1 /*
   2  * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "classfile/metadataOnStackMark.hpp"
  27 #include "classfile/symbolTable.hpp"
  28 #include "code/codeCache.hpp"
  29 #include "gc/g1/concurrentMarkThread.inline.hpp"
  30 #include "gc/g1/g1CollectedHeap.inline.hpp"
  31 #include "gc/g1/g1CollectorState.hpp"
  32 #include "gc/g1/g1ConcurrentMark.inline.hpp"
  33 #include "gc/g1/g1HeapVerifier.hpp"
  34 #include "gc/g1/g1OopClosures.inline.hpp"
  35 #include "gc/g1/g1CardLiveData.inline.hpp"
  36 #include "gc/g1/g1Policy.hpp"
  37 #include "gc/g1/g1StringDedup.hpp"
  38 #include "gc/g1/heapRegion.inline.hpp"
  39 #include "gc/g1/heapRegionRemSet.hpp"
  40 #include "gc/g1/heapRegionSet.inline.hpp"
  41 #include "gc/shared/gcId.hpp"
  42 #include "gc/shared/gcTimer.hpp"
  43 #include "gc/shared/gcTrace.hpp"
  44 #include "gc/shared/gcTraceTime.inline.hpp"
  45 #include "gc/shared/genOopClosures.inline.hpp"
  46 #include "gc/shared/referencePolicy.hpp"
  47 #include "gc/shared/strongRootsScope.hpp"
  48 #include "gc/shared/suspendibleThreadSet.hpp"
  49 #include "gc/shared/taskqueue.inline.hpp"
  50 #include "gc/shared/vmGCOperations.hpp"
  51 #include "gc/shared/weakProcessor.hpp"
  52 #include "logging/log.hpp"
  53 #include "memory/allocation.hpp"
  54 #include "memory/resourceArea.hpp"
  55 #include "oops/oop.inline.hpp"
  56 #include "runtime/atomic.hpp"
  57 #include "runtime/handles.inline.hpp"
  58 #include "runtime/java.hpp"
  59 #include "runtime/prefetch.inline.hpp"
  60 #include "services/memTracker.hpp"
  61 #include "utilities/align.hpp"
  62 #include "utilities/growableArray.hpp"
  63 
  64 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) {
  65   assert(addr < _cm->finger(), "invariant");
  66   assert(addr >= _task->finger(), "invariant");
  67 
  68   // We move that task's local finger along.
  69   _task->move_finger_to(addr);
  70 
  71   _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr)));
  72   // we only partially drain the local queue and global stack
  73   _task->drain_local_queue(true);
  74   _task->drain_global_stack(true);
  75 
  76   // if the has_aborted flag has been raised, we need to bail out of
  77   // the iteration
  78   return !_task->has_aborted();
  79 }
  80 
  81 G1CMMarkStack::G1CMMarkStack() :
  82   _max_chunk_capacity(0),
  83   _base(NULL),
  84   _chunk_capacity(0) {
  85   set_empty();
  86 }
  87 
  88 bool G1CMMarkStack::resize(size_t new_capacity) {
  89   assert(is_empty(), "Only resize when stack is empty.");
  90   assert(new_capacity <= _max_chunk_capacity,
  91          "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity);
  92 
  93   TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC);
  94 
  95   if (new_base == NULL) {
  96     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));
  97     return false;
  98   }
  99   // Release old mapping.
 100   if (_base != NULL) {
 101     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 102   }
 103 
 104   _base = new_base;
 105   _chunk_capacity = new_capacity;
 106   set_empty();
 107 
 108   return true;
 109 }
 110 
 111 size_t G1CMMarkStack::capacity_alignment() {
 112   return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry);
 113 }
 114 
 115 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) {
 116   guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized.");
 117 
 118   size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry);
 119 
 120   _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 121   size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar;
 122 
 123   guarantee(initial_chunk_capacity <= _max_chunk_capacity,
 124             "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT,
 125             _max_chunk_capacity,
 126             initial_chunk_capacity);
 127 
 128   log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT,
 129                 initial_chunk_capacity, _max_chunk_capacity);
 130 
 131   return resize(initial_chunk_capacity);
 132 }
 133 
 134 void G1CMMarkStack::expand() {
 135   if (_chunk_capacity == _max_chunk_capacity) {
 136     log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity);
 137     return;
 138   }
 139   size_t old_capacity = _chunk_capacity;
 140   // Double capacity if possible
 141   size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity);
 142 
 143   if (resize(new_capacity)) {
 144     log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 145                   old_capacity, new_capacity);
 146   } else {
 147     log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks",
 148                     old_capacity, new_capacity);
 149   }
 150 }
 151 
 152 G1CMMarkStack::~G1CMMarkStack() {
 153   if (_base != NULL) {
 154     MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity);
 155   }
 156 }
 157 
 158 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) {
 159   elem->next = *list;
 160   *list = elem;
 161 }
 162 
 163 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) {
 164   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 165   add_chunk_to_list(&_chunk_list, elem);
 166   _chunks_in_chunk_list++;
 167 }
 168 
 169 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) {
 170   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 171   add_chunk_to_list(&_free_list, elem);
 172 }
 173 
 174 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) {
 175   TaskQueueEntryChunk* result = *list;
 176   if (result != NULL) {
 177     *list = (*list)->next;
 178   }
 179   return result;
 180 }
 181 
 182 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() {
 183   MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag);
 184   TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list);
 185   if (result != NULL) {
 186     _chunks_in_chunk_list--;
 187   }
 188   return result;
 189 }
 190 
 191 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() {
 192   MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag);
 193   return remove_chunk_from_list(&_free_list);
 194 }
 195 
 196 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() {
 197   // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code.
 198   // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding
 199   // wraparound of _hwm.
 200   if (_hwm >= _chunk_capacity) {
 201     return NULL;
 202   }
 203 
 204   size_t cur_idx = Atomic::add(1u, &_hwm) - 1;
 205   if (cur_idx >= _chunk_capacity) {
 206     return NULL;
 207   }
 208 
 209   TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk;
 210   result->next = NULL;
 211   return result;
 212 }
 213 
 214 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) {
 215   // Get a new chunk.
 216   TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list();
 217 
 218   if (new_chunk == NULL) {
 219     // Did not get a chunk from the free list. Allocate from backing memory.
 220     new_chunk = allocate_new_chunk();
 221 
 222     if (new_chunk == NULL) {
 223       return false;
 224     }
 225   }
 226 
 227   Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 228 
 229   add_chunk_to_chunk_list(new_chunk);
 230 
 231   return true;
 232 }
 233 
 234 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) {
 235   TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list();
 236 
 237   if (cur == NULL) {
 238     return false;
 239   }
 240 
 241   Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry));
 242 
 243   add_chunk_to_free_list(cur);
 244   return true;
 245 }
 246 
 247 void G1CMMarkStack::set_empty() {
 248   _chunks_in_chunk_list = 0;
 249   _hwm = 0;
 250   _chunk_list = NULL;
 251   _free_list = NULL;
 252 }
 253 
 254 G1CMRootRegions::G1CMRootRegions() :
 255   _cm(NULL), _scan_in_progress(false),
 256   _should_abort(false), _claimed_survivor_index(0) { }
 257 
 258 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) {
 259   _survivors = survivors;
 260   _cm = cm;
 261 }
 262 
 263 void G1CMRootRegions::prepare_for_scan() {
 264   assert(!scan_in_progress(), "pre-condition");
 265 
 266   // Currently, only survivors can be root regions.
 267   _claimed_survivor_index = 0;
 268   _scan_in_progress = _survivors->regions()->is_nonempty();
 269   _should_abort = false;
 270 }
 271 
 272 HeapRegion* G1CMRootRegions::claim_next() {
 273   if (_should_abort) {
 274     // If someone has set the should_abort flag, we return NULL to
 275     // force the caller to bail out of their loop.
 276     return NULL;
 277   }
 278 
 279   // Currently, only survivors can be root regions.
 280   const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions();
 281 
 282   int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1;
 283   if (claimed_index < survivor_regions->length()) {
 284     return survivor_regions->at(claimed_index);
 285   }
 286   return NULL;
 287 }
 288 
 289 uint G1CMRootRegions::num_root_regions() const {
 290   return (uint)_survivors->regions()->length();
 291 }
 292 
 293 void G1CMRootRegions::notify_scan_done() {
 294   MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 295   _scan_in_progress = false;
 296   RootRegionScan_lock->notify_all();
 297 }
 298 
 299 void G1CMRootRegions::cancel_scan() {
 300   notify_scan_done();
 301 }
 302 
 303 void G1CMRootRegions::scan_finished() {
 304   assert(scan_in_progress(), "pre-condition");
 305 
 306   // Currently, only survivors can be root regions.
 307   if (!_should_abort) {
 308     assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index);
 309     assert((uint)_claimed_survivor_index >= _survivors->length(),
 310            "we should have claimed all survivors, claimed index = %u, length = %u",
 311            (uint)_claimed_survivor_index, _survivors->length());
 312   }
 313 
 314   notify_scan_done();
 315 }
 316 
 317 bool G1CMRootRegions::wait_until_scan_finished() {
 318   if (!scan_in_progress()) return false;
 319 
 320   {
 321     MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag);
 322     while (scan_in_progress()) {
 323       RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag);
 324     }
 325   }
 326   return true;
 327 }
 328 
 329 // Returns the maximum number of workers to be used in a concurrent
 330 // phase based on the number of GC workers being used in a STW
 331 // phase.
 332 static uint scale_concurrent_worker_threads(uint num_gc_workers) {
 333   return MAX2((num_gc_workers + 2) / 4, 1U);
 334 }
 335 
 336 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h,
 337                                    G1RegionToSpaceMapper* prev_bitmap_storage,
 338                                    G1RegionToSpaceMapper* next_bitmap_storage) :
 339   // _cm_thread set inside the constructor
 340   _g1h(g1h),
 341   _completed_initialization(false),
 342 
 343   _cleanup_list("Concurrent Mark Cleanup List"),
 344   _mark_bitmap_1(),
 345   _mark_bitmap_2(),
 346   _prev_mark_bitmap(&_mark_bitmap_1),
 347   _next_mark_bitmap(&_mark_bitmap_2),
 348 
 349   _heap_start(_g1h->reserved_region().start()),
 350   _heap_end(_g1h->reserved_region().end()),
 351 
 352   _root_regions(),
 353 
 354   _global_mark_stack(),
 355 
 356   // _finger set in set_non_marking_state
 357 
 358   _max_num_tasks(ParallelGCThreads),
 359   // _num_active_tasks set in set_non_marking_state()
 360   // _tasks set inside the constructor
 361 
 362   _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)),
 363   _terminator(ParallelTaskTerminator((int) _max_num_tasks, _task_queues)),
 364 
 365   _first_overflow_barrier_sync(),
 366   _second_overflow_barrier_sync(),
 367 
 368   _has_overflown(false),
 369   _concurrent(false),
 370   _has_aborted(false),
 371   _restart_for_overflow(false),
 372   _concurrent_marking_in_progress(false),
 373   _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()),
 374   _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()),
 375 
 376   // _verbose_level set below
 377 
 378   _init_times(),
 379   _remark_times(),
 380   _remark_mark_times(),
 381   _remark_weak_ref_times(),
 382   _cleanup_times(),
 383   _total_counting_time(0.0),
 384   _total_rs_scrub_time(0.0),
 385 
 386   _accum_task_vtime(NULL),
 387 
 388   _concurrent_workers(NULL),
 389   _num_concurrent_workers(0),
 390   _max_concurrent_workers(0)
 391 {
 392   _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage);
 393   _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage);
 394 
 395   // Create & start ConcurrentMark thread.
 396   _cm_thread = new ConcurrentMarkThread(this);
 397   if (_cm_thread->osthread() == NULL) {
 398     vm_shutdown_during_initialization("Could not create ConcurrentMarkThread");
 399   }
 400 
 401   assert(CGC_lock != NULL, "CGC_lock must be initialized");
 402 
 403   SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set();
 404   satb_qs.set_buffer_size(G1SATBBufferSize);
 405 
 406   _root_regions.init(_g1h->survivor(), this);
 407 
 408   if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) {
 409     // Calculate the number of concurrent worker threads by scaling
 410     // the number of parallel GC threads.
 411     uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads);
 412     FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num);
 413   }
 414 
 415   assert(ConcGCThreads > 0, "ConcGCThreads have been set.");
 416   if (ConcGCThreads > ParallelGCThreads) {
 417     log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).",
 418                     ConcGCThreads, ParallelGCThreads);
 419     return;
 420   }
 421 
 422   log_debug(gc)("ConcGCThreads: %u", ConcGCThreads);
 423   log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads);
 424 
 425   _num_concurrent_workers = ConcGCThreads;
 426   _max_concurrent_workers = _num_concurrent_workers;
 427 
 428   _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true);
 429   _concurrent_workers->initialize_workers();
 430 
 431   if (FLAG_IS_DEFAULT(MarkStackSize)) {
 432     size_t mark_stack_size =
 433       MIN2(MarkStackSizeMax,
 434           MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE)));
 435     // Verify that the calculated value for MarkStackSize is in range.
 436     // It would be nice to use the private utility routine from Arguments.
 437     if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) {
 438       log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): "
 439                       "must be between 1 and " SIZE_FORMAT,
 440                       mark_stack_size, MarkStackSizeMax);
 441       return;
 442     }
 443     FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size);
 444   } else {
 445     // Verify MarkStackSize is in range.
