1 /*
   2  * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/metadataOnStackMark.hpp"
  27 #include "classfile/symbolTable.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc/g1/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/heapRegion.inline.hpp"
  39 #include "gc/g1/heapRegionRemSet.hpp"
  40 #include "gc/g1/heapRegionSet.inline.hpp"
  41 #include "gc/shared/adaptiveSizePolicy.hpp"
  42 #include "gc/shared/gcId.hpp"
  43 #include "gc/shared/gcTimer.hpp"
  44 #include "gc/shared/gcTrace.hpp"
  45 #include "gc/shared/gcTraceTime.inline.hpp"
  46 #include "gc/shared/genOopClosures.inline.hpp"
  47 #include "gc/shared/referencePolicy.hpp"
  48 #include "gc/shared/strongRootsScope.hpp"
  49 #include "gc/shared/suspendibleThreadSet.hpp"
  50 #include "gc/shared/taskqueue.inline.hpp"
  51 #include "gc/shared/vmGCOperations.hpp"
  52 #include "gc/shared/weakProcessor.hpp"
  53 #include "include/jvm.h"
  54 #include "logging/log.hpp"
  55 #include "memory/allocation.hpp"
  56 #include "memory/resourceArea.hpp"
  57 #include "oops/access.inline.hpp"
  58 #include "oops/oop.inline.hpp"
  59 #include "runtime/atomic.hpp"
  60 #include "runtime/handles.inline.hpp"
  61 #include "runtime/java.hpp"
  62 #include "runtime/prefetch.inline.hpp"
  63 #include "services/memTracker.hpp"
  64 #include "utilities/align.hpp"
  65 #include "utilities/growableArray.hpp"
  66 
  67 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
  68   assert(addr < _cm->finger(), "invariant");
  69   assert(addr >= _task->finger(), "invariant");
  70 
  71   // We move that task's local finger along.
  72   _task->move_finger_to(addr);
  73 
  74   _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
  75   // we only partially drain the local queue and global stack
  76   _task->drain_local_queue(true);
  77   _task->drain_global_stack(true);
  78 
  79   // if the has_aborted flag has been raised, we need to bail out of
  80   // the iteration
  81   return !_task->has_aborted();
  82 }
  83 
  84 G1CMMarkStack::G1CMMarkStack() :
  85   _max_chunk_capacity(0),
  86   _base(NULL),
  87   _chunk_capacity(0) {
  88   set_empty();
  89 }
  90 
  91 bool G1CMMarkStack::resize(size_t new_capacity) {
  92   assert(is_empty(), "Only resize when stack is empty.");
  93   assert(new_capacity <= _max_chunk_capacity,
  94          "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
  95 
  96   TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
  97 
  98   if (new_base == NULL) {
  99     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));
 100     return false;
 101   }
 102   // Release old mapping.
 103   if (_base != NULL) {
 104     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 105   }
 106 
 107   _base = new_base;
 108   _chunk_capacity = new_capacity;
 109   set_empty();
 110 
 111   return true;
 112 }
 113 
 114 size_t G1CMMarkStack::capacity_alignment() {
 115   return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
 116 }
 117 
 118 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
 119   guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
 120 
 121   size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
 122 
 123   _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 124   size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 125 
 126   guarantee(initial_chunk_capacity <= _max_chunk_capacity,
 127             "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
 128             _max_chunk_capacity,
 129             initial_chunk_capacity);
 130 
 131   log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
 132                 initial_chunk_capacity, _max_chunk_capacity);
 133 
 134   return resize(initial_chunk_capacity);
 135 }
 136 
 137 void G1CMMarkStack::expand() {
 138   if (_chunk_capacity == _max_chunk_capacity) {
 139     log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
 140     return;
 141   }
 142   size_t old_capacity = _chunk_capacity;
 143   // Double capacity if possible
 144   size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
 145 
 146   if (resize(new_capacity)) {
 147     log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 148                   old_capacity, new_capacity);
 149   } else {
 150     log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 151                     old_capacity, new_capacity);
 152   }
 153 }
 154 
 155 G1CMMarkStack::~G1CMMarkStack() {
 156   if (_base != NULL) {
 157     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 158   }
 159 }
 160 
 161 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
 162   elem->next = *list;
 163   *list = elem;
 164 }
 165 
 166 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
 167   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 168   add_chunk_to_list(&_chunk_list, elem);
 169   _chunks_in_chunk_list++;
 170 }
 171 
 172 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
 173   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 174   add_chunk_to_list(&_free_list, elem);
 175 }
 176 
 177 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
 178   TaskQueueEntryChunk* result = *list;
 179   if (result != NULL) {
 180     *list = (*list)->next;
 181   }
 182   return result;
 183 }
 184 
 185 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
 186   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 187   TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
 188   if (result != NULL) {
 189     _chunks_in_chunk_list--;
 190   }
 191   return result;
 192 }
 193 
 194 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
 195   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 196   return remove_chunk_from_list(&_free_list);
 197 }
 198 
 199 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
 200   // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
 201   // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
 202   // wraparound of _hwm.
 203   if (_hwm >= _chunk_capacity) {
 204     return NULL;
 205   }
 206 
 207   size_t cur_idx = Atomic::add(1u, &_hwm) - 1;
 208   if (cur_idx >= _chunk_capacity) {
 209     return NULL;
 210   }
 211 
 212   TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
 213   result->next = NULL;
 214   return result;
 215 }
 216 
 217 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
 218   // Get a new chunk.
 219   TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
 220 
 221   if (new_chunk == NULL) {
 222     // Did not get a chunk from the free list. Allocate from backing memory.
 223     new_chunk = allocate_new_chunk();
 224 
 225     if (new_chunk == NULL) {
 226       return false;
 227     }
 228   }
 229 
 230   Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 231 
 232   add_chunk_to_chunk_list(new_chunk);
 233 
 234   return true;
 235 }
 236 
 237 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
 238   TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
 239 
 240   if (cur == NULL) {
 241     return false;
 242   }
 243 
 244   Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 245 
 246   add_chunk_to_free_list(cur);
 247   return true;
 248 }
 249 
 250 void G1CMMarkStack::set_empty() {
 251   _chunks_in_chunk_list = 0;
 252   _hwm = 0;
 253   _chunk_list = NULL;
 254   _free_list = NULL;
 255 }
 256 
 257 G1CMRootRegions::G1CMRootRegions() :
 258   _survivors(NULL), _cm(NULL), _scan_in_progress(false),
 259   _should_abort(false), _claimed_survivor_index(0) { }
 260 
 261 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
 262   _survivors = survivors;
 263   _cm = cm;
 264 }
 265 
 266 void G1CMRootRegions::prepare_for_scan() {
 267   assert(!scan_in_progress(), "pre-condition");
 268 
 269   // Currently, only survivors can be root regions.
 270   _claimed_survivor_index = 0;
 271   _scan_in_progress = _survivors->regions()->is_nonempty();
 272   _should_abort = false;
 273 }
 274 
 275 HeapRegion* G1CMRootRegions::claim_next() {
 276   if (_should_abort) {
 277     // If someone has set the should_abort flag, we return NULL to
 278     // force the caller to bail out of their loop.
 279     return NULL;
 280   }
 281 
 282   // Currently, only survivors can be root regions.
 283   const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
 284 
 285   int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
 286   if (claimed_index < survivor_regions->length()) {
 287     return survivor_regions->at(claimed_index);
 288   }
 289   return NULL;
 290 }
 291 
 292 uint G1CMRootRegions::num_root_regions() const {
 293   return (uint)_survivors->regions()->length();
 294 }
 295 
 296 void G1CMRootRegions::notify_scan_done() {
 297   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 298   _scan_in_progress = false;
 299   RootRegionScan_lock->notify_all();
 300 }
 301 
 302 void G1CMRootRegions::cancel_scan() {
 303   notify_scan_done();
 304 }
 305 
 306 void G1CMRootRegions::scan_finished() {
 307   assert(scan_in_progress(), "pre-condition");
 308 
 309   // Currently, only survivors can be root regions.
 310   if (!_should_abort) {
 311     assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
 312     assert((uint)_claimed_survivor_index >= _survivors->length(),
 313            "we should have claimed all survivors, claimed index = %u, length = %u",
 314            (uint)_claimed_survivor_index, _survivors->length());
 315   }
 316 
 317   notify_scan_done();
 318 }
 319 
 320 bool G1CMRootRegions::wait_until_scan_finished() {
 321   if (!scan_in_progress()) {
 322     return false;
 323   }
 324 
 325   {
 326     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 327     while (scan_in_progress()) {
 328       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 329     }
 330   }
 331   return true;
 332 }
 333 
 334 // Returns the maximum number of workers to be used in a concurrent
 335 // phase based on the number of GC workers being used in a STW
 336 // phase.
 337 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
 338   return MAX2((num_gc_workers + 2) / 4, 1U);
 339 }
 340 
 341 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
 342                                    G1RegionToSpaceMapper* prev_bitmap_storage,
 343                                    G1RegionToSpaceMapper* next_bitmap_storage) :
 344   // _cm_thread set inside the constructor
 345   _g1h(g1h),
 346   _completed_initialization(false),
 347 
 348   _mark_bitmap_1(),
 349   _mark_bitmap_2(),
 350   _prev_mark_bitmap(&_mark_bitmap_1),
 351   _next_mark_bitmap(&_mark_bitmap_2),
 352 
 353   _heap(_g1h->reserved_region()),
 354 
 355   _root_regions(),
 356 
 357   _global_mark_stack(),
 358 
 359   // _finger set in set_non_marking_state
 360 
 361   _worker_id_offset(DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads),
 362   _max_num_tasks(ParallelGCThreads),
 363   // _num_active_tasks set in set_non_marking_state()
 364   // _tasks set inside the constructor
 365 
 366   _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
 367   _terminator(ParallelTaskTerminator((int) _max_num_tasks, _task_queues)),
 368 
 369   _first_overflow_barrier_sync(),
 370   _second_overflow_barrier_sync(),
 371 
 372   _has_overflown(false),
 373   _concurrent(false),
 374   _has_aborted(false),
 375   _restart_for_overflow(false),
 376   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 377   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 378 
 379   // _verbose_level set below
 380 
 381   _init_times(),
 382   _remark_times(),
 383   _remark_mark_times(),
 384   _remark_weak_ref_times(),
 385   _cleanup_times(),
 386   _total_cleanup_time(0.0),
 387 
 388   _accum_task_vtime(NULL),
 389 
 390   _concurrent_workers(NULL),
 391   _num_concurrent_workers(0),
 392   _max_concurrent_workers(0),
 393 
 394   _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)),
 395   _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC))
 396 {
 397   _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 398   _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
 399 
 400   // Create & start ConcurrentMark thread.
 401   _cm_thread = new G1ConcurrentMarkThread(this);
 402   if (_cm_thread->osthread() == NULL) {
 403     vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 404   }
 405 
 406   assert(CGC_lock != NULL, "CGC_lock must be initialized");
 407 
 408   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 409   satb_qs.set_buffer_size(G1SATBBufferSize);
 410 
 411   _root_regions.init(_g1h->survivor(), this);
 412 
 413   if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
 414     // Calculate the number of concurrent worker threads by scaling
 415     // the number of parallel GC threads.
 416     uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
 417     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 418   }
 419 
 420   assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
 421   if (ConcGCThreads > ParallelGCThreads) {
 422     log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
 423                     ConcGCThreads, ParallelGCThreads);
 424     return;
 425   }
 426 
 427   log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset);
 428   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 429 
 430   _num_concurrent_workers = ConcGCThreads;
 431   _max_concurrent_workers = _num_concurrent_workers;
 432 
 433   _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true);
 434   _concurrent_workers->initialize_workers();
 435 
 436   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 437     size_t mark_stack_size =
 438       MIN2(MarkStackSizeMax,
 439           MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
 440     // Verify that the calculated value for MarkStackSize is in range.
 441     // It would be nice to use the private utility routine from Arguments.
 442     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 443       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 444                       "must be between 1 and " SIZE_FORMAT,
 445                       mark_stack_size, MarkStackSizeMax);
 446       return;
 447     }
 448     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 449   } else {
 450     // Verify MarkStackSize is in range.
 451     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 452       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 453         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 454           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 455                           "must be between 1 and " SIZE_FORMAT,
 456                           MarkStackSize, MarkStackSizeMax);
 457           return;
 458         }
 459       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 460         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 461           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 462                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 463                           MarkStackSize, MarkStackSizeMax);
 464           return;
 465         }
 466       }
 467     }
 468   }
 469 
 470   if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
 471     vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
 472   }
 473 
 474   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
 475   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
 476 
 477   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 478   _num_active_tasks = _max_num_tasks;
 479 
 480   for (uint i = 0; i < _max_num_tasks; ++i) {
 481     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 482     task_queue->initialize();
 483     _task_queues->register_queue(i, task_queue);
 484 
 485     _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions());
 486 
 487     _accum_task_vtime[i] = 0.0;
 488   }
 489 
 490   reset_at_marking_complete();
 491   _completed_initialization = true;
 492 }
 493 
 494 void G1ConcurrentMark::reset() {
 495   _has_aborted = false;
 496 
 497   reset_marking_for_restart();
 498 
 499   // Reset all tasks, since different phases will use different number of active
 500   // threads. So, it's easiest to have all of them ready.
