1 /* 2 * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/metadataOnStackMark.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc/g1/g1BarrierSet.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/g1ConcurrentMarkThread.inline.hpp" 34 #include "gc/g1/g1HeapVerifier.hpp" 35 #include "gc/g1/g1OopClosures.inline.hpp" 36 #include "gc/g1/g1Policy.hpp" 37 #include "gc/g1/g1RegionMarkStatsCache.inline.hpp" 38 #include "gc/g1/g1StringDedup.hpp" 39 #include "gc/g1/g1ThreadLocalData.hpp" 40 #include "gc/g1/heapRegion.inline.hpp" 41 #include "gc/g1/heapRegionRemSet.hpp" 42 #include "gc/g1/heapRegionSet.inline.hpp" 43 #include "gc/shared/adaptiveSizePolicy.hpp" 44 #include "gc/shared/gcId.hpp" 45 #include "gc/shared/gcTimer.hpp" 46 #include "gc/shared/gcTrace.hpp" 47 #include "gc/shared/gcTraceTime.inline.hpp" 48 #include "gc/shared/genOopClosures.inline.hpp" 49 #include "gc/shared/referencePolicy.hpp" 50 #include "gc/shared/strongRootsScope.hpp" 51 #include "gc/shared/suspendibleThreadSet.hpp" 52 #include "gc/shared/taskqueue.inline.hpp" 53 #include "gc/shared/vmGCOperations.hpp" 54 #include "gc/shared/weakProcessor.hpp" 55 #include "include/jvm.h" 56 #include "logging/log.hpp" 57 #include "memory/allocation.hpp" 58 #include "memory/resourceArea.hpp" 59 #include "oops/access.inline.hpp" 60 #include "oops/oop.inline.hpp" 61 #include "runtime/atomic.hpp" 62 #include "runtime/handles.inline.hpp" 63 #include "runtime/java.hpp" 64 #include "runtime/prefetch.inline.hpp" 65 #include "services/memTracker.hpp" 66 #include "utilities/align.hpp" 67 #include "utilities/growableArray.hpp" 68 69 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) { 70 assert(addr < _cm->finger(), "invariant"); 71 assert(addr >= _task->finger(), "invariant"); 72 73 // We move that task's local finger along. 74 _task->move_finger_to(addr); 75 76 _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr))); 77 // we only partially drain the local queue and global stack 78 _task->drain_local_queue(true); 79 _task->drain_global_stack(true); 80 81 // if the has_aborted flag has been raised, we need to bail out of 82 // the iteration 83 return !_task->has_aborted(); 84 } 85 86 G1CMMarkStack::G1CMMarkStack() : 87 _max_chunk_capacity(0), 88 _base(NULL), 89 _chunk_capacity(0) { 90 set_empty(); 91 } 92 93 bool G1CMMarkStack::resize(size_t new_capacity) { 94 assert(is_empty(), "Only resize when stack is empty."); 95 assert(new_capacity <= _max_chunk_capacity, 96 "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity); 97 98 TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC); 99 100 if (new_base == NULL) { 101 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)); 102 return false; 103 } 104 // Release old mapping. 105 if (_base != NULL) { 106 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity); 107 } 108 109 _base = new_base; 110 _chunk_capacity = new_capacity; 111 set_empty(); 112 113 return true; 114 } 115 116 size_t G1CMMarkStack::capacity_alignment() { 117 return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry); 118 } 119 120 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) { 121 guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized."); 122 123 size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry); 124 125 _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar; 126 size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar; 127 128 guarantee(initial_chunk_capacity <= _max_chunk_capacity, 129 "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT, 130 _max_chunk_capacity, 131 initial_chunk_capacity); 132 133 log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT, 134 initial_chunk_capacity, _max_chunk_capacity); 135 136 return resize(initial_chunk_capacity); 137 } 138 139 void G1CMMarkStack::expand() { 140 if (_chunk_capacity == _max_chunk_capacity) { 141 log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity); 142 return; 143 } 144 size_t old_capacity = _chunk_capacity; 145 // Double capacity if possible 146 size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity); 147 148 if (resize(new_capacity)) { 149 log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks", 150 old_capacity, new_capacity); 151 } else { 152 log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks", 153 old_capacity, new_capacity); 154 } 155 } 156 157 G1CMMarkStack::~G1CMMarkStack() { 158 if (_base != NULL) { 159 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity); 160 } 161 } 162 163 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) { 164 elem->next = *list; 165 *list = elem; 166 } 167 168 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) { 169 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag); 170 add_chunk_to_list(&_chunk_list, elem); 171 _chunks_in_chunk_list++; 172 } 173 174 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) { 175 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag); 176 add_chunk_to_list(&_free_list, elem); 177 } 178 179 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) { 180 TaskQueueEntryChunk* result = *list; 181 if (result != NULL) { 182 *list = (*list)->next; 183 } 184 return result; 185 } 186 187 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() { 188 MutexLockerEx x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag); 189 TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list); 190 if (result != NULL) { 191 _chunks_in_chunk_list--; 192 } 193 return result; 194 } 195 196 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() { 197 MutexLockerEx x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag); 198 return remove_chunk_from_list(&_free_list); 199 } 200 201 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() { 202 // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code. 203 // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding 204 // wraparound of _hwm. 205 if (_hwm >= _chunk_capacity) { 206 return NULL; 207 } 208 209 size_t cur_idx = Atomic::add(1u, &_hwm) - 1; 210 if (cur_idx >= _chunk_capacity) { 211 return NULL; 212 } 213 214 TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk; 215 result->next = NULL; 216 return result; 217 } 218 219 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) { 220 // Get a new chunk. 221 TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list(); 222 223 if (new_chunk == NULL) { 224 // Did not get a chunk from the free list. Allocate from backing memory. 225 new_chunk = allocate_new_chunk(); 226 227 if (new_chunk == NULL) { 228 return false; 229 } 230 } 231 232 Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry)); 233 234 add_chunk_to_chunk_list(new_chunk); 235 236 return true; 237 } 238 239 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) { 240 TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list(); 241 242 if (cur == NULL) { 243 return false; 244 } 245 246 Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry)); 247 248 add_chunk_to_free_list(cur); 249 return true; 250 } 251 252 void G1CMMarkStack::set_empty() { 253 _chunks_in_chunk_list = 0; 254 _hwm = 0; 255 _chunk_list = NULL; 256 _free_list = NULL; 257 } 258 259 G1CMRootRegions::G1CMRootRegions() : 260 _survivors(NULL), _cm(NULL), _scan_in_progress(false), 261 _should_abort(false), _claimed_survivor_index(0) { } 262 263 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) { 264 _survivors = survivors; 265 _cm = cm; 266 } 267 268 void G1CMRootRegions::prepare_for_scan() { 269 assert(!scan_in_progress(), "pre-condition"); 270 271 // Currently, only survivors can be root regions. 272 _claimed_survivor_index = 0; 273 _scan_in_progress = _survivors->regions()->is_nonempty(); 274 _should_abort = false; 275 } 276 277 HeapRegion* G1CMRootRegions::claim_next() { 278 if (_should_abort) { 279 // If someone has set the should_abort flag, we return NULL to 280 // force the caller to bail out of their loop. 281 return NULL; 282 } 283 284 // Currently, only survivors can be root regions. 285 const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions(); 286 287 int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1; 288 if (claimed_index < survivor_regions->length()) { 289 return survivor_regions->at(claimed_index); 290 } 291 return NULL; 292 } 293 294 uint G1CMRootRegions::num_root_regions() const { 295 return (uint)_survivors->regions()->length(); 296 } 297 298 void G1CMRootRegions::notify_scan_done() { 299 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 300 _scan_in_progress = false; 301 RootRegionScan_lock->notify_all(); 302 } 303 304 void G1CMRootRegions::cancel_scan() { 305 notify_scan_done(); 306 } 307 308 void G1CMRootRegions::scan_finished() { 309 assert(scan_in_progress(), "pre-condition"); 310 311 // Currently, only survivors can be root regions. 312 if (!_should_abort) { 313 assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index); 314 assert((uint)_claimed_survivor_index >= _survivors->length(), 315 "we should have claimed all survivors, claimed index = %u, length = %u", 316 (uint)_claimed_survivor_index, _survivors->length()); 317 } 318 319 notify_scan_done(); 320 } 321 322 bool G1CMRootRegions::wait_until_scan_finished() { 323 if (!scan_in_progress()) { 324 return false; 325 } 326 327 { 328 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 329 while (scan_in_progress()) { 330 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 331 } 332 } 333 return true; 334 } 335 336 // Returns the maximum number of workers to be used in a concurrent 337 // phase based on the number of GC workers being used in a STW 338 // phase. 339 static uint scale_concurrent_worker_threads(uint num_gc_workers) { 340 return MAX2((num_gc_workers + 2) / 4, 1U); 341 } 342 343 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, 344 G1RegionToSpaceMapper* prev_bitmap_storage, 345 G1RegionToSpaceMapper* next_bitmap_storage) : 346 // _cm_thread set inside the constructor 347 _g1h(g1h), 348 _completed_initialization(false), 349 350 _mark_bitmap_1(), 351 _mark_bitmap_2(), 352 _prev_mark_bitmap(&_mark_bitmap_1), 353 _next_mark_bitmap(&_mark_bitmap_2), 354 355 _heap(_g1h->reserved_region()), 356 357 _root_regions(), 358 359 _global_mark_stack(), 360 361 // _finger set in set_non_marking_state 362 363 _worker_id_offset(DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads), 364 _max_num_tasks(ParallelGCThreads), 365 // _num_active_tasks set in set_non_marking_state() 366 // _tasks set inside the constructor 367 368 _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)), 369 _terminator(ParallelTaskTerminator((int) _max_num_tasks, _task_queues)), 370 371 _first_overflow_barrier_sync(), 372 _second_overflow_barrier_sync(), 373 374 _has_overflown(false), 375 _concurrent(false), 376 _has_aborted(false), 377 _restart_for_overflow(false), 378 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 379 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()), 380 381 // _verbose_level set below 382 383 _init_times(), 384 _remark_times(), 385 _remark_mark_times(), 386 _remark_weak_ref_times(), 387 _cleanup_times(), 388 _total_cleanup_time(0.0), 389 390 _accum_task_vtime(NULL), 391 392 _concurrent_workers(NULL), 393 _num_concurrent_workers(0), 394 _max_concurrent_workers(0), 395 396 _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)), 397 _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC)) 398 { 399 _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage); 400 _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage); 401 402 // Create & start ConcurrentMark thread. 403 _cm_thread = new G1ConcurrentMarkThread(this); 404 if (_cm_thread->osthread() == NULL) { 405 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread"); 406 } 407 408 assert(CGC_lock != NULL, "CGC_lock must be initialized"); 409 410 SATBMarkQueueSet& satb_qs = G1BarrierSet::satb_mark_queue_set(); 411 satb_qs.set_buffer_size(G1SATBBufferSize); 412 413 _root_regions.init(_g1h->survivor(), this); 414 415 if (FLAG_IS_DEFAULT(ConcGCThreads) || ConcGCThreads == 0) { 416 // Calculate the number of concurrent worker threads by scaling 417 // the number of parallel GC threads. 418 uint marking_thread_num = scale_concurrent_worker_threads(ParallelGCThreads); 419 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num); 420 } 421 422 assert(ConcGCThreads > 0, "ConcGCThreads have been set."); 423 if (ConcGCThreads > ParallelGCThreads) { 424 log_warning(gc)("More ConcGCThreads (%u) than ParallelGCThreads (%u).", 425 ConcGCThreads, ParallelGCThreads); 426 return; 427 } 428 429 log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset); 430 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 431 432 _num_concurrent_workers = ConcGCThreads; 433 _max_concurrent_workers = _num_concurrent_workers; 434 435 _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true); 436 _concurrent_workers->initialize_workers(); 437 438 if (FLAG_IS_DEFAULT(MarkStackSize)) { 439 size_t mark_stack_size = 440 MIN2(MarkStackSizeMax, 441 MAX2(MarkStackSize, (size_t) (_max_concurrent_workers * TASKQUEUE_SIZE))); 442 // Verify that the calculated value for MarkStackSize is in range. 443 // It would be nice to use the private utility routine from Arguments. 444 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 445 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): " 446 "must be between 1 and " SIZE_FORMAT, 447 mark_stack_size, MarkStackSizeMax); 448 return; 449 } 450 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size); 451 } else { 452 // Verify MarkStackSize is in range. 