(G1CollectedHeap::heap()),
1943     _phase(phase),
1944     _info(info)
1945   { }
1946 
1947   void operator()(G1TaskQueueEntry task_entry) const {
1948     if (task_entry.is_array_slice()) {
1949       guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1950       return;
1951     }
1952     guarantee(oopDesc::is_oop(task_entry.obj()),
1953               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1954               p2i(task_entry.obj()), _phase, _info);
1955     guarantee(!_g1h->is_in_cset(task_entry.obj()),
1956               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1957               p2i(task_entry.obj()), _phase, _info);
1958   }
1959 };
1960 
1961 void G1ConcurrentMark::verify_no_cset_oops() {
1962   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1963   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1964     return;
1965   }
1966 
1967   // Verify entries on the global mark stack
1968   _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1969 
1970   // Verify entries on the task queues
1971   for (uint i = 0; i < _max_num_tasks; ++i) {
1972     G1CMTaskQueue* queue = _task_queues->queue(i);
1973     queue->iterate(VerifyNoCSetOops("Queue", i));
1974   }
1975 
1976   // Verify the global finger
1977   HeapWord* global_finger = finger();
1978   if (global_finger != NULL && global_finger < _heap.end()) {
1979     // Since we always iterate over all regions, we might get a NULL HeapRegion
1980     // here.
1981     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1982     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1983               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1984               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1985   }
1986 
1987   // Verify the task fingers
1988   assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
1989   for (uint i = 0; i < _num_concurrent_workers; ++i) {
1990     G1CMTask* task = _tasks[i];
1991     HeapWord* task_finger = task->finger();
1992     if (task_finger != NULL && task_finger < _heap.end()) {
1993       // See above note on the global finger verification.
1994       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
1995       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
1996                 !task_hr->in_collection_set(),
1997                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
1998                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
1999     }
2000   }
2001 }
2002 #endif // PRODUCT
2003 
2004 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
2005   _g1h->g1_rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
2006 }
2007 
2008 void G1ConcurrentMark::print_stats() {
2009   if (!log_is_enabled(Debug, gc, stats)) {
2010     return;
2011   }
2012   log_debug(gc, stats)("---------------------------------------------------------------------");
2013   for (size_t i = 0; i < _num_active_tasks; ++i) {
2014     _tasks[i]->print_stats();
2015     log_debug(gc, stats)("---------------------------------------------------------------------");
2016   }
2017 }
2018 
2019 void G1ConcurrentMark::concurrent_cycle_abort() {
2020   if (!cm_thread()->during_cycle() || _has_aborted) {
2021     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2022     return;
2023   }
2024 
2025   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2026   // concurrent bitmap clearing.
2027   {
2028     GCTraceTime(Debug, gc) debug("Clear Next Bitmap");
2029     clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
2030   }
2031   // Note we cannot clear the previous marking bitmap here
2032   // since VerifyDuringGC verifies the objects marked during
2033   // a full GC against the previous bitmap.
2034 
2035   // Empty mark stack
2036   reset_marking_for_restart();
2037   for (uint i = 0; i < _max_num_tasks; ++i) {
2038     _tasks[i]->clear_region_fields();
2039   }
2040   _first_overflow_barrier_sync.abort();
2041   _second_overflow_barrier_sync.abort();
2042   _has_aborted = true;
2043 
2044   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2045   satb_mq_set.abandon_partial_marking();
2046   // This can be called either during or outside marking, we'll read
2047   // the expected_active value from the SATB queue set.
2048   satb_mq_set.set_active_all_threads(
2049                                  false, /* new active value */
2050                                  satb_mq_set.is_active() /* expected_active */);
2051 }
2052 
2053 static void print_ms_time_info(const char* prefix, const char* name,
2054                                NumberSeq& ns) {
2055   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2056                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2057   if (ns.num() > 0) {
2058     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2059                            prefix, ns.sd(), ns.maximum());
2060   }
2061 }
2062 
2063 void G1ConcurrentMark::print_summary_info() {
2064   Log(gc, marking) log;
2065   if (!log.is_trace()) {
2066     return;
2067   }
2068 
2069   log.trace(" Concurrent marking:");
2070   print_ms_time_info("  ", "init marks", _init_times);
2071   print_ms_time_info("  ", "remarks", _remark_times);
2072   {
2073     print_ms_time_info("     ", "final marks", _remark_mark_times);
2074     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2075 
2076   }
2077   print_ms_time_info("  ", "cleanups", _cleanup_times);
2078   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2079             _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2080   log.trace("  Total stop_world time = %8.2f s.",
2081             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2082   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2083             cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2084 }
2085 
2086 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2087   _concurrent_workers->print_worker_threads_on(st);
2088 }
2089 
2090 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2091   _concurrent_workers->threads_do(tc);
2092 }
2093 
2094 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2095   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2096                p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2097   _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2098   _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2099 }
2100 
2101 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2102   ReferenceProcessor* result = g1h->ref_processor_cm();
2103   assert(result != NULL, "CM reference processor should not be NULL");
2104   return result;
2105 }
2106 
2107 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2108                                G1CMTask* task)
2109   : MetadataVisitingOopIterateClosure(get_cm_oop_closure_ref_processor(g1h)),
2110     _g1h(g1h), _task(task)
2111 { }
2112 
2113 void G1CMTask::setup_for_region(HeapRegion* hr) {
2114   assert(hr != NULL,
2115         "claim_region() should have filtered out NULL regions");
2116   _curr_region  = hr;
2117   _finger       = hr->bottom();
2118   update_region_limit();
2119 }
2120 
2121 void G1CMTask::update_region_limit() {
2122   HeapRegion* hr            = _curr_region;
2123   HeapWord* bottom          = hr->bottom();
2124   HeapWord* limit           = hr->next_top_at_mark_start();
2125 
2126   if (limit == bottom) {
2127     // The region was collected underneath our feet.