 501   for (uint i = 0; i < _max_num_tasks; ++i) {
 502     _tasks[i]->reset(_next_mark_bitmap);
 503   }
 504 
 505   uint max_regions = _g1h->max_regions();
 506   for (uint i = 0; i < max_regions; i++) {
 507     _top_at_rebuild_starts[i] = NULL;
 508     _region_mark_stats[i].clear();
 509   }
 510 }
 511 
 512 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) {
 513   for (uint j = 0; j < _max_num_tasks; ++j) {
 514     _tasks[j]->clear_mark_stats_cache(region_idx);
 515   }
 516   _top_at_rebuild_starts[region_idx] = NULL;
 517   _region_mark_stats[region_idx].clear();
 518 }
 519 
 520 void G1ConcurrentMark::clear_statistics(HeapRegion* r) {
 521   uint const region_idx = r->hrm_index();
 522   if (r->is_humongous()) {
 523     assert(r->is_starts_humongous(), "Got humongous continues region here");
 524     uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size());
 525     for (uint j = region_idx; j < (region_idx + size_in_regions); j++) {
 526       clear_statistics_in_region(j);
 527     }
 528   } else {
 529     clear_statistics_in_region(region_idx);
 530   }
 531 }
 532 
 533 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) {
 534   if (bitmap->is_marked(addr)) {
 535     bitmap->clear(addr);
 536   }
 537 }
 538 
 539 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) {
 540   assert_at_safepoint_on_vm_thread();
 541 
 542   // Need to clear all mark bits of the humongous object.
 543   clear_mark_if_set(_prev_mark_bitmap, r->bottom());
 544   clear_mark_if_set(_next_mark_bitmap, r->bottom());
 545 
 546   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
 547     return;
 548   }
 549 
 550   // Clear any statistics about the region gathered so far.
 551   clear_statistics(r);
 552 }
 553 
 554 void G1ConcurrentMark::reset_marking_for_restart() {
 555   _global_mark_stack.set_empty();
 556 
 557   // Expand the marking stack, if we have to and if we can.
 558   if (has_overflown()) {
 559     _global_mark_stack.expand();
 560 
 561     uint max_regions = _g1h->max_regions();
 562     for (uint i = 0; i < max_regions; i++) {
 563       _region_mark_stats[i].clear_during_overflow();
 564     }
 565   }
 566 
 567   clear_has_overflown();
 568   _finger = _heap.start();
 569 
 570   for (uint i = 0; i < _max_num_tasks; ++i) {
 571     G1CMTaskQueue* queue = _task_queues->queue(i);
 572     queue->set_empty();
 573   }
 574 }
 575 
 576 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 577   assert(active_tasks <= _max_num_tasks, "we should not have more");
 578 
 579   _num_active_tasks = active_tasks;
 580   // Need to update the three data structures below according to the
 581   // number of active threads for this phase.
 582   _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues);
 583   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 584   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 585 }
 586 
 587 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 588   set_concurrency(active_tasks);
 589 
 590   _concurrent = concurrent;
 591 
 592   if (!concurrent) {
 593     // At this point we should be in a STW phase, and completed marking.
 594     assert_at_safepoint_on_vm_thread();
 595     assert(out_of_regions(),
 596            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 597            p2i(_finger), p2i(_heap.end()));
 598   }
 599 }
 600 
 601 void G1ConcurrentMark::reset_at_marking_complete() {
 602   // We set the global marking state to some default values when we're
 603   // not doing marking.
 604   reset_marking_for_restart();
 605   _num_active_tasks = 0;
 606 }
 607 
 608 G1ConcurrentMark::~G1ConcurrentMark() {
 609   FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts);
 610   FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats);
 611   // The G1ConcurrentMark instance is never freed.
 612   ShouldNotReachHere();
 613 }
 614 
 615 class G1ClearBitMapTask : public AbstractGangTask {
 616 public:
 617   static size_t chunk_size() { return M; }
 618 
 619 private:
 620   // Heap region closure used for clearing the given mark bitmap.
 621   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 622   private:
 623     G1CMBitMap* _bitmap;
 624     G1ConcurrentMark* _cm;
 625   public:
 626     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
 627     }
 628 
 629     virtual bool do_heap_region(HeapRegion* r) {
 630       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 631 
 632       HeapWord* cur = r->bottom();
 633       HeapWord* const end = r->end();
 634 
 635       while (cur < end) {
 636         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 637         _bitmap->clear_range(mr);
 638 
 639         cur += chunk_size_in_words;
 640 
 641         // Abort iteration if after yielding the marking has been aborted.
 642         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 643           return true;
 644         }
 645         // Repeat the asserts from before the start of the closure. We will do them
 646         // as asserts here to minimize their overhead on the product. However, we
 647         // will have them as guarantees at the beginning / end of the bitmap
 648         // clearing to get some checking in the product.
 649         assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
 650         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 651       }
 652       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 653 
 654       return false;
 655     }
 656   };
 657 
 658   G1ClearBitmapHRClosure _cl;
 659   HeapRegionClaimer _hr_claimer;
 660   bool _suspendible; // If the task is suspendible, workers must join the STS.
 661 
 662 public:
 663   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 664     AbstractGangTask("G1 Clear Bitmap"),
 665     _cl(bitmap, suspendible ? cm : NULL),
 666     _hr_claimer(n_workers),
 667     _suspendible(suspendible)
 668   { }
 669 
 670   void work(uint worker_id) {
 671     SuspendibleThreadSetJoiner sts_join(_suspendible);
 672     G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
 673   }
 674 
 675   bool is_complete() {
 676     return _cl.is_complete();
 677   }
 678 };
 679 
 680 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 681   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 682 
 683   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 684   size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 685 
 686   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 687 
 688   G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
 689 
 690   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 691   workers->run_task(&cl, num_workers);
 692   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 693 }
 694 
 695 void G1ConcurrentMark::cleanup_for_next_mark() {
 696   // Make sure that the concurrent mark thread looks to still be in
 697   // the current cycle.
 698   guarantee(cm_thread()->during_cycle(), "invariant");
 699 
 700   // We are finishing up the current cycle by clearing the next
 701   // marking bitmap and getting it ready for the next cycle. During
 702   // this time no other cycle can start. So, let's make sure that this
 703   // is the case.
 704   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 705 
 706   clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
 707 
 708   // Repeat the asserts from above.
 709   guarantee(cm_thread()->during_cycle(), "invariant");
 710   guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant");
 711 }
 712 
 713 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 714   assert_at_safepoint_on_vm_thread();
 715   clear_bitmap(_prev_mark_bitmap, workers, false);
 716 }
 717 
 718 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 719   G1CMBitMap* _bitmap;
 720  public:
 721   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 722   }
 723 
 724   virtual bool do_heap_region(HeapRegion* r) {
 725     // This closure can be called concurrently to the mutator, so we must make sure
 726     // that the result of the getNextMarkedWordAddress() call is compared to the
 727     // value passed to it as limit to detect any found bits.
 728     // end never changes in G1.
 729     HeapWord* end = r->end();
 730     return _bitmap->get_next_marked_addr(r->bottom(), end) != end;
 731   }
 732 };
 733 
 734 bool G1ConcurrentMark::next_mark_bitmap_is_clear() {
 735   CheckBitmapClearHRClosure cl(_next_mark_bitmap);
 736   _g1h->heap_region_iterate(&cl);
 737   return cl.is_complete();
 738 }
 739 
 740 class NoteStartOfMarkHRClosure : public HeapRegionClosure {
 741 public:
 742   bool do_heap_region(HeapRegion* r) {
 743     r->note_start_of_marking();
 744     return false;
 745   }
 746 };
 747 
 748 void G1ConcurrentMark::pre_initial_mark() {
 749   // Initialize marking structures. This has to be done in a STW phase.
 750   reset();
 751 
 752   // For each region note start of marking.
 753   NoteStartOfMarkHRClosure startcl;
 754   _g1h->heap_region_iterate(&startcl);
 755 }
 756 
 757 
 758 void G1ConcurrentMark::post_initial_mark() {
 759   // Start Concurrent Marking weak-reference discovery.
 760   ReferenceProcessor* rp = _g1h->ref_processor_cm();
 761   // enable ("weak") refs discovery
 762   rp->enable_discovery();
 763   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 764 
 765   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 766   // This is the start of  the marking cycle, we're expected all
 767   // threads to have SATB queues with active set to false.
 768   satb_mq_set.set_active_all_threads(true, /* new active value */
 769                                      false /* expected_active */);
 770 
 771   _root_regions.prepare_for_scan();
 772 
 773   // update_g1_committed() will be called at the end of an evac pause
 774   // when marking is on. So, it's also called at the end of the
 775   // initial-mark pause to update the heap end, if the heap expands
 776   // during it. No need to call it here.
 777 }
 778 
 779 /*
 780  * Notice that in the next two methods, we actually leave the STS
 781  * during the barrier sync and join it immediately afterwards. If we
 782  * do not do this, the following deadlock can occur: one thread could
 783  * be in the barrier sync code, waiting for the other thread to also
 784  * sync up, whereas another one could be trying to yield, while also
 785  * waiting for the other threads to sync up too.
 786  *
 787  * Note, however, that this code is also used during remark and in
 788  * this case we should not attempt to leave / enter the STS, otherwise
 789  * we'll either hit an assert (debug / fastdebug) or deadlock
 790  * (product). So we should only leave / enter the STS if we are
 791  * operating concurrently.
 792  *
 793  * Because the thread that does the sync barrier has left the STS, it
 794  * is possible to be suspended for a Full GC or an evacuation pause
 795  * could occur. This is actually safe, since the entering the sync
 796  * barrier is one of the last things do_marking_step() does, and it
 797  * doesn't manipulate any data structures afterwards.
 798  */
 799 
 800 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 801   bool barrier_aborted;
 802   {
 803     SuspendibleThreadSetLeaver sts_leave(concurrent());
 804     barrier_aborted = !_first_overflow_barrier_sync.enter();
 805   }
 806 
 807   // at this point everyone should have synced up and not be doing any
 808   // more work
 809 
 810   if (barrier_aborted) {
 811     // If the barrier aborted we ignore the overflow condition and
 812     // just abort the whole marking phase as quickly as possible.
 813     return;
 814   }
 815 }
 816 
 817 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 818   SuspendibleThreadSetLeaver sts_leave(concurrent());
 819   _second_overflow_barrier_sync.enter();
 820 
 821   // at this point everything should be re-initialized and ready to go
 822 }
 823 
 824 class G1CMConcurrentMarkingTask : public AbstractGangTask {
 825   G1ConcurrentMark*     _cm;
 826 
 827 public:
 828   void work(uint worker_id) {
 829     assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
 830     ResourceMark rm;
 831 
 832     double start_vtime = os::elapsedVTime();
 833 
 834     {
 835       SuspendibleThreadSetJoiner sts_join;
 836 
 837       assert(worker_id < _cm->active_tasks(), "invariant");
 838 
 839       G1CMTask* task = _cm->task(worker_id);
 840       task->record_start_time();
 841       if (!_cm->has_aborted()) {
 842         do {
 843           task->do_marking_step(G1ConcMarkStepDurationMillis,
 844                                 true  /* do_termination */,
 845                                 false /* is_serial*/);
 846 
 847           _cm->do_yield_check();
 848         } while (!_cm->has_aborted() && task->has_aborted());
 849       }
 850       task->record_end_time();
 851       guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
 852     }
 853 
 854     double end_vtime = os::elapsedVTime();
 855     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 856   }
 857 
 858   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) :
 859       AbstractGangTask("Concurrent Mark"), _cm(cm) { }
 860 
 861   ~G1CMConcurrentMarkingTask() { }
 862 };
 863 
 864 uint G1ConcurrentMark::calc_active_marking_workers() {
 865   uint result = 0;
 866   if (!UseDynamicNumberOfGCThreads ||
 867       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 868        !ForceDynamicNumberOfGCThreads)) {
 869     result = _max_concurrent_workers;
 870   } else {
 871     result =
 872       AdaptiveSizePolicy::calc_default_active_workers(_max_concurrent_workers,
 873                                                       1, /* Minimum workers */
 874                                                       _num_concurrent_workers,
 875                                                       Threads::number_of_non_daemon_threads());
 876     // Don't scale the result down by scale_concurrent_workers() because
 877     // that scaling has already gone into "_max_concurrent_workers".
