1 /* 2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/metadataOnStackMark.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc/g1/concurrentMarkThread.inline.hpp" 30 #include "gc/g1/g1CollectedHeap.inline.hpp" 31 #include "gc/g1/g1CollectorState.hpp" 32 #include "gc/g1/g1ConcurrentMark.inline.hpp" 33 #include "gc/g1/g1HeapVerifier.hpp" 34 #include "gc/g1/g1OopClosures.inline.hpp" 35 #include "gc/g1/g1CardLiveData.inline.hpp" 36 #include "gc/g1/g1Policy.hpp" 37 #include "gc/g1/g1StringDedup.hpp" 38 #include "gc/g1/heapRegion.inline.hpp" 39 #include "gc/g1/heapRegionRemSet.hpp" 40 #include "gc/g1/heapRegionSet.inline.hpp" 41 #include "gc/g1/suspendibleThreadSet.hpp" 42 #include "gc/shared/gcId.hpp" 43 #include "gc/shared/gcTimer.hpp" 44 #include "gc/shared/gcTrace.hpp" 45 #include "gc/shared/gcTraceTime.inline.hpp" 46 #include "gc/shared/genOopClosures.inline.hpp" 47 #include "gc/shared/referencePolicy.hpp" 48 #include "gc/shared/strongRootsScope.hpp" 49 #include "gc/shared/taskqueue.inline.hpp" 50 #include "gc/shared/vmGCOperations.hpp" 51 #include "logging/log.hpp" 52 #include "memory/allocation.hpp" 53 #include "memory/resourceArea.hpp" 54 #include "oops/oop.inline.hpp" 55 #include "runtime/atomic.inline.hpp" 56 #include "runtime/handles.inline.hpp" 57 #include "runtime/java.hpp" 58 #include "runtime/prefetch.inline.hpp" 59 #include "services/memTracker.hpp" 60 #include "utilities/growableArray.hpp" 61 62 // Concurrent marking bit map wrapper 63 64 G1CMBitMapRO::G1CMBitMapRO(int shifter) : 65 _bm(), 66 _shifter(shifter) { 67 _bmStartWord = 0; 68 _bmWordSize = 0; 69 } 70 71 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr, 72 const HeapWord* limit) const { 73 // First we must round addr *up* to a possible object boundary. 74 addr = (HeapWord*)align_size_up((intptr_t)addr, 75 HeapWordSize << _shifter); 76 size_t addrOffset = heapWordToOffset(addr); 77 assert(limit != NULL, "limit must not be NULL"); 78 size_t limitOffset = heapWordToOffset(limit); 79 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); 80 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 81 assert(nextAddr >= addr, "get_next_one postcondition"); 82 assert(nextAddr == limit || isMarked(nextAddr), 83 "get_next_one postcondition"); 84 return nextAddr; 85 } 86 87 #ifndef PRODUCT 88 bool G1CMBitMapRO::covers(MemRegion heap_rs) const { 89 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 90 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize, 91 "size inconsistency"); 92 return _bmStartWord == (HeapWord*)(heap_rs.start()) && 93 _bmWordSize == heap_rs.word_size(); 94 } 95 #endif 96 97 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const { 98 _bm.print_on_error(st, prefix); 99 } 100 101 size_t G1CMBitMap::compute_size(size_t heap_size) { 102 return ReservedSpace::allocation_align_size_up(heap_size / mark_distance()); 103 } 104 105 size_t G1CMBitMap::mark_distance() { 106 return MinObjAlignmentInBytes * BitsPerByte; 107 } 108 109 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) { 110 _bmStartWord = heap.start(); 111 _bmWordSize = heap.word_size(); 112 113 _bm = BitMapView((BitMap::bm_word_t*) storage->reserved().start(), _bmWordSize >> _shifter); 114 115 storage->set_mapping_changed_listener(&_listener); 116 } 117 118 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) { 119 if (zero_filled) { 120 return; 121 } 122 // We need to clear the bitmap on commit, removing any existing information. 