453 if (FLAG_IS_CMDLINE(MarkStackSize)) { 454 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 455 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 456 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): " 457 "must be between 1 and " SIZE_FORMAT, 458 MarkStackSize, MarkStackSizeMax); 459 return; 460 } 461 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 462 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 463 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")" 464 " or for MarkStackSizeMax (" SIZE_FORMAT ")", 465 MarkStackSize, MarkStackSizeMax); 466 return; 467 } 468 } 469 } 470 } 471 472 if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) { 473 vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack."); 474 } 475 476 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC); 477 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC); 478 479 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 480 _num_active_tasks = _max_num_tasks; 481 482 for (uint i = 0; i < _max_num_tasks; ++i) { 483 G1CMTaskQueue* task_queue = new G1CMTaskQueue(); 484 task_queue->initialize(); 485 _task_queues->register_queue(i, task_queue); 486 487 _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions()); 488 489 _accum_task_vtime[i] = 0.0; 490 } 491 492 reset_at_marking_complete(); 493 _completed_initialization = true; 494 } 495 496 void G1ConcurrentMark::reset() { 497 _has_aborted = false; 498 499 reset_marking_for_restart(); 500 501 // Reset all tasks, since different phases will use different number of active 502 // threads. So, it's easiest to have all of them ready. 503 for (uint i = 0; i < _max_num_tasks; ++i) { 504 _tasks[i]->reset(_next_mark_bitmap); 505 } 506 507 uint max_regions = _g1h->max_regions(); 508 for (uint i = 0; i < max_regions; i++) { 509 _top_at_rebuild_starts[i] = NULL; 510 _region_mark_stats[i].clear(); 511 } 512 } 513 514 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) { 515 for (uint j = 0; j < _max_num_tasks; ++j) { 516 _tasks[j]->clear_mark_stats_cache(region_idx); 517 } 518 _top_at_rebuild_starts[region_idx] = NULL; 519 _region_mark_stats[region_idx].clear(); 520 } 521 522 void G1ConcurrentMark::clear_statistics(HeapRegion* r) { 523 uint const region_idx = r->hrm_index(); 524 if (r->is_humongous()) { 525 assert(r->is_starts_humongous(), "Got humongous continues region here"); 526 uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size()); 527 for (uint j = region_idx; j < (region_idx + size_in_regions); j++) { 528 clear_statistics_in_region(j); 529 } 530 } else { 531 clear_statistics_in_region(region_idx); 532 } 533 } 534 535 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) { 536 if (bitmap->is_marked(addr)) { 537 bitmap->clear(addr); 538 } 539 } 540 541 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) { 542 assert_at_safepoint_on_vm_thread(); 543 544 // Need to clear all mark bits of the humongous object. 545 clear_mark_if_set(_prev_mark_bitmap, r->bottom()); 546 clear_mark_if_set(_next_mark_bitmap, r->bottom()); 547 548 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) { 549 return; 550 } 551 552 // Clear any statistics about the region gathered so far. 553 clear_statistics(r); 554 } 555 556 void G1ConcurrentMark::reset_marking_for_restart() { 557 _global_mark_stack.set_empty(); 558 559 // Expand the marking stack, if we have to and if we can. 560 if (has_overflown()) { 561 _global_mark_stack.expand(); 562 563 uint max_regions = _g1h->max_regions(); 564 for (uint i = 0; i < max_regions; i++) { 565 _region_mark_stats[i].clear_during_overflow(); 566 } 567 } 568 569 clear_has_overflown(); 570 _finger = _heap.start(); 571 572 for (uint i = 0; i < _max_num_tasks; ++i) { 573 G1CMTaskQueue* queue = _task_queues->queue(i); 574 queue->set_empty(); 575 } 576 } 577 578 void G1ConcurrentMark::set_concurrency(uint active_tasks) { 579 assert(active_tasks <= _max_num_tasks, "we should not have more"); 580 581 _num_active_tasks = active_tasks; 582 // Need to update the three data structures below according to the 583 // number of active threads for this phase. 584 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 585 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 586 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 587 } 588 589 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 590 set_concurrency(active_tasks); 591 592 _concurrent = concurrent; 593 594 if (!concurrent) { 595 // At this point we should be in a STW phase, and completed marking. 596 assert_at_safepoint_on_vm_thread(); 597 assert(out_of_regions(), 598 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT, 599 p2i(_finger), p2i(_heap.end())); 600 } 601 } 602 603 void G1ConcurrentMark::reset_at_marking_complete() { 604 // We set the global marking state to some default values when we're 605 // not doing marking. 606 reset_marking_for_restart(); 607 _num_active_tasks = 0; 608 } 609 610 G1ConcurrentMark::~G1ConcurrentMark() { 611 FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts); 612 FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats); 613 // The G1ConcurrentMark instance is never freed. 614 ShouldNotReachHere(); 615 } 616 617 class G1ClearBitMapTask : public AbstractGangTask { 618 public: 619 static size_t chunk_size() { return M; } 620 621 private: 622 // Heap region closure used for clearing the given mark bitmap. 623 class G1ClearBitmapHRClosure : public HeapRegionClosure { 624 private: 625 G1CMBitMap* _bitmap; 626 G1ConcurrentMark* _cm; 627 public: 628 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) { 629 } 630 631 virtual bool do_heap_region(HeapRegion* r) { 632 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize; 633 634 HeapWord* cur = r->bottom(); 635 HeapWord* const end = r->end(); 636 637 while (cur < end) { 638 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end)); 639 _bitmap->clear_range(mr); 640 641 cur += chunk_size_in_words; 642 643 // Abort iteration if after yielding the marking has been aborted. 644 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) { 645 return true; 646 } 647 // Repeat the asserts from before the start of the closure. We will do them 648 // as asserts here to minimize their overhead on the product. However, we 649 // will have them as guarantees at the beginning / end of the bitmap 650 // clearing to get some checking in the product. 651 assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant"); 652 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant"); 653 } 654 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index()); 655 656 return false; 657 } 658 }; 659 660 G1ClearBitmapHRClosure _cl; 661 HeapRegionClaimer _hr_claimer; 662 bool _suspendible; // If the task is suspendible, workers must join the STS. 663 664 public: 665 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) : 666 AbstractGangTask("G1 Clear Bitmap"), 667 _cl(bitmap, suspendible ? cm : NULL), 668 _hr_claimer(n_workers), 669 _suspendible(suspendible) 670 { } 671 672 void work(uint worker_id) { 673 SuspendibleThreadSetJoiner sts_join(_suspendible); 674 G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id); 675 } 676 677 bool is_complete() { 678 return _cl.is_complete(); 679 } 680 }; 681 682 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) { 683 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint."); 684 685 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor(); 686 size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size(); 687 688 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers()); 689 690 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield); 691 692 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks); 693 workers->run_task(&cl, num_workers); 694 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding."); 695 } 696 697 void G1ConcurrentMark::cleanup_for_next_mark() { 698 // Make sure that the concurrent mark thread looks to still be in 699 // the current cycle. 700 guarantee(cm_thread()->during_cycle(), "invariant"); 701 702 // We are finishing up the current cycle by clearing the next 703 // marking bitmap and getting it ready for the next cycle. During 704 // this time no other cycle can start. So, let's make sure that this 705 // is the case. 706 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant"); 707 708 clear_bitmap(_next_mark_bitmap, _concurrent_workers, true); 709 710 // Repeat the asserts from above. 711 guarantee(cm_thread()->during_cycle(), "invariant"); 712 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant"); 713 } 714 715 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) { 716 assert_at_safepoint_on_vm_thread(); 717 clear_bitmap(_prev_mark_bitmap, workers, false); 718 } 719 720 class CheckBitmapClearHRClosure : public HeapRegionClosure { 721 G1CMBitMap* _bitmap; 722 public: 723 CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) { 724 } 725 726 virtual bool do_heap_region(HeapRegion* r) { 727 // This closure can be called concurrently to the mutator, so we must make sure 728 // that the result of the getNextMarkedWordAddress() call is compared to the 729 // value passed to it as limit to detect any found bits. 730 // end never changes in G1. 731 HeapWord* end = r->end(); 732 return _bitmap->get_next_marked_addr(r->bottom(), end) != end; 733 } 734 }; 735 736 bool G1ConcurrentMark::next_mark_bitmap_is_clear() { 737 CheckBitmapClearHRClosure cl(_next_mark_bitmap); 738 _g1h->heap_region_iterate(&cl); 739 return cl.is_complete(); 740 } 741 742 class NoteStartOfMarkHRClosure : public HeapRegionClosure { 743 public: 744 bool do_heap_region(HeapRegion* r) { 745 r->note_start_of_marking(); 746 return false; 747 } 748 }; 749 750 void G1ConcurrentMark::pre_initial_mark() { 751 // Initialize marking structures. This has to be done in a STW phase. 752 reset(); 753 754 // For each region note start of marking. 755 NoteStartOfMarkHRClosure startcl; 756 _g1h->heap_region_iterate(&startcl); 757 } 758 759 760 void G1ConcurrentMark::post_initial_mark() { 761 // Start Concurrent Marking weak-reference discovery. 762 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 763 // enable ("weak") refs discovery 764 rp->enable_discovery(); 765 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 766 767 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 768 // This is the start of the marking cycle, we're expected all 769 // threads to have SATB queues with active set to false. 770 satb_mq_set.set_active_all_threads(true, /* new active value */ 771 false /* expected_active */); 772 773 _root_regions.prepare_for_scan(); 774 775 // update_g1_committed() will be called at the end of an evac pause 776 // when marking is on. So, it's also called at the end of the 777 // initial-mark pause to update the heap end, if the heap expands 778 // during it. No need to call it here. 779 } 780 781 /* 782 * Notice that in the next two methods, we actually leave the STS 783 * during the barrier sync and join it immediately afterwards. If we 784 * do not do this, the following deadlock can occur: one thread could 785 * be in the barrier sync code, waiting for the other thread to also 786 * sync up, whereas another one could be trying to yield, while also 787 * waiting for the other threads to sync up too. 788 * 789 * Note, however, that this code is also used during remark and in 790 * this case we should not attempt to leave / enter the STS, otherwise 791 * we'll either hit an assert (debug / fastdebug) or deadlock 792 * (product). So we should only leave / enter the STS if we are 793 * operating concurrently. 794 * 795 * Because the thread that does the sync barrier has left the STS, it 796 * is possible to be suspended for a Full GC or an evacuation pause 797 * could occur. This is actually safe, since the entering the sync 798 * barrier is one of the last things do_marking_step() does, and it 799 * doesn't manipulate any data structures afterwards. 800 */ 801 802 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 803 bool barrier_aborted; 804 { 805 SuspendibleThreadSetLeaver sts_leave(concurrent()); 806 barrier_aborted = !_first_overflow_barrier_sync.enter(); 807 } 808 809 // at this point everyone should have synced up and not be doing any 810 // more work 811 812 if (barrier_aborted) { 813 // If the barrier aborted we ignore the overflow condition and 814 // just abort the whole marking phase as quickly as possible. 815 return; 816 } 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 G1ConcurrentMark* _cm; 828 829 public: 830 void work(uint worker_id) { 831 assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread"); 832 ResourceMark rm; 833 834 double start_vtime = os::elapsedVTime(); 835 836 { 837 SuspendibleThreadSetJoiner sts_join; 838 839 assert(worker_id < _cm->active_tasks(), "invariant"); 840 841 G1CMTask* task = _cm->task(worker_id); 842 task->record_start_time(); 843 if (!_cm->has_aborted()) { 844 do { 845 task->do_marking_step(G1ConcMarkStepDurationMillis, 846 true /* do_termination */, 847 false /* is_serial*/); 848 849 _cm->do_yield_check(); 850 } while (!_cm->has_aborted() && task->has_aborted()); 851 } 852 task->record_end_time(); 853 guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant"); 854 } 855 856 double end_vtime = os::elapsedVTime(); 857 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 858 } 859 860 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) : 861 AbstractGangTask("Concurrent Mark"), _cm(cm) { } 862 863 ~G1CMConcurrentMarkingTask() { } 864 }; 865 866 uint G1ConcurrentMark::calc_active_marking_workers() { 867 uint result = 0; 868 if (!UseDynamicNumberOfGCThreads || 869 (!FLAG_IS_DEFAULT(ConcGCThreads) && 870 !ForceDynamicNumberOfGCThreads)) { 871 result = _max_concurrent_workers; 872 } else { 873 result = 874 AdaptiveSizePolicy::calc_default_active_workers(_max_concurrent_workers, 875 1, /* Minimum workers */ 876 _num_concurrent_workers, 877 Threads::number_of_non_daemon_threads()); 878 // Don't scale the result down by scale_concurrent_workers() because 879 // that scaling has already gone into "_max_concurrent_workers". 