 446     if (FLAG_IS_CMDLINE(MarkStackSize)) {
 447       if (FLAG_IS_DEFAULT(MarkStackSizeMax)) {
 448         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 449           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): "
 450                           "must be between 1 and " SIZE_FORMAT,
 451                           MarkStackSize, MarkStackSizeMax);
 452           return;
 453         }
 454       } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) {
 455         if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) {
 456           log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")"
 457                           " or for MarkStackSizeMax (" SIZE_FORMAT ")",
 458                           MarkStackSize, MarkStackSizeMax);
 459           return;
 460         }
 461       }
 462     }
 463   }
 464 
 465   if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) {
 466     vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack.");
 467   }
 468 
 469   _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC);
 470   _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC);
 471 
 472   // so that the assertion in MarkingTaskQueue::task_queue doesn't fail
 473   _num_active_tasks = _max_num_tasks;
 474 
 475   for (uint i = 0; i < _max_num_tasks; ++i) {
 476     G1CMTaskQueue* task_queue = new G1CMTaskQueue();
 477     task_queue->initialize();
 478     _task_queues->register_queue(i, task_queue);
 479 
 480     _tasks[i] = new G1CMTask(i, this, task_queue);
 481 
 482     _accum_task_vtime[i] = 0.0;
 483   }
 484 
 485   set_non_marking_state();
 486   _completed_initialization = true;
 487 }
 488 
 489 void G1ConcurrentMark::reset() {
 490   // Starting values for these two. This should be called in a STW
 491   // phase.
 492   MemRegion reserved = _g1h->g1_reserved();
 493   _heap_start = reserved.start();
 494   _heap_end   = reserved.end();
 495 
 496   // Separated the asserts so that we know which one fires.
 497   assert(_heap_start != NULL, "heap bounds should look ok");
 498   assert(_heap_end != NULL, "heap bounds should look ok");
 499   assert(_heap_start < _heap_end, "heap bounds should look ok");
 500 
 501   // Reset all the marking data structures and any necessary flags
 502   reset_marking_state();
 503 
 504   // We reset all of them, since different phases will use
 505   // different number of active threads. So, it's easiest to have all
 506   // of them ready.
 507   for (uint i = 0; i < _max_num_tasks; ++i) {
 508     _tasks[i]->reset(_next_mark_bitmap);
 509   }
 510 
 511   // we need this to make sure that the flag is on during the evac
 512   // pause with initial mark piggy-backed
 513   set_concurrent_marking_in_progress();
 514 }
 515 
 516 
 517 void G1ConcurrentMark::reset_marking_state() {
 518   _global_mark_stack.set_empty();
 519 
 520   // Expand the marking stack, if we have to and if we can.
 521   if (has_overflown()) {
 522     _global_mark_stack.expand();
 523   }
 524 
 525   clear_has_overflown();
 526   _finger = _heap_start;
 527 
 528   for (uint i = 0; i < _max_num_tasks; ++i) {
 529     G1CMTaskQueue* queue = _task_queues->queue(i);
 530     queue->set_empty();
 531   }
 532 }
 533 
 534 void G1ConcurrentMark::set_concurrency(uint active_tasks) {
 535   assert(active_tasks <= _max_num_tasks, "we should not have more");
 536 
 537   _num_active_tasks = active_tasks;
 538   // Need to update the three data structures below according to the
 539   // number of active threads for this phase.
 540   _terminator   = ParallelTaskTerminator((int) active_tasks, _task_queues);
 541   _first_overflow_barrier_sync.set_n_workers((int) active_tasks);
 542   _second_overflow_barrier_sync.set_n_workers((int) active_tasks);
 543 }
 544 
 545 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) {
 546   set_concurrency(active_tasks);
 547 
 548   _concurrent = concurrent;
 549   // We propagate this to all tasks, not just the active ones.
 550   for (uint i = 0; i < _max_num_tasks; ++i) {
 551     _tasks[i]->set_concurrent(concurrent);
 552   }
 553 
 554   if (concurrent) {
 555     set_concurrent_marking_in_progress();
 556   } else {
 557     // We currently assume that the concurrent flag has been set to
 558     // false before we start remark. At this point we should also be
 559     // in a STW phase.
 560     assert(!concurrent_marking_in_progress(), "invariant");
 561     assert(out_of_regions(),
 562            "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT,
 563            p2i(_finger), p2i(_heap_end));
 564   }
 565 }
 566 
 567 void G1ConcurrentMark::set_non_marking_state() {
 568   // We set the global marking state to some default values when we're
 569   // not doing marking.
 570   reset_marking_state();
 571   _num_active_tasks = 0;
 572   clear_concurrent_marking_in_progress();
 573 }
 574 
 575 G1ConcurrentMark::~G1ConcurrentMark() {
 576   // The G1ConcurrentMark instance is never freed.
 577   ShouldNotReachHere();
 578 }
 579 
 580 class G1ClearBitMapTask : public AbstractGangTask {
 581 public:
 582   static size_t chunk_size() { return M; }
 583 
 584 private:
 585   // Heap region closure used for clearing the given mark bitmap.
 586   class G1ClearBitmapHRClosure : public HeapRegionClosure {
 587   private:
 588     G1CMBitMap* _bitmap;
 589     G1ConcurrentMark* _cm;
 590   public:
 591     G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) {
 592     }
 593 
 594     virtual bool doHeapRegion(HeapRegion* r) {
 595       size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize;
 596 
 597       HeapWord* cur = r->bottom();
 598       HeapWord* const end = r->end();
 599 
 600       while (cur < end) {
 601         MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end));
 602         _bitmap->clear_range(mr);
 603 
 604         cur += chunk_size_in_words;
 605 
 606         // Abort iteration if after yielding the marking has been aborted.
 607         if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) {
 608           return true;
 609         }
 610         // Repeat the asserts from before the start of the closure. We will do them
 611         // as asserts here to minimize their overhead on the product. However, we
 612         // will have them as guarantees at the beginning / end of the bitmap
 613         // clearing to get some checking in the product.
 614         assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant");
 615         assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant");
 616       }
 617       assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index());
 618 
 619       return false;
 620     }
 621   };
 622 
 623   G1ClearBitmapHRClosure _cl;
 624   HeapRegionClaimer _hr_claimer;
 625   bool _suspendible; // If the task is suspendible, workers must join the STS.
 626 
 627 public:
 628   G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) :
 629     AbstractGangTask("G1 Clear Bitmap"),
 630     _cl(bitmap, suspendible ? cm : NULL),
 631     _hr_claimer(n_workers),
 632     _suspendible(suspendible)
 633   { }
 634 
 635   void work(uint worker_id) {
 636     SuspendibleThreadSetJoiner sts_join(_suspendible);
 637     G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id);
 638   }
 639 
 640   bool is_complete() {
 641     return _cl.complete();
 642   }
 643 };
 644 
 645 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) {
 646   assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint.");
 647 
 648   size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor();
 649   size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size();
 650 
 651   uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers());
 652 
 653   G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield);
 654 
 655   log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks);
 656   workers->run_task(&cl, num_workers);
 657   guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding.");
 658 }
 659 
 660 void G1ConcurrentMark::cleanup_for_next_mark() {
 661   // Make sure that the concurrent mark thread looks to still be in
 662   // the current cycle.
 663   guarantee(cm_thread()->during_cycle(), "invariant");
 664 
 665   // We are finishing up the current cycle by clearing the next
 666   // marking bitmap and getting it ready for the next cycle. During
 667   // this time no other cycle can start. So, let's make sure that this
 668   // is the case.
 669   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 670 
 671   clear_bitmap(_next_mark_bitmap, _concurrent_workers, true);
 672 
 673   // Clear the live count data. If the marking has been aborted, the abort()
 674   // call already did that.
 675   if (!has_aborted()) {
 676     clear_live_data(_concurrent_workers);
 677     DEBUG_ONLY(verify_live_data_clear());
 678   }
 679 
 680   // Repeat the asserts from above.
 681   guarantee(cm_thread()->during_cycle(), "invariant");
 682   guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant");
 683 }
 684 
 685 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) {
 686   assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint.");
 687   clear_bitmap(_prev_mark_bitmap, workers, false);
 688 }
 689 
 690 class CheckBitmapClearHRClosure : public HeapRegionClosure {
 691   G1CMBitMap* _bitmap;
 692   bool _error;
 693  public:
 694   CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) {
 695   }
 696 
 697   virtual bool doHeapRegion(HeapRegion* r) {
 698     // This closure can be called concurrently to the mutator, so we must make sure
 699     // that the result of the getNextMarkedWordAddress() call is compared to the
 700     // value passed to it as limit to detect any found bits.
 701     // end never changes in G1.
 702     HeapWord* end = r->end();
 703     return _bitmap->get_next_marked_addr(r->bottom(), end) != end;
 704   }
 705 };
 706 
 707 bool G1ConcurrentMark::next_mark_bitmap_is_clear() {
 708   CheckBitmapClearHRClosure cl(_next_mark_bitmap);
 709   _g1h->heap_region_iterate(&cl);
 710   return cl.complete();
 711 }
 712 
 713 class NoteStartOfMarkHRClosure: public HeapRegionClosure {
 714 public:
 715   bool doHeapRegion(HeapRegion* r) {
 716     r->note_start_of_marking();
 717     return false;
 718   }
 719 };
 720 
 721 void G1ConcurrentMark::checkpoint_roots_initial_pre() {
 722   G1CollectedHeap* g1h = G1CollectedHeap::heap();
 723 
 724   _has_aborted = false;
 725 
 726   // Initialize marking structures. This has to be done in a STW phase.
 727   reset();
 728 
 729   // For each region note start of marking.
 730   NoteStartOfMarkHRClosure startcl;
 731   g1h->heap_region_iterate(&startcl);
 732 }
 733 
 734 
 735 void G1ConcurrentMark::checkpoint_roots_initial_post() {
 736   G1CollectedHeap*   g1h = G1CollectedHeap::heap();
 737 
 738   // Start Concurrent Marking weak-reference discovery.
 739   ReferenceProcessor* rp = g1h->ref_processor_cm();
 740   // enable ("weak") refs discovery
 741   rp->enable_discovery();
 742   rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle
 743 
 744   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
 745   // This is the start of  the marking cycle, we're expected all
 746   // threads to have SATB queues with active set to false.
 747   satb_mq_set.set_active_all_threads(true, /* new active value */
 748                                      false /* expected_active */);
 749 
 750   _root_regions.prepare_for_scan();
 751 
 752   // update_g1_committed() will be called at the end of an evac pause
 753   // when marking is on. So, it's also called at the end of the
 754   // initial-mark pause to update the heap end, if the heap expands
 755   // during it. No need to call it here.
 756 }
 757 
 758 /*
 759  * Notice that in the next two methods, we actually leave the STS
 760  * during the barrier sync and join it immediately afterwards. If we
 761  * do not do this, the following deadlock can occur: one thread could
 762  * be in the barrier sync code, waiting for the other thread to also
 763  * sync up, whereas another one could be trying to yield, while also
 764  * waiting for the other threads to sync up too.
 765  *
 766  * Note, however, that this code is also used during remark and in
 767  * this case we should not attempt to leave / enter the STS, otherwise
 768  * we'll either hit an assert (debug / fastdebug) or deadlock
 769  * (product). So we should only leave / enter the STS if we are
 770  * operating concurrently.
 771  *
 772  * Because the thread that does the sync barrier has left the STS, it
 773  * is possible to be suspended for a Full GC or an evacuation pause
 774  * could occur. This is actually safe, since the entering the sync
 775  * barrier is one of the last things do_marking_step() does, and it
 776  * doesn't manipulate any data structures afterwards.
 777  */
 778 
 779 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) {
 780   bool barrier_aborted;
 781   {
 782     SuspendibleThreadSetLeaver sts_leave(concurrent());
 783     barrier_aborted = !_first_overflow_barrier_sync.enter();
 784   }
 785 
 786   // at this point everyone should have synced up and not be doing any
 787   // more work
 788 
 789   if (barrier_aborted) {
 790     // If the barrier aborted we ignore the overflow condition and
 791     // just abort the whole marking phase as quickly as possible.
 792     return;
 793   }
 794 
 795   // If we're executing the concurrent phase of marking, reset the marking
 796   // state; otherwise the marking state is reset after reference processing,
 797   // during the remark pause.
 798   // If we reset here as a result of an overflow during the remark we will
 799   // see assertion failures from any subsequent set_concurrency_and_phase()
 800   // calls.
 801   if (concurrent()) {
 802     // let the task associated with with worker 0 do this
 803     if (worker_id == 0) {
 804       // task 0 is responsible for clearing the global data structures
 805       // We should be here because of an overflow. During STW we should
 806       // not clear the overflow flag since we rely on it being true when
 807       // we exit this method to abort the pause and restart concurrent
 808       // marking.
 809       reset_marking_state();
 810 
 811       log_info(gc, marking)("Concurrent Mark reset for overflow");
 812     }
 813   }
 814 
 815   // after this, each task should reset its own data structures then
 816   // then go into the second barrier
 817 }
 818 
 819 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) {
 820   SuspendibleThreadSetLeaver sts_leave(concurrent());
 821   _second_overflow_barrier_sync.enter();
 822 
 823   // at this point everything should be re-initialized and ready to go
 824 }
 825 
 826 class G1CMConcurrentMarkingTask: public AbstractGangTask {
 827 private:
 828   G1ConcurrentMark*     _cm;
 829   ConcurrentMarkThread* _cmt;
 830 
 831 public:
 832   void work(uint worker_id) {
 833     assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread");
 834     ResourceMark rm;
 835 
 836     double start_vtime = os::elapsedVTime();
 837 
 838     {
 839       SuspendibleThreadSetJoiner sts_join;
 840 
 841       assert(worker_id < _cm->active_tasks(), "invariant");
 842 
 843       G1CMTask* task = _cm->task(worker_id);
 844       task->record_start_time();
 845       if (!_cm->has_aborted()) {
 846         do {
 847           task->do_marking_step(G1ConcMarkStepDurationMillis,
 848                                 true  /* do_termination */,
 849                                 false /* is_serial*/);
 850 
 851           _cm->do_yield_check();
 852         } while (!_cm->has_aborted() && task->has_aborted());
 853       }
 854       task->record_end_time();
 855       guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant");
 856     }
 857 
 858     double end_vtime = os::elapsedVTime();
 859     _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime);
 860   }
 861 
 862   G1CMConcurrentMarkingTask(G1ConcurrentMark* cm,
 863                             ConcurrentMarkThread* cmt) :
 864       AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { }
 865 
 866   ~G1CMConcurrentMarkingTask() { }
 867 };
 868 
 869 uint G1ConcurrentMark::calc_active_marking_workers() {
 870   uint result = 0;
 871   if (!UseDynamicNumberOfGCThreads ||
 872       (!FLAG_IS_DEFAULT(ConcGCThreads) &&
 873        !ForceDynamicNumberOfGCThreads)) {
 874     result = _max_concurrent_workers;
 875   } else {
 876     result =
 877       AdaptiveSizePolicy::calc_default_active_workers(_max_concurrent_workers,
 878                                                       1, /* Minimum workers */
 879                                                       _num_concurrent_workers,
 880                                                       Threads::number_of_non_daemon_threads());
 881     // Don't scale the result down by scale_concurrent_workers() because
 882     // that scaling has already gone into "_max_concurrent_workers".