2128     // We set the finger to bottom to ensure that the bitmap
2129     // iteration that will follow this will not do anything.
2130     // (this is not a condition that holds when we set the region up,
2131     // as the region is not supposed to be empty in the first place)
2132     _finger = bottom;
2133   } else if (limit >= _region_limit) {
2134     assert(limit >= _finger, "peace of mind");
2135   } else {
2136     assert(limit < _region_limit, "only way to get here");
2137     // This can happen under some pretty unusual circumstances.  An
2138     // evacuation pause empties the region underneath our feet (NTAMS
2139     // at bottom). We then do some allocation in the region (NTAMS
2140     // stays at bottom), followed by the region being used as a GC
2141     // alloc region (NTAMS will move to top() and the objects
2142     // originally below it will be grayed). All objects now marked in
2143     // the region are explicitly grayed, if below the global finger,
2144     // and we do not need in fact to scan anything else. So, we simply
2145     // set _finger to be limit to ensure that the bitmap iteration
2146     // doesn't do anything.
2147     _finger = limit;
2148   }
2149 
2150   _region_limit = limit;
2151 }
2152 
2153 void G1CMTask::giveup_current_region() {
2154   assert(_curr_region != NULL, "invariant");
2155   clear_region_fields();
2156 }
2157 
2158 void G1CMTask::clear_region_fields() {
2159   // Values for these three fields that indicate that we're not
2160   // holding on to a region.
2161   _curr_region   = NULL;
2162   _finger        = NULL;
2163   _region_limit  = NULL;
2164 }
2165 
2166 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2167   if (cm_oop_closure == NULL) {
2168     assert(_cm_oop_closure != NULL, "invariant");
2169   } else {
2170     assert(_cm_oop_closure == NULL, "invariant");
2171   }
2172   _cm_oop_closure = cm_oop_closure;
2173 }
2174 
2175 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2176   guarantee(next_mark_bitmap != NULL, "invariant");
2177   _next_mark_bitmap              = next_mark_bitmap;
2178   clear_region_fields();
2179 
2180   _calls                         = 0;
2181   _elapsed_time_ms               = 0.0;
2182   _termination_time_ms           = 0.0;
2183   _termination_start_time_ms     = 0.0;
2184 
2185   _mark_stats_cache.reset();
2186 }
2187 
2188 bool G1CMTask::should_exit_termination() {
2189   regular_clock_call();
2190   // This is called when we are in the termination protocol. We should
2191   // quit if, for some reason, this task wants to abort or the global
2192   // stack is not empty (this means that we can get work from it).
2193   return !_cm->mark_stack_empty() || has_aborted();
2194 }
2195 
2196 void G1CMTask::reached_limit() {
2197   assert(_words_scanned >= _words_scanned_limit ||
2198          _refs_reached >= _refs_reached_limit ,
2199          "shouldn't have been called otherwise");
2200   regular_clock_call();
2201 }
2202 
2203 void G1CMTask::regular_clock_call() {
2204   if (has_aborted()) {
2205     return;
2206   }
2207 
2208   // First, we need to recalculate the words scanned and refs reached
2209   // limits for the next clock call.
2210   recalculate_limits();
2211 
2212   // During the regular clock call we do the following
2213 
2214   // (1) If an overflow has been flagged, then we abort.
2215   if (_cm->has_overflown()) {
2216     set_has_aborted();
2217     return;
2218   }
2219 
2220   // If we are not concurrent (i.e. we're doing remark) we don't need
2221   // to check anything else. The other steps are only needed during
2222   // the concurrent marking phase.
2223   if (!_cm->concurrent()) {
2224     return;
2225   }
2226 
2227   // (2) If marking has been aborted for Full GC, then we also abort.
2228   if (_cm->has_aborted()) {
2229     set_has_aborted();
2230     return;
2231   }
2232 
2233   double curr_time_ms = os::elapsedVTime() * 1000.0;
2234 
2235   // (4) We check whether we should yield. If we have to, then we abort.
2236   if (SuspendibleThreadSet::should_yield()) {
2237     // We should yield. To do this we abort the task. The caller is
2238     // responsible for yielding.
2239     set_has_aborted();
2240     return;
2241   }
2242 
2243   // (5) We check whether we've reached our time quota. If we have,
2244   // then we abort.
2245   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2246   if (elapsed_time_ms > _time_target_ms) {
2247     set_has_aborted();
2248     _has_timed_out = true;
2249     return;
2250   }
2251 
2252   // (6) Finally, we check whether there are enough completed STAB
2253   // buffers available for processing. If there are, we abort.