880 } 881 assert(result > 0 && result <= _max_concurrent_workers, 882 "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u", 883 _max_concurrent_workers, result); 884 return result; 885 } 886 887 void G1ConcurrentMark::scan_root_region(HeapRegion* hr, uint worker_id) { 888 // Currently, only survivors can be root regions. 889 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 890 G1RootRegionScanClosure cl(_g1h, this, worker_id); 891 892 const uintx interval = PrefetchScanIntervalInBytes; 893 HeapWord* curr = hr->bottom(); 894 const HeapWord* end = hr->top(); 895 while (curr < end) { 896 Prefetch::read(curr, interval); 897 oop obj = oop(curr); 898 int size = obj->oop_iterate_size(&cl); 899 assert(size == obj->size(), "sanity"); 900 curr += size; 901 } 902 } 903 904 class G1CMRootRegionScanTask : public AbstractGangTask { 905 G1ConcurrentMark* _cm; 906 public: 907 G1CMRootRegionScanTask(G1ConcurrentMark* cm) : 908 AbstractGangTask("G1 Root Region Scan"), _cm(cm) { } 909 910 void work(uint worker_id) { 911 assert(Thread::current()->is_ConcurrentGC_thread(), 912 "this should only be done by a conc GC thread"); 913 914 G1CMRootRegions* root_regions = _cm->root_regions(); 915 HeapRegion* hr = root_regions->claim_next(); 916 while (hr != NULL) { 917 _cm->scan_root_region(hr, worker_id); 918 hr = root_regions->claim_next(); 919 } 920 } 921 }; 922 923 void G1ConcurrentMark::scan_root_regions() { 924 // scan_in_progress() will have been set to true only if there was 925 // at least one root region to scan. So, if it's false, we 926 // should not attempt to do any further work. 927 if (root_regions()->scan_in_progress()) { 928 assert(!has_aborted(), "Aborting before root region scanning is finished not supported."); 929 930 _num_concurrent_workers = MIN2(calc_active_marking_workers(), 931 // We distribute work on a per-region basis, so starting 932 // more threads than that is useless. 933 root_regions()->num_root_regions()); 934 assert(_num_concurrent_workers <= _max_concurrent_workers, 935 "Maximum number of marking threads exceeded"); 936 937 G1CMRootRegionScanTask task(this); 938 log_debug(gc, ergo)("Running %s using %u workers for %u work units.", 939 task.name(), _num_concurrent_workers, root_regions()->num_root_regions()); 940 _concurrent_workers->run_task(&task, _num_concurrent_workers); 941 942 // It's possible that has_aborted() is true here without actually 943 // aborting the survivor scan earlier. This is OK as it's 944 // mainly used for sanity checking. 945 root_regions()->scan_finished(); 946 } 947 } 948 949 void G1ConcurrentMark::concurrent_cycle_start() { 950 _gc_timer_cm->register_gc_start(); 951 952 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start()); 953 954 _g1h->trace_heap_before_gc(_gc_tracer_cm); 955 } 956 957 void G1ConcurrentMark::concurrent_cycle_end() { 958 _g1h->collector_state()->set_clearing_next_bitmap(false); 959 960 _g1h->trace_heap_after_gc(_gc_tracer_cm); 961 962 if (has_aborted()) { 963 log_info(gc, marking)("Concurrent Mark Abort"); 964 _gc_tracer_cm->report_concurrent_mode_failure(); 965 } 966 967 _gc_timer_cm->register_gc_end(); 968 969 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 970 } 971 972 void G1ConcurrentMark::mark_from_roots() { 973 _restart_for_overflow = false; 974 975 _num_concurrent_workers = calc_active_marking_workers(); 976 977 uint active_workers = MAX2(1U, _num_concurrent_workers); 978 979 // Setting active workers is not guaranteed since fewer 980 // worker threads may currently exist and more may not be 981 // available. 982 active_workers = _concurrent_workers->update_active_workers(active_workers); 983 log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers()); 984 985 // Parallel task terminator is set in "set_concurrency_and_phase()" 986 set_concurrency_and_phase(active_workers, true /* concurrent */); 987 988 G1CMConcurrentMarkingTask marking_task(this); 989 _concurrent_workers->run_task(&marking_task); 990 print_stats(); 991 } 992 993 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) { 994 G1HeapVerifier* verifier = _g1h->verifier(); 995 996 verifier->verify_region_sets_optional(); 997 998 if (VerifyDuringGC) { 999 GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm); 1000 1001 size_t const BufLen = 512; 1002 char buffer[BufLen]; 1003 1004 jio_snprintf(buffer, BufLen, "During GC (%s)", caller); 1005 verifier->verify(type, vo, buffer); 1006 } 1007 1008 verifier->check_bitmaps(caller); 1009 } 1010 1011 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask { 1012 G1CollectedHeap* _g1h; 1013 G1ConcurrentMark* _cm; 1014 HeapRegionClaimer _hrclaimer; 1015 uint _total_selected_for_rebuild; 1016 1017 G1PrintRegionLivenessInfoClosure _cl; 1018 1019 class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure { 1020 G1CollectedHeap* _g1h; 1021 G1ConcurrentMark* _cm; 1022 1023 G1PrintRegionLivenessInfoClosure* _cl; 1024 1025 uint _num_regions_selected_for_rebuild; // The number of regions actually selected for rebuild. 1026 1027 void update_remset_before_rebuild(HeapRegion * hr) { 1028 G1RemSetTrackingPolicy* tracking_policy = _g1h->g1_policy()->remset_tracker(); 1029 1030 size_t const live_bytes = _cm->liveness(hr->hrm_index()) * HeapWordSize; 1031 bool selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes); 1032 if (selected_for_rebuild) { 1033 _num_regions_selected_for_rebuild++; 1034 } 1035 _cm->update_top_at_rebuild_start(hr); 1036 } 1037 1038 // Distribute the given words across the humongous object starting with hr and 1039 // note end of marking. 1040 void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) { 1041 uint const region_idx = hr->hrm_index(); 1042 size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size(); 1043 uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words); 1044 1045 // "Distributing" zero words means that we only note end of marking for these 1046 // regions. 1047 assert(marked_words == 0 || obj_size_in_words == marked_words, 1048 "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT, 1049 obj_size_in_words, marked_words); 1050 1051 for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) { 1052 HeapRegion* const r = _g1h->region_at(i); 1053 size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words); 1054 1055 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)", 1056 words_to_add, i, r->get_type_str()); 1057 add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize); 1058 marked_words -= words_to_add; 1059 } 1060 assert(marked_words == 0, 1061 SIZE_FORMAT " words left after distributing space across %u regions", 1062 marked_words, num_regions_in_humongous); 1063 } 1064 1065 void update_marked_bytes(HeapRegion* hr) { 1066 uint const region_idx = hr->hrm_index(); 1067 size_t const marked_words = _cm->liveness(region_idx); 1068 // The marking attributes the object's size completely to the humongous starts 1069 // region. We need to distribute this value across the entire set of regions a 1070 // humongous object spans. 1071 if (hr->is_humongous()) { 1072 assert(hr->is_starts_humongous() || marked_words == 0, 1073 "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)", 1074 marked_words, region_idx, hr->get_type_str()); 1075 if (hr->is_starts_humongous()) { 1076 distribute_marked_bytes(hr, marked_words); 1077 } 1078 } else { 1079 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str()); 1080 add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize); 1081 } 1082 } 1083 1084 void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) { 1085 hr->add_to_marked_bytes(marked_bytes); 1086 _cl->do_heap_region(hr); 1087 hr->note_end_of_marking(); 1088 } 1089 1090 public: 1091 G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) : 1092 _g1h(g1h), _cm(cm), _num_regions_selected_for_rebuild(0), _cl(cl) { } 1093 1094 virtual bool do_heap_region(HeapRegion* r) { 1095 update_remset_before_rebuild(r); 1096 update_marked_bytes(r); 1097 1098 return false; 1099 } 1100 1101 uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; } 1102 }; 1103 1104 public: 1105 G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) : 1106 AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"), 1107 _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { } 1108 1109 virtual void work(uint worker_id) { 1110 G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl); 1111 _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id); 1112 Atomic::add(update_cl.num_selected_for_rebuild(), &_total_selected_for_rebuild); 1113 } 1114 1115 uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; } 1116 1117 // Number of regions for which roughly one thread should be spawned for this work. 1118 static const uint RegionsPerThread = 384; 1119 }; 1120 1121 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure { 1122 G1CollectedHeap* _g1h; 1123 public: 1124 G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { } 1125 1126 virtual bool do_heap_region(HeapRegion* r) { 1127 _g1h->g1_policy()->remset_tracker()->update_after_rebuild(r); 1128 return false; 1129 } 1130 }; 1131 1132 void G1ConcurrentMark::remark() { 1133 assert_at_safepoint_on_vm_thread(); 1134 1135 // If a full collection has happened, we should not continue. However we might 1136 // have ended up here as the Remark VM operation has been scheduled already. 1137 if (has_aborted()) { 1138 return; 1139 } 1140 1141 G1Policy* g1p = _g1h->g1_policy(); 1142 g1p->record_concurrent_mark_remark_start(); 1143 1144 double start = os::elapsedTime(); 1145 1146 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before"); 1147 1148 { 1149 GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm); 1150 finalize_marking(); 1151 } 1152 1153 double mark_work_end = os::elapsedTime(); 1154 1155 bool const mark_finished = !has_overflown(); 1156 if (mark_finished) { 1157 weak_refs_work(false /* clear_all_soft_refs */); 1158 1159 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 1160 // We're done with marking. 1161 // This is the end of the marking cycle, we're expected all 1162 // threads to have SATB queues with active set to true. 1163 satb_mq_set.set_active_all_threads(false, /* new active value */ 1164 true /* expected_active */); 1165 1166 { 1167 GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm); 1168 flush_all_task_caches(); 1169 } 1170 1171 // Install newly created mark bitmap as "prev". 1172 swap_mark_bitmaps(); 1173 { 1174 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm); 1175 1176 uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) / 1177 G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread; 1178 uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity); 1179 1180 G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers); 1181 log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions()); 1182 _g1h->workers()->run_task(&cl, num_workers); 1183 1184 log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u", 1185 _g1h->num_regions(), cl.total_selected_for_rebuild()); 1186 } 1187 { 1188 GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm); 1189 reclaim_empty_regions(); 1190 } 1191 1192 // Clean out dead classes 1193 if (ClassUnloadingWithConcurrentMark) { 1194 GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm); 1195 ClassLoaderDataGraph::purge(); 1196 } 1197 1198 compute_new_sizes(); 1199 1200 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after"); 1201 1202 assert(!restart_for_overflow(), "sanity"); 1203 // Completely reset the marking state since marking completed 1204 reset_at_marking_complete(); 1205 } else { 1206 // We overflowed. Restart concurrent marking. 1207 _restart_for_overflow = true; 1208 1209 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow"); 1210 1211 // Clear the marking state because we will be restarting 1212 // marking due to overflowing the global mark stack. 1213 reset_marking_for_restart(); 1214 } 1215 1216 { 1217 GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm); 1218 report_object_count(mark_finished); 1219 } 1220 1221 // Statistics 1222 double now = os::elapsedTime(); 1223 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1224 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1225 _remark_times.add((now - start) * 1000.0); 1226 1227 g1p->record_concurrent_mark_remark_end(); 1228 } 1229 1230 class G1ReclaimEmptyRegionsTask : public AbstractGangTask { 1231 // Per-region work during the Cleanup pause. 