 883   }
 884   assert(result > 0 && result <= _max_concurrent_workers,
 885          "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u",
 886          _max_concurrent_workers, result);
 887   return result;
 888 }
 889 
 890 void G1ConcurrentMark::scan_root_region(HeapRegion* hr) {
 891   // Currently, only survivors can be root regions.
 892   assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant");
 893   G1RootRegionScanClosure cl(_g1h, this);
 894 
 895   const uintx interval = PrefetchScanIntervalInBytes;
 896   HeapWord* curr = hr->bottom();
 897   const HeapWord* end = hr->top();
 898   while (curr < end) {
 899     Prefetch::read(curr, interval);
 900     oop obj = oop(curr);
 901     int size = obj->oop_iterate_size(&cl);
 902     assert(size == obj->size(), "sanity");
 903     curr += size;
 904   }
 905 }
 906 
 907 class G1CMRootRegionScanTask : public AbstractGangTask {
 908 private:
 909   G1ConcurrentMark* _cm;
 910 
 911 public:
 912   G1CMRootRegionScanTask(G1ConcurrentMark* cm) :
 913     AbstractGangTask("G1 Root Region Scan"), _cm(cm) { }
 914 
 915   void work(uint worker_id) {
 916     assert(Thread::current()->is_ConcurrentGC_thread(),
 917            "this should only be done by a conc GC thread");
 918 
 919     G1CMRootRegions* root_regions = _cm->root_regions();
 920     HeapRegion* hr = root_regions->claim_next();
 921     while (hr != NULL) {
 922       _cm->scan_root_region(hr);
 923       hr = root_regions->claim_next();
 924     }
 925   }
 926 };
 927 
 928 void G1ConcurrentMark::scan_root_regions() {
 929   // scan_in_progress() will have been set to true only if there was
 930   // at least one root region to scan. So, if it's false, we
 931   // should not attempt to do any further work.
 932   if (root_regions()->scan_in_progress()) {
 933     assert(!has_aborted(), "Aborting before root region scanning is finished not supported.");
 934 
 935     _num_concurrent_workers = MIN2(calc_active_marking_workers(),
 936                                    // We distribute work on a per-region basis, so starting
 937                                    // more threads than that is useless.
 938                                    root_regions()->num_root_regions());
 939     assert(_num_concurrent_workers <= _max_concurrent_workers,
 940            "Maximum number of marking threads exceeded");
 941 
 942     G1CMRootRegionScanTask task(this);
 943     log_debug(gc, ergo)("Running %s using %u workers for %u work units.",
 944                         task.name(), _num_concurrent_workers, root_regions()->num_root_regions());
 945     _concurrent_workers->run_task(&task, _num_concurrent_workers);
 946 
 947     // It's possible that has_aborted() is true here without actually
 948     // aborting the survivor scan earlier. This is OK as it's
 949     // mainly used for sanity checking.
 950     root_regions()->scan_finished();
 951   }
 952 }
 953 
 954 void G1ConcurrentMark::concurrent_cycle_start() {
 955   _gc_timer_cm->register_gc_start();
 956 
 957   _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start());
 958 
 959   _g1h->trace_heap_before_gc(_gc_tracer_cm);
 960 }
 961 
 962 void G1ConcurrentMark::concurrent_cycle_end() {
 963   _g1h->trace_heap_after_gc(_gc_tracer_cm);
 964 
 965   if (has_aborted()) {
 966     _gc_tracer_cm->report_concurrent_mode_failure();
 967   }
 968 
 969   _gc_timer_cm->register_gc_end();
 970 
 971   _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions());
 972 }
 973 
 974 void G1ConcurrentMark::mark_from_roots() {
 975   // we might be tempted to assert that:
 976   // assert(asynch == !SafepointSynchronize::is_at_safepoint(),
 977   //        "inconsistent argument?");
 978   // However that wouldn't be right, because it's possible that
 979   // a safepoint is indeed in progress as a younger generation
 980   // stop-the-world GC happens even as we mark in this generation.
 981 
 982   _restart_for_overflow = false;
 983 
 984   _num_concurrent_workers = calc_active_marking_workers();
 985 
 986   uint active_workers = MAX2(1U, _num_concurrent_workers);
 987 
 988   // Setting active workers is not guaranteed since fewer
 989   // worker threads may currently exist and more may not be
 990   // available.
 991   active_workers = _concurrent_workers->update_active_workers(active_workers);
 992   log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers());
 993 
 994   // Parallel task terminator is set in "set_concurrency_and_phase()"
 995   set_concurrency_and_phase(active_workers, true /* concurrent */);
 996 
 997   G1CMConcurrentMarkingTask marking_task(this, cm_thread());
 998   _concurrent_workers->run_task(&marking_task);
 999   print_stats();
1000 }
1001 
1002 void G1ConcurrentMark::checkpoint_roots_final(bool clear_all_soft_refs) {
1003   // world is stopped at this checkpoint
1004   assert(SafepointSynchronize::is_at_safepoint(),
1005          "world should be stopped");
1006 
1007   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1008 
1009   // If a full collection has happened, we shouldn't do this.
1010   if (has_aborted()) {
1011     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1012     return;
1013   }
1014 
1015   SvcGCMarker sgcm(SvcGCMarker::OTHER);
1016 
1017   if (VerifyDuringGC) {
1018     g1h->verifier()->verify(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "During GC (before)");
1019   }
1020   g1h->verifier()->check_bitmaps("Remark Start");
1021 
1022   G1Policy* g1p = g1h->g1_policy();
1023   g1p->record_concurrent_mark_remark_start();
1024 
1025   double start = os::elapsedTime();
1026 
1027   checkpoint_roots_final_work();
1028 
1029   double mark_work_end = os::elapsedTime();
1030 
1031   weak_refs_work(clear_all_soft_refs);
1032 
1033   if (has_overflown()) {
1034     // We overflowed.  Restart concurrent marking.
1035     _restart_for_overflow = true;
1036 
1037     // Verify the heap w.r.t. the previous marking bitmap.
1038     if (VerifyDuringGC) {
1039       g1h->verifier()->verify(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "During GC (overflow)");
1040     }
1041 
1042     // Clear the marking state because we will be restarting
1043     // marking due to overflowing the global mark stack.
1044     reset_marking_state();
1045   } else {
1046     SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1047     // We're done with marking.
1048     // This is the end of  the marking cycle, we're expected all
1049     // threads to have SATB queues with active set to true.
1050     satb_mq_set.set_active_all_threads(false, /* new active value */
1051                                        true /* expected_active */);
1052 
1053     if (VerifyDuringGC) {
1054       g1h->verifier()->verify(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UseNextMarking, "During GC (after)");
1055     }
1056     g1h->verifier()->check_bitmaps("Remark End");
1057     assert(!restart_for_overflow(), "sanity");
1058     // Completely reset the marking state since marking completed
1059     set_non_marking_state();
1060   }
1061 
1062   // Statistics
1063   double now = os::elapsedTime();
1064   _remark_mark_times.add((mark_work_end - start) * 1000.0);
1065   _remark_weak_ref_times.add((now - mark_work_end) * 1000.0);
1066   _remark_times.add((now - start) * 1000.0);
1067 
1068   g1p->record_concurrent_mark_remark_end();
1069 
1070   G1CMIsAliveClosure is_alive(g1h);
1071   _gc_tracer_cm->report_object_count_after_gc(&is_alive);
1072 }
1073 
1074 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure {
1075   G1CollectedHeap* _g1;
1076   size_t _freed_bytes;
1077   FreeRegionList* _local_cleanup_list;
1078   uint _old_regions_removed;
1079   uint _humongous_regions_removed;
1080   HRRSCleanupTask* _hrrs_cleanup_task;
1081 
1082 public:
1083   G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1,
1084                              FreeRegionList* local_cleanup_list,
1085                              HRRSCleanupTask* hrrs_cleanup_task) :
1086     _g1(g1),
1087     _freed_bytes(0),
1088     _local_cleanup_list(local_cleanup_list),
1089     _old_regions_removed(0),
1090     _humongous_regions_removed(0),
1091     _hrrs_cleanup_task(hrrs_cleanup_task) { }
1092 
1093   size_t freed_bytes() { return _freed_bytes; }
1094   const uint old_regions_removed() { return _old_regions_removed; }
1095   const uint humongous_regions_removed() { return _humongous_regions_removed; }
1096 
1097   bool doHeapRegion(HeapRegion *hr) {
1098     _g1->reset_gc_time_stamps(hr);
1099     hr->note_end_of_marking();
1100 
1101     if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) {
1102       _freed_bytes += hr->used();
1103       hr->set_containing_set(NULL);
1104       if (hr->is_humongous()) {
1105         _humongous_regions_removed++;
1106         _g1->free_humongous_region(hr, _local_cleanup_list, true /* skip_remset */);
1107       } else {
1108         _old_regions_removed++;
1109         _g1->free_region(hr, _local_cleanup_list, true /* skip_remset */);
1110       }
1111     } else {
1112       hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task);
1113     }
1114 
1115     return false;
1116   }
1117 };
1118 
1119 class G1ParNoteEndTask: public AbstractGangTask {
1120   friend class G1NoteEndOfConcMarkClosure;
1121 
1122 protected:
1123   G1CollectedHeap* _g1h;
1124   FreeRegionList* _cleanup_list;
1125   HeapRegionClaimer _hrclaimer;
1126 
1127 public:
1128   G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) :
1129       AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) {
1130   }
1131 
1132   void work(uint worker_id) {
1133     FreeRegionList local_cleanup_list("Local Cleanup List");
1134     HRRSCleanupTask hrrs_cleanup_task;
1135     G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list,
1136                                            &hrrs_cleanup_task);
1137     _g1h->heap_region_par_iterate_from_worker_offset(&g1_note_end, &_hrclaimer, worker_id);
1138     assert(g1_note_end.complete(), "Shouldn't have yielded!");
1139 
1140     // Now update the lists
1141     _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed());
1142     {
1143       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
1144       _g1h->decrement_summary_bytes(g1_note_end.freed_bytes());
1145 
1146       // If we iterate over the global cleanup list at the end of
1147       // cleanup to do this printing we will not guarantee to only
1148       // generate output for the newly-reclaimed regions (the list
1149       // might not be empty at the beginning of cleanup; we might
1150       // still be working on its previous contents). So we do the
1151       // printing here, before we append the new regions to the global
1152       // cleanup list.
1153 
1154       G1HRPrinter* hr_printer = _g1h->hr_printer();
1155       if (hr_printer->is_active()) {
1156         FreeRegionListIterator iter(&local_cleanup_list);
1157         while (iter.more_available()) {
1158           HeapRegion* hr = iter.get_next();
1159           hr_printer->cleanup(hr);
1160         }
1161       }
1162 
1163       _cleanup_list->add_ordered(&local_cleanup_list);
1164       assert(local_cleanup_list.is_empty(), "post-condition");
1165 
1166       HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task);
1167     }
1168   }
1169 };
1170 
1171 void G1ConcurrentMark::cleanup() {
1172   // world is stopped at this checkpoint
1173   assert(SafepointSynchronize::is_at_safepoint(),
1174          "world should be stopped");
1175   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1176 
1177   // If a full collection has happened, we shouldn't do this.
1178   if (has_aborted()) {
1179     g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused
1180     return;
1181   }
1182 
1183   g1h->verifier()->verify_region_sets_optional();
1184 
1185   if (VerifyDuringGC) {
1186     g1h->verifier()->verify(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "During GC (before)");
1187   }
1188   g1h->verifier()->check_bitmaps("Cleanup Start");
1189 
1190   G1Policy* g1p = g1h->g1_policy();
1191   g1p->record_concurrent_mark_cleanup_start();
1192 
1193   double start = os::elapsedTime();
1194 
1195   HeapRegionRemSet::reset_for_cleanup_tasks();
1196 
1197   {
1198     GCTraceTime(Debug, gc)("Finalize Live Data");
1199     finalize_live_data();
1200   }
1201 
1202   if (VerifyDuringGC) {
1203     GCTraceTime(Debug, gc)("Verify Live Data");
1204     verify_live_data();
1205   }
1206 
1207   g1h->collector_state()->set_mark_in_progress(false);
1208 
1209   double count_end = os::elapsedTime();
1210   double this_final_counting_time = (count_end - start);
1211   _total_counting_time += this_final_counting_time;
1212 
1213   if (log_is_enabled(Trace, gc, liveness)) {
1214     G1PrintRegionLivenessInfoClosure cl("Post-Marking");
1215     _g1h->heap_region_iterate(&cl);
1216   }
1217 
1218   // Install newly created mark bitMap as "prev".
1219   swap_mark_bitmaps();
1220 
1221   g1h->reset_gc_time_stamp();
1222 
1223   uint n_workers = _g1h->workers()->active_workers();
1224 
1225   // Note end of marking in all heap regions.