123 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords); 124 _bm->clear_range(mr); 125 } 126 127 void G1CMBitMap::clear_range(MemRegion mr) { 128 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 129 assert(!mr.is_empty(), "unexpected empty region"); 130 // convert address range into offset range 131 _bm.at_put_range(heapWordToOffset(mr.start()), 132 heapWordToOffset(mr.end()), false); 133 } 134 135 G1CMMarkStack::G1CMMarkStack() : 136 _reserved_space(), 137 _base(NULL), 138 _capacity(0), 139 _saved_index((size_t)AllBits), 140 _should_expand(false) { 141 set_empty(); 142 } 143 144 bool G1CMMarkStack::resize(size_t new_capacity) { 145 assert(is_empty(), "Only resize when stack is empty."); 146 assert(new_capacity <= MarkStackSizeMax, 147 "Trying to resize stack to " SIZE_FORMAT " elements when the maximum is " SIZE_FORMAT, new_capacity, MarkStackSizeMax); 148 149 size_t reservation_size = ReservedSpace::allocation_align_size_up(new_capacity * sizeof(oop)); 150 151 ReservedSpace rs(reservation_size); 152 if (!rs.is_reserved()) { 153 log_warning(gc)("Failed to reserve memory for new overflow mark stack with " SIZE_FORMAT " elements and size " SIZE_FORMAT "B.", new_capacity, reservation_size); 154 return false; 155 } 156 157 VirtualSpace vs; 158 159 if (!vs.initialize(rs, rs.size())) { 160 rs.release(); 161 log_warning(gc)("Failed to commit memory for new overflow mark stack of size " SIZE_FORMAT "B.", rs.size()); 162 return false; 163 } 164 165 assert(vs.committed_size() == rs.size(), "Failed to commit all of the mark stack."); 166 167 // Release old mapping. 168 _reserved_space.release(); 169 170 // Save new mapping for future unmapping. 171 _reserved_space = rs; 172 173 MemTracker::record_virtual_memory_type((address)_reserved_space.base(), mtGC); 174 175 _base = (oop*) vs.low(); 176 _capacity = new_capacity; 177 set_empty(); 178 _should_expand = false; 179 180 return true; 181 } 182 183 bool G1CMMarkStack::allocate(size_t capacity) { 184 return resize(capacity); 185 } 186 187 void G1CMMarkStack::expand() { 188 // Clear expansion flag 189 _should_expand = false; 190 191 if (_capacity == MarkStackSizeMax) { 192 log_debug(gc)("Can not expand overflow mark stack further, already at maximum capacity of " SIZE_FORMAT " elements.", _capacity); 193 return; 194 } 195 size_t old_capacity = _capacity; 196 // Double capacity if possible 197 size_t new_capacity = MIN2(old_capacity * 2, MarkStackSizeMax); 198 199 if (resize(new_capacity)) { 200 log_debug(gc)("Expanded marking stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " elements", 201 old_capacity, new_capacity); 202 } else { 203 log_warning(gc)("Failed to expand marking stack capacity from " SIZE_FORMAT " to " SIZE_FORMAT " elements", 204 old_capacity, new_capacity); 205 } 206 } 207 208 G1CMMarkStack::~G1CMMarkStack() { 209 if (_base != NULL) { 210 _base = NULL; 211 _reserved_space.release(); 212 } 213 } 214 215 void G1CMMarkStack::par_push_arr(oop* buffer, size_t n) { 216 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 217 size_t start = _index; 218 size_t next_index = start + n; 219 if (next_index > _capacity) { 220 _overflow = true; 221 return; 222 } 223 // Otherwise. 224 _index = next_index; 225 for (size_t i = 0; i < n; i++) { 226 size_t ind = start + i; 227 assert(ind < _capacity, "By overflow test above."); 228 _base[ind] = buffer[i]; 229 } 230 } 231 232 bool G1CMMarkStack::par_pop_arr(oop* buffer, size_t max, size_t* n) { 233 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 234 size_t index = _index; 235 if (index == 0) { 236 *n = 0; 237 return false; 238 } else { 239 size_t k = MIN2(max, index); 240 size_t new_ind = index - k; 241 for (size_t j = 0; j < k; j++) { 242 buffer[j] = _base[new_ind + j]; 243 } 244 _index = new_ind; 245 *n = k; 246 return true; 247 } 248 } 249 250 void G1CMMarkStack::note_start_of_gc() { 251 assert(_saved_index == (size_t)AllBits, "note_start_of_gc()/end_of_gc() calls bracketed incorrectly"); 252 _saved_index = _index; 253 } 254 255 void G1CMMarkStack::note_end_of_gc() { 256 guarantee(!stack_modified(), "Saved index " SIZE_FORMAT " must be the same as " SIZE_FORMAT, _saved_index, _index); 257 258 _saved_index = (size_t)AllBits; 259 } 260 261 G1CMRootRegions::G1CMRootRegions() : 262 _cm(NULL), _scan_in_progress(false), 263 _should_abort(false), _claimed_survivor_index(0) { } 264 265 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) { 266 _survivors = survivors; 267 _cm = cm; 268 } 269 270 void G1CMRootRegions::prepare_for_scan() { 271 assert(!scan_in_progress(), "pre-condition"); 272 273 // Currently, only survivors can be root regions. 