1232 class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure { 1233 G1CollectedHeap* _g1h; 1234 size_t _freed_bytes; 1235 FreeRegionList* _local_cleanup_list; 1236 uint _old_regions_removed; 1237 uint _humongous_regions_removed; 1238 HRRSCleanupTask* _hrrs_cleanup_task; 1239 1240 public: 1241 G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h, 1242 FreeRegionList* local_cleanup_list, 1243 HRRSCleanupTask* hrrs_cleanup_task) : 1244 _g1h(g1h), 1245 _freed_bytes(0), 1246 _local_cleanup_list(local_cleanup_list), 1247 _old_regions_removed(0), 1248 _humongous_regions_removed(0), 1249 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1250 1251 size_t freed_bytes() { return _freed_bytes; } 1252 const uint old_regions_removed() { return _old_regions_removed; } 1253 const uint humongous_regions_removed() { return _humongous_regions_removed; } 1254 1255 bool do_heap_region(HeapRegion *hr) { 1256 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) { 1257 _freed_bytes += hr->used(); 1258 hr->set_containing_set(NULL); 1259 if (hr->is_humongous()) { 1260 _humongous_regions_removed++; 1261 _g1h->free_humongous_region(hr, _local_cleanup_list); 1262 } else { 1263 _old_regions_removed++; 1264 _g1h->free_region(hr, _local_cleanup_list, false /* skip_remset */, false /* skip_hcc */, true /* locked */); 1265 } 1266 hr->clear_cardtable(); 1267 _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index()); 1268 log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom())); 1269 } else { 1270 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task); 1271 } 1272 1273 return false; 1274 } 1275 }; 1276 1277 G1CollectedHeap* _g1h; 1278 FreeRegionList* _cleanup_list; 1279 HeapRegionClaimer _hrclaimer; 1280 1281 public: 1282 G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) : 1283 AbstractGangTask("G1 Cleanup"), 1284 _g1h(g1h), 1285 _cleanup_list(cleanup_list), 1286 _hrclaimer(n_workers) { 1287 1288 HeapRegionRemSet::reset_for_cleanup_tasks(); 1289 } 1290 1291 void work(uint worker_id) { 1292 FreeRegionList local_cleanup_list("Local Cleanup List"); 1293 HRRSCleanupTask hrrs_cleanup_task; 1294 G1ReclaimEmptyRegionsClosure cl(_g1h, 1295 &local_cleanup_list, 1296 &hrrs_cleanup_task); 1297 _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id); 1298 assert(cl.is_complete(), "Shouldn't have aborted!"); 1299 1300 // Now update the old/humongous region sets 1301 _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed()); 1302 { 1303 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1304 _g1h->decrement_summary_bytes(cl.freed_bytes()); 1305 1306 _cleanup_list->add_ordered(&local_cleanup_list); 1307 assert(local_cleanup_list.is_empty(), "post-condition"); 1308 1309 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1310 } 1311 } 1312 }; 1313 1314 void G1ConcurrentMark::reclaim_empty_regions() { 1315 WorkGang* workers = _g1h->workers(); 1316 FreeRegionList empty_regions_list("Empty Regions After Mark List"); 1317 1318 G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers()); 1319 workers->run_task(&cl); 1320 1321 if (!empty_regions_list.is_empty()) { 1322 log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length()); 1323 // Now print the empty regions list. 1324 G1HRPrinter* hrp = _g1h->hr_printer(); 1325 if (hrp->is_active()) { 1326 FreeRegionListIterator iter(&empty_regions_list); 1327 while (iter.more_available()) { 1328 HeapRegion* hr = iter.get_next(); 1329 hrp->cleanup(hr); 1330 } 1331 } 1332 // And actually make them available. 1333 _g1h->prepend_to_freelist(&empty_regions_list); 1334 } 1335 } 1336 1337 void G1ConcurrentMark::compute_new_sizes() { 1338 MetaspaceGC::compute_new_size(); 1339 1340 // Cleanup will have freed any regions completely full of garbage. 1341 // Update the soft reference policy with the new heap occupancy. 1342 Universe::update_heap_info_at_gc(); 1343 1344 // We reclaimed old regions so we should calculate the sizes to make 1345 // sure we update the old gen/space data. 1346 _g1h->g1mm()->update_sizes(); 1347 } 1348 1349 void G1ConcurrentMark::cleanup() { 1350 assert_at_safepoint_on_vm_thread(); 1351 1352 // If a full collection has happened, we shouldn't do this. 1353 if (has_aborted()) { 1354 return; 1355 } 1356 1357 G1Policy* g1p = _g1h->g1_policy(); 1358 g1p->record_concurrent_mark_cleanup_start(); 1359 1360 double start = os::elapsedTime(); 1361 1362 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before"); 1363 1364 { 1365 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm); 1366 G1UpdateRemSetTrackingAfterRebuild cl(_g1h); 1367 _g1h->heap_region_iterate(&cl); 1368 } 1369 1370 if (log_is_enabled(Trace, gc, liveness)) { 1371 G1PrintRegionLivenessInfoClosure cl("Post-Cleanup"); 1372 _g1h->heap_region_iterate(&cl); 1373 } 1374 1375 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after"); 1376 1377 // We need to make this be a "collection" so any collection pause that 1378 // races with it goes around and waits for Cleanup to finish. 1379 _g1h->increment_total_collections(); 1380 1381 // Local statistics 1382 double recent_cleanup_time = (os::elapsedTime() - start); 1383 _total_cleanup_time += recent_cleanup_time; 1384 _cleanup_times.add(recent_cleanup_time); 1385 1386 { 1387 GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm); 1388 _g1h->g1_policy()->record_concurrent_mark_cleanup_end(); 1389 } 1390 } 1391 1392 // Supporting Object and Oop closures for reference discovery 1393 // and processing in during marking 1394 1395 bool G1CMIsAliveClosure::do_object_b(oop obj) { 1396 HeapWord* addr = (HeapWord*)obj; 1397 return addr != NULL && 1398 (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_ill(obj)); 1399 } 1400 1401 // 'Keep Alive' oop closure used by both serial parallel reference processing. 1402 // Uses the G1CMTask associated with a worker thread (for serial reference 1403 // processing the G1CMTask for worker 0 is used) to preserve (mark) and 1404 // trace referent objects. 1405 // 1406 // Using the G1CMTask and embedded local queues avoids having the worker 1407 // threads operating on the global mark stack. This reduces the risk 1408 // of overflowing the stack - which we would rather avoid at this late 1409 // state. Also using the tasks' local queues removes the potential 1410 // of the workers interfering with each other that could occur if 1411 // operating on the global stack. 1412 1413 class G1CMKeepAliveAndDrainClosure : public OopClosure { 1414 G1ConcurrentMark* _cm; 1415 G1CMTask* _task; 1416 uint _ref_counter_limit; 1417 uint _ref_counter; 1418 bool _is_serial; 1419 public: 1420 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1421 _cm(cm), _task(task), _is_serial(is_serial), 1422 _ref_counter_limit(G1RefProcDrainInterval) { 1423 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1424 _ref_counter = _ref_counter_limit; 1425 } 1426 1427 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 1428 virtual void do_oop( oop* p) { do_oop_work(p); } 1429 1430 template <class T> void do_oop_work(T* p) { 1431 if (_cm->has_overflown()) { 1432 return; 1433 } 1434 if (!_task->deal_with_reference(p)) { 1435 // We did not add anything to the mark bitmap (or mark stack), so there is 1436 // no point trying to drain it. 1437 return; 1438 } 1439 _ref_counter--; 1440 1441 if (_ref_counter == 0) { 1442 // We have dealt with _ref_counter_limit references, pushing them 1443 // and objects reachable from them on to the local stack (and 1444 // possibly the global stack). Call G1CMTask::do_marking_step() to 1445 // process these entries. 1446 // 1447 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if 1448 // there's nothing more to do (i.e. we're done with the entries that 1449 // were pushed as a result of the G1CMTask::deal_with_reference() calls 1450 // above) or we overflow. 1451 // 1452 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1453 // flag while there may still be some work to do. (See the comment at 1454 // the beginning of G1CMTask::do_marking_step() for those conditions - 1455 // one of which is reaching the specified time target.) It is only 1456 // when G1CMTask::do_marking_step() returns without setting the 1457 // has_aborted() flag that the marking step has completed. 1458 do { 1459 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1460 _task->do_marking_step(mark_step_duration_ms, 1461 false /* do_termination */, 1462 _is_serial); 1463 } while (_task->has_aborted() && !_cm->has_overflown()); 1464 _ref_counter = _ref_counter_limit; 1465 } 1466 } 1467 }; 1468 1469 // 'Drain' oop closure used by both serial and parallel reference processing. 1470 // Uses the G1CMTask associated with a given worker thread (for serial 1471 // reference processing the G1CMtask for worker 0 is used). Calls the 1472 // do_marking_step routine, with an unbelievably large timeout value, 1473 // to drain the marking data structures of the remaining entries 1474 // added by the 'keep alive' oop closure above. 1475 1476 class G1CMDrainMarkingStackClosure : public VoidClosure { 1477 G1ConcurrentMark* _cm; 1478 G1CMTask* _task; 1479 bool _is_serial; 1480 public: 1481 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1482 _cm(cm), _task(task), _is_serial(is_serial) { 1483 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1484 } 1485 1486 void do_void() { 1487 do { 1488 // We call G1CMTask::do_marking_step() to completely drain the local 1489 // and global marking stacks of entries pushed by the 'keep alive' 1490 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 1491 // 1492 // G1CMTask::do_marking_step() is called in a loop, which we'll exit 1493 // if there's nothing more to do (i.e. we've completely drained the 1494 // entries that were pushed as a a result of applying the 'keep alive' 1495 // closure to the entries on the discovered ref lists) or we overflow 1496 // the global marking stack. 1497 // 1498 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1499 // flag while there may still be some work to do. (See the comment at 1500 // the beginning of G1CMTask::do_marking_step() for those conditions - 1501 // one of which is reaching the specified time target.) It is only 1502 // when G1CMTask::do_marking_step() returns without setting the 1503 // has_aborted() flag that the marking step has completed. 1504 1505 _task->do_marking_step(1000000000.0 /* something very large */, 1506 true /* do_termination */, 1507 _is_serial); 1508 } while (_task->has_aborted() && !_cm->has_overflown()); 1509 } 1510 }; 1511 1512 // Implementation of AbstractRefProcTaskExecutor for parallel 1513 // reference processing at the end of G1 concurrent marking 1514 1515 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor { 1516 private: 1517 G1CollectedHeap* _g1h; 1518 G1ConcurrentMark* _cm; 1519 WorkGang* _workers; 1520 uint _active_workers; 1521 1522 public: 1523 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 1524 G1ConcurrentMark* cm, 1525 WorkGang* workers, 1526 uint n_workers) : 1527 _g1h(g1h), _cm(cm), 1528 _workers(workers), _active_workers(n_workers) { } 1529 1530 // Executes the given task using concurrent marking worker threads. 1531 virtual void execute(ProcessTask& task); 1532 virtual void execute(EnqueueTask& task); 1533 }; 1534 1535 class G1CMRefProcTaskProxy : public AbstractGangTask { 1536 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 1537 ProcessTask& _proc_task; 1538 G1CollectedHeap* _g1h; 1539 G1ConcurrentMark* _cm; 1540 1541 public: 1542 G1CMRefProcTaskProxy(ProcessTask& proc_task, 1543 G1CollectedHeap* g1h, 1544 G1ConcurrentMark* cm) : 1545 AbstractGangTask("Process reference objects in parallel"), 1546 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 1547 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1548 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 1549 } 1550 1551 virtual void work(uint worker_id) { 1552 ResourceMark rm; 1553 HandleMark hm; 1554 G1CMTask* task = _cm->task(worker_id); 1555 G1CMIsAliveClosure g1_is_alive(_g1h); 1556 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 1557 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 1558 1559 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 1560 } 1561 }; 1562 1563 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 1564 assert(_workers != NULL, "Need parallel worker threads."); 1565 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1566 1567 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 1568 1569 // We need to reset the concurrency level before each 1570 // proxy task execution, so that the termination protocol 1571 // and overflow handling in G1CMTask::do_marking_step() knows 1572 // how many workers to wait for. 1573 _cm->set_concurrency(_active_workers); 1574 _workers->run_task(&proc_task_proxy); 1575 } 1576 1577 class G1CMRefEnqueueTaskProxy : public AbstractGangTask { 1578 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 1579 EnqueueTask& _enq_task; 1580 1581 public: 1582 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 1583 AbstractGangTask("Enqueue reference objects in parallel"), 1584 _enq_task(enq_task) { } 1585 1586 virtual void work(uint worker_id) { 1587 _enq_task.work(worker_id); 1588 } 1589 }; 1590 1591 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 1592 assert(_workers != NULL, "Need parallel worker threads."); 1593 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1594 1595 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 1596 1597 // Not strictly necessary but... 1598 // 1599 // We need to reset the concurrency level before each 1600 // proxy task execution, so that the termination protocol 1601 // and overflow handling in G1CMTask::do_marking_step() knows 1602 // how many workers to wait for. 1603 _cm->set_concurrency(_active_workers); 1604 _workers->run_task(&enq_task_proxy); 1605 } 1606 1607 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) { 1608 ResourceMark rm; 1609 HandleMark hm; 1610 1611 // Is alive closure. 1612 G1CMIsAliveClosure g1_is_alive(_g1h); 1613 1614 // Inner scope to exclude the cleaning of the string and symbol 1615 // tables from the displayed time. 1616 { 1617 GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm); 1618 1619 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1620 1621 // See the comment in G1CollectedHeap::ref_processing_init() 1622 // about how reference processing currently works in G1. 1623 1624 // Set the soft reference policy 1625 rp->setup_policy(clear_all_soft_refs); 1626 assert(_global_mark_stack.is_empty(), "mark stack should be empty"); 1627 1628 // Instances of the 'Keep Alive' and 'Complete GC' closures used 1629 // in serial reference processing. Note these closures are also 1630 // used for serially processing (by the the current thread) the 1631 // JNI references during parallel reference processing. 1632 // 1633 // These closures do not need to synchronize with the worker 1634 // threads involved in parallel reference processing as these 1635 // instances are executed serially by the current thread (e.g. 1636 // reference processing is not multi-threaded and is thus 1637 // performed by the current thread instead of a gang worker). 1638 // 1639 // The gang tasks involved in parallel reference processing create 1640 // their own instances of these closures, which do their own 1641 // synchronization among themselves. 