1226   G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers);
1227   g1h->workers()->run_task(&g1_par_note_end_task);
1228   g1h->check_gc_time_stamps();
1229 
1230   if (!cleanup_list_is_empty()) {
1231     // The cleanup list is not empty, so we'll have to process it
1232     // concurrently. Notify anyone else that might be wanting free
1233     // regions that there will be more free regions coming soon.
1234     g1h->set_free_regions_coming();
1235   }
1236 
1237   // call below, since it affects the metric by which we sort the heap
1238   // regions.
1239   if (G1ScrubRemSets) {
1240     double rs_scrub_start = os::elapsedTime();
1241     g1h->scrub_rem_set();
1242     _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start);
1243   }
1244 
1245   // this will also free any regions totally full of garbage objects,
1246   // and sort the regions.
1247   g1h->g1_policy()->record_concurrent_mark_cleanup_end();
1248 
1249   // Statistics.
1250   double end = os::elapsedTime();
1251   _cleanup_times.add((end - start) * 1000.0);
1252 
1253   // Clean up will have freed any regions completely full of garbage.
1254   // Update the soft reference policy with the new heap occupancy.
1255   Universe::update_heap_info_at_gc();
1256 
1257   if (VerifyDuringGC) {
1258     g1h->verifier()->verify(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "During GC (after)");
1259   }
1260 
1261   g1h->verifier()->check_bitmaps("Cleanup End");
1262 
1263   g1h->verifier()->verify_region_sets_optional();
1264 
1265   // We need to make this be a "collection" so any collection pause that
1266   // races with it goes around and waits for completeCleanup to finish.
1267   g1h->increment_total_collections();
1268 
1269   // Clean out dead classes and update Metaspace sizes.
1270   if (ClassUnloadingWithConcurrentMark) {
1271     ClassLoaderDataGraph::purge();
1272   }
1273   MetaspaceGC::compute_new_size();
1274 
1275   // We reclaimed old regions so we should calculate the sizes to make
1276   // sure we update the old gen/space data.
1277   g1h->g1mm()->update_sizes();
1278   g1h->allocation_context_stats().update_after_mark();
1279 }
1280 
1281 void G1ConcurrentMark::complete_cleanup() {
1282   if (has_aborted()) return;
1283 
1284   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1285 
1286   _cleanup_list.verify_optional();
1287   FreeRegionList tmp_free_list("Tmp Free List");
1288 
1289   log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1290                                   "cleanup list has %u entries",
1291                                   _cleanup_list.length());
1292 
1293   // No one else should be accessing the _cleanup_list at this point,
1294   // so it is not necessary to take any locks
1295   while (!_cleanup_list.is_empty()) {
1296     HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */);
1297     assert(hr != NULL, "Got NULL from a non-empty list");
1298     hr->par_clear();
1299     tmp_free_list.add_ordered(hr);
1300 
1301     // Instead of adding one region at a time to the secondary_free_list,
1302     // we accumulate them in the local list and move them a few at a
1303     // time. This also cuts down on the number of notify_all() calls
1304     // we do during this process. We'll also append the local list when
1305     // _cleanup_list is empty (which means we just removed the last
1306     // region from the _cleanup_list).
1307     if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) ||
1308         _cleanup_list.is_empty()) {
1309       log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : "
1310                                       "appending %u entries to the secondary_free_list, "
1311                                       "cleanup list still has %u entries",
1312                                       tmp_free_list.length(),
1313                                       _cleanup_list.length());
1314 
1315       {
1316         MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag);
1317         g1h->secondary_free_list_add(&tmp_free_list);
1318         SecondaryFreeList_lock->notify_all();
1319       }
1320 #ifndef PRODUCT
1321       if (G1StressConcRegionFreeing) {
1322         for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) {
1323           os::sleep(Thread::current(), (jlong) 1, false);
1324         }
1325       }
1326 #endif
1327     }
1328   }
1329   assert(tmp_free_list.is_empty(), "post-condition");
1330 }
1331 
1332 // Supporting Object and Oop closures for reference discovery
1333 // and processing in during marking
1334 
1335 bool G1CMIsAliveClosure::do_object_b(oop obj) {
1336   HeapWord* addr = (HeapWord*)obj;
1337   return addr != NULL &&
1338          (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj));
1339 }
1340 
1341 // 'Keep Alive' oop closure used by both serial parallel reference processing.
1342 // Uses the G1CMTask associated with a worker thread (for serial reference
1343 // processing the G1CMTask for worker 0 is used) to preserve (mark) and
1344 // trace referent objects.
1345 //
1346 // Using the G1CMTask and embedded local queues avoids having the worker
1347 // threads operating on the global mark stack. This reduces the risk
1348 // of overflowing the stack - which we would rather avoid at this late
1349 // state. Also using the tasks' local queues removes the potential
1350 // of the workers interfering with each other that could occur if
1351 // operating on the global stack.
1352 
1353 class G1CMKeepAliveAndDrainClosure: public OopClosure {
1354   G1ConcurrentMark* _cm;
1355   G1CMTask*         _task;
1356   int               _ref_counter_limit;
1357   int               _ref_counter;
1358   bool              _is_serial;
1359  public:
1360   G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1361     _cm(cm), _task(task), _is_serial(is_serial),
1362     _ref_counter_limit(G1RefProcDrainInterval) {
1363     assert(_ref_counter_limit > 0, "sanity");
1364     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1365     _ref_counter = _ref_counter_limit;
1366   }
1367 
1368   virtual void do_oop(narrowOop* p) { do_oop_work(p); }
1369   virtual void do_oop(      oop* p) { do_oop_work(p); }
1370 
1371   template <class T> void do_oop_work(T* p) {
1372     if (!_cm->has_overflown()) {
1373       oop obj = oopDesc::load_decode_heap_oop(p);
1374       _task->deal_with_reference(obj);
1375       _ref_counter--;
1376 
1377       if (_ref_counter == 0) {
1378         // We have dealt with _ref_counter_limit references, pushing them
1379         // and objects reachable from them on to the local stack (and
1380         // possibly the global stack). Call G1CMTask::do_marking_step() to
1381         // process these entries.
1382         //
1383         // We call G1CMTask::do_marking_step() in a loop, which we'll exit if
1384         // there's nothing more to do (i.e. we're done with the entries that
1385         // were pushed as a result of the G1CMTask::deal_with_reference() calls
1386         // above) or we overflow.
1387         //
1388         // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1389         // flag while there may still be some work to do. (See the comment at
1390         // the beginning of G1CMTask::do_marking_step() for those conditions -
1391         // one of which is reaching the specified time target.) It is only
1392         // when G1CMTask::do_marking_step() returns without setting the
1393         // has_aborted() flag that the marking step has completed.
1394         do {
1395           double mark_step_duration_ms = G1ConcMarkStepDurationMillis;
1396           _task->do_marking_step(mark_step_duration_ms,
1397                                  false      /* do_termination */,
1398                                  _is_serial);
1399         } while (_task->has_aborted() && !_cm->has_overflown());
1400         _ref_counter = _ref_counter_limit;
1401       }
1402     }
1403   }
1404 };
1405 
1406 // 'Drain' oop closure used by both serial and parallel reference processing.
1407 // Uses the G1CMTask associated with a given worker thread (for serial
1408 // reference processing the G1CMtask for worker 0 is used). Calls the
1409 // do_marking_step routine, with an unbelievably large timeout value,
1410 // to drain the marking data structures of the remaining entries
1411 // added by the 'keep alive' oop closure above.
1412 
1413 class G1CMDrainMarkingStackClosure: public VoidClosure {
1414   G1ConcurrentMark* _cm;
1415   G1CMTask*         _task;
1416   bool              _is_serial;
1417  public:
1418   G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) :
1419     _cm(cm), _task(task), _is_serial(is_serial) {
1420     assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code");
1421   }
1422 
1423   void do_void() {
1424     do {
1425       // We call G1CMTask::do_marking_step() to completely drain the local
1426       // and global marking stacks of entries pushed by the 'keep alive'
1427       // oop closure (an instance of G1CMKeepAliveAndDrainClosure above).
1428       //
1429       // G1CMTask::do_marking_step() is called in a loop, which we'll exit
1430       // if there's nothing more to do (i.e. we've completely drained the
1431       // entries that were pushed as a a result of applying the 'keep alive'
1432       // closure to the entries on the discovered ref lists) or we overflow
1433       // the global marking stack.
1434       //
1435       // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted()
1436       // flag while there may still be some work to do. (See the comment at
1437       // the beginning of G1CMTask::do_marking_step() for those conditions -
1438       // one of which is reaching the specified time target.) It is only
1439       // when G1CMTask::do_marking_step() returns without setting the
1440       // has_aborted() flag that the marking step has completed.
1441 
1442       _task->do_marking_step(1000000000.0 /* something very large */,
1443                              true         /* do_termination */,
1444                              _is_serial);
1445     } while (_task->has_aborted() && !_cm->has_overflown());
1446   }
1447 };
1448 
1449 // Implementation of AbstractRefProcTaskExecutor for parallel
1450 // reference processing at the end of G1 concurrent marking
1451 
1452 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor {
1453 private:
1454   G1CollectedHeap*  _g1h;
1455   G1ConcurrentMark* _cm;
1456   WorkGang*         _workers;
1457   uint              _active_workers;
1458 
1459 public:
1460   G1CMRefProcTaskExecutor(G1CollectedHeap* g1h,
1461                           G1ConcurrentMark* cm,
1462                           WorkGang* workers,
1463                           uint n_workers) :
1464     _g1h(g1h), _cm(cm),
1465     _workers(workers), _active_workers(n_workers) { }
1466 
1467   // Executes the given task using concurrent marking worker threads.
1468   virtual void execute(ProcessTask& task);
1469   virtual void execute(EnqueueTask& task);
1470 };
1471 
1472 class G1CMRefProcTaskProxy: public AbstractGangTask {
1473   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
1474   ProcessTask&      _proc_task;
1475   G1CollectedHeap*  _g1h;
1476   G1ConcurrentMark* _cm;
1477 
1478 public:
1479   G1CMRefProcTaskProxy(ProcessTask& proc_task,
1480                        G1CollectedHeap* g1h,
1481                        G1ConcurrentMark* cm) :
1482     AbstractGangTask("Process reference objects in parallel"),
1483     _proc_task(proc_task), _g1h(g1h), _cm(cm) {
1484     ReferenceProcessor* rp = _g1h->ref_processor_cm();
1485     assert(rp->processing_is_mt(), "shouldn't be here otherwise");
1486   }
1487 
1488   virtual void work(uint worker_id) {
1489     ResourceMark rm;
1490     HandleMark hm;
1491     G1CMTask* task = _cm->task(worker_id);
1492     G1CMIsAliveClosure g1_is_alive(_g1h);
1493     G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */);
1494     G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */);
1495 
1496     _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain);
1497   }
1498 };
1499 
1500 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) {
1501   assert(_workers != NULL, "Need parallel worker threads.");
1502   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1503 
1504   G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm);
1505 
1506   // We need to reset the concurrency level before each
1507   // proxy task execution, so that the termination protocol
1508   // and overflow handling in G1CMTask::do_marking_step() knows
1509   // how many workers to wait for.
1510   _cm->set_concurrency(_active_workers);
1511   _workers->run_task(&proc_task_proxy);
1512 }
1513 
1514 class G1CMRefEnqueueTaskProxy: public AbstractGangTask {
1515   typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask;
1516   EnqueueTask& _enq_task;
1517 
1518 public:
1519   G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) :
1520     AbstractGangTask("Enqueue reference objects in parallel"),
1521     _enq_task(enq_task) { }
1522 
1523   virtual void work(uint worker_id) {
1524     _enq_task.work(worker_id);
1525   }
1526 };
1527 
1528 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) {
1529   assert(_workers != NULL, "Need parallel worker threads.");
1530   assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT");
1531 
1532   G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task);
1533 
1534   // Not strictly necessary but...
1535   //
1536   // We need to reset the concurrency level before each
1537   // proxy task execution, so that the termination protocol
1538   // and overflow handling in G1CMTask::do_marking_step() knows
1539   // how many workers to wait for.
1540   _cm->set_concurrency(_active_workers);
1541   _workers->run_task(&enq_task_proxy);
1542 }
1543 
1544 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) {
1545   if (has_overflown()) {
1546     // Skip processing the discovered references if we have
1547     // overflown the global marking stack. Reference objects
1548     // only get discovered once so it is OK to not
1549     // de-populate the discovered reference lists. We could have,
1550     // but the only benefit would be that, when marking restarts,
1551     // less reference objects are discovered.
1552     return;
1553   }
1554 
1555   ResourceMark rm;
1556   HandleMark   hm;
1557 
1558   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1559 
1560   // Is alive closure.
1561   G1CMIsAliveClosure g1_is_alive(g1h);
1562 
1563   // Inner scope to exclude the cleaning of the string and symbol
1564   // tables from the displayed time.
1565   {
1566     GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm);
1567 
1568     ReferenceProcessor* rp = g1h->ref_processor_cm();
1569 
1570     // See the comment in G1CollectedHeap::ref_processing_init()
1571     // about how reference processing currently works in G1.
1572 
1573     // Set the soft reference policy
1574     rp->setup_policy(clear_all_soft_refs);
1575     assert(_global_mark_stack.is_empty(), "mark stack should be empty");
1576 
1577     // Instances of the 'Keep Alive' and 'Complete GC' closures used
1578     // in serial reference processing. Note these closures are also
1579     // used for serially processing (by the the current thread) the
1580     // JNI references during parallel reference processing.
1581     //
1582     // These closures do not need to synchronize with the worker
1583     // threads involved in parallel reference processing as these
1584     // instances are executed serially by the current thread (e.g.
1585     // reference processing is not multi-threaded and is thus
1586     // performed by the current thread instead of a gang worker).