 878   }
 879   assert(result > 0 && result <= _max_concurrent_workers,
 880          "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
 881          _max_concurrent_workers, result);
 882   return result;
 883 }
 884 
 885 void G1ConcurrentMark::scan_root_region(HeapRegion* hr, uint worker_id) {
 886   // Currently, only survivors can be root regions.
 887   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 888   G1RootRegionScanClosure cl(_g1h, this, worker_id);
 889 
 890   const uintx interval = PrefetchScanIntervalInBytes;
 891   HeapWord* curr = hr->bottom();
 892   const HeapWord* end = hr->top();
 893   while (curr < end) {
 894     Prefetch::read(curr, interval);
 895     oop obj = oop(curr);
 896     int size = obj->oop_iterate_size(&cl);
 897     assert(size == obj->size(), "sanity");
 898     curr += size;
 899   }
 900 }
 901 
 902 class G1CMRootRegionScanTask : public AbstractGangTask {
 903   G1ConcurrentMark* _cm;
 904 public:
 905   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 906     AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
 907 
 908   void work(uint worker_id) {
 909     assert(Thread::current()->is_ConcurrentGC_thread(),
 910            "this should only be done by a conc GC thread");
 911 
 912     G1CMRootRegions* root_regions = _cm->root_regions();
 913     HeapRegion* hr = root_regions->claim_next();
 914     while (hr != NULL) {
 915       _cm->scan_root_region(hr, worker_id);
 916       hr = root_regions->claim_next();
 917     }
 918   }
 919 };
 920 
 921 void G1ConcurrentMark::scan_root_regions() {
 922   // scan_in_progress() will have been set to true only if there was
 923   // at least one root region to scan. So, if it's false, we
 924   // should not attempt to do any further work.
 925   if (root_regions()->scan_in_progress()) {
 926     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 927 
 928     _num_concurrent_workers = MIN2(calc_active_marking_workers(),
 929                                    // We distribute work on a per-region basis, so starting
 930                                    // more threads than that is useless.
 931                                    root_regions()->num_root_regions());
 932     assert(_num_concurrent_workers <= _max_concurrent_workers,
 933            "Maximum number of marking threads exceeded");
 934 
 935     G1CMRootRegionScanTask task(this);
 936     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
 937                         task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
 938     _concurrent_workers->run_task(&task, _num_concurrent_workers);
 939 
 940     // It's possible that has_aborted() is true here without actually
 941     // aborting the survivor scan earlier. This is OK as it's
 942     // mainly used for sanity checking.
 943     root_regions()->scan_finished();
 944   }
 945 }
 946 
 947 void G1ConcurrentMark::concurrent_cycle_start() {
 948   _gc_timer_cm->register_gc_start();
 949 
 950   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
 951 
 952   _g1h->trace_heap_before_gc(_gc_tracer_cm);
 953 }
 954 
 955 void G1ConcurrentMark::concurrent_cycle_end() {
 956   _g1h->collector_state()->set_clearing_next_bitmap(false);
 957 
 958   _g1h->trace_heap_after_gc(_gc_tracer_cm);
 959 
 960   if (has_aborted()) {
 961     log_info(gc, marking)("Concurrent Mark Abort");
 962     _gc_tracer_cm->report_concurrent_mode_failure();
 963   }
 964 
 965   _gc_timer_cm->register_gc_end();
 966 
 967   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
 968 }
 969 
 970 void G1ConcurrentMark::mark_from_roots() {
 971   _restart_for_overflow = false;
 972 
 973   _num_concurrent_workers = calc_active_marking_workers();
 974 
 975   uint active_workers = MAX2(1U, _num_concurrent_workers);
 976 
 977   // Setting active workers is not guaranteed since fewer
 978   // worker threads may currently exist and more may not be
 979   // available.
 980   active_workers = _concurrent_workers->update_active_workers(active_workers);
 981   log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
 982 
 983   // Parallel task terminator is set in "set_concurrency_and_phase()"
 984   set_concurrency_and_phase(active_workers, true /* concurrent */);
 985 
 986   G1CMConcurrentMarkingTask marking_task(this);
 987   _concurrent_workers->run_task(&marking_task);
 988   print_stats();
 989 }
 990 
 991 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) {
 992   G1HeapVerifier* verifier = _g1h->verifier();
 993 
 994   verifier->verify_region_sets_optional();
 995 
 996   if (VerifyDuringGC) {
 997     GCTraceTime(Debug, gc, phases) trace(caller, _gc_timer_cm);
 998 
 999     size_t const BufLen = 512;
1000     char buffer[BufLen];
1001 
1002     jio_snprintf(buffer, BufLen, "During GC (%s)", caller);
1003     verifier->verify(type, vo, buffer);
1004   }
1005 
1006   verifier->check_bitmaps(caller);
1007 }
1008 
1009 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask {
1010   G1CollectedHeap* _g1h;
1011   G1ConcurrentMark* _cm;
1012   HeapRegionClaimer _hrclaimer;
1013   uint _total_selected_for_rebuild;
1014  
1015   G1PrintRegionLivenessInfoClosure _cl;
1016 
1017   class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure {
1018     G1CollectedHeap* _g1h;
1019     G1ConcurrentMark* _cm;
1020 
1021     G1PrintRegionLivenessInfoClosure* _cl;
1022 
1023     uint _num_regions_selected_for_rebuild;  // The number of regions actually selected for rebuild.
1024 
1025     void update_remset_before_rebuild(HeapRegion * hr) {
1026       G1RemSetTrackingPolicy* tracking_policy = _g1h->g1_policy()->remset_tracker();
1027 
1028       size_t const live_bytes = _cm->liveness(hr->hrm_index()) * HeapWordSize;
1029       bool selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes);
1030       if (selected_for_rebuild) {
1031         _num_regions_selected_for_rebuild++;
1032       }
1033       _cm->update_top_at_rebuild_start(hr);
1034     }
1035 
1036     // Distribute the given words across the humongous object starting with hr and
1037     // note end of marking.
1038     void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) {
1039       uint const region_idx = hr->hrm_index();
1040       size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size();
1041       uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words);
1042 
1043       // "Distributing" zero words means that we only note end of marking for these
1044       // regions.
1045       assert(marked_words == 0 || obj_size_in_words == marked_words,
1046              "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT,
1047              obj_size_in_words, marked_words);
1048 
1049       for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) {
1050         HeapRegion* const r = _g1h->region_at(i);
1051         size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words);
1052       
1053         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)", 
1054                                words_to_add, i, r->get_type_str());
1055         add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize);
1056         marked_words -= words_to_add;
1057       }
1058       assert(marked_words == 0,
1059              SIZE_FORMAT " words left after distributing space across %u regions",
1060              marked_words, num_regions_in_humongous);
1061     }
1062 
1063     void update_marked_bytes(HeapRegion* hr) {
1064       uint const region_idx = hr->hrm_index();
1065       size_t const marked_words = _cm->liveness(region_idx);
1066       // The marking attributes the object's size completely to the humongous starts
1067       // region. We need to distribute this value across the entire set of regions a
1068       // humongous object spans.
1069       if (hr->is_humongous()) {
1070         assert(hr->is_starts_humongous() || marked_words == 0,
1071                "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)",
1072                marked_words, region_idx, hr->get_type_str());
1073         if (hr->is_starts_humongous()) {
1074           distribute_marked_bytes(hr, marked_words);
1075         }
1076       } else {
1077         log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str());
1078         add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize);
1079       }
1080     }
1081 
1082     void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) {
1083       hr->add_to_marked_bytes(marked_bytes);
1084       _cl->do_heap_region(hr);
1085       hr->note_end_of_marking();
1086     }
1087 
1088   public:
1089     G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) :
1090       _g1h(g1h), _cm(cm), _num_regions_selected_for_rebuild(0), _cl(cl) { }
1091 
1092     virtual bool do_heap_region(HeapRegion* r) {
1093       update_remset_before_rebuild(r);
1094       update_marked_bytes(r);
1095 
1096       return false;
1097     }
1098 
1099     uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; }
1100   };
1101 
1102 public:
1103   G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) :
1104     AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"),
1105     _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { }
1106 
1107   virtual void work(uint worker_id) {
1108     G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl);
1109     _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id);
1110     Atomic::add(update_cl.num_selected_for_rebuild(), &_total_selected_for_rebuild);
1111   }
1112 
1113   uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; }
1114 
1115   // Number of regions for which roughly one thread should be spawned for this work.
1116   static const uint RegionsPerThread = 384; 
1117 };
1118 
1119 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure {
1120   G1CollectedHeap* _g1h;
1121 public:
1122   G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { }
1123 
1124   virtual bool do_heap_region(HeapRegion* r) {
1125     _g1h->g1_policy()->remset_tracker()->update_after_rebuild(r);
1126     return false;
1127   }
1128 };
1129 
1130 void G1ConcurrentMark::remark() {
1131   assert_at_safepoint_on_vm_thread();
1132 
1133   // If a full collection has happened, we should not continue. However we might
1134   // have ended up here as the Remark VM operation has been scheduled already.
1135   if (has_aborted()) {
1136     return;
1137   }
1138 
1139   G1Policy* g1p = _g1h->g1_policy();
1140   g1p->record_concurrent_mark_remark_start();
1141 
1142   double start = os::elapsedTime();
1143 
1144   verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before");
1145 
1146   {
1147     GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm);
1148     finalize_marking();
1149   }
1150 
1151   double mark_work_end = os::elapsedTime();
1152 
1153   bool const mark_finished = !has_overflown();
1154   if (mark_finished) {
1155     weak_refs_work(false /* clear_all_soft_refs */);
1156 
1157     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1158     // We're done with marking.
1159     // This is the end of the marking cycle, we're expected all
1160     // threads to have SATB queues with active set to true.
1161     satb_mq_set.set_active_all_threads(false, /* new active value */
1162                                        true /* expected_active */);
1163 
1164     {
1165       GCTraceTime(Debug, gc, phases) trace("Flush Task Caches", _gc_timer_cm);
1166       flush_all_task_caches();
1167     }
1168 
1169     // Install newly created mark bitmap as "prev".
1170     swap_mark_bitmaps();
1171     {
1172       GCTraceTime(Debug, gc, phases) trace("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm);
1173 
1174       uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) /
1175                                        G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread;
1176       uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity);
1177 
1178       G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers);
1179       log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions());
1180       _g1h->workers()->run_task(&cl, num_workers);
1181 
1182       log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u",
1183                                       _g1h->num_regions(), cl.total_selected_for_rebuild());
1184     }
1185     {
1186       GCTraceTime(Debug, gc, phases) trace("Reclaim Empty Regions", _gc_timer_cm);
1187       reclaim_empty_regions();
1188     }
1189 
1190     // Clean out dead classes
1191     if (ClassUnloadingWithConcurrentMark) {
1192       GCTraceTime(Debug, gc, phases) trace("Purge Metaspace", _gc_timer_cm);
1193       ClassLoaderDataGraph::purge();
1194     }
1195 
1196     compute_new_sizes();
1197 
1198     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after");
1199 
1200     assert(!restart_for_overflow(), "sanity");
1201     // Completely reset the marking state since marking completed
1202     reset_at_marking_complete();
1203   } else {
1204     // We overflowed.  Restart concurrent marking.
1205     _restart_for_overflow = true;
1206 
1207     verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow");
1208 
1209     // Clear the marking state because we will be restarting
1210     // marking due to overflowing the global mark stack.
1211     reset_marking_for_restart();
1212   }
1213 
1214   {
1215     GCTraceTime(Debug, gc, phases) trace("Report Object Count", _gc_timer_cm);
1216     report_object_count(mark_finished);
1217   }
1218 
1219   // Statistics
1220   double now = os::elapsedTime();
1221   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1222   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1223   _remark_times.add((now - start) * 1000.0);
1224 
1225   g1p->record_concurrent_mark_remark_end();
1226 }
1227 
1228 class G1ReclaimEmptyRegionsTask : public AbstractGangTask {
1229   // Per-region work during the Cleanup pause.