1642 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 1643 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 1644 1645 // We need at least one active thread. If reference processing 1646 // is not multi-threaded we use the current (VMThread) thread, 1647 // otherwise we use the work gang from the G1CollectedHeap and 1648 // we utilize all the worker threads we can. 1649 bool processing_is_mt = rp->processing_is_mt(); 1650 uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U); 1651 active_workers = MAX2(MIN2(active_workers, _max_num_tasks), 1U); 1652 1653 // Parallel processing task executor. 1654 G1CMRefProcTaskExecutor par_task_executor(_g1h, this, 1655 _g1h->workers(), active_workers); 1656 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 1657 1658 // Set the concurrency level. The phase was already set prior to 1659 // executing the remark task. 1660 set_concurrency(active_workers); 1661 1662 // Set the degree of MT processing here. If the discovery was done MT, 1663 // the number of threads involved during discovery could differ from 1664 // the number of active workers. This is OK as long as the discovered 1665 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 1666 rp->set_active_mt_degree(active_workers); 1667 1668 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->num_q()); 1669 1670 // Process the weak references. 1671 const ReferenceProcessorStats& stats = 1672 rp->process_discovered_references(&g1_is_alive, 1673 &g1_keep_alive, 1674 &g1_drain_mark_stack, 1675 executor, 1676 &pt); 1677 _gc_tracer_cm->report_gc_reference_stats(stats); 1678 pt.print_all_references(); 1679 1680 // The do_oop work routines of the keep_alive and drain_marking_stack 1681 // oop closures will set the has_overflown flag if we overflow the 1682 // global marking stack. 1683 1684 assert(has_overflown() || _global_mark_stack.is_empty(), 1685 "Mark stack should be empty (unless it has overflown)"); 1686 1687 assert(rp->num_q() == active_workers, "why not"); 1688 1689 rp->enqueue_discovered_references(executor, &pt); 1690 1691 rp->verify_no_references_recorded(); 1692 1693 pt.print_enqueue_phase(); 1694 1695 assert(!rp->discovery_enabled(), "Post condition"); 1696 } 1697 1698 assert(has_overflown() || _global_mark_stack.is_empty(), 1699 "Mark stack should be empty (unless it has overflown)"); 1700 1701 { 1702 GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm); 1703 WeakProcessor::weak_oops_do(&g1_is_alive, &do_nothing_cl); 1704 } 1705 1706 if (has_overflown()) { 1707 // We can not trust g1_is_alive if the marking stack overflowed 1708 return; 1709 } 1710 1711 assert(_global_mark_stack.is_empty(), "Marking should have completed"); 1712 1713 // Unload Klasses, String, Symbols, Code Cache, etc. 1714 if (ClassUnloadingWithConcurrentMark) { 1715 GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm); 1716 bool purged_classes = SystemDictionary::do_unloading(&g1_is_alive, _gc_timer_cm, false /* Defer cleaning */); 1717 _g1h->complete_cleaning(&g1_is_alive, purged_classes); 1718 } else { 1719 GCTraceTime(Debug, gc, phases) debug("Cleanup", _gc_timer_cm); 1720 // No need to clean string table and symbol table as they are treated as strong roots when 1721 // class unloading is disabled. 1722 _g1h->partial_cleaning(&g1_is_alive, false, false, G1StringDedup::is_enabled()); 1723 } 1724 } 1725 1726 // When sampling object counts, we already swapped the mark bitmaps, so we need to use 1727 // the prev bitmap determining liveness. 1728 class G1ObjectCountIsAliveClosure: public BoolObjectClosure { 1729 G1CollectedHeap* _g1h; 1730 public: 1731 G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { } 1732 1733 bool do_object_b(oop obj) { 1734 HeapWord* addr = (HeapWord*)obj; 1735 return addr != NULL && 1736 (!_g1h->is_in_g1_reserved(addr) || !_g1h->is_obj_dead(obj)); 1737 } 1738 }; 1739 1740 void G1ConcurrentMark::report_object_count(bool mark_completed) { 1741 // Depending on the completion of the marking liveness needs to be determined 1742 // using either the next or prev bitmap. 1743 if (mark_completed) { 1744 G1ObjectCountIsAliveClosure is_alive(_g1h); 1745 _gc_tracer_cm->report_object_count_after_gc(&is_alive); 1746 } else { 1747 G1CMIsAliveClosure is_alive(_g1h); 1748 _gc_tracer_cm->report_object_count_after_gc(&is_alive); 1749 } 1750 } 1751 1752 1753 void G1ConcurrentMark::swap_mark_bitmaps() { 1754 G1CMBitMap* temp = _prev_mark_bitmap; 1755 _prev_mark_bitmap = _next_mark_bitmap; 1756 _next_mark_bitmap = temp; 1757 _g1h->collector_state()->set_clearing_next_bitmap(true); 1758 } 1759 1760 // Closure for marking entries in SATB buffers. 1761 class G1CMSATBBufferClosure : public SATBBufferClosure { 1762 private: 1763 G1CMTask* _task; 1764 G1CollectedHeap* _g1h; 1765 1766 // This is very similar to G1CMTask::deal_with_reference, but with 1767 // more relaxed requirements for the argument, so this must be more 1768 // circumspect about treating the argument as an object. 1769 void do_entry(void* entry) const { 1770 _task->increment_refs_reached(); 1771 oop const obj = static_cast<oop>(entry); 1772 _task->make_reference_grey(obj); 1773 } 1774 1775 public: 1776 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h) 1777 : _task(task), _g1h(g1h) { } 1778 1779 virtual void do_buffer(void** buffer, size_t size) { 1780 for (size_t i = 0; i < size; ++i) { 1781 do_entry(buffer[i]); 1782 } 1783 } 1784 }; 1785 1786 class G1RemarkThreadsClosure : public ThreadClosure { 1787 G1CMSATBBufferClosure _cm_satb_cl; 1788 G1CMOopClosure _cm_cl; 1789 MarkingCodeBlobClosure _code_cl; 1790 int _thread_parity; 1791 1792 public: 1793 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) : 1794 _cm_satb_cl(task, g1h), 1795 _cm_cl(g1h, task), 1796 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations), 1797 _thread_parity(Threads::thread_claim_parity()) {} 1798 1799 void do_thread(Thread* thread) { 1800 if (thread->is_Java_thread()) { 1801 if (thread->claim_oops_do(true, _thread_parity)) { 1802 JavaThread* jt = (JavaThread*)thread; 1803 1804 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking 1805 // however the liveness of oops reachable from nmethods have very complex lifecycles: 1806 // * Alive if on the stack of an executing method 1807 // * Weakly reachable otherwise 1808 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be 1809 // live by the SATB invariant but other oops recorded in nmethods may behave differently. 1810 jt->nmethods_do(&_code_cl); 1811 1812 G1ThreadLocalData::satb_mark_queue(jt).apply_closure_and_empty(&_cm_satb_cl); 1813 } 1814 } else if (thread->is_VM_thread()) { 1815 if (thread->claim_oops_do(true, _thread_parity)) { 1816 G1BarrierSet::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl); 1817 } 1818 } 1819 } 1820 }; 1821 1822 class G1CMRemarkTask : public AbstractGangTask { 1823 G1ConcurrentMark* _cm; 1824 public: 1825 void work(uint worker_id) { 1826 G1CMTask* task = _cm->task(worker_id); 1827 task->record_start_time(); 1828 { 1829 ResourceMark rm; 1830 HandleMark hm; 1831 1832 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task); 1833 Threads::threads_do(&threads_f); 1834 } 1835 1836 do { 1837 task->do_marking_step(1000000000.0 /* something very large */, 1838 true /* do_termination */, 1839 false /* is_serial */); 1840 } while (task->has_aborted() && !_cm->has_overflown()); 1841 // If we overflow, then we do not want to restart. We instead 1842 // want to abort remark and do concurrent marking again. 1843 task->record_end_time(); 1844 } 1845 1846 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) : 1847 AbstractGangTask("Par Remark"), _cm(cm) { 1848 _cm->terminator()->reset_for_reuse(active_workers); 1849 } 1850 }; 1851 1852 void G1ConcurrentMark::finalize_marking() { 1853 ResourceMark rm; 1854 HandleMark hm; 1855 1856 _g1h->ensure_parsability(false); 1857 1858 // this is remark, so we'll use up all active threads 1859 uint active_workers = _g1h->workers()->active_workers(); 1860 set_concurrency_and_phase(active_workers, false /* concurrent */); 1861 // Leave _parallel_marking_threads at it's 1862 // value originally calculated in the G1ConcurrentMark 1863 // constructor and pass values of the active workers 1864 // through the gang in the task. 1865 1866 { 1867 StrongRootsScope srs(active_workers); 1868 1869 G1CMRemarkTask remarkTask(this, active_workers); 1870 // We will start all available threads, even if we decide that the 1871 // active_workers will be fewer. The extra ones will just bail out 1872 // immediately. 1873 _g1h->workers()->run_task(&remarkTask); 1874 } 1875 1876 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 1877 guarantee(has_overflown() || 1878 satb_mq_set.completed_buffers_num() == 0, 1879 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT, 1880 BOOL_TO_STR(has_overflown()), 1881 satb_mq_set.completed_buffers_num()); 1882 1883 print_stats(); 1884 } 1885 1886 void G1ConcurrentMark::flush_all_task_caches() { 1887 size_t hits = 0; 1888 size_t misses = 0; 1889 for (uint i = 0; i < _max_num_tasks; i++) { 1890 Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache(); 1891 hits += stats.first; 1892 misses += stats.second; 1893 } 1894 size_t sum = hits + misses; 1895 log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf", 1896 hits, misses, percent_of(hits, sum)); 1897 } 1898 1899 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) { 1900 _prev_mark_bitmap->clear_range(mr); 1901 } 1902 1903 HeapRegion* 1904 G1ConcurrentMark::claim_region(uint worker_id) { 1905 // "checkpoint" the finger 1906 HeapWord* finger = _finger; 1907 1908 while (finger < _heap.end()) { 1909 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 1910 1911 HeapRegion* curr_region = _g1h->heap_region_containing(finger); 1912 // Make sure that the reads below do not float before loading curr_region. 1913 OrderAccess::loadload(); 1914 // Above heap_region_containing may return NULL as we always scan claim 1915 // until the end of the heap. In this case, just jump to the next region. 1916 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords; 1917 1918 // Is the gap between reading the finger and doing the CAS too long? 1919 HeapWord* res = Atomic::cmpxchg(end, &_finger, finger); 1920 if (res == finger && curr_region != NULL) { 1921 // we succeeded 1922 HeapWord* bottom = curr_region->bottom(); 1923 HeapWord* limit = curr_region->next_top_at_mark_start(); 1924 1925 // notice that _finger == end cannot be guaranteed here since, 1926 // someone else might have moved the finger even further 1927 assert(_finger >= end, "the finger should have moved forward"); 1928 1929 if (limit > bottom) { 1930 return curr_region; 1931 } else { 1932 assert(limit == bottom, 1933 "the region limit should be at bottom"); 1934 // we return NULL and the caller should try calling 1935 // claim_region() again. 1936 return NULL; 1937 } 1938 } else { 1939 assert(_finger > finger, "the finger should have moved forward"); 1940 // read it again 1941 finger = _finger; 1942 } 1943 } 1944 1945 return NULL; 1946 } 1947 1948 #ifndef PRODUCT 1949 class VerifyNoCSetOops { 1950 G1CollectedHeap* _g1h; 1951 const char* _phase; 1952 int _info; 1953 1954 public: 1955 VerifyNoCSetOops(const char* phase, int info = -1) : 1956 _g1h(G1CollectedHeap::heap()), 1957 _phase(phase), 1958 _info(info) 1959 { } 1960 1961 void operator()(G1TaskQueueEntry task_entry) const { 1962 if (task_entry.is_array_slice()) { 1963 guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice())); 1964 return; 1965 } 1966 guarantee(oopDesc::is_oop(task_entry.obj()), 1967 "Non-oop " PTR_FORMAT ", phase: %s, info: %d", 1968 p2i(task_entry.obj()), _phase, _info); 1969 guarantee(!_g1h->is_in_cset(task_entry.obj()), 1970 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d", 1971 p2i(task_entry.obj()), _phase, _info); 1972 } 1973 }; 1974 1975 void G1ConcurrentMark::verify_no_cset_oops() { 1976 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1977 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) { 1978 return; 1979 } 1980 1981 // Verify entries on the global mark stack 1982 _global_mark_stack.iterate(VerifyNoCSetOops("Stack")); 1983 1984 // Verify entries on the task queues 1985 for (uint i = 0; i < _max_num_tasks; ++i) { 1986 G1CMTaskQueue* queue = _task_queues->queue(i); 1987 queue->iterate(VerifyNoCSetOops("Queue", i)); 1988 } 1989 1990 // Verify the global finger 1991 HeapWord* global_finger = finger(); 1992 if (global_finger != NULL && global_finger < _heap.end()) { 1993 // Since we always iterate over all regions, we might get a NULL HeapRegion 1994 // here. 1995 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger); 1996 guarantee(global_hr == NULL || global_finger == global_hr->bottom(), 1997 "global finger: " PTR_FORMAT " region: " HR_FORMAT, 1998 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)); 1999 } 2000 2001 // Verify the task fingers 2002 assert(_num_concurrent_workers <= _max_num_tasks, "sanity"); 2003 for (uint i = 0; i < _num_concurrent_workers; ++i) { 2004 G1CMTask* task = _tasks[i]; 2005 HeapWord* task_finger = task->finger(); 2006 if (task_finger != NULL && task_finger < _heap.end()) { 2007 // See above note on the global finger verification. 2008 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger); 2009 guarantee(task_hr == NULL || task_finger == task_hr->bottom() || 2010 !task_hr->in_collection_set(), 2011 "task finger: " PTR_FORMAT " region: " HR_FORMAT, 2012 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)); 2013 } 2014 } 2015 } 2016 #endif // PRODUCT 2017 2018 void G1ConcurrentMark::rebuild_rem_set_concurrently() { 2019 _g1h->g1_rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset); 2020 } 2021 2022 void G1ConcurrentMark::print_stats() { 2023 if (!log_is_enabled(Debug, gc, stats)) { 2024 return; 2025 } 2026 log_debug(gc, stats)("---------------------------------------------------------------------"); 2027 for (size_t i = 0; i < _num_active_tasks; ++i) { 2028 _tasks[i]->print_stats(); 2029 log_debug(gc, stats)("---------------------------------------------------------------------"); 2030 } 2031 } 2032 2033 void G1ConcurrentMark::concurrent_cycle_abort() { 2034 if (!cm_thread()->during_cycle() || _has_aborted) { 2035 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything. 