2254   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2255   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2256     // we do need to process SATB buffers, we'll abort and restart
2257     // the marking task to do so
2258     set_has_aborted();
2259     return;
2260   }
2261 }
2262 
2263 void G1CMTask::recalculate_limits() {
2264   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2265   _words_scanned_limit      = _real_words_scanned_limit;
2266 
2267   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2268   _refs_reached_limit       = _real_refs_reached_limit;
2269 }
2270 
2271 void G1CMTask::decrease_limits() {
2272   // This is called when we believe that we're going to do an infrequent
2273   // operation which will increase the per byte scanned cost (i.e. move
2274   // entries to/from the global stack). It basically tries to decrease the
2275   // scanning limit so that the clock is called earlier.
2276 
2277   _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2278   _refs_reached_limit  = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2279 }
2280 
2281 void G1CMTask::move_entries_to_global_stack() {
2282   // Local array where we'll store the entries that will be popped
2283   // from the local queue.
2284   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2285 
2286   size_t n = 0;
2287   G1TaskQueueEntry task_entry;
2288   while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2289     buffer[n] = task_entry;
2290     ++n;
2291   }
2292   if (n < G1CMMarkStack::EntriesPerChunk) {
2293     buffer[n] = G1TaskQueueEntry();
2294   }
2295 
2296   if (n > 0) {
2297     if (!_cm->mark_stack_push(buffer)) {
2298       set_has_aborted();
2299     }
2300   }
2301 
2302   // This operation was quite expensive, so decrease the limits.
2303   decrease_limits();
2304 }
2305 
2306 bool G1CMTask::get_entries_from_global_stack() {
2307   // Local array where we'll store the entries that will be popped
2308   // from the global stack.
2309   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2310 
2311   if (!_cm->mark_stack_pop(buffer)) {
2312     return false;
2313   }
2314 
2315   // We did actually pop at least one entry.
2316   for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2317     G1TaskQueueEntry task_entry = buffer[i];
2318     if (task_entry.is_null()) {
2319       break;
2320     }
2321     assert(task_entry.is_array_slice() || oopDesc::is_oop(task_entry.obj()), "Element " PTR_FORMAT " must be an array slice or oop", p2i(task_entry.obj()));
2322     bool success = _task_queue->push(task_entry);
2323     // We only call this when the local queue is empty or under a
2324     // given target limit. So, we do not expect this push to fail.
2325     assert(success, "invariant");
2326   }
2327 
2328   // This operation was quite expensive, so decrease the limits
2329   decrease_limits();
2330   return true;
2331 }
2332 
2333 void G1CMTask::drain_local_queue(bool partially) {
2334   if (has_aborted()) {
2335     return;
2336   }
2337 
2338   // Decide what the target size is, depending whether we're going to
2339   // drain it partially (so that other tasks can steal if they run out
2340   // of things to do) or totally (at the very end).
2341   size_t target_size;
2342   if (partially) {
2343     target_size = MIN2((size_t)_task_queue->max_elems()/3, (size_t)GCDrainStackTargetSize);
2344   } else {
2345     target_size = 0;
2346   }
2347 
2348   if (_task_queue->size() > target_size) {
2349     G1TaskQueueEntry entry;
2350     bool ret = _task_queue->pop_local(entry);
2351     while (ret) {
2352       scan_task_entry(entry);
2353       if (_task_queue->size() <= target_size || has_aborted()) {
2354         ret = false;
2355       } else {
2356         ret = _task_queue->pop_local(entry);
2357       }
2358     }
2359   }
2360 }
2361 
2362 void G1CMTask::drain_global_stack(bool partially) {
2363   if (has_aborted()) {
2364     return;
2365   }
2366 
2367   // We have a policy to drain the local queue before we attempt to
2368   // drain the global stack.
2369   assert(partially || _task_queue->size() == 0, "invariant");
2370 
2371   // Decide what the target size is, depending whether we're going to
2372   // drain it partially (so that other tasks can steal if they run out
2373   // of things to do) or totally (at the very end).
2374   // Notice that when draining the global mark stack partially, due to the racyness
2375   // of the mark stack size update we might in fact drop below the target. But,
2376   // this is not a problem.
2377   // In case of total draining, we simply process until the global mark stack is
2378   // totally empty, disregarding the size counter.
2379   if (partially) {
2380     size_t const target_size = _cm->partial_mark_stack_size_target();
2381     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2382       if (get_entries_from_global_stack()) {
2383         drain_local_queue(partially);
2384       }
2385     }
2386   } else {
2387     while (!has_aborted() && get_entries_from_global_stack()) {
2388       drain_local_queue(partially);
2389     }
2390   }
2391 }
2392 
2393 // SATB Queue has several assumptions on whether to call the par or
2394 // non-par versions of the methods. this is why some of the code is
2395 // replicated. We should really get rid of the single-threaded version
2396 // of the code to simplify things.
2397 void G1CMTask::drain_satb_buffers() {
2398   if (has_aborted()) {
2399     return;
2400   }
2401 
2402   // We set this so that the regular clock knows that we're in the
2403   // middle of draining buffers and doesn't set the abort flag when it
2404   // notices that SATB buffers are available for draining. It'd be
2405   // very counter productive if it did that. :-)
2406   _draining_satb_buffers = true;
2407 
2408   G1CMSATBBufferClosure satb_cl(this, _g1h);
2409   SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set();
2410 
2411   // This keeps claiming and applying the closure to completed buffers
2412   // until we run out of buffers or we need to abort.