1230   class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure {
1231     G1CollectedHeap* _g1h;
1232     size_t _freed_bytes;
1233     FreeRegionList* _local_cleanup_list;
1234     uint _old_regions_removed;
1235     uint _humongous_regions_removed;
1236     HRRSCleanupTask* _hrrs_cleanup_task;
1237 
1238   public:
1239     G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h,
1240                                  FreeRegionList* local_cleanup_list,
1241                                  HRRSCleanupTask* hrrs_cleanup_task) :
1242       _g1h(g1h),
1243       _freed_bytes(0),
1244       _local_cleanup_list(local_cleanup_list),
1245       _old_regions_removed(0),
1246       _humongous_regions_removed(0),
1247       _hrrs_cleanup_task(hrrs_cleanup_task) { }
1248 
1249     size_t freed_bytes() { return _freed_bytes; }
1250     const uint old_regions_removed() { return _old_regions_removed; }
1251     const uint humongous_regions_removed() { return _humongous_regions_removed; }
1252 
1253     bool do_heap_region(HeapRegion *hr) {
1254       if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1255         _freed_bytes += hr->used();
1256         hr->set_containing_set(NULL);
1257         if (hr->is_humongous()) {
1258           _humongous_regions_removed++;
1259           _g1h->free_humongous_region(hr, _local_cleanup_list);
1260         } else {
1261           _old_regions_removed++;
1262           _g1h->free_region(hr, _local_cleanup_list, false /* skip_remset */, false /* skip_hcc */, true /* locked */);
1263         }
1264         hr->clear_cardtable();
1265         _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index());
1266         log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom()));
1267       } else {
1268         hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1269       }
1270 
1271       return false;
1272     }
1273   };
1274 
1275   G1CollectedHeap* _g1h;
1276   FreeRegionList* _cleanup_list;
1277   HeapRegionClaimer _hrclaimer;
1278 
1279 public:
1280   G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1281     AbstractGangTask("G1 Cleanup"),
1282     _g1h(g1h),
1283     _cleanup_list(cleanup_list),
1284     _hrclaimer(n_workers) {
1285 
1286     HeapRegionRemSet::reset_for_cleanup_tasks();
1287   }
1288 
1289   void work(uint worker_id) {
1290     FreeRegionList local_cleanup_list("Local Cleanup List");
1291     HRRSCleanupTask hrrs_cleanup_task;
1292     G1ReclaimEmptyRegionsClosure cl(_g1h,
1293                                     &local_cleanup_list,
1294                                     &hrrs_cleanup_task);
1295     _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id);
1296     assert(cl.is_complete(), "Shouldn't have aborted!");
1297 
1298     // Now update the old/humongous region sets
1299     _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed());
1300     {
1301       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1302       _g1h->decrement_summary_bytes(cl.freed_bytes());
1303 
1304       _cleanup_list->add_ordered(&local_cleanup_list);
1305       assert(local_cleanup_list.is_empty(), "post-condition");
1306 
1307       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1308     }
1309   }
1310 };
1311 
1312 void G1ConcurrentMark::reclaim_empty_regions() {
1313   WorkGang* workers = _g1h->workers();
1314   FreeRegionList empty_regions_list("Empty Regions After Mark List");
1315 
1316   G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers());
1317   workers->run_task(&cl);
1318 
1319   if (!empty_regions_list.is_empty()) {
1320     log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length());
1321     // Now print the empty regions list.
1322     G1HRPrinter* hrp = _g1h->hr_printer();
1323     if (hrp->is_active()) {
1324       FreeRegionListIterator iter(&empty_regions_list);
1325       while (iter.more_available()) {
1326         HeapRegion* hr = iter.get_next();
1327         hrp->cleanup(hr);
1328       }
1329     }
1330     // And actually make them available.
1331     _g1h->prepend_to_freelist(&empty_regions_list);
1332   }
1333 }
1334 
1335 void G1ConcurrentMark::compute_new_sizes() {
1336   MetaspaceGC::compute_new_size();
1337 
1338   // Cleanup will have freed any regions completely full of garbage.
1339   // Update the soft reference policy with the new heap occupancy.
1340   Universe::update_heap_info_at_gc();
1341 
1342   // We reclaimed old regions so we should calculate the sizes to make
1343   // sure we update the old gen/space data.
1344   _g1h->g1mm()->update_sizes();
1345 }
1346 
1347 void G1ConcurrentMark::cleanup() {
1348   assert_at_safepoint_on_vm_thread();
1349 
1350   // If a full collection has happened, we shouldn't do this.
1351   if (has_aborted()) {
1352     return;
1353   }
1354 
1355   G1Policy* g1p = _g1h->g1_policy();
1356   g1p->record_concurrent_mark_cleanup_start();
1357 
1358   double start = os::elapsedTime();
1359 
1360   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before");
1361 
1362   {
1363     GCTraceTime(Debug, gc, phases) trace("Update Remembered Set Tracking After Rebuild", _gc_timer_cm);
1364     G1UpdateRemSetTrackingAfterRebuild cl(_g1h);
1365     _g1h->heap_region_iterate(&cl);
1366   }
1367 
1368   if (log_is_enabled(Trace, gc, liveness)) {
1369     G1PrintRegionLivenessInfoClosure cl("Post-Cleanup");
1370     _g1h->heap_region_iterate(&cl);
1371   }
1372 
1373   verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after");
1374 
1375   // We need to make this be a "collection" so any collection pause that
1376   // races with it goes around and waits for Cleanup to finish.
1377   _g1h->increment_total_collections();
1378 
1379   // Local statistics
1380   double recent_cleanup_time = (os::elapsedTime() - start);
1381   _total_cleanup_time += recent_cleanup_time;
1382   _cleanup_times.add(recent_cleanup_time);
1383 
1384   {
1385     GCTraceTime(Debug, gc, phases) trace("Finalize Concurrent Mark Cleanup", _gc_timer_cm);
1386     _g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1387   }
1388 }
1389 
1390 // Supporting Object and Oop closures for reference discovery
1391 // and processing in during marking
1392 
1393 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1394   HeapWord* addr = (HeapWord*)obj;
1395   return addr != NULL &&
1396          (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_ill(obj));
1397 }
1398 
1399 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1400 // Uses the G1CMTask associated with a worker thread (for serial reference
1401 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1402 // trace referent objects.
1403 //
1404 // Using the G1CMTask and embedded local queues avoids having the worker
1405 // threads operating on the global mark stack. This reduces the risk
1406 // of overflowing the stack - which we would rather avoid at this late
1407 // state. Also using the tasks' local queues removes the potential
1408 // of the workers interfering with each other that could occur if
1409 // operating on the global stack.
1410 
1411 class G1CMKeepAliveAndDrainClosure : public OopClosure {
1412   G1ConcurrentMark* _cm;
1413   G1CMTask*         _task;
1414   uint              _ref_counter_limit;
1415   uint              _ref_counter;
1416   bool              _is_serial;
1417 public:
1418   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1419     _cm(cm), _task(task), _is_serial(is_serial),
1420     _ref_counter_limit(G1RefProcDrainInterval) {
1421     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1422     _ref_counter = _ref_counter_limit;
1423   }
1424 
1425   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1426   virtual void do_oop(      oop* p) { do_oop_work(p); }
1427 
1428   template <class T> void do_oop_work(T* p) {
1429     if (_cm->has_overflown()) {
1430       return;
1431     }
1432     if (!_task->deal_with_reference(p)) {
1433       // We did not add anything to the mark bitmap (or mark stack), so there is
1434       // no point trying to drain it.
1435       return;
1436     }
1437     _ref_counter--;
1438 
1439     if (_ref_counter == 0) {
1440       // We have dealt with _ref_counter_limit references, pushing them
1441       // and objects reachable from them on to the local stack (and
1442       // possibly the global stack). Call G1CMTask::do_marking_step() to
1443       // process these entries.
1444       //
1445       // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1446       // there's nothing more to do (i.e. we're done with the entries that
1447       // were pushed as a result of the G1CMTask::deal_with_reference() calls
1448       // above) or we overflow.
1449       //
1450       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1451       // flag while there may still be some work to do. (See the comment at
1452       // the beginning of G1CMTask::do_marking_step() for those conditions -
1453       // one of which is reaching the specified time target.) It is only
1454       // when G1CMTask::do_marking_step() returns without setting the
1455       // has_aborted() flag that the marking step has completed.
1456       do {
1457         double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1458         _task->do_marking_step(mark_step_duration_ms,
1459                                false      /* do_termination */,
1460                                _is_serial);
1461       } while (_task->has_aborted() && !_cm->has_overflown());
1462       _ref_counter = _ref_counter_limit;
1463     }
1464   }
1465 };
1466 
1467 // 'Drain' oop closure used by both serial and parallel reference processing.
1468 // Uses the G1CMTask associated with a given worker thread (for serial
1469 // reference processing the G1CMtask for worker 0 is used). Calls the
1470 // do_marking_step routine, with an unbelievably large timeout value,
1471 // to drain the marking data structures of the remaining entries
1472 // added by the 'keep alive' oop closure above.
1473 
1474 class G1CMDrainMarkingStackClosure : public VoidClosure {
1475   G1ConcurrentMark* _cm;
1476   G1CMTask*         _task;
1477   bool              _is_serial;
1478  public:
1479   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1480     _cm(cm), _task(task), _is_serial(is_serial) {
1481     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1482   }
1483 
1484   void do_void() {
1485     do {
1486       // We call G1CMTask::do_marking_step() to completely drain the local
1487       // and global marking stacks of entries pushed by the 'keep alive'
1488       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1489       //
1490       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1491       // if there's nothing more to do (i.e. we've completely drained the
1492       // entries that were pushed as a a result of applying the 'keep alive'
1493       // closure to the entries on the discovered ref lists) or we overflow
1494       // the global marking stack.
1495       //
1496       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1497       // flag while there may still be some work to do. (See the comment at
1498       // the beginning of G1CMTask::do_marking_step() for those conditions -
1499       // one of which is reaching the specified time target.) It is only
1500       // when G1CMTask::do_marking_step() returns without setting the
1501       // has_aborted() flag that the marking step has completed.
1502 
1503       _task->do_marking_step(1000000000.0 /* something very large */,
1504                              true         /* do_termination */,
1505                              _is_serial);
1506     } while (_task->has_aborted() && !_cm->has_overflown());
1507   }
1508 };
1509 
1510 // Implementation of AbstractRefProcTaskExecutor for parallel
1511 // reference processing at the end of G1 concurrent marking
1512 
1513 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor {
1514 private:
1515   G1CollectedHeap*  _g1h;
1516   G1ConcurrentMark* _cm;
1517   WorkGang*         _workers;
1518   uint              _active_workers;
1519 
1520 public:
1521   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1522                           G1ConcurrentMark* cm,
1523                           WorkGang* workers,
1524                           uint n_workers) :
1525     _g1h(g1h), _cm(cm),
1526     _workers(workers), _active_workers(n_workers) { }
1527 
1528   // Executes the given task using concurrent marking worker threads.
1529   virtual void execute(ProcessTask& task);
1530   virtual void execute(EnqueueTask& task);
1531 };
1532 
1533 class G1CMRefProcTaskProxy : public AbstractGangTask {
1534   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1535   ProcessTask&      _proc_task;
1536   G1CollectedHeap*  _g1h;
1537   G1ConcurrentMark* _cm;
1538 
1539 public:
1540   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1541                        G1CollectedHeap* g1h,
1542                        G1ConcurrentMark* cm) :
1543     AbstractGangTask("Process reference objects in parallel"),
1544     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1545     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1546     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1547   }
1548 
1549   virtual void work(uint worker_id) {
1550     ResourceMark rm;
1551     HandleMark hm;
1552     G1CMTask* task = _cm->task(worker_id);
1553     G1CMIsAliveClosure g1_is_alive(_g1h);
1554     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1555     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1556 
1557     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1558   }
1559 };
1560 
1561 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1562   assert(_workers != NULL, "Need parallel worker threads.");
1563   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1564 
1565   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1566 
1567   // We need to reset the concurrency level before each
1568   // proxy task execution, so that the termination protocol
1569   // and overflow handling in G1CMTask::do_marking_step() knows
1570   // how many workers to wait for.
1571   _cm->set_concurrency(_active_workers);
1572   _workers->run_task(&proc_task_proxy);
1573 }
1574 
1575 class G1CMRefEnqueueTaskProxy : public AbstractGangTask {
1576   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1577   EnqueueTask& _enq_task;
1578 
1579 public:
1580   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1581     AbstractGangTask("Enqueue reference objects in parallel"),
1582     _enq_task(enq_task) { }
1583 
1584   virtual void work(uint worker_id) {
1585     _enq_task.work(worker_id);
1586   }
1587 };
1588 
1589 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1590   assert(_workers != NULL, "Need parallel worker threads.");
1591   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1592 
1593   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1594 
1595   // Not strictly necessary but...
1596   //
1597   // We need to reset the concurrency level before each
1598   // proxy task execution, so that the termination protocol
1599   // and overflow handling in G1CMTask::do_marking_step() knows
1600   // how many workers to wait for.
1601   _cm->set_concurrency(_active_workers);
1602   _workers->run_task(&enq_task_proxy);
1603 }
1604 
1605 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1606   ResourceMark rm;
1607   HandleMark   hm;
1608 
1609   // Is alive closure.
1610   G1CMIsAliveClosure g1_is_alive(_g1h);
1611 
1612   // Inner scope to exclude the cleaning of the string and symbol
1613   // tables from the displayed time.
1614   {
1615     GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm);
1616 
1617     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1618 
1619     // See the comment in G1CollectedHeap::ref_processing_init()
1620     // about how reference processing currently works in G1.
1621 
1622     // Set the soft reference policy
1623     rp->setup_policy(clear_all_soft_refs);
1624     assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1625 
1626     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1627     // in serial reference processing. Note these closures are also
1628     // used for serially processing (by the the current thread) the
1629     // JNI references during parallel reference processing.