274 _claimed_survivor_index = 0; 275 _scan_in_progress = _survivors->regions()->is_nonempty(); 276 _should_abort = false; 277 } 278 279 HeapRegion* G1CMRootRegions::claim_next() { 280 if (_should_abort) { 281 // If someone has set the should_abort flag, we return NULL to 282 // force the caller to bail out of their loop. 283 return NULL; 284 } 285 286 // Currently, only survivors can be root regions. 287 const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions(); 288 289 int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1; 290 if (claimed_index < survivor_regions->length()) { 291 return survivor_regions->at(claimed_index); 292 } 293 return NULL; 294 } 295 296 uint G1CMRootRegions::num_root_regions() const { 297 return (uint)_survivors->regions()->length(); 298 } 299 300 void G1CMRootRegions::notify_scan_done() { 301 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 302 _scan_in_progress = false; 303 RootRegionScan_lock->notify_all(); 304 } 305 306 void G1CMRootRegions::cancel_scan() { 307 notify_scan_done(); 308 } 309 310 void G1CMRootRegions::scan_finished() { 311 assert(scan_in_progress(), "pre-condition"); 312 313 // Currently, only survivors can be root regions. 314 if (!_should_abort) { 315 assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index); 316 assert((uint)_claimed_survivor_index >= _survivors->length(), 317 "we should have claimed all survivors, claimed index = %u, length = %u", 318 (uint)_claimed_survivor_index, _survivors->length()); 319 } 320 321 notify_scan_done(); 322 } 323 324 bool G1CMRootRegions::wait_until_scan_finished() { 325 if (!scan_in_progress()) return false; 326 327 { 328 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 329 while (scan_in_progress()) { 330 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 331 } 332 } 333 return true; 334 } 335 336 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 337 return MAX2((n_par_threads + 2) / 4, 1U); 338 } 339 340 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) : 341 _g1h(g1h), 342 _markBitMap1(), 343 _markBitMap2(), 344 _parallel_marking_threads(0), 345 _max_parallel_marking_threads(0), 346 _sleep_factor(0.0), 347 _marking_task_overhead(1.0), 348 _cleanup_list("Cleanup List"), 349 350 _prevMarkBitMap(&_markBitMap1), 351 _nextMarkBitMap(&_markBitMap2), 352 353 _global_mark_stack(), 354 // _finger set in set_non_marking_state 355 356 _max_worker_id(ParallelGCThreads), 357 // _active_tasks set in set_non_marking_state 358 // _tasks set inside the constructor 359 _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)), 360 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)), 361 362 _has_overflown(false), 363 _concurrent(false), 364 _has_aborted(false), 365 _restart_for_overflow(false), 366 _concurrent_marking_in_progress(false), 367 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 368 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()), 369 370 // _verbose_level set below 371 372 _init_times(), 373 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 374 _cleanup_times(), 375 _total_counting_time(0.0), 376 _total_rs_scrub_time(0.0), 377 378 _parallel_workers(NULL), 379 380 _completed_initialization(false) { 381 382 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage); 383 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage); 384 385 // Create & start a ConcurrentMark thread. 