2413   while (!has_aborted() &&
2414          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2415     regular_clock_call();
2416   }
2417 
2418   _draining_satb_buffers = false;
2419 
2420   assert(has_aborted() ||
2421          _cm->concurrent() ||
2422          satb_mq_set.completed_buffers_num() == 0, "invariant");
2423 
2424   // again, this was a potentially expensive operation, decrease the
2425   // limits to get the regular clock call early
2426   decrease_limits();
2427 }
2428 
2429 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2430   _mark_stats_cache.reset(region_idx);
2431 }
2432 
2433 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2434   return _mark_stats_cache.evict_all();
2435 }
2436 
2437 void G1CMTask::print_stats() {
2438   log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2439   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2440                        _elapsed_time_ms, _termination_time_ms);
2441   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2442                        _step_times_ms.num(),
2443                        _step_times_ms.avg(),
2444                        _step_times_ms.sd(),
2445                        _step_times_ms.maximum(),
2446                        _step_times_ms.sum());
2447   size_t const hits = _mark_stats_cache.hits();
2448   size_t const misses = _mark_stats_cache.misses();
2449   log_debug(gc, stats)("  Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2450                        hits, misses, percent_of(hits, hits + misses));
2451 }
2452 
2453 bool G1ConcurrentMark::try_stealing(uint worker_id, G1TaskQueueEntry& task_entry) {
2454   return _task_queues->steal(worker_id, task_entry);
2455 }
2456 
2457 /*****************************************************************************
2458 
2459     The do_marking_step(time_target_ms, ...) method is the building
2460     block of the parallel marking framework. It can be called in parallel
2461     with other invocations of do_marking_step() on different tasks
2462     (but only one per task, obviously) and concurrently with the
2463     mutator threads, or during remark, hence it eliminates the need
2464     for two versions of the code. When called during remark, it will
2465     pick up from where the task left off during the concurrent marking
2466     phase. Interestingly, tasks are also claimable during evacuation
2467     pauses too, since do_marking_step() ensures that it aborts before
2468     it needs to yield.
2469 
2470     The data structures that it uses to do marking work are the
2471     following:
2472 
2473       (1) Marking Bitmap. If there are gray objects that appear only
2474       on the bitmap (this happens either when dealing with an overflow
2475       or when the initial marking phase has simply marked the roots
2476       and didn't push them on the stack), then tasks claim heap
2477       regions whose bitmap they then scan to find gray objects. A
2478       global finger indicates where the end of the last claimed region
2479       is. A local finger indicates how far into the region a task has
2480       scanned. The two fingers are used to determine how to gray an
2481       object (i.e. whether simply marking it is OK, as it will be
2482       visited by a task in the future, or whether it needs to be also
2483       pushed on a stack).
2484 
2485       (2) Local Queue. The local queue of the task which is accessed
2486       reasonably efficiently by the task. Other tasks can steal from
2487       it when they run out of work. Throughout the marking phase, a
2488       task attempts to keep its local queue short but not totally
2489       empty, so that entries are available for stealing by other
2490       tasks. Only when there is no more work, a task will totally
2491       drain its local queue.
2492 
2493       (3) Global Mark Stack. This handles local queue overflow. During
2494       marking only sets of entries are moved between it and the local
2495       queues, as access to it requires a mutex and more fine-grain
2496       interaction with it which might cause contention. If it
2497       overflows, then the marking phase should restart and iterate
2498       over the bitmap to identify gray objects. Throughout the marking
2499       phase, tasks attempt to keep the global mark stack at a small
2500       length but not totally empty, so that entries are available for
2501       popping by other tasks. Only when there is no more work, tasks
2502       will totally drain the global mark stack.
2503 
2504       (4) SATB Buffer Queue. This is where completed SATB buffers are
2505       made available. Buffers are regularly removed from this queue
2506       and scanned for roots, so that the queue doesn't get too
2507       long. During remark, all completed buffers are processed, as
2508       well as the filled in parts of any uncompleted buffers.
2509 
2510     The do_marking_step() method tries to abort when the time target
2511     has been reached. There are a few other cases when the
2512     do_marking_step() method also aborts:
2513 
2514       (1) When the marking phase has been aborted (after a Full GC).
2515 
2516       (2) When a global overflow (on the global stack) has been
2517       triggered. Before the task aborts, it will actually sync up with
2518       the other tasks to ensure that all the marking data structures
2519       (local queues, stacks, fingers etc.)  are re-initialized so that
2520       when do_marking_step() completes, the marking phase can
2521       immediately restart.
2522 
2523       (3) When enough completed SATB buffers are available. The
2524       do_marking_step() method only tries to drain SATB buffers right
2525       at the beginning. So, if enough buffers are available, the
2526       marking step aborts and the SATB buffers are processed at
2527       the beginning of the next invocation.