1630     //
1631     // These closures do not need to synchronize with the worker
1632     // threads involved in parallel reference processing as these
1633     // instances are executed serially by the current thread (e.g.
1634     // reference processing is not multi-threaded and is thus
1635     // performed by the current thread instead of a gang worker).
1636     //
1637     // The gang tasks involved in parallel reference processing create
1638     // their own instances of these closures, which do their own
1639     // synchronization among themselves.
1640     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1641     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1642 
1643     // We need at least one active thread. If reference processing
1644     // is not multi-threaded we use the current (VMThread) thread,
1645     // otherwise we use the work gang from the G1CollectedHeap and
1646     // we utilize all the worker threads we can.
1647     bool processing_is_mt = rp->processing_is_mt();
1648     uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U);
1649     active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U);
1650 
1651     // Parallel processing task executor.
1652     G1CMRefProcTaskExecutor par_task_executor(_g1h, this,
1653                                               _g1h->workers(), active_workers);
1654     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1655 
1656     // Set the concurrency level. The phase was already set prior to
1657     // executing the remark task.
1658     set_concurrency(active_workers);
1659 
1660     // Set the degree of MT processing here.  If the discovery was done MT,
1661     // the number of threads involved during discovery could differ from
1662     // the number of active workers.  This is OK as long as the discovered
1663     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1664     rp->set_active_mt_degree(active_workers);
1665 
1666     ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
1667 
1668     // Process the weak references.
1669     const ReferenceProcessorStats& stats =
1670         rp->process_discovered_references(&g1_is_alive,
1671                                           &g1_keep_alive,
1672                                           &g1_drain_mark_stack,
1673                                           executor,
1674                                           &pt);
1675     _gc_tracer_cm->report_gc_reference_stats(stats);
1676     pt.print_all_references();
1677 
1678     // The do_oop work routines of the keep_alive and drain_marking_stack
1679     // oop closures will set the has_overflown flag if we overflow the
1680     // global marking stack.
1681 
1682     assert(has_overflown() || _global_mark_stack.is_empty(),
1683            "Mark stack should be empty (unless it has overflown)");
1684 
1685     assert(rp->num_q() == active_workers, "why not");
1686 
1687     rp->enqueue_discovered_references(executor, &pt);
1688 
1689     rp->verify_no_references_recorded();
1690 
1691     pt.print_enqueue_phase();
1692 
1693     assert(!rp->discovery_enabled(), "Post condition");
1694   }
1695 
1696   assert(has_overflown() || _global_mark_stack.is_empty(),
1697          "Mark stack should be empty (unless it has overflown)");
1698 
1699   {
1700     GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1701     WeakProcessor::weak_oops_do(&g1_is_alive, &do_nothing_cl);
1702   }
1703 
1704   if (has_overflown()) {
1705     // We can not trust g1_is_alive if the marking stack overflowed
1706     return;
1707   }
1708 
1709   assert(_global_mark_stack.is_empty(), "Marking should have completed");
1710 
1711   // Unload Klasses, String, Symbols, Code Cache, etc.
1712   if (ClassUnloadingWithConcurrentMark) {
1713     GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1714     bool purged_classes = SystemDictionary::do_unloading(&g1_is_alive, _gc_timer_cm, false /* Defer cleaning */);
1715     _g1h->complete_cleaning(&g1_is_alive, purged_classes);
1716   } else {
1717     GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1718     // No need to clean string table and symbol table as they are treated as strong roots when
1719     // class unloading is disabled.
1720     _g1h->partial_cleaning(&g1_is_alive, false, false, G1StringDedup::is_enabled());
1721   }
1722 }
1723 
1724 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
1725 // the prev bitmap determining liveness.
1726 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {
1727   G1CollectedHeap* _g1h;
1728 public:
1729   G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { }
1730 
1731   bool do_object_b(oop obj) {
1732     HeapWord* addr = (HeapWord*)obj;
1733     return addr != NULL &&
1734            (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_dead(obj));
1735   }
1736 };
1737 
1738 void G1ConcurrentMark::report_object_count(bool mark_completed) {
1739   // Depending on the completion of the marking liveness needs to be determined
1740   // using either the next or prev bitmap.
1741   if (mark_completed) {
1742     G1ObjectCountIsAliveClosure is_alive(_g1h);
1743     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1744   } else {
1745     G1CMIsAliveClosure is_alive(_g1h);
1746     _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1747   }
1748 }
1749 
1750 
1751 void G1ConcurrentMark::swap_mark_bitmaps() {
1752   G1CMBitMap* temp = _prev_mark_bitmap;
1753   _prev_mark_bitmap = _next_mark_bitmap;
1754   _next_mark_bitmap = temp;
1755   _g1h->collector_state()->set_clearing_next_bitmap(true);
1756 }
1757 
1758 // Closure for marking entries in SATB buffers.
1759 class G1CMSATBBufferClosure : public SATBBufferClosure {
1760 private:
1761   G1CMTask* _task;
1762   G1CollectedHeap* _g1h;
1763 
1764   // This is very similar to G1CMTask::deal_with_reference, but with
1765   // more relaxed requirements for the argument, so this must be more
1766   // circumspect about treating the argument as an object.
1767   void do_entry(void* entry) const {
1768     _task->increment_refs_reached();
1769     oop const obj = static_cast<oop>(entry);
1770     _task->make_reference_grey(obj);
1771   }
1772 
1773 public:
1774   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1775     : _task(task), _g1h(g1h) { }
1776 
1777   virtual void do_buffer(void** buffer, size_t size) {
1778     for (size_t i = 0; i < size; ++i) {
1779       do_entry(buffer[i]);
1780     }
1781   }
1782 };
1783 
1784 class G1RemarkThreadsClosure : public ThreadClosure {
1785   G1CMSATBBufferClosure _cm_satb_cl;
1786   G1CMOopClosure _cm_cl;
1787   MarkingCodeBlobClosure _code_cl;
1788   int _thread_parity;
1789 
1790  public:
1791   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1792     _cm_satb_cl(task, g1h),
1793     _cm_cl(g1h, task),
1794     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1795     _thread_parity(Threads::thread_claim_parity()) {}
1796 
1797   void do_thread(Thread* thread) {
1798     if (thread->is_Java_thread()) {
1799       if (thread->claim_oops_do(true, _thread_parity)) {
1800         JavaThread* jt = (JavaThread*)thread;
1801 
1802         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1803         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1804         // * Alive if on the stack of an executing method
1805         // * Weakly reachable otherwise
1806         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1807         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1808         jt->nmethods_do(&_code_cl);
1809 
1810         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
1811       }
1812     } else if (thread->is_VM_thread()) {
1813       if (thread->claim_oops_do(true, _thread_parity)) {
1814         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1815       }
1816     }
1817   }
1818 };
1819 
1820 class G1CMRemarkTask : public AbstractGangTask {
1821   G1ConcurrentMark* _cm;
1822 public:
1823   void work(uint worker_id) {
1824     G1CMTask* task = _cm->task(worker_id);
1825     task->record_start_time();
1826     {
1827       ResourceMark rm;
1828       HandleMark hm;
1829 
1830       G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1831       Threads::threads_do(&threads_f);
1832     }
1833 
1834     do {
1835       task->do_marking_step(1000000000.0 /* something very large */,
1836                             true         /* do_termination       */,
1837                             false        /* is_serial            */);
1838     } while (task->has_aborted() && !_cm->has_overflown());
1839     // If we overflow, then we do not want to restart. We instead
1840     // want to abort remark and do concurrent marking again.
1841     task->record_end_time();
1842   }
1843 
1844   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1845     AbstractGangTask("Par Remark"), _cm(cm) {
1846     _cm->terminator()->reset_for_reuse(active_workers);
1847   }
1848 };
1849 
1850 void G1ConcurrentMark::finalize_marking() {
1851   ResourceMark rm;
1852   HandleMark   hm;
1853 
1854   _g1h->ensure_parsability(false);
1855 
1856   // this is remark, so we'll use up all active threads
1857   uint active_workers = _g1h->workers()->active_workers();
1858   set_concurrency_and_phase(active_workers, false /* concurrent */);
1859   // Leave _parallel_marking_threads at it's
1860   // value originally calculated in the G1ConcurrentMark
1861   // constructor and pass values of the active workers
1862   // through the gang in the task.
1863 
1864   {
1865     StrongRootsScope srs(active_workers);
1866 
1867     G1CMRemarkTask remarkTask(this, active_workers);
1868     // We will start all available threads, even if we decide that the
1869     // active_workers will be fewer. The extra ones will just bail out
1870     // immediately.
1871     _g1h->workers()->run_task(&remarkTask);
1872   }
1873 
1874   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1875   guarantee(has_overflown() ||
1876             satb_mq_set.completed_buffers_num() == 0,
1877             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1878             BOOL_TO_STR(has_overflown()),
1879             satb_mq_set.completed_buffers_num());
1880 
1881   print_stats();
1882 }
1883 
1884 void G1ConcurrentMark::flush_all_task_caches() {
1885   size_t hits = 0;
1886   size_t misses = 0;
1887   for (uint i = 0; i < _max_num_tasks; i++) {
1888     Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache();
1889     hits += stats.first;
1890     misses += stats.second;
1891   }
1892   size_t sum = hits + misses;
1893   log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf",
1894                        hits, misses, percent_of(hits, sum));
1895 }
1896 
1897 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1898   _prev_mark_bitmap->clear_range(mr);
1899 }
1900 
1901 HeapRegion*
1902 G1ConcurrentMark::claim_region(uint worker_id) {
1903   // "checkpoint" the finger
1904   HeapWord* finger = _finger;
1905 
1906   while (finger < _heap.end()) {
1907     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1908 
1909     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1910     // Make sure that the reads below do not float before loading curr_region.
1911     OrderAccess::loadload();
1912     // Above heap_region_containing may return NULL as we always scan claim
1913     // until the end of the heap. In this case, just jump to the next region.
1914     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1915 
1916     // Is the gap between reading the finger and doing the CAS too long?
1917     HeapWord* res = Atomic::cmpxchg(end, &_finger, finger);
1918     if (res == finger && curr_region != NULL) {
1919       // we succeeded
1920       HeapWord*   bottom        = curr_region->bottom();
1921       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1922 
1923       // notice that _finger == end cannot be guaranteed here since,
1924       // someone else might have moved the finger even further
1925       assert(_finger >= end, "the finger should have moved forward");
1926 
1927       if (limit > bottom) {
1928         return curr_region;
1929       } else {
1930         assert(limit == bottom,
1931                "the region limit should be at bottom");
1932         // we return NULL and the caller should try calling
1933         // claim_region() again.
1934         return NULL;
1935       }
1936     } else {
1937       assert(_finger > finger, "the finger should have moved forward");
1938       // read it again
1939       finger = _finger;
1940     }
1941   }
1942 
1943   return NULL;
1944 }
1945 
1946 #ifndef PRODUCT
1947 class VerifyNoCSetOops {
1948   G1CollectedHeap* _g1h;
1949   const char* _phase;
1950   int _info;
1951 
1952 public:
1953   VerifyNoCSetOops(const char* phase, int info = -1) :
1954     _g1h(G1CollectedHeap::heap()),
1955     _phase(phase),
1956     _info(info)
1957   { }
1958 
1959   void operator()(G1TaskQueueEntry task_entry) const {
1960     if (task_entry.is_array_slice()) {
1961       guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1962       return;
1963     }
1964     guarantee(oopDesc::is_oop(task_entry.obj()),
1965               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1966               p2i(task_entry.obj()), _phase, _info);
1967     guarantee(!_g1h->is_in_cset(task_entry.obj()),
1968               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1969               p2i(task_entry.obj()), _phase, _info);
1970   }
1971 };
1972 
1973 void G1ConcurrentMark::verify_no_cset_oops() {
1974   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1975   if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) {
1976     return;
1977   }
1978 
1979   // Verify entries on the global mark stack
1980   _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1981 
1982   // Verify entries on the task queues
1983   for (uint i = 0; i < _max_num_tasks; ++i) {
1984     G1CMTaskQueue* queue = _task_queues->queue(i);
1985     queue->iterate(VerifyNoCSetOops("Queue", i));
1986   }
1987 
1988   // Verify the global finger
1989   HeapWord* global_finger = finger();
1990   if (global_finger != NULL && global_finger < _heap.end()) {
1991     // Since we always iterate over all regions, we might get a NULL HeapRegion
1992     // here.
1993     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1994     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1995               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1996               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1997   }
1998 
1999   // Verify the task fingers
2000   assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
2001   for (uint i = 0; i < _num_concurrent_workers; ++i) {
2002     G1CMTask* task = _tasks[i];
2003     HeapWord* task_finger = task->finger();
2004     if (task_finger != NULL && task_finger < _heap.end()) {
2005       // See above note on the global finger verification.
