1 /* 2 * Copyright (c) 2001, 2020, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderDataGraph.hpp" 27 #include "code/codeCache.hpp" 28 #include "gc/g1/g1BarrierSet.hpp" 29 #include "gc/g1/g1CollectedHeap.inline.hpp" 30 #include "gc/g1/g1CollectorState.hpp" 31 #include "gc/g1/g1ConcurrentMark.inline.hpp" 32 #include "gc/g1/g1ConcurrentMarkThread.inline.hpp" 33 #include "gc/g1/g1DirtyCardQueue.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/g1Trace.hpp" 41 #include "gc/g1/heapRegion.inline.hpp" 42 #include "gc/g1/heapRegionRemSet.hpp" 43 #include "gc/g1/heapRegionSet.inline.hpp" 44 #include "gc/shared/gcId.hpp" 45 #include "gc/shared/gcTimer.hpp" 46 #include "gc/shared/gcTraceTime.inline.hpp" 47 #include "gc/shared/gcVMOperations.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/taskTerminator.hpp" 53 #include "gc/shared/taskqueue.inline.hpp" 54 #include "gc/shared/weakProcessor.inline.hpp" 55 #include "gc/shared/workerPolicy.hpp" 56 #include "include/jvm.h" 57 #include "logging/log.hpp" 58 #include "memory/allocation.hpp" 59 #include "memory/iterator.hpp" 60 #include "memory/resourceArea.hpp" 61 #include "memory/universe.hpp" 62 #include "oops/access.inline.hpp" 63 #include "oops/oop.inline.hpp" 64 #include "runtime/atomic.hpp" 65 #include "runtime/handles.inline.hpp" 66 #include "runtime/java.hpp" 67 #include "runtime/orderAccess.hpp" 68 #include "runtime/prefetch.inline.hpp" 69 #include "services/memTracker.hpp" 70 #include "utilities/align.hpp" 71 #include "utilities/growableArray.hpp" 72 73 bool G1CMBitMapClosure::do_addr(HeapWord* const addr) { 74 assert(addr < _cm->finger(), "invariant"); 75 assert(addr >= _task->finger(), "invariant"); 76 77 // We move that task's local finger along. 78 _task->move_finger_to(addr); 79 80 _task->scan_task_entry(G1TaskQueueEntry::from_oop(oop(addr))); 81 // we only partially drain the local queue and global stack 82 _task->drain_local_queue(true); 83 _task->drain_global_stack(true); 84 85 // if the has_aborted flag has been raised, we need to bail out of 86 // the iteration 87 return !_task->has_aborted(); 88 } 89 90 G1CMMarkStack::G1CMMarkStack() : 91 _max_chunk_capacity(0), 92 _base(NULL), 93 _chunk_capacity(0) { 94 set_empty(); 95 } 96 97 bool G1CMMarkStack::resize(size_t new_capacity) { 98 assert(is_empty(), "Only resize when stack is empty."); 99 assert(new_capacity <= _max_chunk_capacity, 100 "Trying to resize stack to " SIZE_FORMAT " chunks when the maximum is " SIZE_FORMAT, new_capacity, _max_chunk_capacity); 101 102 TaskQueueEntryChunk* new_base = MmapArrayAllocator<TaskQueueEntryChunk>::allocate_or_null(new_capacity, mtGC); 103 104 if (new_base == NULL) { 105 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)); 106 return false; 107 } 108 // Release old mapping. 109 if (_base != NULL) { 110 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity); 111 } 112 113 _base = new_base; 114 _chunk_capacity = new_capacity; 115 set_empty(); 116 117 return true; 118 } 119 120 size_t G1CMMarkStack::capacity_alignment() { 121 return (size_t)lcm(os::vm_allocation_granularity(), sizeof(TaskQueueEntryChunk)) / sizeof(G1TaskQueueEntry); 122 } 123 124 bool G1CMMarkStack::initialize(size_t initial_capacity, size_t max_capacity) { 125 guarantee(_max_chunk_capacity == 0, "G1CMMarkStack already initialized."); 126 127 size_t const TaskEntryChunkSizeInVoidStar = sizeof(TaskQueueEntryChunk) / sizeof(G1TaskQueueEntry); 128 129 _max_chunk_capacity = align_up(max_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar; 130 size_t initial_chunk_capacity = align_up(initial_capacity, capacity_alignment()) / TaskEntryChunkSizeInVoidStar; 131 132 guarantee(initial_chunk_capacity <= _max_chunk_capacity, 133 "Maximum chunk capacity " SIZE_FORMAT " smaller than initial capacity " SIZE_FORMAT, 134 _max_chunk_capacity, 135 initial_chunk_capacity); 136 137 log_debug(gc)("Initialize mark stack with " SIZE_FORMAT " chunks, maximum " SIZE_FORMAT, 138 initial_chunk_capacity, _max_chunk_capacity); 139 140 return resize(initial_chunk_capacity); 141 } 142 143 void G1CMMarkStack::expand() { 144 if (_chunk_capacity == _max_chunk_capacity) { 145 log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " chunks.", _chunk_capacity); 146 return; 147 } 148 size_t old_capacity = _chunk_capacity; 149 // Double capacity if possible 150 size_t new_capacity = MIN2(old_capacity * 2, _max_chunk_capacity); 151 152 if (resize(new_capacity)) { 153 log_debug(gc)("Expanded mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks", 154 old_capacity, new_capacity); 155 } else { 156 log_warning(gc)("Failed to expand mark stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " chunks", 157 old_capacity, new_capacity); 158 } 159 } 160 161 G1CMMarkStack::~G1CMMarkStack() { 162 if (_base != NULL) { 163 MmapArrayAllocator<TaskQueueEntryChunk>::free(_base, _chunk_capacity); 164 } 165 } 166 167 void G1CMMarkStack::add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem) { 168 elem->next = *list; 169 *list = elem; 170 } 171 172 void G1CMMarkStack::add_chunk_to_chunk_list(TaskQueueEntryChunk* elem) { 173 MutexLocker x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag); 174 add_chunk_to_list(&_chunk_list, elem); 175 _chunks_in_chunk_list++; 176 } 177 178 void G1CMMarkStack::add_chunk_to_free_list(TaskQueueEntryChunk* elem) { 179 MutexLocker x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag); 180 add_chunk_to_list(&_free_list, elem); 181 } 182 183 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_list(TaskQueueEntryChunk* volatile* list) { 184 TaskQueueEntryChunk* result = *list; 185 if (result != NULL) { 186 *list = (*list)->next; 187 } 188 return result; 189 } 190 191 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_chunk_list() { 192 MutexLocker x(MarkStackChunkList_lock, Mutex::_no_safepoint_check_flag); 193 TaskQueueEntryChunk* result = remove_chunk_from_list(&_chunk_list); 194 if (result != NULL) { 195 _chunks_in_chunk_list--; 196 } 197 return result; 198 } 199 200 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::remove_chunk_from_free_list() { 201 MutexLocker x(MarkStackFreeList_lock, Mutex::_no_safepoint_check_flag); 202 return remove_chunk_from_list(&_free_list); 203 } 204 205 G1CMMarkStack::TaskQueueEntryChunk* G1CMMarkStack::allocate_new_chunk() { 206 // This dirty read of _hwm is okay because we only ever increase the _hwm in parallel code. 207 // Further this limits _hwm to a value of _chunk_capacity + #threads, avoiding 208 // wraparound of _hwm. 209 if (_hwm >= _chunk_capacity) { 210 return NULL; 211 } 212 213 size_t cur_idx = Atomic::fetch_and_add(&_hwm, 1u); 214 if (cur_idx >= _chunk_capacity) { 215 return NULL; 216 } 217 218 TaskQueueEntryChunk* result = ::new (&_base[cur_idx]) TaskQueueEntryChunk; 219 result->next = NULL; 220 return result; 221 } 222 223 bool G1CMMarkStack::par_push_chunk(G1TaskQueueEntry* ptr_arr) { 224 // Get a new chunk. 225 TaskQueueEntryChunk* new_chunk = remove_chunk_from_free_list(); 226 227 if (new_chunk == NULL) { 228 // Did not get a chunk from the free list. Allocate from backing memory. 229 new_chunk = allocate_new_chunk(); 230 231 if (new_chunk == NULL) { 232 return false; 233 } 234 } 235 236 Copy::conjoint_memory_atomic(ptr_arr, new_chunk->data, EntriesPerChunk * sizeof(G1TaskQueueEntry)); 237 238 add_chunk_to_chunk_list(new_chunk); 239 240 return true; 241 } 242 243 bool G1CMMarkStack::par_pop_chunk(G1TaskQueueEntry* ptr_arr) { 244 TaskQueueEntryChunk* cur = remove_chunk_from_chunk_list(); 245 246 if (cur == NULL) { 247 return false; 248 } 249 250 Copy::conjoint_memory_atomic(cur->data, ptr_arr, EntriesPerChunk * sizeof(G1TaskQueueEntry)); 251 252 add_chunk_to_free_list(cur); 253 return true; 254 } 255 256 void G1CMMarkStack::set_empty() { 257 _chunks_in_chunk_list = 0; 258 _hwm = 0; 259 _chunk_list = NULL; 260 _free_list = NULL; 261 } 262 263 G1CMRootMemRegions::G1CMRootMemRegions(uint const max_regions) : 264 _root_regions(MemRegion::create_array(max_regions, mtGC)), 265 _max_regions(max_regions), 266 _num_root_regions(0), 267 _claimed_root_regions(0), 268 _scan_in_progress(false), 269 _should_abort(false) { } 270 271 G1CMRootMemRegions::~G1CMRootMemRegions() { 272 MemRegion::destroy_array(_root_regions, _max_regions); 273 } 274 275 void G1CMRootMemRegions::reset() { 276 _num_root_regions = 0; 277 } 278 279 void G1CMRootMemRegions::add(HeapWord* start, HeapWord* end) { 280 assert_at_safepoint(); 281 size_t idx = Atomic::fetch_and_add(&_num_root_regions, 1u); 282 assert(idx < _max_regions, "Trying to add more root MemRegions than there is space " SIZE_FORMAT, _max_regions); 283 assert(start != NULL && end != NULL && start <= end, "Start (" PTR_FORMAT ") should be less or equal to " 284 "end (" PTR_FORMAT ")", p2i(start), p2i(end)); 285 _root_regions[idx].set_start(start); 286 _root_regions[idx].set_end(end); 287 } 288 289 void G1CMRootMemRegions::prepare_for_scan() { 290 assert(!scan_in_progress(), "pre-condition"); 291 292 _scan_in_progress = _num_root_regions > 0; 293 294 _claimed_root_regions = 0; 295 _should_abort = false; 296 } 297 298 const MemRegion* G1CMRootMemRegions::claim_next() { 299 if (_should_abort) { 300 // If someone has set the should_abort flag, we return NULL to 301 // force the caller to bail out of their loop. 302 return NULL; 303 } 304 305 if (_claimed_root_regions >= _num_root_regions) { 306 return NULL; 307 } 308 309 size_t claimed_index = Atomic::fetch_and_add(&_claimed_root_regions, 1u); 310 if (claimed_index < _num_root_regions) { 311 return &_root_regions[claimed_index]; 312 } 313 return NULL; 314 } 315 316 uint G1CMRootMemRegions::num_root_regions() const { 317 return (uint)_num_root_regions; 318 } 319 320 void G1CMRootMemRegions::notify_scan_done() { 321 MutexLocker x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 322 _scan_in_progress = false; 323 RootRegionScan_lock->notify_all(); 324 } 325 326 void G1CMRootMemRegions::cancel_scan() { 327 notify_scan_done(); 328 } 329 330 void G1CMRootMemRegions::scan_finished() { 331 assert(scan_in_progress(), "pre-condition"); 332 333 if (!_should_abort) { 334 assert(_claimed_root_regions >= num_root_regions(), 335 "we should have claimed all root regions, claimed " SIZE_FORMAT ", length = %u", 336 _claimed_root_regions, num_root_regions()); 337 } 338 339 notify_scan_done(); 340 } 341 342 bool G1CMRootMemRegions::wait_until_scan_finished() { 343 if (!scan_in_progress()) { 344 return false; 345 } 346 347 { 348 MonitorLocker ml(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 349 while (scan_in_progress()) { 350 ml.wait(); 351 } 352 } 353 return true; 354 } 355 356 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, 357 G1RegionToSpaceMapper* prev_bitmap_storage, 358 G1RegionToSpaceMapper* next_bitmap_storage) : 359 // _cm_thread set inside the constructor 360 _g1h(g1h), 361 362 _mark_bitmap_1(), 363 _mark_bitmap_2(), 364 _prev_mark_bitmap(&_mark_bitmap_1), 365 _next_mark_bitmap(&_mark_bitmap_2), 366 367 _heap(_g1h->reserved_region()), 368 369 _root_regions(_g1h->max_regions()), 370 371 _global_mark_stack(), 372 373 // _finger set in set_non_marking_state 374 375 _worker_id_offset(G1DirtyCardQueueSet::num_par_ids() + G1ConcRefinementThreads), 376 _max_num_tasks(ParallelGCThreads), 377 // _num_active_tasks set in set_non_marking_state() 378 // _tasks set inside the constructor 379 380 _task_queues(new G1CMTaskQueueSet((int) _max_num_tasks)), 381 _terminator((int) _max_num_tasks, _task_queues), 382 383 _first_overflow_barrier_sync(), 384 _second_overflow_barrier_sync(), 385 386 _has_overflown(false), 387 _concurrent(false), 388 _has_aborted(false), 389 _restart_for_overflow(false), 390 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 391 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()), 392 393 // _verbose_level set below 394 395 _init_times(), 396 _remark_times(), 397 _remark_mark_times(), 398 _remark_weak_ref_times(), 399 _cleanup_times(), 400 _total_cleanup_time(0.0), 401 402 _accum_task_vtime(NULL), 403 404 _concurrent_workers(NULL), 405 _num_concurrent_workers(0), 406 _max_concurrent_workers(0), 407 408 _region_mark_stats(NEW_C_HEAP_ARRAY(G1RegionMarkStats, _g1h->max_regions(), mtGC)), 409 _top_at_rebuild_starts(NEW_C_HEAP_ARRAY(HeapWord*, _g1h->max_regions(), mtGC)) 410 { 411 assert(CGC_lock != NULL, "CGC_lock must be initialized"); 412 413 _mark_bitmap_1.initialize(g1h->reserved_region(), prev_bitmap_storage); 414 _mark_bitmap_2.initialize(g1h->reserved_region(), next_bitmap_storage); 415 416 // Create & start ConcurrentMark thread. 417 _cm_thread = new G1ConcurrentMarkThread(this); 418 if (_cm_thread->osthread() == NULL) { 419 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread"); 420 } 421 422 log_debug(gc)("ConcGCThreads: %u offset %u", ConcGCThreads, _worker_id_offset); 423 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 424 425 _num_concurrent_workers = ConcGCThreads; 426 _max_concurrent_workers = _num_concurrent_workers; 427 428 _concurrent_workers = new WorkGang("G1 Conc", _max_concurrent_workers, false, true); 429 _concurrent_workers->initialize_workers(); 430 431 if (!_global_mark_stack.initialize(MarkStackSize, MarkStackSizeMax)) { 432 vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack."); 433 } 434 435 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_num_tasks, mtGC); 436 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_num_tasks, mtGC); 437 438 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 439 _num_active_tasks = _max_num_tasks; 440 441 for (uint i = 0; i < _max_num_tasks; ++i) { 442 G1CMTaskQueue* task_queue = new G1CMTaskQueue(); 443 task_queue->initialize(); 444 _task_queues->register_queue(i, task_queue); 445 446 _tasks[i] = new G1CMTask(i, this, task_queue, _region_mark_stats, _g1h->max_regions()); 447 448 _accum_task_vtime[i] = 0.0; 449 } 450 451 reset_at_marking_complete(); 452 } 453 454 void G1ConcurrentMark::reset() { 455 _has_aborted = false; 456 457 reset_marking_for_restart(); 458 459 // Reset all tasks, since different phases will use different number of active 460 // threads. So, it's easiest to have all of them ready. 