2036 return; 2037 } 2038 2039 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next 2040 // concurrent bitmap clearing. 2041 { 2042 GCTraceTime(Debug, gc) debug("Clear Next Bitmap"); 2043 clear_bitmap(_next_mark_bitmap, _g1h->workers(), false); 2044 } 2045 // Note we cannot clear the previous marking bitmap here 2046 // since VerifyDuringGC verifies the objects marked during 2047 // a full GC against the previous bitmap. 2048 2049 // Empty mark stack 2050 reset_marking_for_restart(); 2051 for (uint i = 0; i < _max_num_tasks; ++i) { 2052 _tasks[i]->clear_region_fields(); 2053 } 2054 _first_overflow_barrier_sync.abort(); 2055 _second_overflow_barrier_sync.abort(); 2056 _has_aborted = true; 2057 2058 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 2059 satb_mq_set.abandon_partial_marking(); 2060 // This can be called either during or outside marking, we'll read 2061 // the expected_active value from the SATB queue set. 2062 satb_mq_set.set_active_all_threads( 2063 false, /* new active value */ 2064 satb_mq_set.is_active() /* expected_active */); 2065 } 2066 2067 static void print_ms_time_info(const char* prefix, const char* name, 2068 NumberSeq& ns) { 2069 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 2070 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 2071 if (ns.num() > 0) { 2072 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]", 2073 prefix, ns.sd(), ns.maximum()); 2074 } 2075 } 2076 2077 void G1ConcurrentMark::print_summary_info() { 2078 Log(gc, marking) log; 2079 if (!log.is_trace()) { 2080 return; 2081 } 2082 2083 log.trace(" Concurrent marking:"); 2084 print_ms_time_info(" ", "init marks", _init_times); 2085 print_ms_time_info(" ", "remarks", _remark_times); 2086 { 2087 print_ms_time_info(" ", "final marks", _remark_mark_times); 2088 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 2089 2090 } 2091 print_ms_time_info(" ", "cleanups", _cleanup_times); 2092 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).", 2093 _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2094 log.trace(" Total stop_world time = %8.2f s.", 2095 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0); 2096 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).", 2097 cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum()); 2098 } 2099 2100 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const { 2101 _concurrent_workers->print_worker_threads_on(st); 2102 } 2103 2104 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const { 2105 _concurrent_workers->threads_do(tc); 2106 } 2107 2108 void G1ConcurrentMark::print_on_error(outputStream* st) const { 2109 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT, 2110 p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap)); 2111 _prev_mark_bitmap->print_on_error(st, " Prev Bits: "); 2112 _next_mark_bitmap->print_on_error(st, " Next Bits: "); 2113 } 2114 2115 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) { 2116 ReferenceProcessor* result = g1h->ref_processor_cm(); 2117 assert(result != NULL, "CM reference processor should not be NULL"); 2118 return result; 2119 } 2120 2121 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 2122 G1CMTask* task) 2123 : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)), 2124 _g1h(g1h), _task(task) 2125 { } 2126 2127 void G1CMTask::setup_for_region(HeapRegion* hr) { 2128 assert(hr != NULL, 2129 "claim_region() should have filtered out NULL regions"); 2130 _curr_region = hr; 2131 _finger = hr->bottom(); 2132 update_region_limit(); 2133 } 2134 2135 void G1CMTask::update_region_limit() { 2136 HeapRegion* hr = _curr_region; 2137 HeapWord* bottom = hr->bottom(); 2138 HeapWord* limit = hr->next_top_at_mark_start(); 2139 2140 if (limit == bottom) { 2141 // The region was collected underneath our feet. 2142 // We set the finger to bottom to ensure that the bitmap 2143 // iteration that will follow this will not do anything. 2144 // (this is not a condition that holds when we set the region up, 2145 // as the region is not supposed to be empty in the first place) 2146 _finger = bottom; 2147 } else if (limit >= _region_limit) { 2148 assert(limit >= _finger, "peace of mind"); 2149 } else { 2150 assert(limit < _region_limit, "only way to get here"); 2151 // This can happen under some pretty unusual circumstances. An 2152 // evacuation pause empties the region underneath our feet (NTAMS 2153 // at bottom). We then do some allocation in the region (NTAMS 2154 // stays at bottom), followed by the region being used as a GC 2155 // alloc region (NTAMS will move to top() and the objects 2156 // originally below it will be grayed). All objects now marked in 2157 // the region are explicitly grayed, if below the global finger, 2158 // and we do not need in fact to scan anything else. So, we simply 2159 // set _finger to be limit to ensure that the bitmap iteration 2160 // doesn't do anything. 2161 _finger = limit; 2162 } 2163 2164 _region_limit = limit; 2165 } 2166 2167 void G1CMTask::giveup_current_region() { 2168 assert(_curr_region != NULL, "invariant"); 2169 clear_region_fields(); 2170 } 2171 2172 void G1CMTask::clear_region_fields() { 2173 // Values for these three fields that indicate that we're not 2174 // holding on to a region. 2175 _curr_region = NULL; 2176 _finger = NULL; 2177 _region_limit = NULL; 2178 } 2179 2180 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 2181 if (cm_oop_closure == NULL) { 2182 assert(_cm_oop_closure != NULL, "invariant"); 2183 } else { 2184 assert(_cm_oop_closure == NULL, "invariant"); 2185 } 2186 _cm_oop_closure = cm_oop_closure; 2187 } 2188 2189 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) { 2190 guarantee(next_mark_bitmap != NULL, "invariant"); 2191 _next_mark_bitmap = next_mark_bitmap; 2192 clear_region_fields(); 2193 2194 _calls = 0; 2195 _elapsed_time_ms = 0.0; 2196 _termination_time_ms = 0.0; 2197 _termination_start_time_ms = 0.0; 2198 2199 _mark_stats_cache.reset(); 2200 } 2201 2202 bool G1CMTask::should_exit_termination() { 2203 regular_clock_call(); 2204 // This is called when we are in the termination protocol. We should 2205 // quit if, for some reason, this task wants to abort or the global 2206 // stack is not empty (this means that we can get work from it). 2207 return !_cm->mark_stack_empty() || has_aborted(); 2208 } 2209 2210 void G1CMTask::reached_limit() { 2211 assert(_words_scanned >= _words_scanned_limit || 2212 _refs_reached >= _refs_reached_limit , 2213 "shouldn't have been called otherwise"); 2214 regular_clock_call(); 2215 } 2216 2217 void G1CMTask::regular_clock_call() { 2218 if (has_aborted()) { 2219 return; 2220 } 2221 2222 // First, we need to recalculate the words scanned and refs reached 2223 // limits for the next clock call. 2224 recalculate_limits(); 2225 2226 // During the regular clock call we do the following 2227 2228 // (1) If an overflow has been flagged, then we abort. 2229 if (_cm->has_overflown()) { 2230 set_has_aborted(); 2231 return; 2232 } 2233 2234 // If we are not concurrent (i.e. we're doing remark) we don't need 2235 // to check anything else. The other steps are only needed during 2236 // the concurrent marking phase. 2237 if (!_cm->concurrent()) { 2238 return; 2239 } 2240 2241 // (2) If marking has been aborted for Full GC, then we also abort. 2242 if (_cm->has_aborted()) { 2243 set_has_aborted(); 2244 return; 2245 } 2246 2247 double curr_time_ms = os::elapsedVTime() * 1000.0; 2248 2249 // (4) We check whether we should yield. If we have to, then we abort. 2250 if (SuspendibleThreadSet::should_yield()) { 2251 // We should yield. To do this we abort the task. The caller is 2252 // responsible for yielding. 2253 set_has_aborted(); 2254 return; 2255 } 2256 2257 // (5) We check whether we've reached our time quota. If we have, 2258 // then we abort. 2259 double elapsed_time_ms = curr_time_ms - _start_time_ms; 2260 if (elapsed_time_ms > _time_target_ms) { 2261 set_has_aborted(); 2262 _has_timed_out = true; 2263 return; 2264 } 2265 2266 // (6) Finally, we check whether there are enough completed STAB 2267 // buffers available for processing. If there are, we abort. 2268 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 2269 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 2270 // we do need to process SATB buffers, we'll abort and restart 2271 // the marking task to do so 2272 set_has_aborted(); 2273 return; 2274 } 2275 } 2276 2277 void G1CMTask::recalculate_limits() { 2278 _real_words_scanned_limit = _words_scanned + words_scanned_period; 2279 _words_scanned_limit = _real_words_scanned_limit; 2280 2281 _real_refs_reached_limit = _refs_reached + refs_reached_period; 2282 _refs_reached_limit = _real_refs_reached_limit; 2283 } 2284 2285 void G1CMTask::decrease_limits() { 2286 // This is called when we believe that we're going to do an infrequent 2287 // operation which will increase the per byte scanned cost (i.e. move 2288 // entries to/from the global stack). It basically tries to decrease the 2289 // scanning limit so that the clock is called earlier. 2290 2291 _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4; 2292 _refs_reached_limit = _real_refs_reached_limit - 3 * refs_reached_period / 4; 2293 } 2294 2295 void G1CMTask::move_entries_to_global_stack() { 2296 // Local array where we'll store the entries that will be popped 2297 // from the local queue. 2298 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk]; 2299 2300 size_t n = 0; 2301 G1TaskQueueEntry task_entry; 2302 while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) { 2303 buffer[n] = task_entry; 2304 ++n; 2305 } 2306 if (n < G1CMMarkStack::EntriesPerChunk) { 2307 buffer[n] = G1TaskQueueEntry(); 2308 } 2309 2310 if (n > 0) { 2311 if (!_cm->mark_stack_push(buffer)) { 2312 set_has_aborted(); 2313 } 2314 } 2315 2316 // This operation was quite expensive, so decrease the limits. 2317 decrease_limits(); 2318 } 2319 2320 bool G1CMTask::get_entries_from_global_stack() { 2321 // Local array where we'll store the entries that will be popped 2322 // from the global stack. 2323 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk]; 2324 2325 if (!_cm->mark_stack_pop(buffer)) { 2326 return false; 2327 } 2328 2329 // We did actually pop at least one entry. 2330 for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) { 2331 G1TaskQueueEntry task_entry = buffer[i]; 2332 if (task_entry.is_null()) { 2333 break; 2334 } 2335 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())); 2336 bool success = _task_queue->push(task_entry); 2337 // We only call this when the local queue is empty or under a 2338 // given target limit. So, we do not expect this push to fail. 2339 assert(success, "invariant"); 2340 } 2341 2342 // This operation was quite expensive, so decrease the limits 2343 decrease_limits(); 2344 return true; 2345 } 2346 2347 void G1CMTask::drain_local_queue(bool partially) { 2348 if (has_aborted()) { 2349 return; 2350 } 2351 2352 // Decide what the target size is, depending whether we're going to 2353 // drain it partially (so that other tasks can steal if they run out 2354 // of things to do) or totally (at the very end). 2355 size_t target_size; 2356 if (partially) { 2357 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 2358 } else { 2359 target_size = 0; 2360 } 2361 2362 if (_task_queue->size() > target_size) { 2363 G1TaskQueueEntry entry; 2364 bool ret = _task_queue->pop_local(entry); 2365 while (ret) { 2366 scan_task_entry(entry); 2367 if (_task_queue->size() <= target_size || has_aborted()) { 2368 ret = false; 2369 } else { 2370 ret = _task_queue->pop_local(entry); 2371 } 2372 } 2373 } 2374 } 2375 2376 void G1CMTask::drain_global_stack(bool partially) { 2377 if (has_aborted()) { 2378 return; 2379 } 2380 2381 // We have a policy to drain the local queue before we attempt to 2382 // drain the global stack. 2383 assert(partially || _task_queue->size() == 0, "invariant"); 2384 2385 // Decide what the target size is, depending whether we're going to 2386 // drain it partially (so that other tasks can steal if they run out 2387 // of things to do) or totally (at the very end). 2388 // Notice that when draining the global mark stack partially, due to the racyness 2389 // of the mark stack size update we might in fact drop below the target. But, 2390 // this is not a problem. 2391 // In case of total draining, we simply process until the global mark stack is 2392 // totally empty, disregarding the size counter. 2393 if (partially) { 2394 size_t const target_size = _cm->partial_mark_stack_size_target(); 2395 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 2396 if (get_entries_from_global_stack()) { 2397 drain_local_queue(partially); 2398 } 2399 } 2400 } else { 2401 while (!has_aborted() && get_entries_from_global_stack()) { 2402 drain_local_queue(partially); 2403 } 2404 } 2405 } 2406 2407 // SATB Queue has several assumptions on whether to call the par or 2408 // non-par versions of the methods. this is why some of the code is 2409 // replicated. We should really get rid of the single-threaded version 2410 // of the code to simplify things. 2411 void G1CMTask::drain_satb_buffers() { 2412 if (has_aborted()) { 2413 return; 2414 } 2415 2416 // We set this so that the regular clock knows that we're in the 2417 // middle of draining buffers and doesn't set the abort flag when it 2418 // notices that SATB buffers are available for draining. It'd be 2419 // very counter productive if it did that. :-) 2420 _draining_satb_buffers = true; 2421 2422 G1CMSATBBufferClosure satb_cl(this, _g1h); 2423 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 2424 2425 // This keeps claiming and applying the closure to completed buffers 2426 // until we run out of buffers or we need to abort. 