2528 
2529       (4) To yield. when we have to yield then we abort and yield
2530       right at the end of do_marking_step(). This saves us from a lot
2531       of hassle as, by yielding we might allow a Full GC. If this
2532       happens then objects will be compacted underneath our feet, the
2533       heap might shrink, etc. We save checking for this by just
2534       aborting and doing the yield right at the end.
2535 
2536     From the above it follows that the do_marking_step() method should
2537     be called in a loop (or, otherwise, regularly) until it completes.
2538 
2539     If a marking step completes without its has_aborted() flag being
2540     true, it means it has completed the current marking phase (and
2541     also all other marking tasks have done so and have all synced up).
2542 
2543     A method called regular_clock_call() is invoked "regularly" (in
2544     sub ms intervals) throughout marking. It is this clock method that
2545     checks all the abort conditions which were mentioned above and
2546     decides when the task should abort. A work-based scheme is used to
2547     trigger this clock method: when the number of object words the
2548     marking phase has scanned or the number of references the marking
2549     phase has visited reach a given limit. Additional invocations to
2550     the method clock have been planted in a few other strategic places
2551     too. The initial reason for the clock method was to avoid calling
2552     vtime too regularly, as it is quite expensive. So, once it was in
2553     place, it was natural to piggy-back all the other conditions on it
2554     too and not constantly check them throughout the code.
2555 
2556     If do_termination is true then do_marking_step will enter its
2557     termination protocol.
2558 
2559     The value of is_serial must be true when do_marking_step is being
2560     called serially (i.e. by the VMThread) and do_marking_step should
2561     skip any synchronization in the termination and overflow code.
2562     Examples include the serial remark code and the serial reference
2563     processing closures.
2564 
2565     The value of is_serial must be false when do_marking_step is
2566     being called by any of the worker threads in a work gang.
2567     Examples include the concurrent marking code (CMMarkingTask),
2568     the MT remark code, and the MT reference processing closures.
2569 
2570  *****************************************************************************/
2571 
2572 void G1CMTask::do_marking_step(double time_target_ms,
2573                                bool do_termination,
2574                                bool is_serial) {
2575   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2576 
2577   _start_time_ms = os::elapsedVTime() * 1000.0;
2578 
2579   // If do_stealing is true then do_marking_step will attempt to
2580   // steal work from the other G1CMTasks. It only makes sense to
2581   // enable stealing when the termination protocol is enabled
2582   // and do_marking_step() is not being called serially.
2583   bool do_stealing = do_termination && !is_serial;
2584 
2585   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2586   _time_target_ms = time_target_ms - diff_prediction_ms;
2587 
2588   // set up the variables that are used in the work-based scheme to
2589   // call the regular clock method
2590   _words_scanned = 0;
2591   _refs_reached  = 0;
2592   recalculate_limits();
2593 
2594   // clear all flags
2595   clear_has_aborted();
2596   _has_timed_out = false;
2597   _draining_satb_buffers = false;
2598 
2599   ++_calls;
2600 
2601   // Set up the bitmap and oop closures. Anything that uses them is
2602   // eventually called from this method, so it is OK to allocate these
2603   // statically.
2604   G1CMBitMapClosure bitmap_closure(this, _cm);
2605   G1CMOopClosure cm_oop_closure(_g1h, this);
2606   set_cm_oop_closure(&cm_oop_closure);
2607 
2608   if (_cm->has_overflown()) {
2609     // This can happen if the mark stack overflows during a GC pause
2610     // and this task, after a yield point, restarts. We have to abort
2611     // as we need to get into the overflow protocol which happens
2612     // right at the end of this task.
2613     set_has_aborted();
2614   }
2615 
2616   // First drain any available SATB buffers. After this, we will not
2617   // look at SATB buffers before the next invocation of this method.
2618   // If enough completed SATB buffers are queued up, the regular clock
2619   // will abort this task so that it restarts.
2620   drain_satb_buffers();
2621   // ...then partially drain the local queue and the global stack
2622   drain_local_queue(true);
2623   drain_global_stack(true);
2624 
2625   do {
2626     if (!has_aborted() && _curr_region != NULL) {
2627       // This means that we're already holding on to a region.
2628       assert(_finger != NULL, "if region is not NULL, then the finger "
2629              "should not be NULL either");
2630 
2631       // We might have restarted this task after an evacuation pause
2632       // which might have evacuated the region we're holding on to
2633       // underneath our feet. Let's read its limit again to make sure
2634       // that we do not iterate over a region of the heap that
2635       // contains garbage (update_region_limit() will also move
2636       // _finger to the start of the region if it is found empty).
2637       update_region_limit();
2638       // We will start from _finger not from the start of the region,
2639       // as we might be restarting this task after aborting half-way
2640       // through scanning this region. In this case, _finger points to
2641       // the address where we last found a marked object. If this is a
2642       // fresh region, _finger points to start().
2643       MemRegion mr = MemRegion(_finger, _region_limit);
2644 
2645       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2646              "humongous regions should go around loop once only");
2647 
2648       // Some special cases:
2649       // If the memory region is empty, we can just give up the region.
2650       // If the current region is humongous then we only need to check
2651       // the bitmap for the bit associated with the start of the object,
2652       // scan the object if it's live, and give up the region.
2653       // Otherwise, let's iterate over the bitmap of the part of the region
2654       // that is left.