461 for (uint i = 0; i < _max_num_tasks; ++i) { 462 _tasks[i]->reset(_next_mark_bitmap); 463 } 464 465 uint max_regions = _g1h->max_regions(); 466 for (uint i = 0; i < max_regions; i++) { 467 _top_at_rebuild_starts[i] = NULL; 468 _region_mark_stats[i].clear(); 469 } 470 } 471 472 void G1ConcurrentMark::clear_statistics_in_region(uint region_idx) { 473 for (uint j = 0; j < _max_num_tasks; ++j) { 474 _tasks[j]->clear_mark_stats_cache(region_idx); 475 } 476 _top_at_rebuild_starts[region_idx] = NULL; 477 _region_mark_stats[region_idx].clear(); 478 } 479 480 void G1ConcurrentMark::clear_statistics(HeapRegion* r) { 481 uint const region_idx = r->hrm_index(); 482 if (r->is_humongous()) { 483 assert(r->is_starts_humongous(), "Got humongous continues region here"); 484 uint const size_in_regions = (uint)_g1h->humongous_obj_size_in_regions(oop(r->humongous_start_region()->bottom())->size()); 485 for (uint j = region_idx; j < (region_idx + size_in_regions); j++) { 486 clear_statistics_in_region(j); 487 } 488 } else { 489 clear_statistics_in_region(region_idx); 490 } 491 } 492 493 static void clear_mark_if_set(G1CMBitMap* bitmap, HeapWord* addr) { 494 if (bitmap->is_marked(addr)) { 495 bitmap->clear(addr); 496 } 497 } 498 499 void G1ConcurrentMark::humongous_object_eagerly_reclaimed(HeapRegion* r) { 500 assert_at_safepoint_on_vm_thread(); 501 502 // Need to clear all mark bits of the humongous object. 503 clear_mark_if_set(_prev_mark_bitmap, r->bottom()); 504 clear_mark_if_set(_next_mark_bitmap, r->bottom()); 505 506 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) { 507 return; 508 } 509 510 // Clear any statistics about the region gathered so far. 511 clear_statistics(r); 512 } 513 514 void G1ConcurrentMark::reset_marking_for_restart() { 515 _global_mark_stack.set_empty(); 516 517 // Expand the marking stack, if we have to and if we can. 518 if (has_overflown()) { 519 _global_mark_stack.expand(); 520 521 uint max_regions = _g1h->max_regions(); 522 for (uint i = 0; i < max_regions; i++) { 523 _region_mark_stats[i].clear_during_overflow(); 524 } 525 } 526 527 clear_has_overflown(); 528 _finger = _heap.start(); 529 530 for (uint i = 0; i < _max_num_tasks; ++i) { 531 G1CMTaskQueue* queue = _task_queues->queue(i); 532 queue->set_empty(); 533 } 534 } 535 536 void G1ConcurrentMark::set_concurrency(uint active_tasks) { 537 assert(active_tasks <= _max_num_tasks, "we should not have more"); 538 539 _num_active_tasks = active_tasks; 540 // Need to update the three data structures below according to the 541 // number of active threads for this phase. 542 _terminator.reset_for_reuse(active_tasks); 543 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 544 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 545 } 546 547 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 548 set_concurrency(active_tasks); 549 550 _concurrent = concurrent; 551 552 if (!concurrent) { 553 // At this point we should be in a STW phase, and completed marking. 554 assert_at_safepoint_on_vm_thread(); 555 assert(out_of_regions(), 556 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT, 557 p2i(_finger), p2i(_heap.end())); 558 } 559 } 560 561 void G1ConcurrentMark::reset_at_marking_complete() { 562 // We set the global marking state to some default values when we're 563 // not doing marking. 564 reset_marking_for_restart(); 565 _num_active_tasks = 0; 566 } 567 568 G1ConcurrentMark::~G1ConcurrentMark() { 569 FREE_C_HEAP_ARRAY(HeapWord*, _top_at_rebuild_starts); 570 FREE_C_HEAP_ARRAY(G1RegionMarkStats, _region_mark_stats); 571 // The G1ConcurrentMark instance is never freed. 572 ShouldNotReachHere(); 573 } 574 575 class G1ClearBitMapTask : public AbstractGangTask { 576 public: 577 static size_t chunk_size() { return M; } 578 579 private: 580 // Heap region closure used for clearing the given mark bitmap. 581 class G1ClearBitmapHRClosure : public HeapRegionClosure { 582 private: 583 G1CMBitMap* _bitmap; 584 G1ConcurrentMark* _cm; 585 public: 586 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _bitmap(bitmap), _cm(cm) { 587 } 588 589 virtual bool do_heap_region(HeapRegion* r) { 590 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize; 591 592 HeapWord* cur = r->bottom(); 593 HeapWord* const end = r->end(); 594 595 while (cur < end) { 596 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end)); 597 _bitmap->clear_range(mr); 598 599 cur += chunk_size_in_words; 600 601 // Abort iteration if after yielding the marking has been aborted. 602 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) { 603 return true; 604 } 605 // Repeat the asserts from before the start of the closure. We will do them 606 // as asserts here to minimize their overhead on the product. However, we 607 // will have them as guarantees at the beginning / end of the bitmap 608 // clearing to get some checking in the product. 609 assert(_cm == NULL || _cm->cm_thread()->during_cycle(), "invariant"); 610 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_or_rebuild_in_progress(), "invariant"); 611 } 612 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index()); 613 614 return false; 615 } 616 }; 617 618 G1ClearBitmapHRClosure _cl; 619 HeapRegionClaimer _hr_claimer; 620 bool _suspendible; // If the task is suspendible, workers must join the STS. 621 622 public: 623 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) : 624 AbstractGangTask("G1 Clear Bitmap"), 625 _cl(bitmap, suspendible ? cm : NULL), 626 _hr_claimer(n_workers), 627 _suspendible(suspendible) 628 { } 629 630 void work(uint worker_id) { 631 SuspendibleThreadSetJoiner sts_join(_suspendible); 632 G1CollectedHeap::heap()->heap_region_par_iterate_from_worker_offset(&_cl, &_hr_claimer, worker_id); 633 } 634 635 bool is_complete() { 636 return _cl.is_complete(); 637 } 638 }; 639 640 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) { 641 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint."); 642 643 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor(); 644 size_t const num_chunks = align_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size(); 645 646 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers()); 647 648 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield); 649 650 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks); 651 workers->run_task(&cl, num_workers); 652 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding."); 653 } 654 655 void G1ConcurrentMark::cleanup_for_next_mark() { 656 // Make sure that the concurrent mark thread looks to still be in 657 // the current cycle. 658 guarantee(cm_thread()->during_cycle(), "invariant"); 659 660 // We are finishing up the current cycle by clearing the next 661 // marking bitmap and getting it ready for the next cycle. During 662 // this time no other cycle can start. So, let's make sure that this 663 // is the case. 664 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant"); 665 666 clear_bitmap(_next_mark_bitmap, _concurrent_workers, true); 667 668 // Repeat the asserts from above. 669 guarantee(cm_thread()->during_cycle(), "invariant"); 670 guarantee(!_g1h->collector_state()->mark_or_rebuild_in_progress(), "invariant"); 671 } 672 673 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) { 674 assert_at_safepoint_on_vm_thread(); 675 clear_bitmap(_prev_mark_bitmap, workers, false); 676 } 677 678 class NoteStartOfMarkHRClosure : public HeapRegionClosure { 679 public: 680 bool do_heap_region(HeapRegion* r) { 681 r->note_start_of_marking(); 682 return false; 683 } 684 }; 685 686 void G1ConcurrentMark::pre_initial_mark() { 687 assert_at_safepoint_on_vm_thread(); 688 689 // Reset marking state. 690 reset(); 691 692 // For each region note start of marking. 693 NoteStartOfMarkHRClosure startcl; 694 _g1h->heap_region_iterate(&startcl); 695 696 _root_regions.reset(); 697 } 698 699 700 void G1ConcurrentMark::post_initial_mark() { 701 // Start Concurrent Marking weak-reference discovery. 702 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 703 // enable ("weak") refs discovery 704 rp->enable_discovery(); 705 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 706 707 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 708 // This is the start of the marking cycle, we're expected all 709 // threads to have SATB queues with active set to false. 710 satb_mq_set.set_active_all_threads(true, /* new active value */ 711 false /* expected_active */); 712 713 _root_regions.prepare_for_scan(); 714 715 // update_g1_committed() will be called at the end of an evac pause 716 // when marking is on. So, it's also called at the end of the 717 // initial-mark pause to update the heap end, if the heap expands 718 // during it. No need to call it here. 719 } 720 721 /* 722 * Notice that in the next two methods, we actually leave the STS 723 * during the barrier sync and join it immediately afterwards. If we 724 * do not do this, the following deadlock can occur: one thread could 725 * be in the barrier sync code, waiting for the other thread to also 726 * sync up, whereas another one could be trying to yield, while also 727 * waiting for the other threads to sync up too. 728 * 729 * Note, however, that this code is also used during remark and in 730 * this case we should not attempt to leave / enter the STS, otherwise 731 * we'll either hit an assert (debug / fastdebug) or deadlock 732 * (product). So we should only leave / enter the STS if we are 733 * operating concurrently. 734 * 735 * Because the thread that does the sync barrier has left the STS, it 736 * is possible to be suspended for a Full GC or an evacuation pause 737 * could occur. This is actually safe, since the entering the sync 738 * barrier is one of the last things do_marking_step() does, and it 739 * doesn't manipulate any data structures afterwards. 740 */ 741 742 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 743 bool barrier_aborted; 744 { 745 SuspendibleThreadSetLeaver sts_leave(concurrent()); 746 barrier_aborted = !_first_overflow_barrier_sync.enter(); 747 } 748 749 // at this point everyone should have synced up and not be doing any 750 // more work 751 752 if (barrier_aborted) { 753 // If the barrier aborted we ignore the overflow condition and 754 // just abort the whole marking phase as quickly as possible. 755 return; 756 } 757 } 758 759 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 760 SuspendibleThreadSetLeaver sts_leave(concurrent()); 761 _second_overflow_barrier_sync.enter(); 762 763 // at this point everything should be re-initialized and ready to go 764 } 765 766 class G1CMConcurrentMarkingTask : public AbstractGangTask { 767 G1ConcurrentMark* _cm; 768 769 public: 770 void work(uint worker_id) { 771 assert(Thread::current()->is_ConcurrentGC_thread(), "Not a concurrent GC thread"); 772 ResourceMark rm; 773 774 double start_vtime = os::elapsedVTime(); 775 776 { 777 SuspendibleThreadSetJoiner sts_join; 778 779 assert(worker_id < _cm->active_tasks(), "invariant"); 780 781 G1CMTask* task = _cm->task(worker_id); 782 task->record_start_time(); 783 if (!_cm->has_aborted()) { 784 do { 785 task->do_marking_step(G1ConcMarkStepDurationMillis, 786 true /* do_termination */, 787 false /* is_serial*/); 788 789 _cm->do_yield_check(); 790 } while (!_cm->has_aborted() && task->has_aborted()); 791 } 792 task->record_end_time(); 793 guarantee(!task->has_aborted() || _cm->has_aborted(), "invariant"); 794 } 795 796 double end_vtime = os::elapsedVTime(); 797 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 798 } 799 800 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm) : 801 AbstractGangTask("Concurrent Mark"), _cm(cm) { } 802 803 ~G1CMConcurrentMarkingTask() { } 804 }; 805 806 uint G1ConcurrentMark::calc_active_marking_workers() { 807 uint result = 0; 808 if (!UseDynamicNumberOfGCThreads || !FLAG_IS_DEFAULT(ConcGCThreads)) { 809 result = _max_concurrent_workers; 810 } else { 811 result = 812 WorkerPolicy::calc_default_active_workers(_max_concurrent_workers, 813 1, /* Minimum workers */ 814 _num_concurrent_workers, 815 Threads::number_of_non_daemon_threads()); 816 // Don't scale the result down by scale_concurrent_workers() because 817 // that scaling has already gone into "_max_concurrent_workers". 818 } 819 assert(result > 0 && result <= _max_concurrent_workers, 820 "Calculated number of marking workers must be larger than zero and at most the maximum %u, but is %u", 821 _max_concurrent_workers, result); 822 return result; 823 } 824 825 void G1ConcurrentMark::scan_root_region(const MemRegion* region, uint worker_id) { 826 #ifdef ASSERT 827 HeapWord* last = region->last(); 828 HeapRegion* hr = _g1h->heap_region_containing(last); 829 assert(hr->is_old() || hr->next_top_at_mark_start() == hr->bottom(), 830 "Root regions must be old or survivor/eden but region %u is %s", hr->hrm_index(), hr->get_type_str()); 831 assert(hr->next_top_at_mark_start() == region->start(), 832 "MemRegion start should be equal to nTAMS"); 833 #endif 834 835 G1RootRegionScanClosure cl(_g1h, this, worker_id); 836 837 const uintx interval = PrefetchScanIntervalInBytes; 838 HeapWord* curr = region->start(); 839 const HeapWord* end = region->end(); 840 while (curr < end) { 841 Prefetch::read(curr, interval); 842 oop obj = oop(curr); 843 int size = obj->oop_iterate_size(&cl); 844 assert(size == obj->size(), "sanity"); 845 curr += size; 846 } 847 } 848 849 class G1CMRootRegionScanTask : public AbstractGangTask { 850 G1ConcurrentMark* _cm; 851 public: 852 G1CMRootRegionScanTask(G1ConcurrentMark* cm) : 853 AbstractGangTask("G1 Root Region Scan"), _cm(cm) { } 854 855 void work(uint worker_id) { 856 assert(Thread::current()->is_ConcurrentGC_thread(), 857 "this should only be done by a conc GC thread"); 858 859 G1CMRootMemRegions* root_regions = _cm->root_regions(); 860 const MemRegion* region = root_regions->claim_next(); 861 while (region != NULL) { 862 _cm->scan_root_region(region, worker_id); 863 region = root_regions->claim_next(); 864 } 865 } 866 }; 867 868 void G1ConcurrentMark::scan_root_regions() { 869 // scan_in_progress() will have been set to true only if there was 870 // at least one root region to scan. So, if it's false, we 871 // should not attempt to do any further work. 872 if (root_regions()->scan_in_progress()) { 873 assert(!has_aborted(), "Aborting before root region scanning is finished not supported."); 874 875 _num_concurrent_workers = MIN2(calc_active_marking_workers(), 876 // We distribute work on a per-region basis, so starting 877 // more threads than that is useless. 