1587     //
1588     // The gang tasks involved in parallel reference processing create
1589     // their own instances of these closures, which do their own
1590     // synchronization among themselves.
1591     G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */);
1592     G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */);
1593 
1594     // We need at least one active thread. If reference processing
1595     // is not multi-threaded we use the current (VMThread) thread,
1596     // otherwise we use the work gang from the G1CollectedHeap and
1597     // we utilize all the worker threads we can.
1598     bool processing_is_mt = rp->processing_is_mt();
1599     uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U);
1600     active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U);
1601 
1602     // Parallel processing task executor.
1603     G1CMRefProcTaskExecutor par_task_executor(g1h, this,
1604                                               g1h->workers(), active_workers);
1605     AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL);
1606 
1607     // Set the concurrency level. The phase was already set prior to
1608     // executing the remark task.
1609     set_concurrency(active_workers);
1610 
1611     // Set the degree of MT processing here.  If the discovery was done MT,
1612     // the number of threads involved during discovery could differ from
1613     // the number of active workers.  This is OK as long as the discovered
1614     // Reference lists are balanced (see balance_all_queues() and balance_queues()).
1615     rp->set_active_mt_degree(active_workers);
1616 
1617     ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q());
1618 
1619     // Process the weak references.
1620     const ReferenceProcessorStats& stats =
1621         rp->process_discovered_references(&g1_is_alive,
1622                                           &g1_keep_alive,
1623                                           &g1_drain_mark_stack,
1624                                           executor,
1625                                           &pt);
1626     _gc_tracer_cm->report_gc_reference_stats(stats);
1627     pt.print_all_references();
1628 
1629     // The do_oop work routines of the keep_alive and drain_marking_stack
1630     // oop closures will set the has_overflown flag if we overflow the
1631     // global marking stack.
1632 
1633     assert(has_overflown() || _global_mark_stack.is_empty(),
1634             "Mark stack should be empty (unless it has overflown)");
1635 
1636     assert(rp->num_q() == active_workers, "why not");
1637 
1638     rp->enqueue_discovered_references(executor, &pt);
1639 
1640     rp->verify_no_references_recorded();
1641 
1642     pt.print_enqueue_phase();
1643 
1644     assert(!rp->discovery_enabled(), "Post condition");
1645   }
1646 
1647   assert(has_overflown() || _global_mark_stack.is_empty(),
1648           "Mark stack should be empty (unless it has overflown)");
1649 
1650   {
1651     GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm);
1652     WeakProcessor::weak_oops_do(&g1_is_alive, &do_nothing_cl);
1653   }
1654 
1655   if (has_overflown()) {
1656     // We can not trust g1_is_alive if the marking stack overflowed
1657     return;
1658   }
1659 
1660   assert(_global_mark_stack.is_empty(), "Marking should have completed");
1661 
1662   // Unload Klasses, String, Symbols, Code Cache, etc.
1663   if (ClassUnloadingWithConcurrentMark) {
1664     GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm);
1665     bool purged_classes = SystemDictionary::do_unloading(&g1_is_alive, _gc_timer_cm, false /* Defer cleaning */);
1666     g1h->complete_cleaning(&g1_is_alive, purged_classes);
1667   } else {
1668     GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm);
1669     // No need to clean string table and symbol table as they are treated as strong roots when
1670     // class unloading is disabled.
1671     g1h->partial_cleaning(&g1_is_alive, false, false, G1StringDedup::is_enabled());
1672 
1673   }
1674 }
1675 
1676 void G1ConcurrentMark::swap_mark_bitmaps() {
1677   G1CMBitMap* temp = _prev_mark_bitmap;
1678   _prev_mark_bitmap = _next_mark_bitmap;
1679   _next_mark_bitmap = temp;
1680 }
1681 
1682 // Closure for marking entries in SATB buffers.
1683 class G1CMSATBBufferClosure : public SATBBufferClosure {
1684 private:
1685   G1CMTask* _task;
1686   G1CollectedHeap* _g1h;
1687 
1688   // This is very similar to G1CMTask::deal_with_reference, but with
1689   // more relaxed requirements for the argument, so this must be more
1690   // circumspect about treating the argument as an object.
1691   void do_entry(void* entry) const {
1692     _task->increment_refs_reached();
1693     oop const obj = static_cast<oop>(entry);
1694     _task->make_reference_grey(obj);
1695   }
1696 
1697 public:
1698   G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h)
1699     : _task(task), _g1h(g1h) { }
1700 
1701   virtual void do_buffer(void** buffer, size_t size) {
1702     for (size_t i = 0; i < size; ++i) {
1703       do_entry(buffer[i]);
1704     }
1705   }
1706 };
1707 
1708 class G1RemarkThreadsClosure : public ThreadClosure {
1709   G1CMSATBBufferClosure _cm_satb_cl;
1710   G1CMOopClosure _cm_cl;
1711   MarkingCodeBlobClosure _code_cl;
1712   int _thread_parity;
1713 
1714  public:
1715   G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) :
1716     _cm_satb_cl(task, g1h),
1717     _cm_cl(g1h, g1h->concurrent_mark(), task),
1718     _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations),
1719     _thread_parity(Threads::thread_claim_parity()) {}
1720 
1721   void do_thread(Thread* thread) {
1722     if (thread->is_Java_thread()) {
1723       if (thread->claim_oops_do(true, _thread_parity)) {
1724         JavaThread* jt = (JavaThread*)thread;
1725 
1726         // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking
1727         // however the liveness of oops reachable from nmethods have very complex lifecycles:
1728         // * Alive if on the stack of an executing method
1729         // * Weakly reachable otherwise
1730         // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be
1731         // live by the SATB invariant but other oops recorded in nmethods may behave differently.
1732         jt->nmethods_do(&_code_cl);
1733 
1734         jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl);
1735       }
1736     } else if (thread->is_VM_thread()) {
1737       if (thread->claim_oops_do(true, _thread_parity)) {
1738         JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl);
1739       }
1740     }
1741   }
1742 };
1743 
1744 class G1CMRemarkTask: public AbstractGangTask {
1745 private:
1746   G1ConcurrentMark* _cm;
1747 public:
1748   void work(uint worker_id) {
1749     G1CMTask* task = _cm->task(worker_id);
1750     task->record_start_time();
1751     {
1752       ResourceMark rm;
1753       HandleMark hm;
1754 
1755       G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task);
1756       Threads::threads_do(&threads_f);
1757     }
1758 
1759     do {
1760       task->do_marking_step(1000000000.0 /* something very large */,
1761                             true         /* do_termination       */,
1762                             false        /* is_serial            */);
1763     } while (task->has_aborted() && !_cm->has_overflown());
1764     // If we overflow, then we do not want to restart. We instead
1765     // want to abort remark and do concurrent marking again.
1766     task->record_end_time();
1767   }
1768 
1769   G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) :
1770     AbstractGangTask("Par Remark"), _cm(cm) {
1771     _cm->terminator()->reset_for_reuse(active_workers);
1772   }
1773 };
1774 
1775 void G1ConcurrentMark::checkpoint_roots_final_work() {
1776   ResourceMark rm;
1777   HandleMark   hm;
1778   G1CollectedHeap* g1h = G1CollectedHeap::heap();
1779 
1780   GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm);
1781 
1782   g1h->ensure_parsability(false);
1783 
1784   // this is remark, so we'll use up all active threads
1785   uint active_workers = g1h->workers()->active_workers();
1786   set_concurrency_and_phase(active_workers, false /* concurrent */);
1787   // Leave _parallel_marking_threads at it's
1788   // value originally calculated in the G1ConcurrentMark
1789   // constructor and pass values of the active workers
1790   // through the gang in the task.
1791 
1792   {
1793     StrongRootsScope srs(active_workers);
1794 
1795     G1CMRemarkTask remarkTask(this, active_workers);
1796     // We will start all available threads, even if we decide that the
1797     // active_workers will be fewer. The extra ones will just bail out
1798     // immediately.
1799     g1h->workers()->run_task(&remarkTask);
1800   }
1801 
1802   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
1803   guarantee(has_overflown() ||
1804             satb_mq_set.completed_buffers_num() == 0,
1805             "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT,
1806             BOOL_TO_STR(has_overflown()),
1807             satb_mq_set.completed_buffers_num());
1808 
1809   print_stats();
1810 }
1811 
1812 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) {
1813   _prev_mark_bitmap->clear_range(mr);
1814 }
1815 
1816 HeapRegion*
1817 G1ConcurrentMark::claim_region(uint worker_id) {
1818   // "checkpoint" the finger
1819   HeapWord* finger = _finger;
1820 
1821   // _heap_end will not change underneath our feet; it only changes at
1822   // yield points.
1823   while (finger < _heap_end) {
1824     assert(_g1h->is_in_g1_reserved(finger), "invariant");
1825 
1826     HeapRegion* curr_region = _g1h->heap_region_containing(finger);
1827     // Make sure that the reads below do not float before loading curr_region.
1828     OrderAccess::loadload();
1829     // Above heap_region_containing may return NULL as we always scan claim
1830     // until the end of the heap. In this case, just jump to the next region.
1831     HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords;
1832 
1833     // Is the gap between reading the finger and doing the CAS too long?
1834     HeapWord* res = Atomic::cmpxchg(end, &_finger, finger);
1835     if (res == finger && curr_region != NULL) {
1836       // we succeeded
1837       HeapWord*   bottom        = curr_region->bottom();
1838       HeapWord*   limit         = curr_region->next_top_at_mark_start();
1839 
1840       // notice that _finger == end cannot be guaranteed here since,
1841       // someone else might have moved the finger even further
1842       assert(_finger >= end, "the finger should have moved forward");
1843 
1844       if (limit > bottom) {
1845         return curr_region;
1846       } else {
1847         assert(limit == bottom,
1848                "the region limit should be at bottom");
1849         // we return NULL and the caller should try calling
1850         // claim_region() again.
1851         return NULL;
1852       }
1853     } else {
1854       assert(_finger > finger, "the finger should have moved forward");
1855       // read it again
1856       finger = _finger;
1857     }
1858   }
1859 
1860   return NULL;
1861 }
1862 
1863 #ifndef PRODUCT
1864 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC {
1865 private:
1866   G1CollectedHeap* _g1h;
1867   const char* _phase;
1868   int _info;
1869 
1870 public:
1871   VerifyNoCSetOops(const char* phase, int info = -1) :
1872     _g1h(G1CollectedHeap::heap()),
1873     _phase(phase),
1874     _info(info)
1875   { }
1876 
1877   void operator()(G1TaskQueueEntry task_entry) const {
1878     if (task_entry.is_array_slice()) {
1879       guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice()));
1880       return;
1881     }
1882     guarantee(oopDesc::is_oop(task_entry.obj()),
1883               "Non-oop " PTR_FORMAT ", phase: %s, info: %d",
1884               p2i(task_entry.obj()), _phase, _info);
1885     guarantee(!_g1h->is_in_cset(task_entry.obj()),
1886               "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d",
1887               p2i(task_entry.obj()), _phase, _info);
1888   }
1889 };
1890 
1891 void G1ConcurrentMark::verify_no_cset_oops() {
1892   assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint");
1893   if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) {
1894     return;
1895   }
1896 
1897   // Verify entries on the global mark stack
1898   _global_mark_stack.iterate(VerifyNoCSetOops("Stack"));
1899 
1900   // Verify entries on the task queues
1901   for (uint i = 0; i < _max_num_tasks; ++i) {
1902     G1CMTaskQueue* queue = _task_queues->queue(i);
1903     queue->iterate(VerifyNoCSetOops("Queue", i));
1904   }
1905 
1906   // Verify the global finger
1907   HeapWord* global_finger = finger();
1908   if (global_finger != NULL && global_finger < _heap_end) {
1909     // Since we always iterate over all regions, we might get a NULL HeapRegion
1910     // here.
1911     HeapRegion* global_hr = _g1h->heap_region_containing(global_finger);
1912     guarantee(global_hr == NULL || global_finger == global_hr->bottom(),
1913               "global finger: " PTR_FORMAT " region: " HR_FORMAT,
1914               p2i(global_finger), HR_FORMAT_PARAMS(global_hr));
1915   }
1916 
1917   // Verify the task fingers
1918   assert(_num_concurrent_workers <= _max_num_tasks, "sanity");
1919   for (uint i = 0; i < _num_concurrent_workers; ++i) {
1920     G1CMTask* task = _tasks[i];
1921     HeapWord* task_finger = task->finger();
1922     if (task_finger != NULL && task_finger < _heap_end) {
1923       // See above note on the global finger verification.
1924       HeapRegion* task_hr = _g1h->heap_region_containing(task_finger);
1925       guarantee(task_hr == NULL || task_finger == task_hr->bottom() ||
1926                 !task_hr->in_collection_set(),
1927                 "task finger: " PTR_FORMAT " region: " HR_FORMAT,
1928                 p2i(task_finger), HR_FORMAT_PARAMS(task_hr));
1929     }
1930   }
1931 }
1932 #endif // PRODUCT
1933 void G1ConcurrentMark::create_live_data() {
1934   _g1h->g1_rem_set()->create_card_live_data(_concurrent_workers, _next_mark_bitmap);
1935 }
1936 
1937 void G1ConcurrentMark::finalize_live_data() {
1938   _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _next_mark_bitmap);
1939 }
1940 
1941 void G1ConcurrentMark::verify_live_data() {
1942   _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _next_mark_bitmap);
1943 }
1944 
1945 void G1ConcurrentMark::clear_live_data(WorkGang* workers) {
1946   _g1h->g1_rem_set()->clear_card_live_data(workers);
1947 }
1948 
1949 #ifdef ASSERT
1950 void G1ConcurrentMark::verify_live_data_clear() {
1951   _g1h->g1_rem_set()->verify_card_live_data_is_clear();
1952 }
1953 #endif
1954 
1955 void G1ConcurrentMark::print_stats() {
1956   if (!log_is_enabled(Debug, gc, stats)) {
1957     return;
1958   }
1959   log_debug(gc, stats)("---------------------------------------------------------------------");
1960   for (size_t i = 0; i < _num_active_tasks; ++i) {
1961     _tasks[i]->print_stats();
1962     log_debug(gc, stats)("---------------------------------------------------------------------");
1963   }
1964 }
1965 
1966 void G1ConcurrentMark::abort() {
1967   if (!cm_thread()->during_cycle() || _has_aborted) {
1968     // We haven't started a concurrent cycle or we have already aborted it. No need to do anything.