2655       // If the iteration is successful, give up the region.
2656       if (mr.is_empty()) {
2657         giveup_current_region();
2658         regular_clock_call();
2659       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2660         if (_next_mark_bitmap->is_marked(mr.start())) {
2661           // The object is marked - apply the closure
2662           bitmap_closure.do_addr(mr.start());
2663         }
2664         // Even if this task aborted while scanning the humongous object
2665         // we can (and should) give up the current region.
2666         giveup_current_region();
2667         regular_clock_call();
2668       } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2669         giveup_current_region();
2670         regular_clock_call();
2671       } else {
2672         assert(has_aborted(), "currently the only way to do so");
2673         // The only way to abort the bitmap iteration is to return
2674         // false from the do_bit() method. However, inside the
2675         // do_bit() method we move the _finger to point to the
2676         // object currently being looked at. So, if we bail out, we
2677         // have definitely set _finger to something non-null.
2678         assert(_finger != NULL, "invariant");
2679 
2680         // Region iteration was actually aborted. So now _finger
2681         // points to the address of the object we last scanned. If we
2682         // leave it there, when we restart this task, we will rescan
2683         // the object. It is easy to avoid this. We move the finger by
2684         // enough to point to the next possible object header.
2685         assert(_finger < _region_limit, "invariant");
2686         HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2687         // Check if bitmap iteration was aborted while scanning the last object
2688         if (new_finger >= _region_limit) {
2689           giveup_current_region();
2690         } else {
2691           move_finger_to(new_finger);
2692         }
2693       }
2694     }
2695     // At this point we have either completed iterating over the
2696     // region we were holding on to, or we have aborted.
2697 
2698     // We then partially drain the local queue and the global stack.
2699     // (Do we really need this?)
2700     drain_local_queue(true);
2701     drain_global_stack(true);
2702 
2703     // Read the note on the claim_region() method on why it might
2704     // return NULL with potentially more regions available for
2705     // claiming and why we have to check out_of_regions() to determine
2706     // whether we're done or not.
2707     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2708       // We are going to try to claim a new region. We should have
2709       // given up on the previous one.
2710       // Separated the asserts so that we know which one fires.
2711       assert(_curr_region  == NULL, "invariant");
2712       assert(_finger       == NULL, "invariant");
2713       assert(_region_limit == NULL, "invariant");
2714       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2715       if (claimed_region != NULL) {
2716         // Yes, we managed to claim one
2717         setup_for_region(claimed_region);
2718         assert(_curr_region == claimed_region, "invariant");
2719       }
2720       // It is important to call the regular clock here. It might take
2721       // a while to claim a region if, for example, we hit a large
2722       // block of empty regions. So we need to call the regular clock
2723       // method once round the loop to make sure it's called
2724       // frequently enough.
2725       regular_clock_call();
2726     }
2727 
2728     if (!has_aborted() && _curr_region == NULL) {
2729       assert(_cm->out_of_regions(),
2730              "at this point we should be out of regions");
2731     }
2732   } while ( _curr_region != NULL && !has_aborted());
2733 
2734   if (!has_aborted()) {
2735     // We cannot check whether the global stack is empty, since other
2736     // tasks might be pushing objects to it concurrently.
2737     assert(_cm->out_of_regions(),
2738            "at this point we should be out of regions");
2739     // Try to reduce the number of available SATB buffers so that
2740     // remark has less work to do.
2741     drain_satb_buffers();
2742   }
2743 
2744   // Since we've done everything else, we can now totally drain the
2745   // local queue and global stack.
2746   drain_local_queue(false);
2747   drain_global_stack(false);
2748 
2749   // Attempt at work stealing from other task's queues.
2750   if (do_stealing && !has_aborted()) {
2751     // We have not aborted. This means that we have finished all that
2752     // we could. Let's try to do some stealing...
2753 
2754     // We cannot check whether the global stack is empty, since other
2755     // tasks might be pushing objects to it concurrently.
2756     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2757            "only way to reach here");
2758     while (!has_aborted()) {
2759       G1TaskQueueEntry entry;
2760       if (_cm->try_stealing(_worker_id, entry)) {
2761         scan_task_entry(entry);
2762 
2763         // And since we're towards the end, let's totally drain the
2764         // local queue and global stack.
2765         drain_local_queue(false);
2766         drain_global_stack(false);
2767       } else {
2768         break;
2769       }
2770     }
2771   }
2772 
2773   // We still haven't aborted. Now, let's try to get into the
2774   // termination protocol.
2775   if (do_termination && !has_aborted()) {
2776     // We cannot check whether the global stack is empty, since other
2777     // tasks might be concurrently pushing objects on it.
2778     // Separated the asserts so that we know which one fires.
2779     assert(_cm->out_of_regions(), "only way to reach here");
2780     assert(_task_queue->size() == 0, "only way to reach here");
2781     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2782 
2783     // The G1CMTask class also extends the TerminatorTerminator class,
2784     // hence its should_exit_termination() method will also decide
2785     // whether to exit the termination protocol or not.
2786     bool finished = (is_serial ||
2787                      _cm->terminator()->offer_termination(this));
2788     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2789     _termination_time_ms +=
2790       termination_end_time_ms - _termination_start_time_ms;
2791 
2792     if (finished) {
2793       // We're all done.