2427 while (!has_aborted() && 2428 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) { 2429 regular_clock_call(); 2430 } 2431 2432 _draining_satb_buffers = false; 2433 2434 assert(has_aborted() || 2435 _cm->concurrent() || 2436 satb_mq_set.completed_buffers_num() == 0, "invariant"); 2437 2438 // again, this was a potentially expensive operation, decrease the 2439 // limits to get the regular clock call early 2440 decrease_limits(); 2441 } 2442 2443 void G1CMTask::clear_mark_stats_cache(uint region_idx) { 2444 _mark_stats_cache.reset(region_idx); 2445 } 2446 2447 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() { 2448 return _mark_stats_cache.evict_all(); 2449 } 2450 2451 void G1CMTask::print_stats() { 2452 log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls); 2453 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 2454 _elapsed_time_ms, _termination_time_ms); 2455 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms", 2456 _step_times_ms.num(), 2457 _step_times_ms.avg(), 2458 _step_times_ms.sd(), 2459 _step_times_ms.maximum(), 2460 _step_times_ms.sum()); 2461 size_t const hits = _mark_stats_cache.hits(); 2462 size_t const misses = _mark_stats_cache.misses(); 2463 log_debug(gc, stats)(" Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f", 2464 hits, misses, percent_of(hits, hits + misses)); 2465 } 2466 2467 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry) { 2468 return _task_queues->steal(worker_id, hash_seed, task_entry); 2469 } 2470 2471 /***************************************************************************** 2472 2473 The do_marking_step(time_target_ms, ...) method is the building 2474 block of the parallel marking framework. It can be called in parallel 2475 with other invocations of do_marking_step() on different tasks 2476 (but only one per task, obviously) and concurrently with the 2477 mutator threads, or during remark, hence it eliminates the need 2478 for two versions of the code. When called during remark, it will 2479 pick up from where the task left off during the concurrent marking 2480 phase. Interestingly, tasks are also claimable during evacuation 2481 pauses too, since do_marking_step() ensures that it aborts before 2482 it needs to yield. 2483 2484 The data structures that it uses to do marking work are the 2485 following: 2486 2487 (1) Marking Bitmap. If there are gray objects that appear only 2488 on the bitmap (this happens either when dealing with an overflow 2489 or when the initial marking phase has simply marked the roots 2490 and didn't push them on the stack), then tasks claim heap 2491 regions whose bitmap they then scan to find gray objects. A 2492 global finger indicates where the end of the last claimed region 2493 is. A local finger indicates how far into the region a task has 2494 scanned. The two fingers are used to determine how to gray an 2495 object (i.e. whether simply marking it is OK, as it will be 2496 visited by a task in the future, or whether it needs to be also 2497 pushed on a stack). 2498 2499 (2) Local Queue. The local queue of the task which is accessed 2500 reasonably efficiently by the task. Other tasks can steal from 2501 it when they run out of work. Throughout the marking phase, a 2502 task attempts to keep its local queue short but not totally 2503 empty, so that entries are available for stealing by other 2504 tasks. Only when there is no more work, a task will totally 2505 drain its local queue. 2506 2507 (3) Global Mark Stack. This handles local queue overflow. During 2508 marking only sets of entries are moved between it and the local 2509 queues, as access to it requires a mutex and more fine-grain 2510 interaction with it which might cause contention. If it 2511 overflows, then the marking phase should restart and iterate 2512 over the bitmap to identify gray objects. Throughout the marking 2513 phase, tasks attempt to keep the global mark stack at a small 2514 length but not totally empty, so that entries are available for 2515 popping by other tasks. Only when there is no more work, tasks 2516 will totally drain the global mark stack. 2517 2518 (4) SATB Buffer Queue. This is where completed SATB buffers are 2519 made available. Buffers are regularly removed from this queue 2520 and scanned for roots, so that the queue doesn't get too 2521 long. During remark, all completed buffers are processed, as 2522 well as the filled in parts of any uncompleted buffers. 2523 2524 The do_marking_step() method tries to abort when the time target 2525 has been reached. There are a few other cases when the 2526 do_marking_step() method also aborts: 2527 2528 (1) When the marking phase has been aborted (after a Full GC). 2529 2530 (2) When a global overflow (on the global stack) has been 2531 triggered. Before the task aborts, it will actually sync up with 2532 the other tasks to ensure that all the marking data structures 2533 (local queues, stacks, fingers etc.) are re-initialized so that 2534 when do_marking_step() completes, the marking phase can 2535 immediately restart. 2536 2537 (3) When enough completed SATB buffers are available. The 2538 do_marking_step() method only tries to drain SATB buffers right 2539 at the beginning. So, if enough buffers are available, the 2540 marking step aborts and the SATB buffers are processed at 2541 the beginning of the next invocation. 2542 2543 (4) To yield. when we have to yield then we abort and yield 2544 right at the end of do_marking_step(). This saves us from a lot 2545 of hassle as, by yielding we might allow a Full GC. If this 2546 happens then objects will be compacted underneath our feet, the 2547 heap might shrink, etc. We save checking for this by just 2548 aborting and doing the yield right at the end. 2549 2550 From the above it follows that the do_marking_step() method should 2551 be called in a loop (or, otherwise, regularly) until it completes. 2552 2553 If a marking step completes without its has_aborted() flag being 2554 true, it means it has completed the current marking phase (and 2555 also all other marking tasks have done so and have all synced up). 2556 2557 A method called regular_clock_call() is invoked "regularly" (in 2558 sub ms intervals) throughout marking. It is this clock method that 2559 checks all the abort conditions which were mentioned above and 2560 decides when the task should abort. A work-based scheme is used to 2561 trigger this clock method: when the number of object words the 2562 marking phase has scanned or the number of references the marking 2563 phase has visited reach a given limit. Additional invocations to 2564 the method clock have been planted in a few other strategic places 2565 too. The initial reason for the clock method was to avoid calling 2566 vtime too regularly, as it is quite expensive. So, once it was in 2567 place, it was natural to piggy-back all the other conditions on it 2568 too and not constantly check them throughout the code. 2569 2570 If do_termination is true then do_marking_step will enter its 2571 termination protocol. 2572 2573 The value of is_serial must be true when do_marking_step is being 2574 called serially (i.e. by the VMThread) and do_marking_step should 2575 skip any synchronization in the termination and overflow code. 2576 Examples include the serial remark code and the serial reference 2577 processing closures. 2578 2579 The value of is_serial must be false when do_marking_step is 2580 being called by any of the worker threads in a work gang. 2581 Examples include the concurrent marking code (CMMarkingTask), 2582 the MT remark code, and the MT reference processing closures. 2583 2584 *****************************************************************************/ 2585 2586 void G1CMTask::do_marking_step(double time_target_ms, 2587 bool do_termination, 2588 bool is_serial) { 2589 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 2590 2591 _start_time_ms = os::elapsedVTime() * 1000.0; 2592 2593 // If do_stealing is true then do_marking_step will attempt to 2594 // steal work from the other G1CMTasks. It only makes sense to 2595 // enable stealing when the termination protocol is enabled 2596 // and do_marking_step() is not being called serially. 2597 bool do_stealing = do_termination && !is_serial; 2598 2599 double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms); 2600 _time_target_ms = time_target_ms - diff_prediction_ms; 2601 2602 // set up the variables that are used in the work-based scheme to 2603 // call the regular clock method 2604 _words_scanned = 0; 2605 _refs_reached = 0; 2606 recalculate_limits(); 2607 2608 // clear all flags 2609 clear_has_aborted(); 2610 _has_timed_out = false; 2611 _draining_satb_buffers = false; 2612 2613 ++_calls; 2614 2615 // Set up the bitmap and oop closures. Anything that uses them is 2616 // eventually called from this method, so it is OK to allocate these 2617 // statically. 2618 G1CMBitMapClosure bitmap_closure(this, _cm); 2619 G1CMOopClosure cm_oop_closure(_g1h, this); 2620 set_cm_oop_closure(&cm_oop_closure); 2621 2622 if (_cm->has_overflown()) { 2623 // This can happen if the mark stack overflows during a GC pause 2624 // and this task, after a yield point, restarts. We have to abort 2625 // as we need to get into the overflow protocol which happens 2626 // right at the end of this task. 2627 set_has_aborted(); 2628 } 2629 2630 // First drain any available SATB buffers. After this, we will not 2631 // look at SATB buffers before the next invocation of this method. 2632 // If enough completed SATB buffers are queued up, the regular clock 2633 // will abort this task so that it restarts. 2634 drain_satb_buffers(); 2635 // ...then partially drain the local queue and the global stack 2636 drain_local_queue(true); 2637 drain_global_stack(true); 2638 2639 do { 2640 if (!has_aborted() && _curr_region != NULL) { 2641 // This means that we're already holding on to a region. 2642 assert(_finger != NULL, "if region is not NULL, then the finger " 2643 "should not be NULL either"); 2644 2645 // We might have restarted this task after an evacuation pause 2646 // which might have evacuated the region we're holding on to 2647 // underneath our feet. Let's read its limit again to make sure 2648 // that we do not iterate over a region of the heap that 2649 // contains garbage (update_region_limit() will also move 2650 // _finger to the start of the region if it is found empty). 2651 update_region_limit(); 2652 // We will start from _finger not from the start of the region, 2653 // as we might be restarting this task after aborting half-way 2654 // through scanning this region. In this case, _finger points to 2655 // the address where we last found a marked object. If this is a 2656 // fresh region, _finger points to start(). 2657 MemRegion mr = MemRegion(_finger, _region_limit); 2658 2659 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(), 2660 "humongous regions should go around loop once only"); 2661 2662 // Some special cases: 2663 // If the memory region is empty, we can just give up the region. 2664 // If the current region is humongous then we only need to check 2665 // the bitmap for the bit associated with the start of the object, 2666 // scan the object if it's live, and give up the region. 2667 // Otherwise, let's iterate over the bitmap of the part of the region 2668 // that is left. 2669 // If the iteration is successful, give up the region. 2670 if (mr.is_empty()) { 2671 giveup_current_region(); 2672 regular_clock_call(); 2673 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) { 2674 if (_next_mark_bitmap->is_marked(mr.start())) { 2675 // The object is marked - apply the closure 2676 bitmap_closure.do_addr(mr.start()); 2677 } 2678 // Even if this task aborted while scanning the humongous object 2679 // we can (and should) give up the current region. 2680 giveup_current_region(); 2681 regular_clock_call(); 2682 } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) { 2683 giveup_current_region(); 2684 regular_clock_call(); 2685 } else { 2686 assert(has_aborted(), "currently the only way to do so"); 2687 // The only way to abort the bitmap iteration is to return 2688 // false from the do_bit() method. However, inside the 2689 // do_bit() method we move the _finger to point to the 2690 // object currently being looked at. So, if we bail out, we 2691 // have definitely set _finger to something non-null. 2692 assert(_finger != NULL, "invariant"); 2693 2694 // Region iteration was actually aborted. So now _finger 2695 // points to the address of the object we last scanned. If we 2696 // leave it there, when we restart this task, we will rescan 2697 // the object. It is easy to avoid this. We move the finger by 2698 // enough to point to the next possible object header. 2699 assert(_finger < _region_limit, "invariant"); 2700 HeapWord* const new_finger = _finger + ((oop)_finger)->size(); 2701 // Check if bitmap iteration was aborted while scanning the last object 2702 if (new_finger >= _region_limit) { 2703 giveup_current_region(); 2704 } else { 2705 move_finger_to(new_finger); 2706 } 2707 } 2708 } 2709 // At this point we have either completed iterating over the 2710 // region we were holding on to, or we have aborted. 2711 2712 // We then partially drain the local queue and the global stack. 2713 // (Do we really need this?) 2714 drain_local_queue(true); 2715 drain_global_stack(true); 2716 2717 // Read the note on the claim_region() method on why it might 2718 // return NULL with potentially more regions available for 2719 // claiming and why we have to check out_of_regions() to determine 2720 // whether we're done or not. 2721 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 2722 // We are going to try to claim a new region. We should have 2723 // given up on the previous one. 2724 // Separated the asserts so that we know which one fires. 