386 _cmThread = new ConcurrentMarkThread(this); 387 assert(cmThread() != NULL, "CM Thread should have been created"); 388 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 389 if (_cmThread->osthread() == NULL) { 390 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread"); 391 } 392 393 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 394 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency"); 395 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency"); 396 397 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 398 satb_qs.set_buffer_size(G1SATBBufferSize); 399 400 _root_regions.init(_g1h->survivor(), this); 401 402 if (ConcGCThreads > ParallelGCThreads) { 403 log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).", 404 ConcGCThreads, ParallelGCThreads); 405 return; 406 } 407 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 408 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 409 // if both are set 410 _sleep_factor = 0.0; 411 _marking_task_overhead = 1.0; 412 } else if (G1MarkingOverheadPercent > 0) { 413 // We will calculate the number of parallel marking threads based 414 // on a target overhead with respect to the soft real-time goal 415 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 416 double overall_cm_overhead = 417 (double) MaxGCPauseMillis * marking_overhead / 418 (double) GCPauseIntervalMillis; 419 double cpu_ratio = 1.0 / os::initial_active_processor_count(); 420 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 421 double marking_task_overhead = 422 overall_cm_overhead / marking_thread_num * os::initial_active_processor_count(); 423 double sleep_factor = 424 (1.0 - marking_task_overhead) / marking_task_overhead; 425 426 FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num); 427 _sleep_factor = sleep_factor; 428 _marking_task_overhead = marking_task_overhead; 429 } else { 430 // Calculate the number of parallel marking threads by scaling 431 // the number of parallel GC threads. 432 uint marking_thread_num = scale_parallel_threads(ParallelGCThreads); 433 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num); 434 _sleep_factor = 0.0; 435 _marking_task_overhead = 1.0; 436 } 437 438 assert(ConcGCThreads > 0, "Should have been set"); 439 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 440 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 441 _parallel_marking_threads = ConcGCThreads; 442 _max_parallel_marking_threads = _parallel_marking_threads; 443 444 _parallel_workers = new WorkGang("G1 Marker", 445 _max_parallel_marking_threads, false, true); 446 if (_parallel_workers == NULL) { 447 vm_exit_during_initialization("Failed necessary allocation."); 448 } else { 449 _parallel_workers->initialize_workers(); 450 } 451 452 if (FLAG_IS_DEFAULT(MarkStackSize)) { 453 size_t mark_stack_size = 454 MIN2(MarkStackSizeMax, 455 MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE))); 456 // Verify that the calculated value for MarkStackSize is in range. 457 // It would be nice to use the private utility routine from Arguments. 458 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 459 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): " 460 "must be between 1 and " SIZE_FORMAT, 461 mark_stack_size, MarkStackSizeMax); 462 return; 463 } 464 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size); 465 } else { 466 // Verify MarkStackSize is in range. 467 if (FLAG_IS_CMDLINE(MarkStackSize)) { 468 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 469 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 470 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): " 471 "must be between 1 and " SIZE_FORMAT, 472 MarkStackSize, MarkStackSizeMax); 473 return; 474 } 475 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 476 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 477 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")" 478 " or for MarkStackSizeMax (" SIZE_FORMAT ")", 479 MarkStackSize, MarkStackSizeMax); 480 return; 481 } 482 } 483 } 484 } 485 486 if (!