878 root_regions()->num_root_regions()); 879 assert(_num_concurrent_workers <= _max_concurrent_workers, 880 "Maximum number of marking threads exceeded"); 881 882 G1CMRootRegionScanTask task(this); 883 log_debug(gc, ergo)("Running %s using %u workers for %u work units.", 884 task.name(), _num_concurrent_workers, root_regions()->num_root_regions()); 885 _concurrent_workers->run_task(&task, _num_concurrent_workers); 886 887 // It's possible that has_aborted() is true here without actually 888 // aborting the survivor scan earlier. This is OK as it's 889 // mainly used for sanity checking. 890 root_regions()->scan_finished(); 891 } 892 } 893 894 void G1ConcurrentMark::concurrent_cycle_start() { 895 _gc_timer_cm->register_gc_start(); 896 897 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start()); 898 899 _g1h->trace_heap_before_gc(_gc_tracer_cm); 900 } 901 902 void G1ConcurrentMark::concurrent_cycle_end() { 903 _g1h->collector_state()->set_clearing_next_bitmap(false); 904 905 _g1h->trace_heap_after_gc(_gc_tracer_cm); 906 907 if (has_aborted()) { 908 log_info(gc, marking)("Concurrent Mark Abort"); 909 _gc_tracer_cm->report_concurrent_mode_failure(); 910 } 911 912 _gc_timer_cm->register_gc_end(); 913 914 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 915 } 916 917 void G1ConcurrentMark::mark_from_roots() { 918 _restart_for_overflow = false; 919 920 _num_concurrent_workers = calc_active_marking_workers(); 921 922 uint active_workers = MAX2(1U, _num_concurrent_workers); 923 924 // Setting active workers is not guaranteed since fewer 925 // worker threads may currently exist and more may not be 926 // available. 927 active_workers = _concurrent_workers->update_active_workers(active_workers); 928 log_info(gc, task)("Using %u workers of %u for marking", active_workers, _concurrent_workers->total_workers()); 929 930 // Parallel task terminator is set in "set_concurrency_and_phase()" 931 set_concurrency_and_phase(active_workers, true /* concurrent */); 932 933 G1CMConcurrentMarkingTask marking_task(this); 934 _concurrent_workers->run_task(&marking_task); 935 print_stats(); 936 } 937 938 void G1ConcurrentMark::verify_during_pause(G1HeapVerifier::G1VerifyType type, VerifyOption vo, const char* caller) { 939 G1HeapVerifier* verifier = _g1h->verifier(); 940 941 verifier->verify_region_sets_optional(); 942 943 if (VerifyDuringGC) { 944 GCTraceTime(Debug, gc, phases) debug(caller, _gc_timer_cm); 945 946 size_t const BufLen = 512; 947 char buffer[BufLen]; 948 949 jio_snprintf(buffer, BufLen, "During GC (%s)", caller); 950 verifier->verify(type, vo, buffer); 951 } 952 953 verifier->check_bitmaps(caller); 954 } 955 956 class G1UpdateRemSetTrackingBeforeRebuildTask : public AbstractGangTask { 957 G1CollectedHeap* _g1h; 958 G1ConcurrentMark* _cm; 959 HeapRegionClaimer _hrclaimer; 960 uint volatile _total_selected_for_rebuild; 961 962 G1PrintRegionLivenessInfoClosure _cl; 963 964 class G1UpdateRemSetTrackingBeforeRebuild : public HeapRegionClosure { 965 G1CollectedHeap* _g1h; 966 G1ConcurrentMark* _cm; 967 968 G1PrintRegionLivenessInfoClosure* _cl; 969 970 uint _num_regions_selected_for_rebuild; // The number of regions actually selected for rebuild. 971 972 void update_remset_before_rebuild(HeapRegion* hr) { 973 G1RemSetTrackingPolicy* tracking_policy = _g1h->policy()->remset_tracker(); 974 975 bool selected_for_rebuild; 976 if (hr->is_humongous()) { 977 bool const is_live = _cm->liveness(hr->humongous_start_region()->hrm_index()) > 0; 978 selected_for_rebuild = tracking_policy->update_humongous_before_rebuild(hr, is_live); 979 } else { 980 size_t const live_bytes = _cm->liveness(hr->hrm_index()); 981 selected_for_rebuild = tracking_policy->update_before_rebuild(hr, live_bytes); 982 } 983 if (selected_for_rebuild) { 984 _num_regions_selected_for_rebuild++; 985 } 986 _cm->update_top_at_rebuild_start(hr); 987 } 988 989 // Distribute the given words across the humongous object starting with hr and 990 // note end of marking. 991 void distribute_marked_bytes(HeapRegion* hr, size_t marked_words) { 992 uint const region_idx = hr->hrm_index(); 993 size_t const obj_size_in_words = (size_t)oop(hr->bottom())->size(); 994 uint const num_regions_in_humongous = (uint)G1CollectedHeap::humongous_obj_size_in_regions(obj_size_in_words); 995 996 // "Distributing" zero words means that we only note end of marking for these 997 // regions. 998 assert(marked_words == 0 || obj_size_in_words == marked_words, 999 "Marked words should either be 0 or the same as humongous object (" SIZE_FORMAT ") but is " SIZE_FORMAT, 1000 obj_size_in_words, marked_words); 1001 1002 for (uint i = region_idx; i < (region_idx + num_regions_in_humongous); i++) { 1003 HeapRegion* const r = _g1h->region_at(i); 1004 size_t const words_to_add = MIN2(HeapRegion::GrainWords, marked_words); 1005 1006 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to humongous region %u (%s)", 1007 words_to_add, i, r->get_type_str()); 1008 add_marked_bytes_and_note_end(r, words_to_add * HeapWordSize); 1009 marked_words -= words_to_add; 1010 } 1011 assert(marked_words == 0, 1012 SIZE_FORMAT " words left after distributing space across %u regions", 1013 marked_words, num_regions_in_humongous); 1014 } 1015 1016 void update_marked_bytes(HeapRegion* hr) { 1017 uint const region_idx = hr->hrm_index(); 1018 size_t const marked_words = _cm->liveness(region_idx); 1019 // The marking attributes the object's size completely to the humongous starts 1020 // region. We need to distribute this value across the entire set of regions a 1021 // humongous object spans. 1022 if (hr->is_humongous()) { 1023 assert(hr->is_starts_humongous() || marked_words == 0, 1024 "Should not have marked words " SIZE_FORMAT " in non-starts humongous region %u (%s)", 1025 marked_words, region_idx, hr->get_type_str()); 1026 if (hr->is_starts_humongous()) { 1027 distribute_marked_bytes(hr, marked_words); 1028 } 1029 } else { 1030 log_trace(gc, marking)("Adding " SIZE_FORMAT " words to region %u (%s)", marked_words, region_idx, hr->get_type_str()); 1031 add_marked_bytes_and_note_end(hr, marked_words * HeapWordSize); 1032 } 1033 } 1034 1035 void add_marked_bytes_and_note_end(HeapRegion* hr, size_t marked_bytes) { 1036 hr->add_to_marked_bytes(marked_bytes); 1037 _cl->do_heap_region(hr); 1038 hr->note_end_of_marking(); 1039 } 1040 1041 public: 1042 G1UpdateRemSetTrackingBeforeRebuild(G1CollectedHeap* g1h, G1ConcurrentMark* cm, G1PrintRegionLivenessInfoClosure* cl) : 1043 _g1h(g1h), _cm(cm), _cl(cl), _num_regions_selected_for_rebuild(0) { } 1044 1045 virtual bool do_heap_region(HeapRegion* r) { 1046 update_remset_before_rebuild(r); 1047 update_marked_bytes(r); 1048 1049 return false; 1050 } 1051 1052 uint num_selected_for_rebuild() const { return _num_regions_selected_for_rebuild; } 1053 }; 1054 1055 public: 1056 G1UpdateRemSetTrackingBeforeRebuildTask(G1CollectedHeap* g1h, G1ConcurrentMark* cm, uint num_workers) : 1057 AbstractGangTask("G1 Update RemSet Tracking Before Rebuild"), 1058 _g1h(g1h), _cm(cm), _hrclaimer(num_workers), _total_selected_for_rebuild(0), _cl("Post-Marking") { } 1059 1060 virtual void work(uint worker_id) { 1061 G1UpdateRemSetTrackingBeforeRebuild update_cl(_g1h, _cm, &_cl); 1062 _g1h->heap_region_par_iterate_from_worker_offset(&update_cl, &_hrclaimer, worker_id); 1063 Atomic::add(&_total_selected_for_rebuild, update_cl.num_selected_for_rebuild()); 1064 } 1065 1066 uint total_selected_for_rebuild() const { return _total_selected_for_rebuild; } 1067 1068 // Number of regions for which roughly one thread should be spawned for this work. 1069 static const uint RegionsPerThread = 384; 1070 }; 1071 1072 class G1UpdateRemSetTrackingAfterRebuild : public HeapRegionClosure { 1073 G1CollectedHeap* _g1h; 1074 public: 1075 G1UpdateRemSetTrackingAfterRebuild(G1CollectedHeap* g1h) : _g1h(g1h) { } 1076 1077 virtual bool do_heap_region(HeapRegion* r) { 1078 _g1h->policy()->remset_tracker()->update_after_rebuild(r); 1079 return false; 1080 } 1081 }; 1082 1083 void G1ConcurrentMark::remark() { 1084 assert_at_safepoint_on_vm_thread(); 1085 1086 // If a full collection has happened, we should not continue. However we might 1087 // have ended up here as the Remark VM operation has been scheduled already. 1088 if (has_aborted()) { 1089 return; 1090 } 1091 1092 G1Policy* policy = _g1h->policy(); 1093 policy->record_concurrent_mark_remark_start(); 1094 1095 double start = os::elapsedTime(); 1096 1097 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark before"); 1098 1099 { 1100 GCTraceTime(Debug, gc, phases) debug("Finalize Marking", _gc_timer_cm); 1101 finalize_marking(); 1102 } 1103 1104 double mark_work_end = os::elapsedTime(); 1105 1106 bool const mark_finished = !has_overflown(); 1107 if (mark_finished) { 1108 weak_refs_work(false /* clear_all_soft_refs */); 1109 1110 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 1111 // We're done with marking. 1112 // This is the end of the marking cycle, we're expected all 1113 // threads to have SATB queues with active set to true. 1114 satb_mq_set.set_active_all_threads(false, /* new active value */ 1115 true /* expected_active */); 1116 1117 { 1118 GCTraceTime(Debug, gc, phases) debug("Flush Task Caches", _gc_timer_cm); 1119 flush_all_task_caches(); 1120 } 1121 1122 // Install newly created mark bitmap as "prev". 1123 swap_mark_bitmaps(); 1124 { 1125 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking Before Rebuild", _gc_timer_cm); 1126 1127 uint const workers_by_capacity = (_g1h->num_regions() + G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread - 1) / 1128 G1UpdateRemSetTrackingBeforeRebuildTask::RegionsPerThread; 1129 uint const num_workers = MIN2(_g1h->workers()->active_workers(), workers_by_capacity); 1130 1131 G1UpdateRemSetTrackingBeforeRebuildTask cl(_g1h, this, num_workers); 1132 log_debug(gc,ergo)("Running %s using %u workers for %u regions in heap", cl.name(), num_workers, _g1h->num_regions()); 1133 _g1h->workers()->run_task(&cl, num_workers); 1134 1135 log_debug(gc, remset, tracking)("Remembered Set Tracking update regions total %u, selected %u", 1136 _g1h->num_regions(), cl.total_selected_for_rebuild()); 1137 } 1138 { 1139 GCTraceTime(Debug, gc, phases) debug("Reclaim Empty Regions", _gc_timer_cm); 1140 reclaim_empty_regions(); 1141 } 1142 1143 // Clean out dead classes 1144 if (ClassUnloadingWithConcurrentMark) { 1145 GCTraceTime(Debug, gc, phases) debug("Purge Metaspace", _gc_timer_cm); 1146 ClassLoaderDataGraph::purge(); 1147 } 1148 1149 compute_new_sizes(); 1150 1151 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark after"); 1152 1153 assert(!restart_for_overflow(), "sanity"); 1154 // Completely reset the marking state since marking completed 1155 reset_at_marking_complete(); 1156 } else { 1157 // We overflowed. Restart concurrent marking. 1158 _restart_for_overflow = true; 1159 1160 verify_during_pause(G1HeapVerifier::G1VerifyRemark, VerifyOption_G1UsePrevMarking, "Remark overflow"); 1161 1162 // Clear the marking state because we will be restarting 1163 // marking due to overflowing the global mark stack. 1164 reset_marking_for_restart(); 1165 } 1166 1167 { 1168 GCTraceTime(Debug, gc, phases) debug("Report Object Count", _gc_timer_cm); 1169 report_object_count(mark_finished); 1170 } 1171 1172 // Statistics 1173 double now = os::elapsedTime(); 1174 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1175 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1176 _remark_times.add((now - start) * 1000.0); 1177 1178 policy->record_concurrent_mark_remark_end(); 1179 } 1180 1181 class G1ReclaimEmptyRegionsTask : public AbstractGangTask { 1182 // Per-region work during the Cleanup pause. 