2725 assert(_curr_region == NULL, "invariant"); 2726 assert(_finger == NULL, "invariant"); 2727 assert(_region_limit == NULL, "invariant"); 2728 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 2729 if (claimed_region != NULL) { 2730 // Yes, we managed to claim one 2731 setup_for_region(claimed_region); 2732 assert(_curr_region == claimed_region, "invariant"); 2733 } 2734 // It is important to call the regular clock here. It might take 2735 // a while to claim a region if, for example, we hit a large 2736 // block of empty regions. So we need to call the regular clock 2737 // method once round the loop to make sure it's called 2738 // frequently enough. 2739 regular_clock_call(); 2740 } 2741 2742 if (!has_aborted() && _curr_region == NULL) { 2743 assert(_cm->out_of_regions(), 2744 "at this point we should be out of regions"); 2745 } 2746 } while ( _curr_region != NULL && !has_aborted()); 2747 2748 if (!has_aborted()) { 2749 // We cannot check whether the global stack is empty, since other 2750 // tasks might be pushing objects to it concurrently. 2751 assert(_cm->out_of_regions(), 2752 "at this point we should be out of regions"); 2753 // Try to reduce the number of available SATB buffers so that 2754 // remark has less work to do. 2755 drain_satb_buffers(); 2756 } 2757 2758 // Since we've done everything else, we can now totally drain the 2759 // local queue and global stack. 2760 drain_local_queue(false); 2761 drain_global_stack(false); 2762 2763 // Attempt at work stealing from other task's queues. 2764 if (do_stealing && !has_aborted()) { 2765 // We have not aborted. This means that we have finished all that 2766 // we could. Let's try to do some stealing... 2767 2768 // We cannot check whether the global stack is empty, since other 2769 // tasks might be pushing objects to it concurrently. 2770 assert(_cm->out_of_regions() && _task_queue->size() == 0, 2771 "only way to reach here"); 2772 while (!has_aborted()) { 2773 G1TaskQueueEntry entry; 2774 if (_cm->try_stealing(_worker_id, &_hash_seed, entry)) { 2775 scan_task_entry(entry); 2776 2777 // And since we're towards the end, let's totally drain the 2778 // local queue and global stack. 2779 drain_local_queue(false); 2780 drain_global_stack(false); 2781 } else { 2782 break; 2783 } 2784 } 2785 } 2786 2787 // We still haven't aborted. Now, let's try to get into the 2788 // termination protocol. 2789 if (do_termination && !has_aborted()) { 2790 // We cannot check whether the global stack is empty, since other 2791 // tasks might be concurrently pushing objects on it. 2792 // Separated the asserts so that we know which one fires. 2793 assert(_cm->out_of_regions(), "only way to reach here"); 2794 assert(_task_queue->size() == 0, "only way to reach here"); 2795 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 2796 2797 // The G1CMTask class also extends the TerminatorTerminator class, 2798 // hence its should_exit_termination() method will also decide 2799 // whether to exit the termination protocol or not. 2800 bool finished = (is_serial || 2801 _cm->terminator()->offer_termination(this)); 2802 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 2803 _termination_time_ms += 2804 termination_end_time_ms - _termination_start_time_ms; 2805 2806 if (finished) { 2807 // We're all done. 2808 2809 // We can now guarantee that the global stack is empty, since 2810 // all other tasks have finished. We separated the guarantees so 2811 // that, if a condition is false, we can immediately find out 2812 // which one. 2813 guarantee(_cm->out_of_regions(), "only way to reach here"); 2814 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 2815 guarantee(_task_queue->size() == 0, "only way to reach here"); 2816 guarantee(!_cm->has_overflown(), "only way to reach here"); 2817 } else { 2818 // Apparently there's more work to do. Let's abort this task. It 2819 // will restart it and we can hopefully find more things to do. 2820 set_has_aborted(); 2821 } 2822 } 2823 2824 // Mainly for debugging purposes to make sure that a pointer to the 2825 // closure which was statically allocated in this frame doesn't 2826 // escape it by accident. 2827 set_cm_oop_closure(NULL); 2828 double end_time_ms = os::elapsedVTime() * 1000.0; 2829 double elapsed_time_ms = end_time_ms - _start_time_ms; 2830 // Update the step history. 2831 _step_times_ms.add(elapsed_time_ms); 2832 2833 if (has_aborted()) { 2834 // The task was aborted for some reason. 2835 if (_has_timed_out) { 2836 double diff_ms = elapsed_time_ms - _time_target_ms; 2837 // Keep statistics of how well we did with respect to hitting 2838 // our target only if we actually timed out (if we aborted for 2839 // other reasons, then the results might get skewed). 2840 _marking_step_diffs_ms.add(diff_ms); 2841 } 2842 2843 if (_cm->has_overflown()) { 2844 // This is the interesting one. We aborted because a global 2845 // overflow was raised. This means we have to restart the 2846 // marking phase and start iterating over regions. However, in 2847 // order to do this we have to make sure that all tasks stop 2848 // what they are doing and re-initialize in a safe manner. We 2849 // will achieve this with the use of two barrier sync points. 2850 2851 if (!is_serial) { 2852 // We only need to enter the sync barrier if being called 2853 // from a parallel context 2854 _cm->enter_first_sync_barrier(_worker_id); 2855 2856 // When we exit this sync barrier we know that all tasks have 2857 // stopped doing marking work. So, it's now safe to 2858 // re-initialize our data structures. 2859 } 2860 2861 clear_region_fields(); 2862 flush_mark_stats_cache(); 2863 2864 if (!is_serial) { 2865 // If we're executing the concurrent phase of marking, reset the marking 2866 // state; otherwise the marking state is reset after reference processing, 2867 // during the remark pause. 2868 // If we reset here as a result of an overflow during the remark we will 2869 // see assertion failures from any subsequent set_concurrency_and_phase() 2870 // calls. 2871 if (_cm->concurrent() && _worker_id == 0) { 2872 // Worker 0 is responsible for clearing the global data structures because 2873 // of an overflow. During STW we should not clear the overflow flag (in 2874 // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit 2875 // method to abort the pause and restart concurrent marking. 2876 _cm->reset_marking_for_restart(); 2877 2878 log_info(gc, marking)("Concurrent Mark reset for overflow"); 2879 } 2880 2881 // ...and enter the second barrier. 2882 _cm->enter_second_sync_barrier(_worker_id); 2883 } 2884 // At this point, if we're during the concurrent phase of 2885 // marking, everything has been re-initialized and we're 2886 // ready to restart. 2887 } 2888 } 2889 } 2890 2891 G1CMTask::G1CMTask(uint worker_id, 2892 G1ConcurrentMark* cm, 2893 G1CMTaskQueue* task_queue, 2894 G1RegionMarkStats* mark_stats, 2895 uint max_regions) : 2896 _objArray_processor(this), 2897 _worker_id(worker_id), 2898 _g1h(G1CollectedHeap::heap()), 2899 _cm(cm), 2900 _next_mark_bitmap(NULL), 2901 _task_queue(task_queue), 2902 _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize), 2903 _calls(0), 2904 _time_target_ms(0.0), 2905 _start_time_ms(0.0), 2906 _cm_oop_closure(NULL), 2907 _curr_region(NULL), 2908 _finger(NULL), 2909 _region_limit(NULL), 2910 _words_scanned(0), 2911 _words_scanned_limit(0), 2912 _real_words_scanned_limit(0), 2913 _refs_reached(0), 2914 _refs_reached_limit(0), 2915 _real_refs_reached_limit(0), 2916 _hash_seed(17), 2917 _has_aborted(false), 2918 _has_timed_out(false), 2919 _draining_satb_buffers(false), 2920 _step_times_ms(), 2921 _elapsed_time_ms(0.0), 2922 _termination_time_ms(0.0), 2923 _termination_start_time_ms(0.0), 2924 _marking_step_diffs_ms() 2925 { 2926 guarantee(task_queue != NULL, "invariant"); 2927 2928 _marking_step_diffs_ms.add(0.5); 2929 } 2930 2931 // These are formatting macros that are used below to ensure 2932 // consistent formatting. The *_H_* versions are used to format the 2933 // header for a particular value and they should be kept consistent 2934 // with the corresponding macro. Also note that most of the macros add 2935 // the necessary white space (as a prefix) which makes them a bit 2936 // easier to compose. 2937 2938 // All the output lines are prefixed with this string to be able to 2939 // identify them easily in a large log file. 2940 #define G1PPRL_LINE_PREFIX "###" 2941 2942 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT 2943 #ifdef _LP64 2944 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 2945 #else // _LP64 2946 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 2947 #endif // _LP64 2948 2949 // For per-region info 2950 #define G1PPRL_TYPE_FORMAT " %-4s" 2951 #define G1PPRL_TYPE_H_FORMAT " %4s" 2952 #define G1PPRL_STATE_FORMAT " %-5s" 2953 #define G1PPRL_STATE_H_FORMAT " %5s" 2954 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9) 2955 #define G1PPRL_BYTE_H_FORMAT " %9s" 2956 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 2957 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 2958 2959 // For summary info 2960 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT 2961 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT 2962 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB" 2963 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%" 2964 2965 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) : 2966 _total_used_bytes(0), _total_capacity_bytes(0), 2967 _total_prev_live_bytes(0), _total_next_live_bytes(0), 2968 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) 2969 { 2970 if (!log_is_enabled(Trace, gc, liveness)) { 2971 return; 2972 } 2973 2974 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2975 MemRegion g1_reserved = g1h->g1_reserved(); 2976 double now = os::elapsedTime(); 2977 2978 // Print the header of the output. 2979 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 2980 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP" 2981 G1PPRL_SUM_ADDR_FORMAT("reserved") 2982 G1PPRL_SUM_BYTE_FORMAT("region-size"), 2983 p2i(g1_reserved.start()), p2i(g1_reserved.end()), 2984 HeapRegion::GrainBytes); 2985 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 2986 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 2987 G1PPRL_TYPE_H_FORMAT 2988 G1PPRL_ADDR_BASE_H_FORMAT 2989 G1PPRL_BYTE_H_FORMAT 2990 G1PPRL_BYTE_H_FORMAT 2991 G1PPRL_BYTE_H_FORMAT 2992 G1PPRL_DOUBLE_H_FORMAT 2993 G1PPRL_BYTE_H_FORMAT 2994 G1PPRL_STATE_H_FORMAT 2995 G1PPRL_BYTE_H_FORMAT, 2996 "type", "address-range", 2997 "used", "prev-live", "next-live", "gc-eff", 2998 "remset", "state", "code-roots"); 2999 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3000 G1PPRL_TYPE_H_FORMAT 3001 G1PPRL_ADDR_BASE_H_FORMAT 3002 G1PPRL_BYTE_H_FORMAT 3003 G1PPRL_BYTE_H_FORMAT 3004 G1PPRL_BYTE_H_FORMAT 3005 G1PPRL_DOUBLE_H_FORMAT 3006 G1PPRL_BYTE_H_FORMAT 3007 G1PPRL_STATE_H_FORMAT 3008 G1PPRL_BYTE_H_FORMAT, 3009 "", "", 3010 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", 3011 "(bytes)", "", "(bytes)"); 3012 } 3013 3014 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) { 3015 if (!log_is_enabled(Trace, gc, liveness)) { 3016 return false; 3017 } 3018 3019 const char* type = r->get_type_str(); 3020 HeapWord* bottom = r->bottom(); 3021 HeapWord* end = r->end(); 3022 size_t capacity_bytes = r->capacity(); 3023 size_t used_bytes = r->used(); 3024 size_t prev_live_bytes = r->live_bytes(); 3025 size_t next_live_bytes = r->next_live_bytes(); 3026 double gc_eff = r->gc_efficiency(); 3027 size_t remset_bytes = r->rem_set()->mem_size(); 3028 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size(); 3029 const char* remset_type = r->rem_set()->get_short_state_str(); 3030 3031 _total_used_bytes += used_bytes; 3032 _total_capacity_bytes += capacity_bytes; 3033 _total_prev_live_bytes += prev_live_bytes; 3034 _total_next_live_bytes += next_live_bytes; 3035 _total_remset_bytes += remset_bytes; 3036 _total_strong_code_roots_bytes += strong_code_roots_bytes; 3037 3038 // Print a line for this particular region. 3039 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3040 G1PPRL_TYPE_FORMAT 3041 G1PPRL_ADDR_BASE_FORMAT 3042 G1PPRL_BYTE_FORMAT 3043 G1PPRL_BYTE_FORMAT 3044 G1PPRL_BYTE_FORMAT 3045 G1PPRL_DOUBLE_FORMAT 3046 G1PPRL_BYTE_FORMAT 3047 G1PPRL_STATE_FORMAT 3048 G1PPRL_BYTE_FORMAT, 3049 type, p2i(bottom), p2i(end), 3050 used_bytes, prev_live_bytes, next_live_bytes, gc_eff, 3051 remset_bytes, remset_type, strong_code_roots_bytes); 3052 3053 return false; 3054 } 3055 3056 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 3057 if (!log_is_enabled(Trace, gc, liveness)) { 3058 return; 3059 } 3060 3061 // add static memory usages to remembered set sizes 3062 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size(); 3063 // Print the footer of the output. 3064 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 3065 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3066 " SUMMARY" 3067 G1PPRL_SUM_MB_FORMAT("capacity") 3068 G1PPRL_SUM_MB_PERC_FORMAT("used") 3069 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 3070 G1PPRL_SUM_MB_PERC_FORMAT("next-live") 3071 G1PPRL_SUM_MB_FORMAT("remset") 3072 G1PPRL_SUM_MB_FORMAT("code-roots"), 3073 bytes_to_mb(_total_capacity_bytes), 3074 bytes_to_mb(_total_used_bytes), 3075 percent_of(_total_used_bytes, _total_capacity_bytes), 3076 bytes_to_mb(_total_prev_live_bytes), 3077 percent_of(_total_prev_live_bytes, _total_capacity_bytes), 3078 bytes_to_mb(_total_next_live_bytes), 3079 percent_of(_total_next_live_bytes, _total_capacity_bytes), 3080 bytes_to_mb(_total_remset_bytes), 3081 bytes_to_mb(_total_strong_code_roots_bytes)); 3082 }