2794 
2795       // We can now guarantee that the global stack is empty, since
2796       // all other tasks have finished. We separated the guarantees so
2797       // that, if a condition is false, we can immediately find out
2798       // which one.
2799       guarantee(_cm->out_of_regions(), "only way to reach here");
2800       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2801       guarantee(_task_queue->size() == 0, "only way to reach here");
2802       guarantee(!_cm->has_overflown(), "only way to reach here");
2803     } else {
2804       // Apparently there's more work to do. Let's abort this task. It
2805       // will restart it and we can hopefully find more things to do.
2806       set_has_aborted();
2807     }
2808   }
2809 
2810   // Mainly for debugging purposes to make sure that a pointer to the
2811   // closure which was statically allocated in this frame doesn't
2812   // escape it by accident.
2813   set_cm_oop_closure(NULL);
2814   double end_time_ms = os::elapsedVTime() * 1000.0;
2815   double elapsed_time_ms = end_time_ms - _start_time_ms;
2816   // Update the step history.
2817   _step_times_ms.add(elapsed_time_ms);
2818 
2819   if (has_aborted()) {
2820     // The task was aborted for some reason.
2821     if (_has_timed_out) {
2822       double diff_ms = elapsed_time_ms - _time_target_ms;
2823       // Keep statistics of how well we did with respect to hitting
2824       // our target only if we actually timed out (if we aborted for
2825       // other reasons, then the results might get skewed).
2826       _marking_step_diffs_ms.add(diff_ms);
2827     }
2828 
2829     if (_cm->has_overflown()) {
2830       // This is the interesting one. We aborted because a global
2831       // overflow was raised. This means we have to restart the
2832       // marking phase and start iterating over regions. However, in
2833       // order to do this we have to make sure that all tasks stop
2834       // what they are doing and re-initialize in a safe manner. We
2835       // will achieve this with the use of two barrier sync points.
2836 
2837       if (!is_serial) {
2838         // We only need to enter the sync barrier if being called
2839         // from a parallel context
2840         _cm->enter_first_sync_barrier(_worker_id);
2841 
2842         // When we exit this sync barrier we know that all tasks have
2843         // stopped doing marking work. So, it's now safe to
2844         // re-initialize our data structures.
2845       }
2846 
2847       clear_region_fields();
2848       flush_mark_stats_cache();
2849 
2850       if (!is_serial) {
2851         // If we're executing the concurrent phase of marking, reset the marking
2852         // state; otherwise the marking state is reset after reference processing,
2853         // during the remark pause.
2854         // If we reset here as a result of an overflow during the remark we will
2855         // see assertion failures from any subsequent set_concurrency_and_phase()
2856         // calls.
2857         if (_cm->concurrent() && _worker_id == 0) {
2858           // Worker 0 is responsible for clearing the global data structures because
2859           // of an overflow. During STW we should not clear the overflow flag (in
2860           // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2861           // method to abort the pause and restart concurrent marking.
2862           _cm->reset_marking_for_restart();
2863 
2864           log_info(gc, marking)("Concurrent Mark reset for overflow");
2865         }
2866 
2867         // ...and enter the second barrier.
2868         _cm->enter_second_sync_barrier(_worker_id);
2869       }
2870       // At this point, if we're during the concurrent phase of
2871       // marking, everything has been re-initialized and we're
2872       // ready to restart.
2873     }
2874   }
2875 }
2876 
2877 G1CMTask::G1CMTask(uint worker_id,
2878                    G1ConcurrentMark* cm,
2879                    G1CMTaskQueue* task_queue,
2880                    G1RegionMarkStats* mark_stats,
2881                    uint max_regions) :
2882   _objArray_processor(this),
2883   _worker_id(worker_id),
2884   _g1h(G1CollectedHeap::heap()),
2885   _cm(cm),
2886   _next_mark_bitmap(NULL),
2887   _task_queue(task_queue),
2888   _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize),
2889   _calls(0),
2890   _time_target_ms(0.0),
2891   _start_time_ms(0.0),
2892   _cm_oop_closure(NULL),
2893   _curr_region(NULL),
2894   _finger(NULL),
2895   _region_limit(NULL),
2896   _words_scanned(0),
2897   _words_scanned_limit(0),
2898   _real_words_scanned_limit(0),
2899   _refs_reached(0),
2900   _refs_reached_limit(0),
2901   _real_refs_reached_limit(0),
2902   _has_aborted(false),
2903   _has_timed_out(false),
2904   _draining_satb_buffers(false),
2905   _step_times_ms(),
2906   _elapsed_time_ms(0.0),
2907   _termination_time_ms(0.0),
2908   _termination_start_time_ms(0.0),
2909   _marking_step_diffs_ms()
2910 {
2911   guarantee(task_queue != NULL, "invariant");
2912 
2913   _marking_step_diffs_ms.add(0.5);
2914 }
2915 
2916 // These are formatting macros that are used below to ensure
2917 // consistent formatting. The *_H_* versions are used to format the
2918 // header for a particular value and they should be kept consistent
2919 // with the corresponding macro. Also note that most of the macros add
2920 // the necessary white space (as a prefix) which makes them a bit
2921 // easier to compose.