1183 class G1ReclaimEmptyRegionsClosure : public HeapRegionClosure { 1184 G1CollectedHeap* _g1h; 1185 size_t _freed_bytes; 1186 FreeRegionList* _local_cleanup_list; 1187 uint _old_regions_removed; 1188 uint _humongous_regions_removed; 1189 1190 public: 1191 G1ReclaimEmptyRegionsClosure(G1CollectedHeap* g1h, 1192 FreeRegionList* local_cleanup_list) : 1193 _g1h(g1h), 1194 _freed_bytes(0), 1195 _local_cleanup_list(local_cleanup_list), 1196 _old_regions_removed(0), 1197 _humongous_regions_removed(0) { } 1198 1199 size_t freed_bytes() { return _freed_bytes; } 1200 const uint old_regions_removed() { return _old_regions_removed; } 1201 const uint humongous_regions_removed() { return _humongous_regions_removed; } 1202 1203 bool do_heap_region(HeapRegion *hr) { 1204 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young() && !hr->is_archive()) { 1205 _freed_bytes += hr->used(); 1206 hr->set_containing_set(NULL); 1207 if (hr->is_humongous()) { 1208 _humongous_regions_removed++; 1209 _g1h->free_humongous_region(hr, _local_cleanup_list); 1210 } else { 1211 _old_regions_removed++; 1212 _g1h->free_region(hr, _local_cleanup_list); 1213 } 1214 hr->clear_cardtable(); 1215 _g1h->concurrent_mark()->clear_statistics_in_region(hr->hrm_index()); 1216 log_trace(gc)("Reclaimed empty region %u (%s) bot " PTR_FORMAT, hr->hrm_index(), hr->get_short_type_str(), p2i(hr->bottom())); 1217 } 1218 1219 return false; 1220 } 1221 }; 1222 1223 G1CollectedHeap* _g1h; 1224 FreeRegionList* _cleanup_list; 1225 HeapRegionClaimer _hrclaimer; 1226 1227 public: 1228 G1ReclaimEmptyRegionsTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) : 1229 AbstractGangTask("G1 Cleanup"), 1230 _g1h(g1h), 1231 _cleanup_list(cleanup_list), 1232 _hrclaimer(n_workers) { 1233 } 1234 1235 void work(uint worker_id) { 1236 FreeRegionList local_cleanup_list("Local Cleanup List"); 1237 G1ReclaimEmptyRegionsClosure cl(_g1h, &local_cleanup_list); 1238 _g1h->heap_region_par_iterate_from_worker_offset(&cl, &_hrclaimer, worker_id); 1239 assert(cl.is_complete(), "Shouldn't have aborted!"); 1240 1241 // Now update the old/humongous region sets 1242 _g1h->remove_from_old_sets(cl.old_regions_removed(), cl.humongous_regions_removed()); 1243 { 1244 MutexLocker x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1245 _g1h->decrement_summary_bytes(cl.freed_bytes()); 1246 1247 _cleanup_list->add_ordered(&local_cleanup_list); 1248 assert(local_cleanup_list.is_empty(), "post-condition"); 1249 } 1250 } 1251 }; 1252 1253 void G1ConcurrentMark::reclaim_empty_regions() { 1254 WorkGang* workers = _g1h->workers(); 1255 FreeRegionList empty_regions_list("Empty Regions After Mark List"); 1256 1257 G1ReclaimEmptyRegionsTask cl(_g1h, &empty_regions_list, workers->active_workers()); 1258 workers->run_task(&cl); 1259 1260 if (!empty_regions_list.is_empty()) { 1261 log_debug(gc)("Reclaimed %u empty regions", empty_regions_list.length()); 1262 // Now print the empty regions list. 1263 G1HRPrinter* hrp = _g1h->hr_printer(); 1264 if (hrp->is_active()) { 1265 FreeRegionListIterator iter(&empty_regions_list); 1266 while (iter.more_available()) { 1267 HeapRegion* hr = iter.get_next(); 1268 hrp->cleanup(hr); 1269 } 1270 } 1271 // And actually make them available. 1272 _g1h->prepend_to_freelist(&empty_regions_list); 1273 } 1274 } 1275 1276 void G1ConcurrentMark::compute_new_sizes() { 1277 MetaspaceGC::compute_new_size(); 1278 1279 // Cleanup will have freed any regions completely full of garbage. 1280 // Update the soft reference policy with the new heap occupancy. 1281 Universe::update_heap_info_at_gc(); 1282 1283 // We reclaimed old regions so we should calculate the sizes to make 1284 // sure we update the old gen/space data. 1285 _g1h->g1mm()->update_sizes(); 1286 } 1287 1288 void G1ConcurrentMark::cleanup() { 1289 assert_at_safepoint_on_vm_thread(); 1290 1291 // If a full collection has happened, we shouldn't do this. 1292 if (has_aborted()) { 1293 return; 1294 } 1295 1296 G1Policy* policy = _g1h->policy(); 1297 policy->record_concurrent_mark_cleanup_start(); 1298 1299 double start = os::elapsedTime(); 1300 1301 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup before"); 1302 1303 { 1304 GCTraceTime(Debug, gc, phases) debug("Update Remembered Set Tracking After Rebuild", _gc_timer_cm); 1305 G1UpdateRemSetTrackingAfterRebuild cl(_g1h); 1306 _g1h->heap_region_iterate(&cl); 1307 } 1308 1309 if (log_is_enabled(Trace, gc, liveness)) { 1310 G1PrintRegionLivenessInfoClosure cl("Post-Cleanup"); 1311 _g1h->heap_region_iterate(&cl); 1312 } 1313 1314 verify_during_pause(G1HeapVerifier::G1VerifyCleanup, VerifyOption_G1UsePrevMarking, "Cleanup after"); 1315 1316 // We need to make this be a "collection" so any collection pause that 1317 // races with it goes around and waits for Cleanup to finish. 1318 _g1h->increment_total_collections(); 1319 1320 // Local statistics 1321 double recent_cleanup_time = (os::elapsedTime() - start); 1322 _total_cleanup_time += recent_cleanup_time; 1323 _cleanup_times.add(recent_cleanup_time); 1324 1325 { 1326 GCTraceTime(Debug, gc, phases) debug("Finalize Concurrent Mark Cleanup", _gc_timer_cm); 1327 policy->record_concurrent_mark_cleanup_end(); 1328 } 1329 } 1330 1331 // 'Keep Alive' oop closure used by both serial parallel reference processing. 1332 // Uses the G1CMTask associated with a worker thread (for serial reference 1333 // processing the G1CMTask for worker 0 is used) to preserve (mark) and 1334 // trace referent objects. 1335 // 1336 // Using the G1CMTask and embedded local queues avoids having the worker 1337 // threads operating on the global mark stack. This reduces the risk 1338 // of overflowing the stack - which we would rather avoid at this late 1339 // state. Also using the tasks' local queues removes the potential 1340 // of the workers interfering with each other that could occur if 1341 // operating on the global stack. 1342 1343 class G1CMKeepAliveAndDrainClosure : public OopClosure { 1344 G1ConcurrentMark* _cm; 1345 G1CMTask* _task; 1346 uint _ref_counter_limit; 1347 uint _ref_counter; 1348 bool _is_serial; 1349 public: 1350 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1351 _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval), 1352 _ref_counter(_ref_counter_limit), _is_serial(is_serial) { 1353 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1354 } 1355 1356 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 1357 virtual void do_oop( oop* p) { do_oop_work(p); } 1358 1359 template <class T> void do_oop_work(T* p) { 1360 if (_cm->has_overflown()) { 1361 return; 1362 } 1363 if (!_task->deal_with_reference(p)) { 1364 // We did not add anything to the mark bitmap (or mark stack), so there is 1365 // no point trying to drain it. 1366 return; 1367 } 1368 _ref_counter--; 1369 1370 if (_ref_counter == 0) { 1371 // We have dealt with _ref_counter_limit references, pushing them 1372 // and objects reachable from them on to the local stack (and 1373 // possibly the global stack). Call G1CMTask::do_marking_step() to 1374 // process these entries. 1375 // 1376 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if 1377 // there's nothing more to do (i.e. we're done with the entries that 1378 // were pushed as a result of the G1CMTask::deal_with_reference() calls 1379 // above) or we overflow. 1380 // 1381 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1382 // flag while there may still be some work to do. (See the comment at 1383 // the beginning of G1CMTask::do_marking_step() for those conditions - 1384 // one of which is reaching the specified time target.) It is only 1385 // when G1CMTask::do_marking_step() returns without setting the 1386 // has_aborted() flag that the marking step has completed. 1387 do { 1388 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1389 _task->do_marking_step(mark_step_duration_ms, 1390 false /* do_termination */, 1391 _is_serial); 1392 } while (_task->has_aborted() && !_cm->has_overflown()); 1393 _ref_counter = _ref_counter_limit; 1394 } 1395 } 1396 }; 1397 1398 // 'Drain' oop closure used by both serial and parallel reference processing. 1399 // Uses the G1CMTask associated with a given worker thread (for serial 1400 // reference processing the G1CMtask for worker 0 is used). Calls the 1401 // do_marking_step routine, with an unbelievably large timeout value, 1402 // to drain the marking data structures of the remaining entries 1403 // added by the 'keep alive' oop closure above. 1404 1405 class G1CMDrainMarkingStackClosure : public VoidClosure { 1406 G1ConcurrentMark* _cm; 1407 G1CMTask* _task; 1408 bool _is_serial; 1409 public: 1410 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1411 _cm(cm), _task(task), _is_serial(is_serial) { 1412 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1413 } 1414 1415 void do_void() { 1416 do { 1417 // We call G1CMTask::do_marking_step() to completely drain the local 1418 // and global marking stacks of entries pushed by the 'keep alive' 1419 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 1420 // 1421 // G1CMTask::do_marking_step() is called in a loop, which we'll exit 1422 // if there's nothing more to do (i.e. we've completely drained the 1423 // entries that were pushed as a a result of applying the 'keep alive' 1424 // closure to the entries on the discovered ref lists) or we overflow 1425 // the global marking stack. 1426 // 1427 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1428 // flag while there may still be some work to do. (See the comment at 1429 // the beginning of G1CMTask::do_marking_step() for those conditions - 1430 // one of which is reaching the specified time target.) It is only 1431 // when G1CMTask::do_marking_step() returns without setting the 1432 // has_aborted() flag that the marking step has completed. 1433 1434 _task->do_marking_step(1000000000.0 /* something very large */, 1435 true /* do_termination */, 1436 _is_serial); 1437 } while (_task->has_aborted() && !_cm->has_overflown()); 1438 } 1439 }; 1440 1441 // Implementation of AbstractRefProcTaskExecutor for parallel 1442 // reference processing at the end of G1 concurrent marking 1443 1444 class G1CMRefProcTaskExecutor : public AbstractRefProcTaskExecutor { 1445 private: 1446 G1CollectedHeap* _g1h; 1447 G1ConcurrentMark* _cm; 1448 WorkGang* _workers; 1449 uint _active_workers; 1450 1451 public: 1452 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 1453 G1ConcurrentMark* cm, 1454 WorkGang* workers, 1455 uint n_workers) : 1456 _g1h(g1h), _cm(cm), 1457 _workers(workers), _active_workers(n_workers) { } 1458 1459 virtual void execute(ProcessTask& task, uint ergo_workers); 1460 }; 1461 1462 class G1CMRefProcTaskProxy : public AbstractGangTask { 1463 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 1464 ProcessTask& _proc_task; 1465 G1CollectedHeap* _g1h; 1466 G1ConcurrentMark* _cm; 1467 1468 public: 1469 G1CMRefProcTaskProxy(ProcessTask& proc_task, 1470 G1CollectedHeap* g1h, 1471 G1ConcurrentMark* cm) : 1472 AbstractGangTask("Process reference objects in parallel"), 1473 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 1474 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1475 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 1476 } 1477 1478 virtual void work(uint worker_id) { 1479 ResourceMark rm; 1480 HandleMark hm; 1481 G1CMTask* task = _cm->task(worker_id); 1482 G1CMIsAliveClosure g1_is_alive(_g1h); 1483 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 1484 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 1485 1486 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 1487 } 1488 }; 1489 1490 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task, uint ergo_workers) { 1491 assert(_workers != NULL, "Need parallel worker threads."); 1492 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1493 assert(_workers->active_workers() >= ergo_workers, 1494 "Ergonomically chosen workers(%u) should be less than or equal to active workers(%u)", 1495 ergo_workers, _workers->active_workers()); 1496 1497 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 1498 1499 // We need to reset the concurrency level before each 1500 // proxy task execution, so that the termination protocol 1501 // and overflow handling in G1CMTask::do_marking_step() knows 1502 // how many workers to wait for. 1503 _cm->set_concurrency(ergo_workers); 1504 _workers->run_task(&proc_task_proxy, ergo_workers); 1505 } 1506 1507 void G1ConcurrentMark::weak_refs_work(bool clear_all_soft_refs) { 1508 ResourceMark rm; 1509 HandleMark hm; 1510 1511 // Is alive closure. 1512 G1CMIsAliveClosure g1_is_alive(_g1h); 1513 1514 // Inner scope to exclude the cleaning of the string table 1515 // from the displayed time. 1516 { 1517 GCTraceTime(Debug, gc, phases) debug("Reference Processing", _gc_timer_cm); 1518 1519 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1520 1521 // See the comment in G1CollectedHeap::ref_processing_init() 1522 // about how reference processing currently works in G1. 1523 1524 // Set the soft reference policy 1525 rp->setup_policy(clear_all_soft_refs); 1526 assert(_global_mark_stack.is_empty(), "mark stack should be empty"); 1527 1528 // Instances of the 'Keep Alive' and 'Complete GC' closures used 1529 // in serial reference processing. Note these closures are also 1530 // used for serially processing (by the the current thread) the 1531 // JNI references during parallel reference processing. 1532 // 1533 // These closures do not need to synchronize with the worker 1534 // threads involved in parallel reference processing as these 1535 // instances are executed serially by the current thread (e.g. 1536 // reference processing is not multi-threaded and is thus 1537 // performed by the current thread instead of a gang worker). 1538 // 1539 // The gang tasks involved in parallel reference processing create 1540 // their own instances of these closures, which do their own 1541 // synchronization among themselves. 1542 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 1543 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 1544 1545 // We need at least one active thread. If reference processing 1546 // is not multi-threaded we use the current (VMThread) thread, 1547 // otherwise we use the work gang from the G1CollectedHeap and 1548 // we utilize all the worker threads we can. 1549 bool processing_is_mt = rp->processing_is_mt(); 1550 uint active_workers = (processing_is_mt ? _g1h->workers()->active_workers() : 1U); 1551 active_workers = clamp(active_workers, 1u, _max_num_tasks); 1552 1553 // Parallel processing task executor. 1554 G1CMRefProcTaskExecutor par_task_executor(_g1h, this, 1555 _g1h->workers(), active_workers); 1556 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 1557 1558 // Set the concurrency level. The phase was already set prior to 1559 // executing the remark task. 1560 set_concurrency(active_workers); 1561 1562 // Set the degree of MT processing here. If the discovery was done MT, 1563 // the number of threads involved during discovery could differ from 1564 // the number of active workers. This is OK as long as the discovered 1565 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 1566 rp->set_active_mt_degree(active_workers); 1567 1568 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 1569 1570 // Process the weak references. 