2006       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
2007       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
2008                 !task_hr->in_collection_set(),
2009                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
2010                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
2011     }
2012   }
2013 }
2014 #endif // PRODUCT
2015 
2016 void G1ConcurrentMark::rebuild_rem_set_concurrently() {
2017   _g1h->g1_rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset);
2018 }
2019 
2020 void G1ConcurrentMark::print_stats() {
2021   if (!log_is_enabled(Debug, gc, stats)) {
2022     return;
2023   }
2024   log_debug(gc, stats)("---------------------------------------------------------------------");
2025   for (size_t i = 0; i < _num_active_tasks; ++i) {
2026     _tasks[i]->print_stats();
2027     log_debug(gc, stats)("---------------------------------------------------------------------");
2028   }
2029 }
2030 
2031 void G1ConcurrentMark::concurrent_cycle_abort() {
2032   if (!cm_thread()->during_cycle() || _has_aborted) {
2033     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
2034     return;
2035   }
2036 
2037   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
2038   // concurrent bitmap clearing.
2039   {
2040     GCTraceTime(Debug, gc)("Clear Next Bitmap");
2041     clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
2042   }
2043   // Note we cannot clear the previous marking bitmap here
2044   // since VerifyDuringGC verifies the objects marked during
2045   // a full GC against the previous bitmap.
2046 
2047   // Empty mark stack
2048   reset_marking_for_restart();
2049   for (uint i = 0; i < _max_num_tasks; ++i) {
2050     _tasks[i]->clear_region_fields();
2051   }
2052   _first_overflow_barrier_sync.abort();
2053   _second_overflow_barrier_sync.abort();
2054   _has_aborted = true;
2055 
2056   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2057   satb_mq_set.abandon_partial_marking();
2058   // This can be called either during or outside marking, we'll read
2059   // the expected_active value from the SATB queue set.
2060   satb_mq_set.set_active_all_threads(
2061                                  false, /* new active value */
2062                                  satb_mq_set.is_active() /* expected_active */);
2063 }
2064 
2065 static void print_ms_time_info(const char* prefix, const char* name,
2066                                NumberSeq& ns) {
2067   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2068                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2069   if (ns.num() > 0) {
2070     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2071                            prefix, ns.sd(), ns.maximum());
2072   }
2073 }
2074 
2075 void G1ConcurrentMark::print_summary_info() {
2076   Log(gc, marking) log;
2077   if (!log.is_trace()) {
2078     return;
2079   }
2080 
2081   log.trace(" Concurrent marking:");
2082   print_ms_time_info("  ", "init marks", _init_times);
2083   print_ms_time_info("  ", "remarks", _remark_times);
2084   {
2085     print_ms_time_info("     ", "final marks", _remark_mark_times);
2086     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2087 
2088   }
2089   print_ms_time_info("  ", "cleanups", _cleanup_times);
2090   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2091             _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2092   log.trace("  Total stop_world time = %8.2f s.",
2093             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2094   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2095             cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2096 }
2097 
2098 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2099   _concurrent_workers->print_worker_threads_on(st);
2100 }
2101 
2102 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2103   _concurrent_workers->threads_do(tc);
2104 }
2105 
2106 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2107   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2108                p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2109   _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2110   _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2111 }
2112 
2113 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2114   ReferenceProcessor* result = g1h->ref_processor_cm();
2115   assert(result != NULL, "CM reference processor should not be NULL");
2116   return result;
2117 }
2118 
2119 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2120                                G1CMTask* task)
2121   : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2122     _g1h(g1h), _task(task)
2123 { }
2124 
2125 void G1CMTask::setup_for_region(HeapRegion* hr) {
2126   assert(hr != NULL,
2127         "claim_region() should have filtered out NULL regions");
2128   _curr_region  = hr;
2129   _finger       = hr->bottom();
2130   update_region_limit();
2131 }
2132 
2133 void G1CMTask::update_region_limit() {
2134   HeapRegion* hr            = _curr_region;
2135   HeapWord* bottom          = hr->bottom();
2136   HeapWord* limit           = hr->next_top_at_mark_start();
2137 
2138   if (limit == bottom) {
2139     // The region was collected underneath our feet.
2140     // We set the finger to bottom to ensure that the bitmap
2141     // iteration that will follow this will not do anything.
2142     // (this is not a condition that holds when we set the region up,
2143     // as the region is not supposed to be empty in the first place)
2144     _finger = bottom;
2145   } else if (limit >= _region_limit) {
2146     assert(limit >= _finger, "peace of mind");
2147   } else {
2148     assert(limit < _region_limit, "only way to get here");
2149     // This can happen under some pretty unusual circumstances.  An
2150     // evacuation pause empties the region underneath our feet (NTAMS
2151     // at bottom). We then do some allocation in the region (NTAMS
2152     // stays at bottom), followed by the region being used as a GC
2153     // alloc region (NTAMS will move to top() and the objects
2154     // originally below it will be grayed). All objects now marked in
2155     // the region are explicitly grayed, if below the global finger,
2156     // and we do not need in fact to scan anything else. So, we simply
2157     // set _finger to be limit to ensure that the bitmap iteration
2158     // doesn't do anything.
2159     _finger = limit;
2160   }
2161 
2162   _region_limit = limit;
2163 }
2164 
2165 void G1CMTask::giveup_current_region() {
2166   assert(_curr_region != NULL, "invariant");
2167   clear_region_fields();
2168 }
2169 
2170 void G1CMTask::clear_region_fields() {
2171   // Values for these three fields that indicate that we're not
2172   // holding on to a region.
2173   _curr_region   = NULL;
2174   _finger        = NULL;
2175   _region_limit  = NULL;
2176 }
2177 
2178 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2179   if (cm_oop_closure == NULL) {
2180     assert(_cm_oop_closure != NULL, "invariant");
2181   } else {
2182     assert(_cm_oop_closure == NULL, "invariant");
2183   }
2184   _cm_oop_closure = cm_oop_closure;
2185 }
2186 
2187 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2188   guarantee(next_mark_bitmap != NULL, "invariant");
2189   _next_mark_bitmap              = next_mark_bitmap;
2190   clear_region_fields();
2191 
2192   _calls                         = 0;
2193   _elapsed_time_ms               = 0.0;
2194   _termination_time_ms           = 0.0;
2195   _termination_start_time_ms     = 0.0;
2196 
2197   _mark_stats_cache.reset();
2198 }
2199 
2200 bool G1CMTask::should_exit_termination() {
2201   regular_clock_call();
2202   // This is called when we are in the termination protocol. We should
2203   // quit if, for some reason, this task wants to abort or the global
2204   // stack is not empty (this means that we can get work from it).
2205   return !_cm->mark_stack_empty() || has_aborted();
2206 }
2207 
2208 void G1CMTask::reached_limit() {
2209   assert(_words_scanned >= _words_scanned_limit ||
2210          _refs_reached >= _refs_reached_limit ,
2211          "shouldn't have been called otherwise");
2212   regular_clock_call();
2213 }
2214 
2215 void G1CMTask::regular_clock_call() {
2216   if (has_aborted()) {
2217     return;
2218   }
2219 
2220   // First, we need to recalculate the words scanned and refs reached
2221   // limits for the next clock call.
2222   recalculate_limits();
2223 
2224   // During the regular clock call we do the following
2225 
2226   // (1) If an overflow has been flagged, then we abort.
2227   if (_cm->has_overflown()) {
2228     set_has_aborted();
2229     return;
2230   }
2231 
2232   // If we are not concurrent (i.e. we're doing remark) we don't need
2233   // to check anything else. The other steps are only needed during
2234   // the concurrent marking phase.
2235   if (!_cm->concurrent()) {
2236     return;
2237   }
2238 
2239   // (2) If marking has been aborted for Full GC, then we also abort.
2240   if (_cm->has_aborted()) {
2241     set_has_aborted();
2242     return;
2243   }
2244 
2245   double curr_time_ms = os::elapsedVTime() * 1000.0;
2246 
2247   // (4) We check whether we should yield. If we have to, then we abort.
2248   if (SuspendibleThreadSet::should_yield()) {
2249     // We should yield. To do this we abort the task. The caller is
2250     // responsible for yielding.
2251     set_has_aborted();
2252     return;
2253   }
2254 
2255   // (5) We check whether we've reached our time quota. If we have,
2256   // then we abort.
2257   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2258   if (elapsed_time_ms > _time_target_ms) {
2259     set_has_aborted();
2260     _has_timed_out = true;
2261     return;
2262   }
2263 
2264   // (6) Finally, we check whether there are enough completed STAB
2265   // buffers available for processing. If there are, we abort.
2266   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2267   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2268     // we do need to process SATB buffers, we'll abort and restart
2269     // the marking task to do so
2270     set_has_aborted();
2271     return;
2272   }
2273 }
2274 
2275 void G1CMTask::recalculate_limits() {
2276   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2277   _words_scanned_limit      = _real_words_scanned_limit;
2278 
2279   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2280   _refs_reached_limit       = _real_refs_reached_limit;
2281 }
2282 
2283 void G1CMTask::decrease_limits() {
2284   // This is called when we believe that we're going to do an infrequent
2285   // operation which will increase the per byte scanned cost (i.e. move
2286   // entries to/from the global stack). It basically tries to decrease the
2287   // scanning limit so that the clock is called earlier.
2288 
2289   _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2290   _refs_reached_limit  = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2291 }
2292 
2293 void G1CMTask::move_entries_to_global_stack() {
2294   // Local array where we'll store the entries that will be popped
2295   // from the local queue.
2296   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2297 
2298   size_t n = 0;
2299   G1TaskQueueEntry task_entry;
2300   while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2301     buffer[n] = task_entry;
2302     ++n;
2303   }
2304   if (n < G1CMMarkStack::EntriesPerChunk) {
2305     buffer[n] = G1TaskQueueEntry();
2306   }
2307 
2308   if (n > 0) {
2309     if (!_cm->mark_stack_push(buffer)) {
2310       set_has_aborted();
2311     }
2312   }
2313 
2314   // This operation was quite expensive, so decrease the limits.
2315   decrease_limits();
2316 }
2317 
2318 bool G1CMTask::get_entries_from_global_stack() {
2319   // Local array where we'll store the entries that will be popped
2320   // from the global stack.
2321   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2322 
2323   if (!_cm->mark_stack_pop(buffer)) {
2324     return false;
2325   }
2326 
2327   // We did actually pop at least one entry.
2328   for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2329     G1TaskQueueEntry task_entry = buffer[i];
2330     if (task_entry.is_null()) {
2331       break;
2332     }
2333     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()));
2334     bool success = _task_queue->push(task_entry);
2335     // We only call this when the local queue is empty or under a
2336     // given target limit. So, we do not expect this push to fail.
2337     assert(success, "invariant");
2338   }
2339 
2340   // This operation was quite expensive, so decrease the limits
2341   decrease_limits();
2342   return true;
2343 }
2344 
2345 void G1CMTask::drain_local_queue(bool partially) {
2346   if (has_aborted()) {
2347     return;
2348   }
2349 
2350   // Decide what the target size is, depending whether we're going to
2351   // drain it partially (so that other tasks can steal if they run out
2352   // of things to do) or totally (at the very end).
2353   size_t target_size;
2354   if (partially) {
2355     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2356   } else {
2357     target_size = 0;
2358   }
2359 
2360   if (_task_queue->size() > target_size) {
2361     G1TaskQueueEntry entry;
2362     bool ret = _task_queue->pop_local(entry);
2363     while (ret) {
2364       scan_task_entry(entry);
2365       if (_task_queue->size() <= target_size || has_aborted()) {
2366         ret = false;
2367       } else {
2368         ret = _task_queue->pop_local(entry);
2369       }
2370     }
2371   }
2372 }
2373 
2374 void G1CMTask::drain_global_stack(bool partially) {
2375   if (has_aborted()) {
2376     return;
2377   }
2378 
2379   // We have a policy to drain the local queue before we attempt to
2380   // drain the global stack.
2381   assert(partially || _task_queue->size() == 0, "invariant");
2382 
2383   // Decide what the target size is, depending whether we're going to
2384   // drain it partially (so that other tasks can steal if they run out
2385   // of things to do) or totally (at the very end).
2386   // Notice that when draining the global mark stack partially, due to the racyness
2387   // of the mark stack size update we might in fact drop below the target. But,
2388   // this is not a problem.
2389   // In case of total draining, we simply process until the global mark stack is
2390   // totally empty, disregarding the size counter.
2391   if (partially) {
2392     size_t const target_size = _cm->partial_mark_stack_size_target();
2393     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2394       if (get_entries_from_global_stack()) {
2395         drain_local_queue(partially);
2396       }
2397     }
2398   } else {
2399     while (!has_aborted() && get_entries_from_global_stack()) {
2400       drain_local_queue(partially);
2401     }
2402   }
2403 }
2404 
2405 // SATB Queue has several assumptions on whether to call the par or
2406 // non-par versions of the methods. this is why some of the code is
2407 // replicated. We should really get rid of the single-threaded version
2408 // of the code to simplify things.