1969     return;
1970   }
1971 
1972   // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next
1973   // concurrent bitmap clearing.
1974   {
1975     GCTraceTime(Debug, gc)("Clear Next Bitmap");
1976     clear_bitmap(_next_mark_bitmap, _g1h->workers(), false);
1977   }
1978   // Note we cannot clear the previous marking bitmap here
1979   // since VerifyDuringGC verifies the objects marked during
1980   // a full GC against the previous bitmap.
1981 
1982   {
1983     GCTraceTime(Debug, gc)("Clear Live Data");
1984     clear_live_data(_g1h->workers());
1985   }
1986   DEBUG_ONLY({
1987     GCTraceTime(Debug, gc)("Verify Live Data Clear");
1988     verify_live_data_clear();
1989   })
1990   // Empty mark stack
1991   reset_marking_state();
1992   for (uint i = 0; i < _max_num_tasks; ++i) {
1993     _tasks[i]->clear_region_fields();
1994   }
1995   _first_overflow_barrier_sync.abort();
1996   _second_overflow_barrier_sync.abort();
1997   _has_aborted = true;
1998 
1999   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2000   satb_mq_set.abandon_partial_marking();
2001   // This can be called either during or outside marking, we'll read
2002   // the expected_active value from the SATB queue set.
2003   satb_mq_set.set_active_all_threads(
2004                                  false, /* new active value */
2005                                  satb_mq_set.is_active() /* expected_active */);
2006 }
2007 
2008 static void print_ms_time_info(const char* prefix, const char* name,
2009                                NumberSeq& ns) {
2010   log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).",
2011                          prefix, ns.num(), name, ns.sum()/1000.0, ns.avg());
2012   if (ns.num() > 0) {
2013     log_trace(gc, marking)("%s         [std. dev = %8.2f ms, max = %8.2f ms]",
2014                            prefix, ns.sd(), ns.maximum());
2015   }
2016 }
2017 
2018 void G1ConcurrentMark::print_summary_info() {
2019   Log(gc, marking) log;
2020   if (!log.is_trace()) {
2021     return;
2022   }
2023 
2024   log.trace(" Concurrent marking:");
2025   print_ms_time_info("  ", "init marks", _init_times);
2026   print_ms_time_info("  ", "remarks", _remark_times);
2027   {
2028     print_ms_time_info("     ", "final marks", _remark_mark_times);
2029     print_ms_time_info("     ", "weak refs", _remark_weak_ref_times);
2030 
2031   }
2032   print_ms_time_info("  ", "cleanups", _cleanup_times);
2033   log.trace("    Finalize live data total time = %8.2f s (avg = %8.2f ms).",
2034             _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2035   if (G1ScrubRemSets) {
2036     log.trace("    RS scrub total time = %8.2f s (avg = %8.2f ms).",
2037               _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0));
2038   }
2039   log.trace("  Total stop_world time = %8.2f s.",
2040             (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0);
2041   log.trace("  Total concurrent time = %8.2f s (%8.2f s marking).",
2042             cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum());
2043 }
2044 
2045 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const {
2046   _concurrent_workers->print_worker_threads_on(st);
2047 }
2048 
2049 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const {
2050   _concurrent_workers->threads_do(tc);
2051 }
2052 
2053 void G1ConcurrentMark::print_on_error(outputStream* st) const {
2054   st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT,
2055                p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap));
2056   _prev_mark_bitmap->print_on_error(st, " Prev Bits: ");
2057   _next_mark_bitmap->print_on_error(st, " Next Bits: ");
2058 }
2059 
2060 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) {
2061   ReferenceProcessor* result = g1h->ref_processor_cm();
2062   assert(result != NULL, "CM reference processor should not be NULL");
2063   return result;
2064 }
2065 
2066 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h,
2067                                G1ConcurrentMark* cm,
2068                                G1CMTask* task)
2069   : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)),
2070     _g1h(g1h), _cm(cm), _task(task)
2071 { }
2072 
2073 void G1CMTask::setup_for_region(HeapRegion* hr) {
2074   assert(hr != NULL,
2075         "claim_region() should have filtered out NULL regions");
2076   _curr_region  = hr;
2077   _finger       = hr->bottom();
2078   update_region_limit();
2079 }
2080 
2081 void G1CMTask::update_region_limit() {
2082   HeapRegion* hr            = _curr_region;
2083   HeapWord* bottom          = hr->bottom();
2084   HeapWord* limit           = hr->next_top_at_mark_start();
2085 
2086   if (limit == bottom) {
2087     // The region was collected underneath our feet.
2088     // We set the finger to bottom to ensure that the bitmap
2089     // iteration that will follow this will not do anything.
2090     // (this is not a condition that holds when we set the region up,
2091     // as the region is not supposed to be empty in the first place)
2092     _finger = bottom;
2093   } else if (limit >= _region_limit) {
2094     assert(limit >= _finger, "peace of mind");
2095   } else {
2096     assert(limit < _region_limit, "only way to get here");
2097     // This can happen under some pretty unusual circumstances.  An
2098     // evacuation pause empties the region underneath our feet (NTAMS
2099     // at bottom). We then do some allocation in the region (NTAMS
2100     // stays at bottom), followed by the region being used as a GC
2101     // alloc region (NTAMS will move to top() and the objects
2102     // originally below it will be grayed). All objects now marked in
2103     // the region are explicitly grayed, if below the global finger,
2104     // and we do not need in fact to scan anything else. So, we simply
2105     // set _finger to be limit to ensure that the bitmap iteration
2106     // doesn't do anything.
2107     _finger = limit;
2108   }
2109 
2110   _region_limit = limit;
2111 }
2112 
2113 void G1CMTask::giveup_current_region() {
2114   assert(_curr_region != NULL, "invariant");
2115   clear_region_fields();
2116 }
2117 
2118 void G1CMTask::clear_region_fields() {
2119   // Values for these three fields that indicate that we're not
2120   // holding on to a region.
2121   _curr_region   = NULL;
2122   _finger        = NULL;
2123   _region_limit  = NULL;
2124 }
2125 
2126 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) {
2127   if (cm_oop_closure == NULL) {
2128     assert(_cm_oop_closure != NULL, "invariant");
2129   } else {
2130     assert(_cm_oop_closure == NULL, "invariant");
2131   }
2132   _cm_oop_closure = cm_oop_closure;
2133 }
2134 
2135 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) {
2136   guarantee(next_mark_bitmap != NULL, "invariant");
2137   _next_mark_bitmap              = next_mark_bitmap;
2138   clear_region_fields();
2139 
2140   _calls                         = 0;
2141   _elapsed_time_ms               = 0.0;
2142   _termination_time_ms           = 0.0;
2143   _termination_start_time_ms     = 0.0;
2144 }
2145 
2146 bool G1CMTask::should_exit_termination() {
2147   regular_clock_call();
2148   // This is called when we are in the termination protocol. We should
2149   // quit if, for some reason, this task wants to abort or the global
2150   // stack is not empty (this means that we can get work from it).
2151   return !_cm->mark_stack_empty() || has_aborted();
2152 }
2153 
2154 void G1CMTask::reached_limit() {
2155   assert(_words_scanned >= _words_scanned_limit ||
2156          _refs_reached >= _refs_reached_limit ,
2157          "shouldn't have been called otherwise");
2158   regular_clock_call();
2159 }
2160 
2161 void G1CMTask::regular_clock_call() {
2162   if (has_aborted()) return;
2163 
2164   // First, we need to recalculate the words scanned and refs reached
2165   // limits for the next clock call.
2166   recalculate_limits();
2167 
2168   // During the regular clock call we do the following
2169 
2170   // (1) If an overflow has been flagged, then we abort.
2171   if (_cm->has_overflown()) {
2172     set_has_aborted();
2173     return;
2174   }
2175 
2176   // If we are not concurrent (i.e. we're doing remark) we don't need
2177   // to check anything else. The other steps are only needed during
2178   // the concurrent marking phase.
2179   if (!_concurrent) {
2180     return;
2181   }
2182 
2183   // (2) If marking has been aborted for Full GC, then we also abort.
2184   if (_cm->has_aborted()) {
2185     set_has_aborted();
2186     return;
2187   }
2188 
2189   double curr_time_ms = os::elapsedVTime() * 1000.0;
2190 
2191   // (4) We check whether we should yield. If we have to, then we abort.
2192   if (SuspendibleThreadSet::should_yield()) {
2193     // We should yield. To do this we abort the task. The caller is
2194     // responsible for yielding.
2195     set_has_aborted();
2196     return;
2197   }
2198 
2199   // (5) We check whether we've reached our time quota. If we have,
2200   // then we abort.
2201   double elapsed_time_ms = curr_time_ms - _start_time_ms;
2202   if (elapsed_time_ms > _time_target_ms) {
2203     set_has_aborted();
2204     _has_timed_out = true;
2205     return;
2206   }
2207 
2208   // (6) Finally, we check whether there are enough completed STAB
2209   // buffers available for processing. If there are, we abort.
2210   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2211   if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) {
2212     // we do need to process SATB buffers, we'll abort and restart
2213     // the marking task to do so
2214     set_has_aborted();
2215     return;
2216   }
2217 }
2218 
2219 void G1CMTask::recalculate_limits() {
2220   _real_words_scanned_limit = _words_scanned + words_scanned_period;
2221   _words_scanned_limit      = _real_words_scanned_limit;
2222 
2223   _real_refs_reached_limit  = _refs_reached  + refs_reached_period;
2224   _refs_reached_limit       = _real_refs_reached_limit;
2225 }
2226 
2227 void G1CMTask::decrease_limits() {
2228   // This is called when we believe that we're going to do an infrequent
2229   // operation which will increase the per byte scanned cost (i.e. move
2230   // entries to/from the global stack). It basically tries to decrease the
2231   // scanning limit so that the clock is called earlier.
2232 
2233   _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4;
2234   _refs_reached_limit  = _real_refs_reached_limit - 3 * refs_reached_period / 4;
2235 }
2236 
2237 void G1CMTask::move_entries_to_global_stack() {
2238   // Local array where we'll store the entries that will be popped
2239   // from the local queue.
2240   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2241 
2242   size_t n = 0;
2243   G1TaskQueueEntry task_entry;
2244   while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) {
2245     buffer[n] = task_entry;
2246     ++n;
2247   }
2248   if (n < G1CMMarkStack::EntriesPerChunk) {
2249     buffer[n] = G1TaskQueueEntry();
2250   }
2251 
2252   if (n > 0) {
2253     if (!_cm->mark_stack_push(buffer)) {
2254       set_has_aborted();
2255     }
2256   }
2257 
2258   // This operation was quite expensive, so decrease the limits.
2259   decrease_limits();
2260 }
2261 
2262 bool G1CMTask::get_entries_from_global_stack() {
2263   // Local array where we'll store the entries that will be popped
2264   // from the global stack.
2265   G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk];
2266 
2267   if (!_cm->mark_stack_pop(buffer)) {
2268     return false;
2269   }
2270 
2271   // We did actually pop at least one entry.
2272   for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) {
2273     G1TaskQueueEntry task_entry = buffer[i];
2274     if (task_entry.is_null()) {
2275       break;
2276     }
2277     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()));
2278     bool success = _task_queue->push(task_entry);
2279     // We only call this when the local queue is empty or under a
2280     // given target limit. So, we do not expect this push to fail.
2281     assert(success, "invariant");
2282   }
2283 
2284   // This operation was quite expensive, so decrease the limits
2285   decrease_limits();
2286   return true;
2287 }
2288 
2289 void G1CMTask::drain_local_queue(bool partially) {
2290   if (has_aborted()) {
2291     return;
2292   }
2293 
2294   // Decide what the target size is, depending whether we're going to
2295   // drain it partially (so that other tasks can steal if they run out
2296   // of things to do) or totally (at the very end).
2297   size_t target_size;
2298   if (partially) {
2299     target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize);
2300   } else {
2301     target_size = 0;
2302   }
2303 
2304   if (_task_queue->size() > target_size) {
2305     G1TaskQueueEntry entry;
2306     bool ret = _task_queue->pop_local(entry);
2307     while (ret) {
2308       scan_task_entry(entry);
2309       if (_task_queue->size() <= target_size || has_aborted()) {
2310         ret = false;
2311       } else {
2312         ret = _task_queue->pop_local(entry);
2313       }
2314     }
2315   }
2316 }
2317 
2318 void G1CMTask::drain_global_stack(bool partially) {
2319   if (has_aborted()) return;
2320 
2321   // We have a policy to drain the local queue before we attempt to
2322   // drain the global stack.
2323   assert(partially || _task_queue->size() == 0, "invariant");
2324 
2325   // Decide what the target size is, depending whether we're going to
2326   // drain it partially (so that other tasks can steal if they run out
2327   // of things to do) or totally (at the very end).
2328   // Notice that when draining the global mark stack partially, due to the racyness
2329   // of the mark stack size update we might in fact drop below the target. But,
2330   // this is not a problem.
2331   // In case of total draining, we simply process until the global mark stack is
2332   // totally empty, disregarding the size counter.