1571 const ReferenceProcessorStats& stats = 1572 rp->process_discovered_references(&g1_is_alive, 1573 &g1_keep_alive, 1574 &g1_drain_mark_stack, 1575 executor, 1576 &pt); 1577 _gc_tracer_cm->report_gc_reference_stats(stats); 1578 pt.print_all_references(); 1579 1580 // The do_oop work routines of the keep_alive and drain_marking_stack 1581 // oop closures will set the has_overflown flag if we overflow the 1582 // global marking stack. 1583 1584 assert(has_overflown() || _global_mark_stack.is_empty(), 1585 "Mark stack should be empty (unless it has overflown)"); 1586 1587 assert(rp->num_queues() == active_workers, "why not"); 1588 1589 rp->verify_no_references_recorded(); 1590 assert(!rp->discovery_enabled(), "Post condition"); 1591 } 1592 1593 if (has_overflown()) { 1594 // We can not trust g1_is_alive and the contents of the heap if the marking stack 1595 // overflowed while processing references. Exit the VM. 1596 fatal("Overflow during reference processing, can not continue. Please " 1597 "increase MarkStackSizeMax (current value: " SIZE_FORMAT ") and " 1598 "restart.", MarkStackSizeMax); 1599 return; 1600 } 1601 1602 assert(_global_mark_stack.is_empty(), "Marking should have completed"); 1603 1604 { 1605 GCTraceTime(Debug, gc, phases) debug("Weak Processing", _gc_timer_cm); 1606 WeakProcessor::weak_oops_do(_g1h->workers(), &g1_is_alive, &do_nothing_cl, 1); 1607 } 1608 1609 // Unload Klasses, String, Code Cache, etc. 1610 if (ClassUnloadingWithConcurrentMark) { 1611 GCTraceTime(Debug, gc, phases) debug("Class Unloading", _gc_timer_cm); 1612 bool purged_classes = SystemDictionary::do_unloading(_gc_timer_cm); 1613 _g1h->complete_cleaning(&g1_is_alive, purged_classes); 1614 } else if (StringDedup::is_enabled()) { 1615 GCTraceTime(Debug, gc, phases) debug("String Deduplication", _gc_timer_cm); 1616 _g1h->string_dedup_cleaning(&g1_is_alive, NULL); 1617 } 1618 } 1619 1620 class G1PrecleanYieldClosure : public YieldClosure { 1621 G1ConcurrentMark* _cm; 1622 1623 public: 1624 G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { } 1625 1626 virtual bool should_return() { 1627 return _cm->has_aborted(); 1628 } 1629 1630 virtual bool should_return_fine_grain() { 1631 _cm->do_yield_check(); 1632 return _cm->has_aborted(); 1633 } 1634 }; 1635 1636 void G1ConcurrentMark::preclean() { 1637 assert(G1UseReferencePrecleaning, "Precleaning must be enabled."); 1638 1639 SuspendibleThreadSetJoiner joiner; 1640 1641 G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */); 1642 G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */); 1643 1644 set_concurrency_and_phase(1, true); 1645 1646 G1PrecleanYieldClosure yield_cl(this); 1647 1648 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1649 // Precleaning is single threaded. Temporarily disable MT discovery. 1650 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 1651 rp->preclean_discovered_references(rp->is_alive_non_header(), 1652 &keep_alive, 1653 &drain_mark_stack, 1654 &yield_cl, 1655 _gc_timer_cm); 1656 } 1657 1658 // When sampling object counts, we already swapped the mark bitmaps, so we need to use 1659 // the prev bitmap determining liveness. 1660 class G1ObjectCountIsAliveClosure: public BoolObjectClosure { 1661 G1CollectedHeap* _g1h; 1662 public: 1663 G1ObjectCountIsAliveClosure(G1CollectedHeap* g1h) : _g1h(g1h) { } 1664 1665 bool do_object_b(oop obj) { 1666 return obj != NULL && 1667 (!_g1h->is_in_g1_reserved(obj) || !_g1h->is_obj_dead(obj)); 1668 } 1669 }; 1670 1671 void G1ConcurrentMark::report_object_count(bool mark_completed) { 1672 // Depending on the completion of the marking liveness needs to be determined 1673 // using either the next or prev bitmap. 1674 if (mark_completed) { 1675 G1ObjectCountIsAliveClosure is_alive(_g1h); 1676 _gc_tracer_cm->report_object_count_after_gc(&is_alive); 1677 } else { 1678 G1CMIsAliveClosure is_alive(_g1h); 1679 _gc_tracer_cm->report_object_count_after_gc(&is_alive); 1680 } 1681 } 1682 1683 1684 void G1ConcurrentMark::swap_mark_bitmaps() { 1685 G1CMBitMap* temp = _prev_mark_bitmap; 1686 _prev_mark_bitmap = _next_mark_bitmap; 1687 _next_mark_bitmap = temp; 1688 _g1h->collector_state()->set_clearing_next_bitmap(true); 1689 } 1690 1691 // Closure for marking entries in SATB buffers. 1692 class G1CMSATBBufferClosure : public SATBBufferClosure { 1693 private: 1694 G1CMTask* _task; 1695 G1CollectedHeap* _g1h; 1696 1697 // This is very similar to G1CMTask::deal_with_reference, but with 1698 // more relaxed requirements for the argument, so this must be more 1699 // circumspect about treating the argument as an object. 1700 void do_entry(void* entry) const { 1701 _task->increment_refs_reached(); 1702 oop const obj = static_cast<oop>(entry); 1703 _task->make_reference_grey(obj); 1704 } 1705 1706 public: 1707 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h) 1708 : _task(task), _g1h(g1h) { } 1709 1710 virtual void do_buffer(void** buffer, size_t size) { 1711 for (size_t i = 0; i < size; ++i) { 1712 do_entry(buffer[i]); 1713 } 1714 } 1715 }; 1716 1717 class G1RemarkThreadsClosure : public ThreadClosure { 1718 G1CMSATBBufferClosure _cm_satb_cl; 1719 G1CMOopClosure _cm_cl; 1720 MarkingCodeBlobClosure _code_cl; 1721 uintx _claim_token; 1722 1723 public: 1724 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) : 1725 _cm_satb_cl(task, g1h), 1726 _cm_cl(g1h, task), 1727 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations), 1728 _claim_token(Threads::thread_claim_token()) {} 1729 1730 void do_thread(Thread* thread) { 1731 if (thread->claim_threads_do(true, _claim_token)) { 1732 SATBMarkQueue& queue = G1ThreadLocalData::satb_mark_queue(thread); 1733 queue.apply_closure_and_empty(&_cm_satb_cl); 1734 if (thread->is_Java_thread()) { 1735 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking 1736 // however the liveness of oops reachable from nmethods have very complex lifecycles: 1737 // * Alive if on the stack of an executing method 1738 // * Weakly reachable otherwise 1739 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be 1740 // live by the SATB invariant but other oops recorded in nmethods may behave differently. 1741 JavaThread* jt = (JavaThread*)thread; 1742 jt->nmethods_do(&_code_cl); 1743 } 1744 } 1745 } 1746 }; 1747 1748 class G1CMRemarkTask : public AbstractGangTask { 1749 G1ConcurrentMark* _cm; 1750 public: 1751 void work(uint worker_id) { 1752 G1CMTask* task = _cm->task(worker_id); 1753 task->record_start_time(); 1754 { 1755 ResourceMark rm; 1756 HandleMark hm; 1757 1758 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task); 1759 Threads::threads_do(&threads_f); 1760 } 1761 1762 do { 1763 task->do_marking_step(1000000000.0 /* something very large */, 1764 true /* do_termination */, 1765 false /* is_serial */); 1766 } while (task->has_aborted() && !_cm->has_overflown()); 1767 // If we overflow, then we do not want to restart. We instead 1768 // want to abort remark and do concurrent marking again. 1769 task->record_end_time(); 1770 } 1771 1772 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) : 1773 AbstractGangTask("Par Remark"), _cm(cm) { 1774 _cm->terminator()->reset_for_reuse(active_workers); 1775 } 1776 }; 1777 1778 void G1ConcurrentMark::finalize_marking() { 1779 ResourceMark rm; 1780 HandleMark hm; 1781 1782 _g1h->ensure_parsability(false); 1783 1784 // this is remark, so we'll use up all active threads 1785 uint active_workers = _g1h->workers()->active_workers(); 1786 set_concurrency_and_phase(active_workers, false /* concurrent */); 1787 // Leave _parallel_marking_threads at it's 1788 // value originally calculated in the G1ConcurrentMark 1789 // constructor and pass values of the active workers 1790 // through the gang in the task. 1791 1792 { 1793 StrongRootsScope srs(active_workers); 1794 1795 G1CMRemarkTask remarkTask(this, active_workers); 1796 // We will start all available threads, even if we decide that the 1797 // active_workers will be fewer. The extra ones will just bail out 1798 // immediately. 1799 _g1h->workers()->run_task(&remarkTask); 1800 } 1801 1802 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 1803 guarantee(has_overflown() || 1804 satb_mq_set.completed_buffers_num() == 0, 1805 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT, 1806 BOOL_TO_STR(has_overflown()), 1807 satb_mq_set.completed_buffers_num()); 1808 1809 print_stats(); 1810 } 1811 1812 void G1ConcurrentMark::flush_all_task_caches() { 1813 size_t hits = 0; 1814 size_t misses = 0; 1815 for (uint i = 0; i < _max_num_tasks; i++) { 1816 Pair<size_t, size_t> stats = _tasks[i]->flush_mark_stats_cache(); 1817 hits += stats.first; 1818 misses += stats.second; 1819 } 1820 size_t sum = hits + misses; 1821 log_debug(gc, stats)("Mark stats cache hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %1.3lf", 1822 hits, misses, percent_of(hits, sum)); 1823 } 1824 1825 void G1ConcurrentMark::clear_range_in_prev_bitmap(MemRegion mr) { 1826 _prev_mark_bitmap->clear_range(mr); 1827 } 1828 1829 HeapRegion* 1830 G1ConcurrentMark::claim_region(uint worker_id) { 1831 // "checkpoint" the finger 1832 HeapWord* finger = _finger; 1833 1834 while (finger < _heap.end()) { 1835 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 1836 1837 HeapRegion* curr_region = _g1h->heap_region_containing(finger); 1838 // Make sure that the reads below do not float before loading curr_region. 1839 OrderAccess::loadload(); 1840 // Above heap_region_containing may return NULL as we always scan claim 1841 // until the end of the heap. In this case, just jump to the next region. 1842 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords; 1843 1844 // Is the gap between reading the finger and doing the CAS too long? 1845 HeapWord* res = Atomic::cmpxchg(&_finger, finger, end); 1846 if (res == finger && curr_region != NULL) { 1847 // we succeeded 1848 HeapWord* bottom = curr_region->bottom(); 1849 HeapWord* limit = curr_region->next_top_at_mark_start(); 1850 1851 // notice that _finger == end cannot be guaranteed here since, 1852 // someone else might have moved the finger even further 1853 assert(_finger >= end, "the finger should have moved forward"); 1854 1855 if (limit > bottom) { 1856 return curr_region; 1857 } else { 1858 assert(limit == bottom, 1859 "the region limit should be at bottom"); 1860 // we return NULL and the caller should try calling 1861 // claim_region() again. 1862 return NULL; 1863 } 1864 } else { 1865 assert(_finger > finger, "the finger should have moved forward"); 1866 // read it again 1867 finger = _finger; 1868 } 1869 } 1870 1871 return NULL; 1872 } 1873 1874 #ifndef PRODUCT 1875 class VerifyNoCSetOops { 1876 G1CollectedHeap* _g1h; 1877 const char* _phase; 1878 int _info; 1879 1880 public: 1881 VerifyNoCSetOops(const char* phase, int info = -1) : 1882 _g1h(G1CollectedHeap::heap()), 1883 _phase(phase), 1884 _info(info) 1885 { } 1886 1887 void operator()(G1TaskQueueEntry task_entry) const { 1888 if (task_entry.is_array_slice()) { 1889 guarantee(_g1h->is_in_reserved(task_entry.slice()), "Slice " PTR_FORMAT " must be in heap.", p2i(task_entry.slice())); 1890 return; 1891 } 1892 guarantee(oopDesc::is_oop(task_entry.obj()), 1893 "Non-oop " PTR_FORMAT ", phase: %s, info: %d", 1894 p2i(task_entry.obj()), _phase, _info); 1895 HeapRegion* r = _g1h->heap_region_containing(task_entry.obj()); 1896 guarantee(!(r->in_collection_set() || r->has_index_in_opt_cset()), 1897 "obj " PTR_FORMAT " from %s (%d) in region %u in (optional) collection set", 1898 p2i(task_entry.obj()), _phase, _info, r->hrm_index()); 1899 } 1900 }; 1901 1902 void G1ConcurrentMark::verify_no_collection_set_oops() { 1903 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1904 if (!_g1h->collector_state()->mark_or_rebuild_in_progress()) { 1905 return; 1906 } 1907 1908 // Verify entries on the global mark stack 1909 _global_mark_stack.iterate(VerifyNoCSetOops("Stack")); 1910 1911 // Verify entries on the task queues 1912 for (uint i = 0; i < _max_num_tasks; ++i) { 1913 G1CMTaskQueue* queue = _task_queues->queue(i); 1914 queue->iterate(VerifyNoCSetOops("Queue", i)); 1915 } 1916 1917 // Verify the global finger 1918 HeapWord* global_finger = finger(); 1919 if (global_finger != NULL && global_finger < _heap.end()) { 1920 // Since we always iterate over all regions, we might get a NULL HeapRegion 1921 // here. 1922 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger); 1923 guarantee(global_hr == NULL || global_finger == global_hr->bottom(), 1924 "global finger: " PTR_FORMAT " region: " HR_FORMAT, 1925 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)); 1926 } 1927 1928 // Verify the task fingers 1929 assert(_num_concurrent_workers <= _max_num_tasks, "sanity"); 1930 for (uint i = 0; i < _num_concurrent_workers; ++i) { 1931 G1CMTask* task = _tasks[i]; 1932 HeapWord* task_finger = task->finger(); 1933 if (task_finger != NULL && task_finger < _heap.end()) { 1934 // See above note on the global finger verification. 1935 HeapRegion* r = _g1h->heap_region_containing(task_finger); 1936 guarantee(r == NULL || task_finger == r->bottom() || 1937 !r->in_collection_set() || !r->has_index_in_opt_cset(), 1938 "task finger: " PTR_FORMAT " region: " HR_FORMAT, 1939 p2i(task_finger), HR_FORMAT_PARAMS(r)); 1940 } 1941 } 1942 } 1943 #endif // PRODUCT 1944 1945 void G1ConcurrentMark::rebuild_rem_set_concurrently() { 1946 _g1h->rem_set()->rebuild_rem_set(this, _concurrent_workers, _worker_id_offset); 1947 } 1948 1949 void G1ConcurrentMark::print_stats() { 1950 if (!log_is_enabled(Debug, gc, stats)) { 1951 return; 1952 } 1953 log_debug(gc, stats)("---------------------------------------------------------------------"); 1954 for (size_t i = 0; i < _num_active_tasks; ++i) { 1955 _tasks[i]->print_stats(); 1956 log_debug(gc, stats)("---------------------------------------------------------------------"); 1957 } 1958 } 1959 1960 void G1ConcurrentMark::concurrent_cycle_abort() { 1961 if (!cm_thread()->during_cycle() || _has_aborted) { 1962 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything. 