2409 void G1CMTask::drain_satb_buffers() {
2410   if (has_aborted()) {
2411     return;
2412   }
2413 
2414   // We set this so that the regular clock knows that we're in the
2415   // middle of draining buffers and doesn't set the abort flag when it
2416   // notices that SATB buffers are available for draining. It'd be
2417   // very counter productive if it did that. :-)
2418   _draining_satb_buffers = true;
2419 
2420   G1CMSATBBufferClosure satb_cl(this, _g1h);
2421   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2422 
2423   // This keeps claiming and applying the closure to completed buffers
2424   // until we run out of buffers or we need to abort.
2425   while (!has_aborted() &&
2426          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2427     regular_clock_call();
2428   }
2429 
2430   _draining_satb_buffers = false;
2431 
2432   assert(has_aborted() ||
2433          _cm->concurrent() ||
2434          satb_mq_set.completed_buffers_num() == 0, "invariant");
2435 
2436   // again, this was a potentially expensive operation, decrease the
2437   // limits to get the regular clock call early
2438   decrease_limits();
2439 }
2440 
2441 void G1CMTask::clear_mark_stats_cache(uint region_idx) {
2442   _mark_stats_cache.reset(region_idx);
2443 }
2444 
2445 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() {
2446   return _mark_stats_cache.evict_all();
2447 }
2448 
2449 void G1CMTask::print_stats() {
2450   log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls);
2451   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2452                        _elapsed_time_ms, _termination_time_ms);
2453   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms",
2454                        _step_times_ms.num(),
2455                        _step_times_ms.avg(),
2456                        _step_times_ms.sd(),
2457                        _step_times_ms.maximum(),
2458                        _step_times_ms.sum());
2459   size_t const hits = _mark_stats_cache.hits();
2460   size_t const misses = _mark_stats_cache.misses();
2461   log_debug(gc, stats)("  Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f",
2462                        hits, misses, percent_of(hits, hits + misses));
2463 }
2464 
2465 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry) {
2466   return _task_queues->steal(worker_id, hash_seed, task_entry);
2467 }
2468 
2469 /*****************************************************************************
2470 
2471     The do_marking_step(time_target_ms, ...) method is the building
2472     block of the parallel marking framework. It can be called in parallel
2473     with other invocations of do_marking_step() on different tasks
2474     (but only one per task, obviously) and concurrently with the
2475     mutator threads, or during remark, hence it eliminates the need
2476     for two versions of the code. When called during remark, it will
2477     pick up from where the task left off during the concurrent marking
2478     phase. Interestingly, tasks are also claimable during evacuation
2479     pauses too, since do_marking_step() ensures that it aborts before
2480     it needs to yield.
2481 
2482     The data structures that it uses to do marking work are the
2483     following:
2484 
2485       (1) Marking Bitmap. If there are gray objects that appear only
2486       on the bitmap (this happens either when dealing with an overflow
2487       or when the initial marking phase has simply marked the roots
2488       and didn't push them on the stack), then tasks claim heap
2489       regions whose bitmap they then scan to find gray objects. A
2490       global finger indicates where the end of the last claimed region
2491       is. A local finger indicates how far into the region a task has
2492       scanned. The two fingers are used to determine how to gray an
2493       object (i.e. whether simply marking it is OK, as it will be
2494       visited by a task in the future, or whether it needs to be also
2495       pushed on a stack).
2496 
2497       (2) Local Queue. The local queue of the task which is accessed
2498       reasonably efficiently by the task. Other tasks can steal from
2499       it when they run out of work. Throughout the marking phase, a
2500       task attempts to keep its local queue short but not totally
2501       empty, so that entries are available for stealing by other
2502       tasks. Only when there is no more work, a task will totally
2503       drain its local queue.
2504 
2505       (3) Global Mark Stack. This handles local queue overflow. During
2506       marking only sets of entries are moved between it and the local
2507       queues, as access to it requires a mutex and more fine-grain
2508       interaction with it which might cause contention. If it
2509       overflows, then the marking phase should restart and iterate
2510       over the bitmap to identify gray objects. Throughout the marking
2511       phase, tasks attempt to keep the global mark stack at a small
2512       length but not totally empty, so that entries are available for
2513       popping by other tasks. Only when there is no more work, tasks
2514       will totally drain the global mark stack.
2515 
2516       (4) SATB Buffer Queue. This is where completed SATB buffers are
2517       made available. Buffers are regularly removed from this queue
2518       and scanned for roots, so that the queue doesn't get too
2519       long. During remark, all completed buffers are processed, as
2520       well as the filled in parts of any uncompleted buffers.
2521 
2522     The do_marking_step() method tries to abort when the time target
2523     has been reached. There are a few other cases when the
2524     do_marking_step() method also aborts:
2525 
2526       (1) When the marking phase has been aborted (after a Full GC).
2527 
2528       (2) When a global overflow (on the global stack) has been
2529       triggered. Before the task aborts, it will actually sync up with
2530       the other tasks to ensure that all the marking data structures
2531       (local queues, stacks, fingers etc.)  are re-initialized so that
2532       when do_marking_step() completes, the marking phase can
2533       immediately restart.
2534 
2535       (3) When enough completed SATB buffers are available. The
2536       do_marking_step() method only tries to drain SATB buffers right
2537       at the beginning. So, if enough buffers are available, the
2538       marking step aborts and the SATB buffers are processed at
2539       the beginning of the next invocation.
2540 
2541       (4) To yield. when we have to yield then we abort and yield
2542       right at the end of do_marking_step(). This saves us from a lot
2543       of hassle as, by yielding we might allow a Full GC. If this
2544       happens then objects will be compacted underneath our feet, the
2545       heap might shrink, etc. We save checking for this by just
2546       aborting and doing the yield right at the end.
2547 
2548     From the above it follows that the do_marking_step() method should
2549     be called in a loop (or, otherwise, regularly) until it completes.
2550 
2551     If a marking step completes without its has_aborted() flag being
2552     true, it means it has completed the current marking phase (and
2553     also all other marking tasks have done so and have all synced up).
2554 
2555     A method called regular_clock_call() is invoked "regularly" (in
2556     sub ms intervals) throughout marking. It is this clock method that
2557     checks all the abort conditions which were mentioned above and
2558     decides when the task should abort. A work-based scheme is used to
2559     trigger this clock method: when the number of object words the
2560     marking phase has scanned or the number of references the marking
2561     phase has visited reach a given limit. Additional invocations to
2562     the method clock have been planted in a few other strategic places
2563     too. The initial reason for the clock method was to avoid calling
2564     vtime too regularly, as it is quite expensive. So, once it was in
2565     place, it was natural to piggy-back all the other conditions on it
2566     too and not constantly check them throughout the code.
2567 
2568     If do_termination is true then do_marking_step will enter its
2569     termination protocol.
2570 
2571     The value of is_serial must be true when do_marking_step is being
2572     called serially (i.e. by the VMThread) and do_marking_step should
2573     skip any synchronization in the termination and overflow code.
2574     Examples include the serial remark code and the serial reference
2575     processing closures.
2576 
2577     The value of is_serial must be false when do_marking_step is
2578     being called by any of the worker threads in a work gang.
2579     Examples include the concurrent marking code (CMMarkingTask),
2580     the MT remark code, and the MT reference processing closures.
2581 
2582  *****************************************************************************/
2583 
2584 void G1CMTask::do_marking_step(double time_target_ms,
2585                                bool do_termination,
2586                                bool is_serial) {
2587   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2588 
2589   _start_time_ms = os::elapsedVTime() * 1000.0;
2590 
2591   // If do_stealing is true then do_marking_step will attempt to
2592   // steal work from the other G1CMTasks. It only makes sense to
2593   // enable stealing when the termination protocol is enabled
2594   // and do_marking_step() is not being called serially.
2595   bool do_stealing = do_termination && !is_serial;
2596 
2597   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2598   _time_target_ms = time_target_ms - diff_prediction_ms;
2599 
2600   // set up the variables that are used in the work-based scheme to
2601   // call the regular clock method
2602   _words_scanned = 0;
2603   _refs_reached  = 0;
2604   recalculate_limits();
2605 
2606   // clear all flags
2607   clear_has_aborted();
2608   _has_timed_out = false;
2609   _draining_satb_buffers = false;
2610 
2611   ++_calls;
2612 
2613   // Set up the bitmap and oop closures. Anything that uses them is
2614   // eventually called from this method, so it is OK to allocate these
2615   // statically.
2616   G1CMBitMapClosure bitmap_closure(this, _cm);
2617   G1CMOopClosure cm_oop_closure(_g1h, this);
2618   set_cm_oop_closure(&cm_oop_closure);
2619 
2620   if (_cm->has_overflown()) {
2621     // This can happen if the mark stack overflows during a GC pause
2622     // and this task, after a yield point, restarts. We have to abort
2623     // as we need to get into the overflow protocol which happens
2624     // right at the end of this task.
2625     set_has_aborted();
2626   }
2627 
2628   // First drain any available SATB buffers. After this, we will not
2629   // look at SATB buffers before the next invocation of this method.
2630   // If enough completed SATB buffers are queued up, the regular clock
2631   // will abort this task so that it restarts.
2632   drain_satb_buffers();
2633   // ...then partially drain the local queue and the global stack
2634   drain_local_queue(true);
2635   drain_global_stack(true);
2636 
2637   do {
2638     if (!has_aborted() && _curr_region != NULL) {
2639       // This means that we're already holding on to a region.
2640       assert(_finger != NULL, "if region is not NULL, then the finger "
2641              "should not be NULL either");
2642 
2643       // We might have restarted this task after an evacuation pause
2644       // which might have evacuated the region we're holding on to
2645       // underneath our feet. Let's read its limit again to make sure
2646       // that we do not iterate over a region of the heap that
2647       // contains garbage (update_region_limit() will also move
2648       // _finger to the start of the region if it is found empty).
2649       update_region_limit();
2650       // We will start from _finger not from the start of the region,
2651       // as we might be restarting this task after aborting half-way
2652       // through scanning this region. In this case, _finger points to
2653       // the address where we last found a marked object. If this is a
2654       // fresh region, _finger points to start().
2655       MemRegion mr = MemRegion(_finger, _region_limit);
2656 
2657       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2658              "humongous regions should go around loop once only");
2659 
2660       // Some special cases:
2661       // If the memory region is empty, we can just give up the region.
2662       // If the current region is humongous then we only need to check
2663       // the bitmap for the bit associated with the start of the object,
2664       // scan the object if it's live, and give up the region.
2665       // Otherwise, let's iterate over the bitmap of the part of the region
2666       // that is left.
2667       // If the iteration is successful, give up the region.
2668       if (mr.is_empty()) {
2669         giveup_current_region();
2670         regular_clock_call();
2671       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2672         if (_next_mark_bitmap->is_marked(mr.start())) {
2673           // The object is marked - apply the closure
2674           bitmap_closure.do_addr(mr.start());
2675         }
2676         // Even if this task aborted while scanning the humongous object
2677         // we can (and should) give up the current region.
2678         giveup_current_region();
2679         regular_clock_call();
2680       } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2681         giveup_current_region();
2682         regular_clock_call();
2683       } else {
2684         assert(has_aborted(), "currently the only way to do so");
2685         // The only way to abort the bitmap iteration is to return
2686         // false from the do_bit() method. However, inside the
2687         // do_bit() method we move the _finger to point to the
2688         // object currently being looked at. So, if we bail out, we
2689         // have definitely set _finger to something non-null.
2690         assert(_finger != NULL, "invariant");
2691 
2692         // Region iteration was actually aborted. So now _finger
2693         // points to the address of the object we last scanned. If we
2694         // leave it there, when we restart this task, we will rescan
2695         // the object. It is easy to avoid this. We move the finger by
2696         // enough to point to the next possible object header.
2697         assert(_finger < _region_limit, "invariant");
2698         HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2699         // Check if bitmap iteration was aborted while scanning the last object
2700         if (new_finger >= _region_limit) {
2701           giveup_current_region();
2702         } else {
2703           move_finger_to(new_finger);
2704         }
2705       }
2706     }
2707     // At this point we have either completed iterating over the
2708     // region we were holding on to, or we have aborted.
2709 
2710     // We then partially drain the local queue and the global stack.
2711     // (Do we really need this?)
2712     drain_local_queue(true);
2713     drain_global_stack(true);
2714 
2715     // Read the note on the claim_region() method on why it might
2716     // return NULL with potentially more regions available for
2717     // claiming and why we have to check out_of_regions() to determine
2718     // whether we're done or not.
2719     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2720       // We are going to try to claim a new region. We should have
2721       // given up on the previous one.
2722       // Separated the asserts so that we know which one fires.