2333   if (partially) {
2334     size_t const target_size = _cm->partial_mark_stack_size_target();
2335     while (!has_aborted() && _cm->mark_stack_size() > target_size) {
2336       if (get_entries_from_global_stack()) {
2337         drain_local_queue(partially);
2338       }
2339     }
2340   } else {
2341     while (!has_aborted() && get_entries_from_global_stack()) {
2342       drain_local_queue(partially);
2343     }
2344   }
2345 }
2346 
2347 // SATB Queue has several assumptions on whether to call the par or
2348 // non-par versions of the methods. this is why some of the code is
2349 // replicated. We should really get rid of the single-threaded version
2350 // of the code to simplify things.
2351 void G1CMTask::drain_satb_buffers() {
2352   if (has_aborted()) return;
2353 
2354   // We set this so that the regular clock knows that we're in the
2355   // middle of draining buffers and doesn't set the abort flag when it
2356   // notices that SATB buffers are available for draining. It'd be
2357   // very counter productive if it did that. :-)
2358   _draining_satb_buffers = true;
2359 
2360   G1CMSATBBufferClosure satb_cl(this, _g1h);
2361   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
2362 
2363   // This keeps claiming and applying the closure to completed buffers
2364   // until we run out of buffers or we need to abort.
2365   while (!has_aborted() &&
2366          satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) {
2367     regular_clock_call();
2368   }
2369 
2370   _draining_satb_buffers = false;
2371 
2372   assert(has_aborted() ||
2373          _concurrent ||
2374          satb_mq_set.completed_buffers_num() == 0, "invariant");
2375 
2376   // again, this was a potentially expensive operation, decrease the
2377   // limits to get the regular clock call early
2378   decrease_limits();
2379 }
2380 
2381 void G1CMTask::print_stats() {
2382   log_debug(gc, stats)("Marking Stats, task = %u, calls = %u",
2383                        _worker_id, _calls);
2384   log_debug(gc, stats)("  Elapsed time = %1.2lfms, Termination time = %1.2lfms",
2385                        _elapsed_time_ms, _termination_time_ms);
2386   log_debug(gc, stats)("  Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms",
2387                        _step_times_ms.num(), _step_times_ms.avg(),
2388                        _step_times_ms.sd());
2389   log_debug(gc, stats)("                    max = %1.2lfms, total = %1.2lfms",
2390                        _step_times_ms.maximum(), _step_times_ms.sum());
2391 }
2392 
2393 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry) {
2394   return _task_queues->steal(worker_id, hash_seed, task_entry);
2395 }
2396 
2397 /*****************************************************************************
2398 
2399     The do_marking_step(time_target_ms, ...) method is the building
2400     block of the parallel marking framework. It can be called in parallel
2401     with other invocations of do_marking_step() on different tasks
2402     (but only one per task, obviously) and concurrently with the
2403     mutator threads, or during remark, hence it eliminates the need
2404     for two versions of the code. When called during remark, it will
2405     pick up from where the task left off during the concurrent marking
2406     phase. Interestingly, tasks are also claimable during evacuation
2407     pauses too, since do_marking_step() ensures that it aborts before
2408     it needs to yield.
2409 
2410     The data structures that it uses to do marking work are the
2411     following:
2412 
2413       (1) Marking Bitmap. If there are gray objects that appear only
2414       on the bitmap (this happens either when dealing with an overflow
2415       or when the initial marking phase has simply marked the roots
2416       and didn't push them on the stack), then tasks claim heap
2417       regions whose bitmap they then scan to find gray objects. A
2418       global finger indicates where the end of the last claimed region
2419       is. A local finger indicates how far into the region a task has
2420       scanned. The two fingers are used to determine how to gray an
2421       object (i.e. whether simply marking it is OK, as it will be
2422       visited by a task in the future, or whether it needs to be also
2423       pushed on a stack).
2424 
2425       (2) Local Queue. The local queue of the task which is accessed
2426       reasonably efficiently by the task. Other tasks can steal from
2427       it when they run out of work. Throughout the marking phase, a
2428       task attempts to keep its local queue short but not totally
2429       empty, so that entries are available for stealing by other
2430       tasks. Only when there is no more work, a task will totally
2431       drain its local queue.
2432 
2433       (3) Global Mark Stack. This handles local queue overflow. During
2434       marking only sets of entries are moved between it and the local
2435       queues, as access to it requires a mutex and more fine-grain
2436       interaction with it which might cause contention. If it
2437       overflows, then the marking phase should restart and iterate
2438       over the bitmap to identify gray objects. Throughout the marking
2439       phase, tasks attempt to keep the global mark stack at a small
2440       length but not totally empty, so that entries are available for
2441       popping by other tasks. Only when there is no more work, tasks
2442       will totally drain the global mark stack.
2443 
2444       (4) SATB Buffer Queue. This is where completed SATB buffers are
2445       made available. Buffers are regularly removed from this queue
2446       and scanned for roots, so that the queue doesn't get too
2447       long. During remark, all completed buffers are processed, as
2448       well as the filled in parts of any uncompleted buffers.
2449 
2450     The do_marking_step() method tries to abort when the time target
2451     has been reached. There are a few other cases when the
2452     do_marking_step() method also aborts:
2453 
2454       (1) When the marking phase has been aborted (after a Full GC).
2455 
2456       (2) When a global overflow (on the global stack) has been
2457       triggered. Before the task aborts, it will actually sync up with
2458       the other tasks to ensure that all the marking data structures
2459       (local queues, stacks, fingers etc.)  are re-initialized so that
2460       when do_marking_step() completes, the marking phase can
2461       immediately restart.
2462 
2463       (3) When enough completed SATB buffers are available. The
2464       do_marking_step() method only tries to drain SATB buffers right
2465       at the beginning. So, if enough buffers are available, the
2466       marking step aborts and the SATB buffers are processed at
2467       the beginning of the next invocation.
2468 
2469       (4) To yield. when we have to yield then we abort and yield
2470       right at the end of do_marking_step(). This saves us from a lot
2471       of hassle as, by yielding we might allow a Full GC. If this
2472       happens then objects will be compacted underneath our feet, the
2473       heap might shrink, etc. We save checking for this by just
2474       aborting and doing the yield right at the end.
2475 
2476     From the above it follows that the do_marking_step() method should
2477     be called in a loop (or, otherwise, regularly) until it completes.
2478 
2479     If a marking step completes without its has_aborted() flag being
2480     true, it means it has completed the current marking phase (and
2481     also all other marking tasks have done so and have all synced up).
2482 
2483     A method called regular_clock_call() is invoked "regularly" (in
2484     sub ms intervals) throughout marking. It is this clock method that
2485     checks all the abort conditions which were mentioned above and
2486     decides when the task should abort. A work-based scheme is used to
2487     trigger this clock method: when the number of object words the
2488     marking phase has scanned or the number of references the marking
2489     phase has visited reach a given limit. Additional invocations to
2490     the method clock have been planted in a few other strategic places
2491     too. The initial reason for the clock method was to avoid calling
2492     vtime too regularly, as it is quite expensive. So, once it was in
2493     place, it was natural to piggy-back all the other conditions on it
2494     too and not constantly check them throughout the code.
2495 
2496     If do_termination is true then do_marking_step will enter its
2497     termination protocol.
2498 
2499     The value of is_serial must be true when do_marking_step is being
2500     called serially (i.e. by the VMThread) and do_marking_step should
2501     skip any synchronization in the termination and overflow code.
2502     Examples include the serial remark code and the serial reference
2503     processing closures.
2504 
2505     The value of is_serial must be false when do_marking_step is
2506     being called by any of the worker threads in a work gang.
2507     Examples include the concurrent marking code (CMMarkingTask),
2508     the MT remark code, and the MT reference processing closures.
2509 
2510  *****************************************************************************/
2511 
2512 void G1CMTask::do_marking_step(double time_target_ms,
2513                                bool do_termination,
2514                                bool is_serial) {
2515   assert(time_target_ms >= 1.0, "minimum granularity is 1ms");
2516   assert(_concurrent == _cm->concurrent(), "they should be the same");
2517 
2518   _start_time_ms = os::elapsedVTime() * 1000.0;
2519 
2520   // If do_stealing is true then do_marking_step will attempt to
2521   // steal work from the other G1CMTasks. It only makes sense to
2522   // enable stealing when the termination protocol is enabled
2523   // and do_marking_step() is not being called serially.
2524   bool do_stealing = do_termination && !is_serial;
2525 
2526   double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms);
2527   _time_target_ms = time_target_ms - diff_prediction_ms;
2528 
2529   // set up the variables that are used in the work-based scheme to
2530   // call the regular clock method
2531   _words_scanned = 0;
2532   _refs_reached  = 0;
2533   recalculate_limits();
2534 
2535   // clear all flags
2536   clear_has_aborted();
2537   _has_timed_out = false;
2538   _draining_satb_buffers = false;
2539 
2540   ++_calls;
2541 
2542   // Set up the bitmap and oop closures. Anything that uses them is
2543   // eventually called from this method, so it is OK to allocate these
2544   // statically.
2545   G1CMBitMapClosure bitmap_closure(this, _cm);
2546   G1CMOopClosure    cm_oop_closure(_g1h, _cm, this);
2547   set_cm_oop_closure(&cm_oop_closure);
2548 
2549   if (_cm->has_overflown()) {
2550     // This can happen if the mark stack overflows during a GC pause
2551     // and this task, after a yield point, restarts. We have to abort
2552     // as we need to get into the overflow protocol which happens
2553     // right at the end of this task.
2554     set_has_aborted();
2555   }
2556 
2557   // First drain any available SATB buffers. After this, we will not
2558   // look at SATB buffers before the next invocation of this method.
2559   // If enough completed SATB buffers are queued up, the regular clock
2560   // will abort this task so that it restarts.
2561   drain_satb_buffers();
2562   // ...then partially drain the local queue and the global stack
2563   drain_local_queue(true);
2564   drain_global_stack(true);
2565 
2566   do {
2567     if (!has_aborted() && _curr_region != NULL) {
2568       // This means that we're already holding on to a region.
2569       assert(_finger != NULL, "if region is not NULL, then the finger "
2570              "should not be NULL either");
2571 
2572       // We might have restarted this task after an evacuation pause
2573       // which might have evacuated the region we're holding on to
2574       // underneath our feet. Let's read its limit again to make sure
2575       // that we do not iterate over a region of the heap that
2576       // contains garbage (update_region_limit() will also move
2577       // _finger to the start of the region if it is found empty).
2578       update_region_limit();
2579       // We will start from _finger not from the start of the region,
2580       // as we might be restarting this task after aborting half-way
2581       // through scanning this region. In this case, _finger points to
2582       // the address where we last found a marked object. If this is a
2583       // fresh region, _finger points to start().
2584       MemRegion mr = MemRegion(_finger, _region_limit);
2585 
2586       assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(),
2587              "humongous regions should go around loop once only");
2588 
2589       // Some special cases:
2590       // If the memory region is empty, we can just give up the region.
2591       // If the current region is humongous then we only need to check
2592       // the bitmap for the bit associated with the start of the object,
2593       // scan the object if it's live, and give up the region.
2594       // Otherwise, let's iterate over the bitmap of the part of the region
2595       // that is left.
2596       // If the iteration is successful, give up the region.
2597       if (mr.is_empty()) {
2598         giveup_current_region();
2599         regular_clock_call();
2600       } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) {
2601         if (_next_mark_bitmap->is_marked(mr.start())) {
2602           // The object is marked - apply the closure
2603           bitmap_closure.do_addr(mr.start());
2604         }
2605         // Even if this task aborted while scanning the humongous object
2606         // we can (and should) give up the current region.
2607         giveup_current_region();
2608         regular_clock_call();
2609       } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) {
2610         giveup_current_region();
2611         regular_clock_call();
2612       } else {
2613         assert(has_aborted(), "currently the only way to do so");
2614         // The only way to abort the bitmap iteration is to return
2615         // false from the do_bit() method. However, inside the
2616         // do_bit() method we move the _finger to point to the
2617         // object currently being looked at. So, if we bail out, we
2618         // have definitely set _finger to something non-null.
2619         assert(_finger != NULL, "invariant");
2620 
2621         // Region iteration was actually aborted. So now _finger
2622         // points to the address of the object we last scanned. If we
2623         // leave it there, when we restart this task, we will rescan
2624         // the object. It is easy to avoid this. We move the finger by
2625         // enough to point to the next possible object header.
2626         assert(_finger < _region_limit, "invariant");
2627         HeapWord* const new_finger = _finger + ((oop)_finger)->size();
2628         // Check if bitmap iteration was aborted while scanning the last object
2629         if (new_finger >= _region_limit) {
2630           giveup_current_region();
2631         } else {
2632           move_finger_to(new_finger);
2633         }
2634       }
2635     }
2636     // At this point we have either completed iterating over the
2637     // region we were holding on to, or we have aborted.
2638 
2639     // We then partially drain the local queue and the global stack.
2640     // (Do we really need this?)
2641     drain_local_queue(true);
2642     drain_global_stack(true);
2643 
2644     // Read the note on the claim_region() method on why it might
2645     // return NULL with potentially more regions available for
2646     // claiming and why we have to check out_of_regions() to determine
2647     // whether we're done or not.
2648     while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) {
2649       // We are going to try to claim a new region. We should have
2650       // given up on the previous one.
2651       // Separated the asserts so that we know which one fires.
2652       assert(_curr_region  == NULL, "invariant");
2653       assert(_finger       == NULL, "invariant");
2654       assert(_region_limit == NULL, "invariant");
2655       HeapRegion* claimed_region = _cm->claim_region(_worker_id);
2656       if (claimed_region != NULL) {
2657         // Yes, we managed to claim one
2658         setup_for_region(claimed_region);
2659         assert(_curr_region == claimed_region, "invariant");
2660       }
2661       // It is important to call the regular clock here. It might take
2662       // a while to claim a region if, for example, we hit a large
2663       // block of empty regions. So we need to call the regular clock
2664       // method once round the loop to make sure it's called
2665       // frequently enough.