1963 return; 1964 } 1965 1966 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next 1967 // concurrent bitmap clearing. 1968 { 1969 GCTraceTime(Debug, gc) debug("Clear Next Bitmap"); 1970 clear_bitmap(_next_mark_bitmap, _g1h->workers(), false); 1971 } 1972 // Note we cannot clear the previous marking bitmap here 1973 // since VerifyDuringGC verifies the objects marked during 1974 // a full GC against the previous bitmap. 1975 1976 // Empty mark stack 1977 reset_marking_for_restart(); 1978 for (uint i = 0; i < _max_num_tasks; ++i) { 1979 _tasks[i]->clear_region_fields(); 1980 } 1981 _first_overflow_barrier_sync.abort(); 1982 _second_overflow_barrier_sync.abort(); 1983 _has_aborted = true; 1984 1985 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 1986 satb_mq_set.abandon_partial_marking(); 1987 // This can be called either during or outside marking, we'll read 1988 // the expected_active value from the SATB queue set. 1989 satb_mq_set.set_active_all_threads( 1990 false, /* new active value */ 1991 satb_mq_set.is_active() /* expected_active */); 1992 } 1993 1994 static void print_ms_time_info(const char* prefix, const char* name, 1995 NumberSeq& ns) { 1996 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 1997 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 1998 if (ns.num() > 0) { 1999 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]", 2000 prefix, ns.sd(), ns.maximum()); 2001 } 2002 } 2003 2004 void G1ConcurrentMark::print_summary_info() { 2005 Log(gc, marking) log; 2006 if (!log.is_trace()) { 2007 return; 2008 } 2009 2010 log.trace(" Concurrent marking:"); 2011 print_ms_time_info(" ", "init marks", _init_times); 2012 print_ms_time_info(" ", "remarks", _remark_times); 2013 { 2014 print_ms_time_info(" ", "final marks", _remark_mark_times); 2015 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 2016 2017 } 2018 print_ms_time_info(" ", "cleanups", _cleanup_times); 2019 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).", 2020 _total_cleanup_time, (_cleanup_times.num() > 0 ? _total_cleanup_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2021 log.trace(" Total stop_world time = %8.2f s.", 2022 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0); 2023 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).", 2024 cm_thread()->vtime_accum(), cm_thread()->vtime_mark_accum()); 2025 } 2026 2027 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const { 2028 _concurrent_workers->threads_do(tc); 2029 } 2030 2031 void G1ConcurrentMark::print_on_error(outputStream* st) const { 2032 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT, 2033 p2i(_prev_mark_bitmap), p2i(_next_mark_bitmap)); 2034 _prev_mark_bitmap->print_on_error(st, " Prev Bits: "); 2035 _next_mark_bitmap->print_on_error(st, " Next Bits: "); 2036 } 2037 2038 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) { 2039 ReferenceProcessor* result = g1h->ref_processor_cm(); 2040 assert(result != NULL, "CM reference processor should not be NULL"); 2041 return result; 2042 } 2043 2044 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 2045 G1CMTask* task) 2046 : MetadataVisitingOopIterateClosure(get_cm_oop_closure_ref_processor(g1h)), 2047 _g1h(g1h), _task(task) 2048 { } 2049 2050 void G1CMTask::setup_for_region(HeapRegion* hr) { 2051 assert(hr != NULL, 2052 "claim_region() should have filtered out NULL regions"); 2053 _curr_region = hr; 2054 _finger = hr->bottom(); 2055 update_region_limit(); 2056 } 2057 2058 void G1CMTask::update_region_limit() { 2059 HeapRegion* hr = _curr_region; 2060 HeapWord* bottom = hr->bottom(); 2061 HeapWord* limit = hr->next_top_at_mark_start(); 2062 2063 if (limit == bottom) { 2064 // The region was collected underneath our feet. 2065 // We set the finger to bottom to ensure that the bitmap 2066 // iteration that will follow this will not do anything. 2067 // (this is not a condition that holds when we set the region up, 2068 // as the region is not supposed to be empty in the first place) 2069 _finger = bottom; 2070 } else if (limit >= _region_limit) { 2071 assert(limit >= _finger, "peace of mind"); 2072 } else { 2073 assert(limit < _region_limit, "only way to get here"); 2074 // This can happen under some pretty unusual circumstances. An 2075 // evacuation pause empties the region underneath our feet (NTAMS 2076 // at bottom). We then do some allocation in the region (NTAMS 2077 // stays at bottom), followed by the region being used as a GC 2078 // alloc region (NTAMS will move to top() and the objects 2079 // originally below it will be grayed). All objects now marked in 2080 // the region are explicitly grayed, if below the global finger, 2081 // and we do not need in fact to scan anything else. So, we simply 2082 // set _finger to be limit to ensure that the bitmap iteration 2083 // doesn't do anything. 2084 _finger = limit; 2085 } 2086 2087 _region_limit = limit; 2088 } 2089 2090 void G1CMTask::giveup_current_region() { 2091 assert(_curr_region != NULL, "invariant"); 2092 clear_region_fields(); 2093 } 2094 2095 void G1CMTask::clear_region_fields() { 2096 // Values for these three fields that indicate that we're not 2097 // holding on to a region. 2098 _curr_region = NULL; 2099 _finger = NULL; 2100 _region_limit = NULL; 2101 } 2102 2103 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 2104 if (cm_oop_closure == NULL) { 2105 assert(_cm_oop_closure != NULL, "invariant"); 2106 } else { 2107 assert(_cm_oop_closure == NULL, "invariant"); 2108 } 2109 _cm_oop_closure = cm_oop_closure; 2110 } 2111 2112 void G1CMTask::reset(G1CMBitMap* next_mark_bitmap) { 2113 guarantee(next_mark_bitmap != NULL, "invariant"); 2114 _next_mark_bitmap = next_mark_bitmap; 2115 clear_region_fields(); 2116 2117 _calls = 0; 2118 _elapsed_time_ms = 0.0; 2119 _termination_time_ms = 0.0; 2120 _termination_start_time_ms = 0.0; 2121 2122 _mark_stats_cache.reset(); 2123 } 2124 2125 bool G1CMTask::should_exit_termination() { 2126 if (!regular_clock_call()) { 2127 return true; 2128 } 2129 2130 // This is called when we are in the termination protocol. We should 2131 // quit if, for some reason, this task wants to abort or the global 2132 // stack is not empty (this means that we can get work from it). 2133 return !_cm->mark_stack_empty() || has_aborted(); 2134 } 2135 2136 void G1CMTask::reached_limit() { 2137 assert(_words_scanned >= _words_scanned_limit || 2138 _refs_reached >= _refs_reached_limit , 2139 "shouldn't have been called otherwise"); 2140 abort_marking_if_regular_check_fail(); 2141 } 2142 2143 bool G1CMTask::regular_clock_call() { 2144 if (has_aborted()) { 2145 return false; 2146 } 2147 2148 // First, we need to recalculate the words scanned and refs reached 2149 // limits for the next clock call. 2150 recalculate_limits(); 2151 2152 // During the regular clock call we do the following 2153 2154 // (1) If an overflow has been flagged, then we abort. 2155 if (_cm->has_overflown()) { 2156 return false; 2157 } 2158 2159 // If we are not concurrent (i.e. we're doing remark) we don't need 2160 // to check anything else. The other steps are only needed during 2161 // the concurrent marking phase. 2162 if (!_cm->concurrent()) { 2163 return true; 2164 } 2165 2166 // (2) If marking has been aborted for Full GC, then we also abort. 2167 if (_cm->has_aborted()) { 2168 return false; 2169 } 2170 2171 double curr_time_ms = os::elapsedVTime() * 1000.0; 2172 2173 // (4) We check whether we should yield. If we have to, then we abort. 2174 if (SuspendibleThreadSet::should_yield()) { 2175 // We should yield. To do this we abort the task. The caller is 2176 // responsible for yielding. 2177 return false; 2178 } 2179 2180 // (5) We check whether we've reached our time quota. If we have, 2181 // then we abort. 2182 double elapsed_time_ms = curr_time_ms - _start_time_ms; 2183 if (elapsed_time_ms > _time_target_ms) { 2184 _has_timed_out = true; 2185 return false; 2186 } 2187 2188 // (6) Finally, we check whether there are enough completed STAB 2189 // buffers available for processing. If there are, we abort. 2190 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 2191 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 2192 // we do need to process SATB buffers, we'll abort and restart 2193 // the marking task to do so 2194 return false; 2195 } 2196 return true; 2197 } 2198 2199 void G1CMTask::recalculate_limits() { 2200 _real_words_scanned_limit = _words_scanned + words_scanned_period; 2201 _words_scanned_limit = _real_words_scanned_limit; 2202 2203 _real_refs_reached_limit = _refs_reached + refs_reached_period; 2204 _refs_reached_limit = _real_refs_reached_limit; 2205 } 2206 2207 void G1CMTask::decrease_limits() { 2208 // This is called when we believe that we're going to do an infrequent 2209 // operation which will increase the per byte scanned cost (i.e. move 2210 // entries to/from the global stack). It basically tries to decrease the 2211 // scanning limit so that the clock is called earlier. 2212 2213 _words_scanned_limit = _real_words_scanned_limit - 3 * words_scanned_period / 4; 2214 _refs_reached_limit = _real_refs_reached_limit - 3 * refs_reached_period / 4; 2215 } 2216 2217 void G1CMTask::move_entries_to_global_stack() { 2218 // Local array where we'll store the entries that will be popped 2219 // from the local queue. 2220 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk]; 2221 2222 size_t n = 0; 2223 G1TaskQueueEntry task_entry; 2224 while (n < G1CMMarkStack::EntriesPerChunk && _task_queue->pop_local(task_entry)) { 2225 buffer[n] = task_entry; 2226 ++n; 2227 } 2228 if (n < G1CMMarkStack::EntriesPerChunk) { 2229 buffer[n] = G1TaskQueueEntry(); 2230 } 2231 2232 if (n > 0) { 2233 if (!_cm->mark_stack_push(buffer)) { 2234 set_has_aborted(); 2235 } 2236 } 2237 2238 // This operation was quite expensive, so decrease the limits. 2239 decrease_limits(); 2240 } 2241 2242 bool G1CMTask::get_entries_from_global_stack() { 2243 // Local array where we'll store the entries that will be popped 2244 // from the global stack. 2245 G1TaskQueueEntry buffer[G1CMMarkStack::EntriesPerChunk]; 2246 2247 if (!_cm->mark_stack_pop(buffer)) { 2248 return false; 2249 } 2250 2251 // We did actually pop at least one entry. 2252 for (size_t i = 0; i < G1CMMarkStack::EntriesPerChunk; ++i) { 2253 G1TaskQueueEntry task_entry = buffer[i]; 2254 if (task_entry.is_null()) { 2255 break; 2256 } 2257 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())); 2258 bool success = _task_queue->push(task_entry); 2259 // We only call this when the local queue is empty or under a 2260 // given target limit. So, we do not expect this push to fail. 2261 assert(success, "invariant"); 2262 } 2263 2264 // This operation was quite expensive, so decrease the limits 2265 decrease_limits(); 2266 return true; 2267 } 2268 2269 void G1CMTask::drain_local_queue(bool partially) { 2270 if (has_aborted()) { 2271 return; 2272 } 2273 2274 // Decide what the target size is, depending whether we're going to 2275 // drain it partially (so that other tasks can steal if they run out 2276 // of things to do) or totally (at the very end). 2277 size_t target_size; 2278 if (partially) { 2279 target_size = MIN2((size_t)_task_queue->max_elems()/3, (size_t)GCDrainStackTargetSize); 2280 } else { 2281 target_size = 0; 2282 } 2283 2284 if (_task_queue->size() > target_size) { 2285 G1TaskQueueEntry entry; 2286 bool ret = _task_queue->pop_local(entry); 2287 while (ret) { 2288 scan_task_entry(entry); 2289 if (_task_queue->size() <= target_size || has_aborted()) { 2290 ret = false; 2291 } else { 2292 ret = _task_queue->pop_local(entry); 2293 } 2294 } 2295 } 2296 } 2297 2298 void G1CMTask::drain_global_stack(bool partially) { 2299 if (has_aborted()) { 2300 return; 2301 } 2302 2303 // We have a policy to drain the local queue before we attempt to 2304 // drain the global stack. 2305 assert(partially || _task_queue->size() == 0, "invariant"); 2306 2307 // Decide what the target size is, depending whether we're going to 2308 // drain it partially (so that other tasks can steal if they run out 2309 // of things to do) or totally (at the very end). 2310 // Notice that when draining the global mark stack partially, due to the racyness 2311 // of the mark stack size update we might in fact drop below the target. But, 2312 // this is not a problem. 2313 // In case of total draining, we simply process until the global mark stack is 2314 // totally empty, disregarding the size counter. 2315 if (partially) { 2316 size_t const target_size = _cm->partial_mark_stack_size_target(); 2317 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 2318 if (get_entries_from_global_stack()) { 2319 drain_local_queue(partially); 2320 } 2321 } 2322 } else { 2323 while (!has_aborted() && get_entries_from_global_stack()) { 2324 drain_local_queue(partially); 2325 } 2326 } 2327 } 2328 2329 // SATB Queue has several assumptions on whether to call the par or 2330 // non-par versions of the methods. this is why some of the code is 2331 // replicated. We should really get rid of the single-threaded version 2332 // of the code to simplify things. 2333 void G1CMTask::drain_satb_buffers() { 2334 if (has_aborted()) { 2335 return; 2336 } 2337 2338 // We set this so that the regular clock knows that we're in the 2339 // middle of draining buffers and doesn't set the abort flag when it 2340 // notices that SATB buffers are available for draining. It'd be 2341 // very counter productive if it did that. :-) 2342 _draining_satb_buffers = true; 2343 2344 G1CMSATBBufferClosure satb_cl(this, _g1h); 2345 SATBMarkQueueSet& satb_mq_set = G1BarrierSet::satb_mark_queue_set(); 2346 2347 // This keeps claiming and applying the closure to completed buffers 2348 // until we run out of buffers or we need to abort. 2349 while (!has_aborted() && 2350 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) { 2351 abort_marking_if_regular_check_fail(); 2352 } 2353 2354 // Can't assert qset is empty here, even if not aborted. If concurrent, 2355 // some other thread might be adding to the queue. If not concurrent, 2356 // some other thread might have won the race for the last buffer, but 2357 // has not yet decremented the count. 