2922 
2923 // All the output lines are prefixed with this string to be able to
2924 // identify them easily in a large log file.
2925 #define G1PPRL_LINE_PREFIX            "###"
2926 
2927 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2928 #ifdef _LP64
2929 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2930 #else // _LP64
2931 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2932 #endif // _LP64
2933 
2934 // For per-region info
2935 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2936 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2937 #define G1PPRL_STATE_FORMAT           "   %-5s"
2938 #define G1PPRL_STATE_H_FORMAT         "   %5s"
2939 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2940 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2941 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2942 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2943 
2944 // For summary info
2945 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2946 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2947 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2948 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2949 
2950 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2951   _total_used_bytes(0), _total_capacity_bytes(0),
2952   _total_prev_live_bytes(0), _total_next_live_bytes(0),
2953   _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2954 {
2955   if (!log_is_enabled(Trace, gc, liveness)) {
2956     return;
2957   }
2958 
2959   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2960   MemRegion g1_reserved = g1h->g1_reserved();
2961   double now = os::elapsedTime();
2962 
2963   // Print the header of the output.
2964   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2965   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2966                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2967                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2968                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2969                           HeapRegion::GrainBytes);
2970   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2971   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2972                           G1PPRL_TYPE_H_FORMAT
2973                           G1PPRL_ADDR_BASE_H_FORMAT
2974                           G1PPRL_BYTE_H_FORMAT
2975                           G1PPRL_BYTE_H_FORMAT
2976                           G1PPRL_BYTE_H_FORMAT
2977                           G1PPRL_DOUBLE_H_FORMAT
2978                           G1PPRL_BYTE_H_FORMAT
2979                           G1PPRL_STATE_H_FORMAT
2980                           G1PPRL_BYTE_H_FORMAT,
2981                           "type", "address-range",
2982                           "used", "prev-live", "next-live", "gc-eff",
2983                           "remset", "state", "code-roots");
2984   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2985                           G1PPRL_TYPE_H_FORMAT
2986                           G1PPRL_ADDR_BASE_H_FORMAT
2987                           G1PPRL_BYTE_H_FORMAT
2988                           G1PPRL_BYTE_H_FORMAT
2989                           G1PPRL_BYTE_H_FORMAT
2990                           G1PPRL_DOUBLE_H_FORMAT
2991                           G1PPRL_BYTE_H_FORMAT
2992                           G1PPRL_STATE_H_FORMAT
2993                           G1PPRL_BYTE_H_FORMAT,
2994                           "", "",
2995                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
2996                           "(bytes)", "", "(bytes)");
2997 }
2998 
2999 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
3000   if (!log_is_enabled(Trace, gc, liveness)) {
3001     return false;
3002   }
3003 
3004   const char* type       = r->get_type_str();
3005   HeapWord* bottom       = r->bottom();
3006   HeapWord* end          = r->end();
3007   size_t capacity_bytes  = r->capacity();
3008   size_t used_bytes      = r->used();
3009   size_t prev_live_bytes = r->live_bytes();
3010   size_t next_live_bytes = r->next_live_bytes();
3011   double gc_eff          = r->gc_efficiency();
3012   size_t remset_bytes    = r->rem_set()->mem_size();
3013   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3014   const char* remset_type = r->rem_set()->get_short_state_str();
3015 
3016   _total_used_bytes      += used_bytes;
3017   _total_capacity_bytes  += capacity_bytes;
3018   _total_prev_live_bytes += prev_live_bytes;
3019   _total_next_live_bytes += next_live_bytes;
3020   _total_remset_bytes    += remset_bytes;
3021   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3022 
3023   // Print a line for this particular region.
3024   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3025                           G1PPRL_TYPE_FORMAT
3026                           G1PPRL_ADDR_BASE_FORMAT
3027                           G1PPRL_BYTE_FORMAT
3028                           G1PPRL_BYTE_FORMAT
3029                           G1PPRL_BYTE_FORMAT
3030                           G1PPRL_DOUBLE_FORMAT
3031                           G1PPRL_BYTE_FORMAT
3032                           G1PPRL_STATE_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, remset_type, strong_code_roots_bytes);
3037 
3038   return false;
3039 }
3040 
3041 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3042   if (!log_is_enabled(Trace, gc, liveness)) {
3043     return;
3044   }
3045 
3046   // add static memory usages to remembered set sizes
3047   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3048   // Print the footer of the output.
3049   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3050   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3051                          " SUMMARY"
3052                          G1PPRL_SUM_MB_FORMAT("capacity")
3053                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3054                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3055                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3056                          G1PPRL_SUM_MB_FORMAT("remset")
3057                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3058                          bytes_to_mb(_total_capacity_bytes),
3059                          bytes_to_mb(_total_used_bytes),
3060                          percent_of(_total_used_bytes, _total_capacity_bytes),
3061                          bytes_to_mb(_total_prev_live_bytes),
3062                          percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3063                          bytes_to_mb(_total_next_live_bytes),
3064                          percent_of(_total_next_live_bytes, _total_capacity_bytes),
3065                          bytes_to_mb(_total_remset_bytes),
3066                          bytes_to_mb(_total_strong_code_roots_bytes));
3067 }