2723       assert(_curr_region  == NULL, "invariant");
2724       assert(_finger       == NULL, "invariant");
2725       assert(_region_limit == NULL, "invariant");
2726       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2727       if (claimed_region != NULL) {
2728         // Yes, we managed to claim one
2729         setup_for_region(claimed_region);
2730         assert(_curr_region == claimed_region, "invariant");
2731       }
2732       // It is important to call the regular clock here. It might take
2733       // a while to claim a region if, for example, we hit a large
2734       // block of empty regions. So we need to call the regular clock
2735       // method once round the loop to make sure it's called
2736       // frequently enough.
2737       regular_clock_call();
2738     }
2739 
2740     if (!has_aborted() && _curr_region == NULL) {
2741       assert(_cm->out_of_regions(),
2742              "at this point we should be out of regions");
2743     }
2744   } while ( _curr_region != NULL && !has_aborted());
2745 
2746   if (!has_aborted()) {
2747     // We cannot check whether the global stack is empty, since other
2748     // tasks might be pushing objects to it concurrently.
2749     assert(_cm->out_of_regions(),
2750            "at this point we should be out of regions");
2751     // Try to reduce the number of available SATB buffers so that
2752     // remark has less work to do.
2753     drain_satb_buffers();
2754   }
2755 
2756   // Since we've done everything else, we can now totally drain the
2757   // local queue and global stack.
2758   drain_local_queue(false);
2759   drain_global_stack(false);
2760 
2761   // Attempt at work stealing from other task's queues.
2762   if (do_stealing && !has_aborted()) {
2763     // We have not aborted. This means that we have finished all that
2764     // we could. Let's try to do some stealing...
2765 
2766     // We cannot check whether the global stack is empty, since other
2767     // tasks might be pushing objects to it concurrently.
2768     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2769            "only way to reach here");
2770     while (!has_aborted()) {
2771       G1TaskQueueEntry entry;
2772       if (_cm->try_stealing(_worker_id, &_hash_seed, entry)) {
2773         scan_task_entry(entry);
2774 
2775         // And since we're towards the end, let's totally drain the
2776         // local queue and global stack.
2777         drain_local_queue(false);
2778         drain_global_stack(false);
2779       } else {
2780         break;
2781       }
2782     }
2783   }
2784 
2785   // We still haven't aborted. Now, let's try to get into the
2786   // termination protocol.
2787   if (do_termination && !has_aborted()) {
2788     // We cannot check whether the global stack is empty, since other
2789     // tasks might be concurrently pushing objects on it.
2790     // Separated the asserts so that we know which one fires.
2791     assert(_cm->out_of_regions(), "only way to reach here");
2792     assert(_task_queue->size() == 0, "only way to reach here");
2793     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2794 
2795     // The G1CMTask class also extends the TerminatorTerminator class,
2796     // hence its should_exit_termination() method will also decide
2797     // whether to exit the termination protocol or not.
2798     bool finished = (is_serial ||
2799                      _cm->terminator()->offer_termination(this));
2800     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2801     _termination_time_ms +=
2802       termination_end_time_ms - _termination_start_time_ms;
2803 
2804     if (finished) {
2805       // We're all done.
2806 
2807       // We can now guarantee that the global stack is empty, since
2808       // all other tasks have finished. We separated the guarantees so
2809       // that, if a condition is false, we can immediately find out
2810       // which one.
2811       guarantee(_cm->out_of_regions(), "only way to reach here");
2812       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2813       guarantee(_task_queue->size() == 0, "only way to reach here");
2814       guarantee(!_cm->has_overflown(), "only way to reach here");
2815     } else {
2816       // Apparently there's more work to do. Let's abort this task. It
2817       // will restart it and we can hopefully find more things to do.
2818       set_has_aborted();
2819     }
2820   }
2821 
2822   // Mainly for debugging purposes to make sure that a pointer to the
2823   // closure which was statically allocated in this frame doesn't
2824   // escape it by accident.
2825   set_cm_oop_closure(NULL);
2826   double end_time_ms = os::elapsedVTime() * 1000.0;
2827   double elapsed_time_ms = end_time_ms - _start_time_ms;
2828   // Update the step history.
2829   _step_times_ms.add(elapsed_time_ms);
2830 
2831   if (has_aborted()) {
2832     // The task was aborted for some reason.
2833     if (_has_timed_out) {
2834       double diff_ms = elapsed_time_ms - _time_target_ms;
2835       // Keep statistics of how well we did with respect to hitting
2836       // our target only if we actually timed out (if we aborted for
2837       // other reasons, then the results might get skewed).
2838       _marking_step_diffs_ms.add(diff_ms);
2839     }
2840 
2841     if (_cm->has_overflown()) {
2842       // This is the interesting one. We aborted because a global
2843       // overflow was raised. This means we have to restart the
2844       // marking phase and start iterating over regions. However, in
2845       // order to do this we have to make sure that all tasks stop
2846       // what they are doing and re-initialize in a safe manner. We
2847       // will achieve this with the use of two barrier sync points.
2848 
2849       if (!is_serial) {
2850         // We only need to enter the sync barrier if being called
2851         // from a parallel context
2852         _cm->enter_first_sync_barrier(_worker_id);
2853 
2854         // When we exit this sync barrier we know that all tasks have
2855         // stopped doing marking work. So, it's now safe to
2856         // re-initialize our data structures.
2857       }
2858 
2859       clear_region_fields();
2860       flush_mark_stats_cache();
2861 
2862       if (!is_serial) {
2863         // If we're executing the concurrent phase of marking, reset the marking
2864         // state; otherwise the marking state is reset after reference processing,
2865         // during the remark pause.
2866         // If we reset here as a result of an overflow during the remark we will
2867         // see assertion failures from any subsequent set_concurrency_and_phase()
2868         // calls.
2869         if (_cm->concurrent() && _worker_id == 0) {
2870           // Worker 0 is responsible for clearing the global data structures because
2871           // of an overflow. During STW we should not clear the overflow flag (in
2872           // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit
2873           // method to abort the pause and restart concurrent marking.
2874           _cm->reset_marking_for_restart();
2875 
2876           log_info(gc, marking)("Concurrent Mark reset for overflow");
2877         }
2878 
2879         // ...and enter the second barrier.
2880         _cm->enter_second_sync_barrier(_worker_id);
2881       }
2882       // At this point, if we're during the concurrent phase of
2883       // marking, everything has been re-initialized and we're
2884       // ready to restart.
2885     }
2886   }
2887 }
2888 
2889 G1CMTask::G1CMTask(uint worker_id,
2890                    G1ConcurrentMark* cm,
2891                    G1CMTaskQueue* task_queue,
2892                    G1RegionMarkStats* mark_stats,
2893                    uint max_regions) :
2894   _objArray_processor(this),
2895   _worker_id(worker_id),
2896   _g1h(G1CollectedHeap::heap()),
2897   _cm(cm),
2898   _next_mark_bitmap(NULL),
2899   _task_queue(task_queue),
2900   _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize),
2901   _calls(0),
2902   _time_target_ms(0.0),
2903   _start_time_ms(0.0),
2904   _cm_oop_closure(NULL),
2905   _curr_region(NULL),
2906   _finger(NULL),
2907   _region_limit(NULL),
2908   _words_scanned(0),
2909   _words_scanned_limit(0),
2910   _real_words_scanned_limit(0),
2911   _refs_reached(0),
2912   _refs_reached_limit(0),
2913   _real_refs_reached_limit(0),
2914   _hash_seed(17),
2915   _has_aborted(false),
2916   _has_timed_out(false),
2917   _draining_satb_buffers(false),
2918   _step_times_ms(),
2919   _elapsed_time_ms(0.0),
2920   _termination_time_ms(0.0),
2921   _termination_start_time_ms(0.0),
2922   _marking_step_diffs_ms()
2923 {
2924   guarantee(task_queue != NULL, "invariant");
2925 
2926   _marking_step_diffs_ms.add(0.5);
2927 }
2928 
2929 // These are formatting macros that are used below to ensure
2930 // consistent formatting. The *_H_* versions are used to format the
2931 // header for a particular value and they should be kept consistent
2932 // with the corresponding macro. Also note that most of the macros add
2933 // the necessary white space (as a prefix) which makes them a bit
2934 // easier to compose.
2935 
2936 // All the output lines are prefixed with this string to be able to
2937 // identify them easily in a large log file.
2938 #define G1PPRL_LINE_PREFIX            "###"
2939 
2940 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2941 #ifdef _LP64
2942 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2943 #else // _LP64
2944 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2945 #endif // _LP64
2946 
2947 // For per-region info
2948 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2949 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2950 #define G1PPRL_STATE_FORMAT           "   %-5s"
2951 #define G1PPRL_STATE_H_FORMAT         "   %5s"
2952 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2953 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2954 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2955 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2956 
2957 // For summary info
2958 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2959 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2960 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2961 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2962 
2963 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2964   _total_used_bytes(0), _total_capacity_bytes(0),
2965   _total_prev_live_bytes(0), _total_next_live_bytes(0),
2966   _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2967 {
2968   if (!log_is_enabled(Trace, gc, liveness)) {
2969     return;
2970   }
2971 
2972   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2973   MemRegion g1_reserved = g1h->g1_reserved();
2974   double now = os::elapsedTime();
2975 
2976   // Print the header of the output.
2977   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2978   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2979                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2980                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2981                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2982                           HeapRegion::GrainBytes);
2983   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
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                           "type", "address-range",
2995                           "used", "prev-live", "next-live", "gc-eff",
2996                           "remset", "state", "code-roots");
2997   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2998                           G1PPRL_TYPE_H_FORMAT
2999                           G1PPRL_ADDR_BASE_H_FORMAT
3000                           G1PPRL_BYTE_H_FORMAT
3001                           G1PPRL_BYTE_H_FORMAT
3002                           G1PPRL_BYTE_H_FORMAT
3003                           G1PPRL_DOUBLE_H_FORMAT
3004                           G1PPRL_BYTE_H_FORMAT
3005                           G1PPRL_STATE_H_FORMAT
3006                           G1PPRL_BYTE_H_FORMAT,
3007                           "", "",
3008                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
3009                           "(bytes)", "", "(bytes)");
3010 }
3011 
3012 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) {
3013   if (!log_is_enabled(Trace, gc, liveness)) {
3014     return false;
3015   }
3016 
3017   const char* type       = r->get_type_str();
3018   HeapWord* bottom       = r->bottom();
3019   HeapWord* end          = r->end();
3020   size_t capacity_bytes  = r->capacity();
3021   size_t used_bytes      = r->used();
3022   size_t prev_live_bytes = r->live_bytes();
3023   size_t next_live_bytes = r->next_live_bytes();
3024   double gc_eff          = r->gc_efficiency();
3025   size_t remset_bytes    = r->rem_set()->mem_size();
3026   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
3027   const char* remset_type = r->rem_set()->get_short_state_str();
3028 
3029   _total_used_bytes      += used_bytes;
3030   _total_capacity_bytes  += capacity_bytes;
3031   _total_prev_live_bytes += prev_live_bytes;
3032   _total_next_live_bytes += next_live_bytes;
3033   _total_remset_bytes    += remset_bytes;
3034   _total_strong_code_roots_bytes += strong_code_roots_bytes;
3035 
3036   // Print a line for this particular region.
3037   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3038                           G1PPRL_TYPE_FORMAT
3039                           G1PPRL_ADDR_BASE_FORMAT
3040                           G1PPRL_BYTE_FORMAT
3041                           G1PPRL_BYTE_FORMAT
3042                           G1PPRL_BYTE_FORMAT
3043                           G1PPRL_DOUBLE_FORMAT
3044                           G1PPRL_BYTE_FORMAT
3045                           G1PPRL_STATE_FORMAT
3046                           G1PPRL_BYTE_FORMAT,
3047                           type, p2i(bottom), p2i(end),
3048                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
3049                           remset_bytes, remset_type, strong_code_roots_bytes);
3050 
3051   return false;
3052 }
3053 
3054 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
3055   if (!log_is_enabled(Trace, gc, liveness)) {
3056     return;
3057   }
3058 
3059   // add static memory usages to remembered set sizes
3060   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
3061   // Print the footer of the output.
3062   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
3063   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
3064                          " SUMMARY"
3065                          G1PPRL_SUM_MB_FORMAT("capacity")
3066                          G1PPRL_SUM_MB_PERC_FORMAT("used")
3067                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
3068                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
3069                          G1PPRL_SUM_MB_FORMAT("remset")
3070                          G1PPRL_SUM_MB_FORMAT("code-roots"),
3071                          bytes_to_mb(_total_capacity_bytes),
3072                          bytes_to_mb(_total_used_bytes),
3073                          percent_of(_total_used_bytes, _total_capacity_bytes),
3074                          bytes_to_mb(_total_prev_live_bytes),
3075                          percent_of(_total_prev_live_bytes, _total_capacity_bytes),
3076                          bytes_to_mb(_total_next_live_bytes),
3077                          percent_of(_total_next_live_bytes, _total_capacity_bytes),
3078                          bytes_to_mb(_total_remset_bytes),
3079                          bytes_to_mb(_total_strong_code_roots_bytes));
3080 }