2666       regular_clock_call();
2667     }
2668 
2669     if (!has_aborted() && _curr_region == NULL) {
2670       assert(_cm->out_of_regions(),
2671              "at this point we should be out of regions");
2672     }
2673   } while ( _curr_region != NULL && !has_aborted());
2674 
2675   if (!has_aborted()) {
2676     // We cannot check whether the global stack is empty, since other
2677     // tasks might be pushing objects to it concurrently.
2678     assert(_cm->out_of_regions(),
2679            "at this point we should be out of regions");
2680     // Try to reduce the number of available SATB buffers so that
2681     // remark has less work to do.
2682     drain_satb_buffers();
2683   }
2684 
2685   // Since we've done everything else, we can now totally drain the
2686   // local queue and global stack.
2687   drain_local_queue(false);
2688   drain_global_stack(false);
2689 
2690   // Attempt at work stealing from other task's queues.
2691   if (do_stealing && !has_aborted()) {
2692     // We have not aborted. This means that we have finished all that
2693     // we could. Let's try to do some stealing...
2694 
2695     // We cannot check whether the global stack is empty, since other
2696     // tasks might be pushing objects to it concurrently.
2697     assert(_cm->out_of_regions() && _task_queue->size() == 0,
2698            "only way to reach here");
2699     while (!has_aborted()) {
2700       G1TaskQueueEntry entry;
2701       if (_cm->try_stealing(_worker_id, &_hash_seed, entry)) {
2702         scan_task_entry(entry);
2703 
2704         // And since we're towards the end, let's totally drain the
2705         // local queue and global stack.
2706         drain_local_queue(false);
2707         drain_global_stack(false);
2708       } else {
2709         break;
2710       }
2711     }
2712   }
2713 
2714   // We still haven't aborted. Now, let's try to get into the
2715   // termination protocol.
2716   if (do_termination && !has_aborted()) {
2717     // We cannot check whether the global stack is empty, since other
2718     // tasks might be concurrently pushing objects on it.
2719     // Separated the asserts so that we know which one fires.
2720     assert(_cm->out_of_regions(), "only way to reach here");
2721     assert(_task_queue->size() == 0, "only way to reach here");
2722     _termination_start_time_ms = os::elapsedVTime() * 1000.0;
2723 
2724     // The G1CMTask class also extends the TerminatorTerminator class,
2725     // hence its should_exit_termination() method will also decide
2726     // whether to exit the termination protocol or not.
2727     bool finished = (is_serial ||
2728                      _cm->terminator()->offer_termination(this));
2729     double termination_end_time_ms = os::elapsedVTime() * 1000.0;
2730     _termination_time_ms +=
2731       termination_end_time_ms - _termination_start_time_ms;
2732 
2733     if (finished) {
2734       // We're all done.
2735 
2736       if (_worker_id == 0) {
2737         // Let's allow task 0 to do this
2738         if (_concurrent) {
2739           assert(_cm->concurrent_marking_in_progress(), "invariant");
2740           // We need to set this to false before the next
2741           // safepoint. This way we ensure that the marking phase
2742           // doesn't observe any more heap expansions.
2743           _cm->clear_concurrent_marking_in_progress();
2744         }
2745       }
2746 
2747       // We can now guarantee that the global stack is empty, since
2748       // all other tasks have finished. We separated the guarantees so
2749       // that, if a condition is false, we can immediately find out
2750       // which one.
2751       guarantee(_cm->out_of_regions(), "only way to reach here");
2752       guarantee(_cm->mark_stack_empty(), "only way to reach here");
2753       guarantee(_task_queue->size() == 0, "only way to reach here");
2754       guarantee(!_cm->has_overflown(), "only way to reach here");
2755     } else {
2756       // Apparently there's more work to do. Let's abort this task. It
2757       // will restart it and we can hopefully find more things to do.
2758       set_has_aborted();
2759     }
2760   }
2761 
2762   // Mainly for debugging purposes to make sure that a pointer to the
2763   // closure which was statically allocated in this frame doesn't
2764   // escape it by accident.
2765   set_cm_oop_closure(NULL);
2766   double end_time_ms = os::elapsedVTime() * 1000.0;
2767   double elapsed_time_ms = end_time_ms - _start_time_ms;
2768   // Update the step history.
2769   _step_times_ms.add(elapsed_time_ms);
2770 
2771   if (has_aborted()) {
2772     // The task was aborted for some reason.
2773     if (_has_timed_out) {
2774       double diff_ms = elapsed_time_ms - _time_target_ms;
2775       // Keep statistics of how well we did with respect to hitting
2776       // our target only if we actually timed out (if we aborted for
2777       // other reasons, then the results might get skewed).
2778       _marking_step_diffs_ms.add(diff_ms);
2779     }
2780 
2781     if (_cm->has_overflown()) {
2782       // This is the interesting one. We aborted because a global
2783       // overflow was raised. This means we have to restart the
2784       // marking phase and start iterating over regions. However, in
2785       // order to do this we have to make sure that all tasks stop
2786       // what they are doing and re-initialize in a safe manner. We
2787       // will achieve this with the use of two barrier sync points.
2788 
2789       if (!is_serial) {
2790         // We only need to enter the sync barrier if being called
2791         // from a parallel context
2792         _cm->enter_first_sync_barrier(_worker_id);
2793 
2794         // When we exit this sync barrier we know that all tasks have
2795         // stopped doing marking work. So, it's now safe to
2796         // re-initialize our data structures. At the end of this method,
2797         // task 0 will clear the global data structures.
2798       }
2799 
2800       // We clear the local state of this task...
2801       clear_region_fields();
2802 
2803       if (!is_serial) {
2804         // ...and enter the second barrier.
2805         _cm->enter_second_sync_barrier(_worker_id);
2806       }
2807       // At this point, if we're during the concurrent phase of
2808       // marking, everything has been re-initialized and we're
2809       // ready to restart.
2810     }
2811   }
2812 }
2813 
2814 G1CMTask::G1CMTask(uint worker_id, G1ConcurrentMark* cm, G1CMTaskQueue* task_queue) :
2815   _objArray_processor(this),
2816   _worker_id(worker_id),
2817   _g1h(G1CollectedHeap::heap()),
2818   _cm(cm),
2819   _next_mark_bitmap(NULL),
2820   _task_queue(task_queue),
2821   _calls(0),
2822   _time_target_ms(0.0),
2823   _start_time_ms(0.0),
2824   _cm_oop_closure(NULL),
2825   _curr_region(NULL),
2826   _finger(NULL),
2827   _region_limit(NULL),
2828   _words_scanned(0),
2829   _words_scanned_limit(0),
2830   _real_words_scanned_limit(0),
2831   _refs_reached(0),
2832   _refs_reached_limit(0),
2833   _real_refs_reached_limit(0),
2834   _hash_seed(17),
2835   _has_aborted(false),
2836   _has_timed_out(false),
2837   _draining_satb_buffers(false),
2838   _step_times_ms(),
2839   _elapsed_time_ms(0.0),
2840   _termination_time_ms(0.0),
2841   _termination_start_time_ms(0.0),
2842   _concurrent(false),
2843   _marking_step_diffs_ms()
2844 {
2845   guarantee(task_queue != NULL, "invariant");
2846 
2847   _marking_step_diffs_ms.add(0.5);
2848 }
2849 
2850 // These are formatting macros that are used below to ensure
2851 // consistent formatting. The *_H_* versions are used to format the
2852 // header for a particular value and they should be kept consistent
2853 // with the corresponding macro. Also note that most of the macros add
2854 // the necessary white space (as a prefix) which makes them a bit
2855 // easier to compose.
2856 
2857 // All the output lines are prefixed with this string to be able to
2858 // identify them easily in a large log file.
2859 #define G1PPRL_LINE_PREFIX            "###"
2860 
2861 #define G1PPRL_ADDR_BASE_FORMAT    " " PTR_FORMAT "-" PTR_FORMAT
2862 #ifdef _LP64
2863 #define G1PPRL_ADDR_BASE_H_FORMAT  " %37s"
2864 #else // _LP64
2865 #define G1PPRL_ADDR_BASE_H_FORMAT  " %21s"
2866 #endif // _LP64
2867 
2868 // For per-region info
2869 #define G1PPRL_TYPE_FORMAT            "   %-4s"
2870 #define G1PPRL_TYPE_H_FORMAT          "   %4s"
2871 #define G1PPRL_BYTE_FORMAT            "  " SIZE_FORMAT_W(9)
2872 #define G1PPRL_BYTE_H_FORMAT          "  %9s"
2873 #define G1PPRL_DOUBLE_FORMAT          "  %14.1f"
2874 #define G1PPRL_DOUBLE_H_FORMAT        "  %14s"
2875 
2876 // For summary info
2877 #define G1PPRL_SUM_ADDR_FORMAT(tag)    "  " tag ":" G1PPRL_ADDR_BASE_FORMAT
2878 #define G1PPRL_SUM_BYTE_FORMAT(tag)    "  " tag ": " SIZE_FORMAT
2879 #define G1PPRL_SUM_MB_FORMAT(tag)      "  " tag ": %1.2f MB"
2880 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%"
2881 
2882 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) :
2883   _total_used_bytes(0), _total_capacity_bytes(0),
2884   _total_prev_live_bytes(0), _total_next_live_bytes(0),
2885   _total_remset_bytes(0), _total_strong_code_roots_bytes(0)
2886 {
2887   G1CollectedHeap* g1h = G1CollectedHeap::heap();
2888   MemRegion g1_reserved = g1h->g1_reserved();
2889   double now = os::elapsedTime();
2890 
2891   // Print the header of the output.
2892   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now);
2893   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP"
2894                           G1PPRL_SUM_ADDR_FORMAT("reserved")
2895                           G1PPRL_SUM_BYTE_FORMAT("region-size"),
2896                           p2i(g1_reserved.start()), p2i(g1_reserved.end()),
2897                           HeapRegion::GrainBytes);
2898   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2899   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2900                           G1PPRL_TYPE_H_FORMAT
2901                           G1PPRL_ADDR_BASE_H_FORMAT
2902                           G1PPRL_BYTE_H_FORMAT
2903                           G1PPRL_BYTE_H_FORMAT
2904                           G1PPRL_BYTE_H_FORMAT
2905                           G1PPRL_DOUBLE_H_FORMAT
2906                           G1PPRL_BYTE_H_FORMAT
2907                           G1PPRL_BYTE_H_FORMAT,
2908                           "type", "address-range",
2909                           "used", "prev-live", "next-live", "gc-eff",
2910                           "remset", "code-roots");
2911   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2912                           G1PPRL_TYPE_H_FORMAT
2913                           G1PPRL_ADDR_BASE_H_FORMAT
2914                           G1PPRL_BYTE_H_FORMAT
2915                           G1PPRL_BYTE_H_FORMAT
2916                           G1PPRL_BYTE_H_FORMAT
2917                           G1PPRL_DOUBLE_H_FORMAT
2918                           G1PPRL_BYTE_H_FORMAT
2919                           G1PPRL_BYTE_H_FORMAT,
2920                           "", "",
2921                           "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)",
2922                           "(bytes)", "(bytes)");
2923 }
2924 
2925 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) {
2926   const char* type       = r->get_type_str();
2927   HeapWord* bottom       = r->bottom();
2928   HeapWord* end          = r->end();
2929   size_t capacity_bytes  = r->capacity();
2930   size_t used_bytes      = r->used();
2931   size_t prev_live_bytes = r->live_bytes();
2932   size_t next_live_bytes = r->next_live_bytes();
2933   double gc_eff          = r->gc_efficiency();
2934   size_t remset_bytes    = r->rem_set()->mem_size();
2935   size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size();
2936 
2937   _total_used_bytes      += used_bytes;
2938   _total_capacity_bytes  += capacity_bytes;
2939   _total_prev_live_bytes += prev_live_bytes;
2940   _total_next_live_bytes += next_live_bytes;
2941   _total_remset_bytes    += remset_bytes;
2942   _total_strong_code_roots_bytes += strong_code_roots_bytes;
2943 
2944   // Print a line for this particular region.
2945   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2946                           G1PPRL_TYPE_FORMAT
2947                           G1PPRL_ADDR_BASE_FORMAT
2948                           G1PPRL_BYTE_FORMAT
2949                           G1PPRL_BYTE_FORMAT
2950                           G1PPRL_BYTE_FORMAT
2951                           G1PPRL_DOUBLE_FORMAT
2952                           G1PPRL_BYTE_FORMAT
2953                           G1PPRL_BYTE_FORMAT,
2954                           type, p2i(bottom), p2i(end),
2955                           used_bytes, prev_live_bytes, next_live_bytes, gc_eff,
2956                           remset_bytes, strong_code_roots_bytes);
2957 
2958   return false;
2959 }
2960 
2961 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() {
2962   // add static memory usages to remembered set sizes
2963   _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size();
2964   // Print the footer of the output.
2965   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX);
2966   log_trace(gc, liveness)(G1PPRL_LINE_PREFIX
2967                          " SUMMARY"
2968                          G1PPRL_SUM_MB_FORMAT("capacity")
2969                          G1PPRL_SUM_MB_PERC_FORMAT("used")
2970                          G1PPRL_SUM_MB_PERC_FORMAT("prev-live")
2971                          G1PPRL_SUM_MB_PERC_FORMAT("next-live")
2972                          G1PPRL_SUM_MB_FORMAT("remset")
2973                          G1PPRL_SUM_MB_FORMAT("code-roots"),
2974                          bytes_to_mb(_total_capacity_bytes),
2975                          bytes_to_mb(_total_used_bytes),
2976                          percent_of(_total_used_bytes, _total_capacity_bytes),
2977                          bytes_to_mb(_total_prev_live_bytes),
2978                          percent_of(_total_prev_live_bytes, _total_capacity_bytes),
2979                          bytes_to_mb(_total_next_live_bytes),
2980                          percent_of(_total_next_live_bytes, _total_capacity_bytes),
2981                          bytes_to_mb(_total_remset_bytes),
2982                          bytes_to_mb(_total_strong_code_roots_bytes));
2983 }