2358 2359 _draining_satb_buffers = false; 2360 2361 // again, this was a potentially expensive operation, decrease the 2362 // limits to get the regular clock call early 2363 decrease_limits(); 2364 } 2365 2366 void G1CMTask::clear_mark_stats_cache(uint region_idx) { 2367 _mark_stats_cache.reset(region_idx); 2368 } 2369 2370 Pair<size_t, size_t> G1CMTask::flush_mark_stats_cache() { 2371 return _mark_stats_cache.evict_all(); 2372 } 2373 2374 void G1CMTask::print_stats() { 2375 log_debug(gc, stats)("Marking Stats, task = %u, calls = %u", _worker_id, _calls); 2376 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 2377 _elapsed_time_ms, _termination_time_ms); 2378 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms max = %1.2lfms, total = %1.2lfms", 2379 _step_times_ms.num(), 2380 _step_times_ms.avg(), 2381 _step_times_ms.sd(), 2382 _step_times_ms.maximum(), 2383 _step_times_ms.sum()); 2384 size_t const hits = _mark_stats_cache.hits(); 2385 size_t const misses = _mark_stats_cache.misses(); 2386 log_debug(gc, stats)(" Mark Stats Cache: hits " SIZE_FORMAT " misses " SIZE_FORMAT " ratio %.3f", 2387 hits, misses, percent_of(hits, hits + misses)); 2388 } 2389 2390 bool G1ConcurrentMark::try_stealing(uint worker_id, G1TaskQueueEntry& task_entry) { 2391 return _task_queues->steal(worker_id, task_entry); 2392 } 2393 2394 /***************************************************************************** 2395 2396 The do_marking_step(time_target_ms, ...) method is the building 2397 block of the parallel marking framework. It can be called in parallel 2398 with other invocations of do_marking_step() on different tasks 2399 (but only one per task, obviously) and concurrently with the 2400 mutator threads, or during remark, hence it eliminates the need 2401 for two versions of the code. When called during remark, it will 2402 pick up from where the task left off during the concurrent marking 2403 phase. Interestingly, tasks are also claimable during evacuation 2404 pauses too, since do_marking_step() ensures that it aborts before 2405 it needs to yield. 2406 2407 The data structures that it uses to do marking work are the 2408 following: 2409 2410 (1) Marking Bitmap. If there are gray objects that appear only 2411 on the bitmap (this happens either when dealing with an overflow 2412 or when the initial marking phase has simply marked the roots 2413 and didn't push them on the stack), then tasks claim heap 2414 regions whose bitmap they then scan to find gray objects. A 2415 global finger indicates where the end of the last claimed region 2416 is. A local finger indicates how far into the region a task has 2417 scanned. The two fingers are used to determine how to gray an 2418 object (i.e. whether simply marking it is OK, as it will be 2419 visited by a task in the future, or whether it needs to be also 2420 pushed on a stack). 2421 2422 (2) Local Queue. The local queue of the task which is accessed 2423 reasonably efficiently by the task. Other tasks can steal from 2424 it when they run out of work. Throughout the marking phase, a 2425 task attempts to keep its local queue short but not totally 2426 empty, so that entries are available for stealing by other 2427 tasks. Only when there is no more work, a task will totally 2428 drain its local queue. 2429 2430 (3) Global Mark Stack. This handles local queue overflow. During 2431 marking only sets of entries are moved between it and the local 2432 queues, as access to it requires a mutex and more fine-grain 2433 interaction with it which might cause contention. If it 2434 overflows, then the marking phase should restart and iterate 2435 over the bitmap to identify gray objects. Throughout the marking 2436 phase, tasks attempt to keep the global mark stack at a small 2437 length but not totally empty, so that entries are available for 2438 popping by other tasks. Only when there is no more work, tasks 2439 will totally drain the global mark stack. 2440 2441 (4) SATB Buffer Queue. This is where completed SATB buffers are 2442 made available. Buffers are regularly removed from this queue 2443 and scanned for roots, so that the queue doesn't get too 2444 long. During remark, all completed buffers are processed, as 2445 well as the filled in parts of any uncompleted buffers. 2446 2447 The do_marking_step() method tries to abort when the time target 2448 has been reached. There are a few other cases when the 2449 do_marking_step() method also aborts: 2450 2451 (1) When the marking phase has been aborted (after a Full GC). 2452 2453 (2) When a global overflow (on the global stack) has been 2454 triggered. Before the task aborts, it will actually sync up with 2455 the other tasks to ensure that all the marking data structures 2456 (local queues, stacks, fingers etc.) are re-initialized so that 2457 when do_marking_step() completes, the marking phase can 2458 immediately restart. 2459 2460 (3) When enough completed SATB buffers are available. The 2461 do_marking_step() method only tries to drain SATB buffers right 2462 at the beginning. So, if enough buffers are available, the 2463 marking step aborts and the SATB buffers are processed at 2464 the beginning of the next invocation. 2465 2466 (4) To yield. when we have to yield then we abort and yield 2467 right at the end of do_marking_step(). This saves us from a lot 2468 of hassle as, by yielding we might allow a Full GC. If this 2469 happens then objects will be compacted underneath our feet, the 2470 heap might shrink, etc. We save checking for this by just 2471 aborting and doing the yield right at the end. 2472 2473 From the above it follows that the do_marking_step() method should 2474 be called in a loop (or, otherwise, regularly) until it completes. 2475 2476 If a marking step completes without its has_aborted() flag being 2477 true, it means it has completed the current marking phase (and 2478 also all other marking tasks have done so and have all synced up). 2479 2480 A method called regular_clock_call() is invoked "regularly" (in 2481 sub ms intervals) throughout marking. It is this clock method that 2482 checks all the abort conditions which were mentioned above and 2483 decides when the task should abort. A work-based scheme is used to 2484 trigger this clock method: when the number of object words the 2485 marking phase has scanned or the number of references the marking 2486 phase has visited reach a given limit. Additional invocations to 2487 the method clock have been planted in a few other strategic places 2488 too. The initial reason for the clock method was to avoid calling 2489 vtime too regularly, as it is quite expensive. So, once it was in 2490 place, it was natural to piggy-back all the other conditions on it 2491 too and not constantly check them throughout the code. 2492 2493 If do_termination is true then do_marking_step will enter its 2494 termination protocol. 2495 2496 The value of is_serial must be true when do_marking_step is being 2497 called serially (i.e. by the VMThread) and do_marking_step should 2498 skip any synchronization in the termination and overflow code. 2499 Examples include the serial remark code and the serial reference 2500 processing closures. 2501 2502 The value of is_serial must be false when do_marking_step is 2503 being called by any of the worker threads in a work gang. 2504 Examples include the concurrent marking code (CMMarkingTask), 2505 the MT remark code, and the MT reference processing closures. 2506 2507 *****************************************************************************/ 2508 2509 void G1CMTask::do_marking_step(double time_target_ms, 2510 bool do_termination, 2511 bool is_serial) { 2512 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 2513 2514 _start_time_ms = os::elapsedVTime() * 1000.0; 2515 2516 // If do_stealing is true then do_marking_step will attempt to 2517 // steal work from the other G1CMTasks. It only makes sense to 2518 // enable stealing when the termination protocol is enabled 2519 // and do_marking_step() is not being called serially. 2520 bool do_stealing = do_termination && !is_serial; 2521 2522 G1Predictions const& predictor = _g1h->policy()->predictor(); 2523 double diff_prediction_ms = predictor.predict_zero_bounded(&_marking_step_diff_ms); 2524 _time_target_ms = time_target_ms - diff_prediction_ms; 2525 2526 // set up the variables that are used in the work-based scheme to 2527 // call the regular clock method 2528 _words_scanned = 0; 2529 _refs_reached = 0; 2530 recalculate_limits(); 2531 2532 // clear all flags 2533 clear_has_aborted(); 2534 _has_timed_out = false; 2535 _draining_satb_buffers = false; 2536 2537 ++_calls; 2538 2539 // Set up the bitmap and oop closures. Anything that uses them is 2540 // eventually called from this method, so it is OK to allocate these 2541 // statically. 2542 G1CMBitMapClosure bitmap_closure(this, _cm); 2543 G1CMOopClosure cm_oop_closure(_g1h, this); 2544 set_cm_oop_closure(&cm_oop_closure); 2545 2546 if (_cm->has_overflown()) { 2547 // This can happen if the mark stack overflows during a GC pause 2548 // and this task, after a yield point, restarts. We have to abort 2549 // as we need to get into the overflow protocol which happens 2550 // right at the end of this task. 2551 set_has_aborted(); 2552 } 2553 2554 // First drain any available SATB buffers. After this, we will not 2555 // look at SATB buffers before the next invocation of this method. 2556 // If enough completed SATB buffers are queued up, the regular clock 2557 // will abort this task so that it restarts. 2558 drain_satb_buffers(); 2559 // ...then partially drain the local queue and the global stack 2560 drain_local_queue(true); 2561 drain_global_stack(true); 2562 2563 do { 2564 if (!has_aborted() && _curr_region != NULL) { 2565 // This means that we're already holding on to a region. 2566 assert(_finger != NULL, "if region is not NULL, then the finger " 2567 "should not be NULL either"); 2568 2569 // We might have restarted this task after an evacuation pause 2570 // which might have evacuated the region we're holding on to 2571 // underneath our feet. Let's read its limit again to make sure 2572 // that we do not iterate over a region of the heap that 2573 // contains garbage (update_region_limit() will also move 2574 // _finger to the start of the region if it is found empty). 2575 update_region_limit(); 2576 // We will start from _finger not from the start of the region, 2577 // as we might be restarting this task after aborting half-way 2578 // through scanning this region. In this case, _finger points to 2579 // the address where we last found a marked object. If this is a 2580 // fresh region, _finger points to start(). 2581 MemRegion mr = MemRegion(_finger, _region_limit); 2582 2583 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(), 2584 "humongous regions should go around loop once only"); 2585 2586 // Some special cases: 2587 // If the memory region is empty, we can just give up the region. 2588 // If the current region is humongous then we only need to check 2589 // the bitmap for the bit associated with the start of the object, 2590 // scan the object if it's live, and give up the region. 2591 // Otherwise, let's iterate over the bitmap of the part of the region 2592 // that is left. 2593 // If the iteration is successful, give up the region. 2594 if (mr.is_empty()) { 2595 giveup_current_region(); 2596 abort_marking_if_regular_check_fail(); 2597 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) { 2598 if (_next_mark_bitmap->is_marked(mr.start())) { 2599 // The object is marked - apply the closure 2600 bitmap_closure.do_addr(mr.start()); 2601 } 2602 // Even if this task aborted while scanning the humongous object 2603 // we can (and should) give up the current region. 2604 giveup_current_region(); 2605 abort_marking_if_regular_check_fail(); 2606 } else if (_next_mark_bitmap->iterate(&bitmap_closure, mr)) { 2607 giveup_current_region(); 2608 abort_marking_if_regular_check_fail(); 2609 } else { 2610 assert(has_aborted(), "currently the only way to do so"); 2611 // The only way to abort the bitmap iteration is to return 2612 // false from the do_bit() method. However, inside the 2613 // do_bit() method we move the _finger to point to the 2614 // object currently being looked at. So, if we bail out, we 2615 // have definitely set _finger to something non-null. 2616 assert(_finger != NULL, "invariant"); 2617 2618 // Region iteration was actually aborted. So now _finger 2619 // points to the address of the object we last scanned. If we 2620 // leave it there, when we restart this task, we will rescan 2621 // the object. It is easy to avoid this. We move the finger by 2622 // enough to point to the next possible object header. 2623 assert(_finger < _region_limit, "invariant"); 2624 HeapWord* const new_finger = _finger + ((oop)_finger)->size(); 2625 // Check if bitmap iteration was aborted while scanning the last object 2626 if (new_finger >= _region_limit) { 2627 giveup_current_region(); 2628 } else { 2629 move_finger_to(new_finger); 2630 } 2631 } 2632 } 2633 // At this point we have either completed iterating over the 2634 // region we were holding on to, or we have aborted. 2635 2636 // We then partially drain the local queue and the global stack. 2637 // (Do we really need this?) 2638 drain_local_queue(true); 2639 drain_global_stack(true); 2640 2641 // Read the note on the claim_region() method on why it might 2642 // return NULL with potentially more regions available for 2643 // claiming and why we have to check out_of_regions() to determine 2644 // whether we're done or not. 2645 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 2646 // We are going to try to claim a new region. We should have 2647 // given up on the previous one. 2648 // Separated the asserts so that we know which one fires. 2649 assert(_curr_region == NULL, "invariant"); 2650 assert(_finger == NULL, "invariant"); 2651 assert(_region_limit == NULL, "invariant"); 2652 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 2653 if (claimed_region != NULL) { 2654 // Yes, we managed to claim one 2655 setup_for_region(claimed_region); 2656 assert(_curr_region == claimed_region, "invariant"); 2657 } 2658 // It is important to call the regular clock here. It might take 2659 // a while to claim a region if, for example, we hit a large 2660 // block of empty regions. So we need to call the regular clock 2661 // method once round the loop to make sure it's called 2662 // frequently enough. 2663 abort_marking_if_regular_check_fail(); 2664 } 2665 2666 if (!has_aborted() && _curr_region == NULL) { 2667 assert(_cm->out_of_regions(), 2668 "at this point we should be out of regions"); 2669 } 2670 } while ( _curr_region != NULL && !has_aborted()); 2671 2672 if (!has_aborted()) { 2673 // We cannot check whether the global stack is empty, since other 2674 // tasks might be pushing objects to it concurrently. 2675 assert(_cm->out_of_regions(), 2676 "at this point we should be out of regions"); 2677 // Try to reduce the number of available SATB buffers so that 2678 // remark has less work to do. 2679 drain_satb_buffers(); 2680 } 2681 2682 // Since we've done everything else, we can now totally drain the 2683 // local queue and global stack. 2684 drain_local_queue(false); 2685 drain_global_stack(false); 2686 2687 // Attempt at work stealing from other task's queues. 2688 if (do_stealing && !has_aborted()) { 2689 // We have not aborted. This means that we have finished all that 2690 // we could. Let's try to do some stealing... 2691 2692 // We cannot check whether the global stack is empty, since other 2693 // tasks might be pushing objects to it concurrently. 2694 assert(_cm->out_of_regions() && _task_queue->size() == 0, 2695 "only way to reach here"); 2696 while (!has_aborted()) { 2697 G1TaskQueueEntry entry; 2698 if (_cm->try_stealing(_worker_id, entry)) { 2699 scan_task_entry(entry); 2700 2701 // And since we're towards the end, let's totally drain the 2702 // local queue and global stack. 2703 drain_local_queue(false); 2704 drain_global_stack(false); 2705 } else { 2706 break; 2707 } 2708 } 2709 } 2710 2711 // We still haven't aborted. Now, let's try to get into the 2712 // termination protocol. 2713 if (do_termination && !has_aborted()) { 2714 // We cannot check whether the global stack is empty, since other 2715 // tasks might be concurrently pushing objects on it. 2716 // Separated the asserts so that we know which one fires. 2717 assert(_cm->out_of_regions(), "only way to reach here"); 2718 assert(_task_queue->size() == 0, "only way to reach here"); 2719 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 2720 2721 // The G1CMTask class also extends the TerminatorTerminator class, 2722 // hence its should_exit_termination() method will also decide 2723 // whether to exit the termination protocol or not. 2724 bool finished = (is_serial || 2725 _cm->terminator()->offer_termination(this)); 2726 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 2727 _termination_time_ms += 2728 termination_end_time_ms - _termination_start_time_ms; 2729 2730 if (finished) { 2731 // We're all done. 2732 2733 // We can now guarantee that the global stack is empty, since 2734 // all other tasks have finished. We separated the guarantees so 2735 // that, if a condition is false, we can immediately find out 2736 // which one. 2737 guarantee(_cm->out_of_regions(), "only way to reach here"); 2738 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 2739 guarantee(_task_queue->size() == 0, "only way to reach here"); 2740 guarantee(!_cm->has_overflown(), "only way to reach here"); 2741 guarantee(!has_aborted(), "should never happen if termination has completed"); 2742 } else { 2743 // Apparently there's more work to do. Let's abort this task. It 2744 // will restart it and we can hopefully find more things to do. 2745 set_has_aborted(); 2746 } 2747 } 2748 2749 // Mainly for debugging purposes to make sure that a pointer to the 2750 // closure which was statically allocated in this frame doesn't 2751 // escape it by accident. 2752 set_cm_oop_closure(NULL); 2753 double end_time_ms = os::elapsedVTime() * 1000.0; 2754 double elapsed_time_ms = end_time_ms - _start_time_ms; 2755 // Update the step history. 2756 _step_times_ms.add(elapsed_time_ms); 2757 2758 if (has_aborted()) { 2759 // The task was aborted for some reason. 2760 if (_has_timed_out) { 2761 double diff_ms = elapsed_time_ms - _time_target_ms; 2762 // Keep statistics of how well we did with respect to hitting 2763 // our target only if we actually timed out (if we aborted for 2764 // other reasons, then the results might get skewed). 2765 _marking_step_diff_ms.add(diff_ms); 2766 } 2767 2768 if (_cm->has_overflown()) { 2769 // This is the interesting one. We aborted because a global 2770 // overflow was raised. This means we have to restart the 2771 // marking phase and start iterating over regions. However, in 2772 // order to do this we have to make sure that all tasks stop 2773 // what they are doing and re-initialize in a safe manner. We 2774 // will achieve this with the use of two barrier sync points. 2775 2776 if (!is_serial) { 2777 // We only need to enter the sync barrier if being called 2778 // from a parallel context 2779 _cm->enter_first_sync_barrier(_worker_id); 2780 2781 // When we exit this sync barrier we know that all tasks have 2782 // stopped doing marking work. So, it's now safe to 2783 // re-initialize our data structures. 2784 } 2785 2786 clear_region_fields(); 2787 flush_mark_stats_cache(); 2788 2789 if (!is_serial) { 2790 // If we're executing the concurrent phase of marking, reset the marking 2791 // state; otherwise the marking state is reset after reference processing, 2792 // during the remark pause. 2793 // If we reset here as a result of an overflow during the remark we will 2794 // see assertion failures from any subsequent set_concurrency_and_phase() 2795 // calls. 2796 if (_cm->concurrent() && _worker_id == 0) { 2797 // Worker 0 is responsible for clearing the global data structures because 2798 // of an overflow. During STW we should not clear the overflow flag (in 2799 // G1ConcurrentMark::reset_marking_state()) since we rely on it being true when we exit 2800 // method to abort the pause and restart concurrent marking. 2801 _cm->reset_marking_for_restart(); 2802 2803 log_info(gc, marking)("Concurrent Mark reset for overflow"); 2804 } 2805 2806 // ...and enter the second barrier. 2807 _cm->enter_second_sync_barrier(_worker_id); 2808 } 2809 // At this point, if we're during the concurrent phase of 2810 // marking, everything has been re-initialized and we're 2811 // ready to restart. 2812 } 2813 } 2814 } 2815 2816 G1CMTask::G1CMTask(uint worker_id, 2817 G1ConcurrentMark* cm, 2818 G1CMTaskQueue* task_queue, 2819 G1RegionMarkStats* mark_stats, 2820 uint max_regions) : 2821 _objArray_processor(this), 2822 _worker_id(worker_id), 2823 _g1h(G1CollectedHeap::heap()), 2824 _cm(cm), 2825 _next_mark_bitmap(NULL), 2826 _task_queue(task_queue), 2827 _mark_stats_cache(mark_stats, max_regions, RegionMarkStatsCacheSize), 2828 _calls(0), 2829 _time_target_ms(0.0), 2830 _start_time_ms(0.0), 2831 _cm_oop_closure(NULL), 2832 _curr_region(NULL), 2833 _finger(NULL), 2834 _region_limit(NULL), 2835 _words_scanned(0), 2836 _words_scanned_limit(0), 2837 _real_words_scanned_limit(0), 2838 _refs_reached(0), 2839 _refs_reached_limit(0), 2840 _real_refs_reached_limit(0), 2841 _has_aborted(false), 2842 _has_timed_out(false), 2843 _draining_satb_buffers(false), 2844 _step_times_ms(), 2845 _elapsed_time_ms(0.0), 2846 _termination_time_ms(0.0), 2847 _termination_start_time_ms(0.0), 2848 _marking_step_diff_ms() 2849 { 2850 guarantee(task_queue != NULL, "invariant"); 2851 2852 _marking_step_diff_ms.add(0.5); 2853 } 2854 2855 // These are formatting macros that are used below to ensure 2856 // consistent formatting. The *_H_* versions are used to format the 2857 // header for a particular value and they should be kept consistent 2858 // with the corresponding macro. Also note that most of the macros add 2859 // the necessary white space (as a prefix) which makes them a bit 2860 // easier to compose. 2861 2862 // All the output lines are prefixed with this string to be able to 2863 // identify them easily in a large log file. 2864 #define G1PPRL_LINE_PREFIX "###" 2865 2866 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT 2867 #ifdef _LP64 2868 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 2869 #else // _LP64 2870 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 2871 #endif // _LP64 2872 2873 // For per-region info 2874 #define G1PPRL_TYPE_FORMAT " %-4s" 2875 #define G1PPRL_TYPE_H_FORMAT " %4s" 2876 #define G1PPRL_STATE_FORMAT " %-5s" 2877 #define G1PPRL_STATE_H_FORMAT " %5s" 2878 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9) 2879 #define G1PPRL_BYTE_H_FORMAT " %9s" 2880 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 2881 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 2882 2883 // For summary info 2884 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT 2885 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT 2886 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB" 2887 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%" 2888 2889 G1PrintRegionLivenessInfoClosure::G1PrintRegionLivenessInfoClosure(const char* phase_name) : 2890 _total_used_bytes(0), _total_capacity_bytes(0), 2891 _total_prev_live_bytes(0), _total_next_live_bytes(0), 2892 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) 2893 { 2894 if (!log_is_enabled(Trace, gc, liveness)) { 2895 return; 2896 } 2897 2898 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2899 MemRegion g1_reserved = g1h->g1_reserved(); 2900 double now = os::elapsedTime(); 2901 2902 // Print the header of the output. 2903 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 2904 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP" 2905 G1PPRL_SUM_ADDR_FORMAT("reserved") 2906 G1PPRL_SUM_BYTE_FORMAT("region-size"), 2907 p2i(g1_reserved.start()), p2i(g1_reserved.end()), 2908 HeapRegion::GrainBytes); 2909 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 2910 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 2911 G1PPRL_TYPE_H_FORMAT 2912 G1PPRL_ADDR_BASE_H_FORMAT 2913 G1PPRL_BYTE_H_FORMAT 2914 G1PPRL_BYTE_H_FORMAT 2915 G1PPRL_BYTE_H_FORMAT 2916 G1PPRL_DOUBLE_H_FORMAT 2917 G1PPRL_BYTE_H_FORMAT 2918 G1PPRL_STATE_H_FORMAT 2919 G1PPRL_BYTE_H_FORMAT, 2920 "type", "address-range", 2921 "used", "prev-live", "next-live", "gc-eff", 2922 "remset", "state", "code-roots"); 2923 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 2924 G1PPRL_TYPE_H_FORMAT 2925 G1PPRL_ADDR_BASE_H_FORMAT 2926 G1PPRL_BYTE_H_FORMAT 2927 G1PPRL_BYTE_H_FORMAT 2928 G1PPRL_BYTE_H_FORMAT 2929 G1PPRL_DOUBLE_H_FORMAT 2930 G1PPRL_BYTE_H_FORMAT 2931 G1PPRL_STATE_H_FORMAT 2932 G1PPRL_BYTE_H_FORMAT, 2933 "", "", 2934 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", 2935 "(bytes)", "", "(bytes)"); 2936 } 2937 2938 bool G1PrintRegionLivenessInfoClosure::do_heap_region(HeapRegion* r) { 2939 if (!log_is_enabled(Trace, gc, liveness)) { 2940 return false; 2941 } 2942 2943 const char* type = r->get_type_str(); 2944 HeapWord* bottom = r->bottom(); 2945 HeapWord* end = r->end(); 2946 size_t capacity_bytes = r->capacity(); 2947 size_t used_bytes = r->used(); 2948 size_t prev_live_bytes = r->live_bytes(); 2949 size_t next_live_bytes = r->next_live_bytes(); 2950 double gc_eff = r->gc_efficiency(); 2951 size_t remset_bytes = r->rem_set()->mem_size(); 2952 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size(); 2953 const char* remset_type = r->rem_set()->get_short_state_str(); 2954 2955 _total_used_bytes += used_bytes; 2956 _total_capacity_bytes += capacity_bytes; 2957 _total_prev_live_bytes += prev_live_bytes; 2958 _total_next_live_bytes += next_live_bytes; 2959 _total_remset_bytes += remset_bytes; 2960 _total_strong_code_roots_bytes += strong_code_roots_bytes; 2961 2962 // Print a line for this particular region. 2963 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 2964 G1PPRL_TYPE_FORMAT 2965 G1PPRL_ADDR_BASE_FORMAT 2966 G1PPRL_BYTE_FORMAT 2967 G1PPRL_BYTE_FORMAT 2968 G1PPRL_BYTE_FORMAT 2969 G1PPRL_DOUBLE_FORMAT 2970 G1PPRL_BYTE_FORMAT 2971 G1PPRL_STATE_FORMAT 2972 G1PPRL_BYTE_FORMAT, 2973 type, p2i(bottom), p2i(end), 2974 used_bytes, prev_live_bytes, next_live_bytes, gc_eff, 2975 remset_bytes, remset_type, strong_code_roots_bytes); 2976 2977 return false; 2978 } 2979 2980 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 2981 if (!log_is_enabled(Trace, gc, liveness)) { 2982 return; 2983 } 2984 2985 // add static memory usages to remembered set sizes 2986 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size(); 2987 // Print the footer of the output. 2988 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 2989 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 2990 " SUMMARY" 2991 G1PPRL_SUM_MB_FORMAT("capacity") 2992 G1PPRL_SUM_MB_PERC_FORMAT("used") 2993 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 2994 G1PPRL_SUM_MB_PERC_FORMAT("next-live") 2995 G1PPRL_SUM_MB_FORMAT("remset") 2996 G1PPRL_SUM_MB_FORMAT("code-roots"), 2997 bytes_to_mb(_total_capacity_bytes), 2998 bytes_to_mb(_total_used_bytes), 2999 percent_of(_total_used_bytes, _total_capacity_bytes), 3000 bytes_to_mb(_total_prev_live_bytes), 3001 percent_of(_total_prev_live_bytes, _total_capacity_bytes), 3002 bytes_to_mb(_total_next_live_bytes), 3003 percent_of(_total_next_live_bytes, _total_capacity_bytes), 3004 bytes_to_mb(_total_remset_bytes), 3005 bytes_to_mb(_total_strong_code_roots_bytes)); 3006 }