_global_mark_stack.allocate(MarkStackSize)) { 487 vm_exit_during_initialization("Failed to allocate initial concurrent mark overflow mark stack."); 488 return; 489 } 490 491 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC); 492 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC); 493 494 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 495 _active_tasks = _max_worker_id; 496 497 for (uint i = 0; i < _max_worker_id; ++i) { 498 G1CMTaskQueue* task_queue = new G1CMTaskQueue(); 499 task_queue->initialize(); 500 _task_queues->register_queue(i, task_queue); 501 502 _tasks[i] = new G1CMTask(i, this, task_queue, _task_queues); 503 504 _accum_task_vtime[i] = 0.0; 505 } 506 507 // so that the call below can read a sensible value 508 _heap_start = g1h->reserved_region().start(); 509 set_non_marking_state(); 510 _completed_initialization = true; 511 } 512 513 void G1ConcurrentMark::reset() { 514 // Starting values for these two. This should be called in a STW 515 // phase. 516 MemRegion reserved = _g1h->g1_reserved(); 517 _heap_start = reserved.start(); 518 _heap_end = reserved.end(); 519 520 // Separated the asserts so that we know which one fires. 521 assert(_heap_start != NULL, "heap bounds should look ok"); 522 assert(_heap_end != NULL, "heap bounds should look ok"); 523 assert(_heap_start < _heap_end, "heap bounds should look ok"); 524 525 // Reset all the marking data structures and any necessary flags 526 reset_marking_state(); 527 528 // We do reset all of them, since different phases will use 529 // different number of active threads. So, it's easiest to have all 530 // of them ready. 531 for (uint i = 0; i < _max_worker_id; ++i) { 532 _tasks[i]->reset(_nextMarkBitMap); 533 } 534 535 // we need this to make sure that the flag is on during the evac 536 // pause with initial mark piggy-backed 537 set_concurrent_marking_in_progress(); 538 } 539 540 541 void G1ConcurrentMark::reset_marking_state(bool clear_overflow) { 542 _global_mark_stack.set_should_expand(has_overflown()); 543 _global_mark_stack.set_empty(); // Also clears the overflow stack's overflow flag 544 if (clear_overflow) { 545 clear_has_overflown(); 546 } else { 547 assert(has_overflown(), "pre-condition"); 548 } 549 _finger = _heap_start; 550 551 for (uint i = 0; i < _max_worker_id; ++i) { 552 G1CMTaskQueue* queue = _task_queues->queue(i); 553 queue->set_empty(); 554 } 555 } 556 557 void G1ConcurrentMark::set_concurrency(uint active_tasks) { 558 assert(active_tasks <= _max_worker_id, "we should not have more"); 559 560 _active_tasks = active_tasks; 561 // Need to update the three data structures below according to the 562 // number of active threads for this phase. 563 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 564 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 565 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 566 } 567 568 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 569 set_concurrency(active_tasks); 570 571 _concurrent = concurrent; 572 // We propagate this to all tasks, not just the active ones. 573 for (uint i = 0; i < _max_worker_id; ++i) 574 _tasks[i]->set_concurrent(concurrent); 575 576 if (concurrent) { 577 set_concurrent_marking_in_progress(); 578 } else { 579 // We currently assume that the concurrent flag has been set to 580 // false before we start remark. At this point we should also be 581 // in a STW phase. 582 assert(!concurrent_marking_in_progress(), "invariant"); 583 assert(out_of_regions(), 584 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT, 585 p2i(_finger), p2i(_heap_end)); 586 } 587 } 588 589 void G1ConcurrentMark::set_non_marking_state() { 590 // We set the global marking state to some default values when we're 591 // not doing marking. 592 reset_marking_state(); 593 _active_tasks = 0; 594 clear_concurrent_marking_in_progress(); 595 } 596 597 G1ConcurrentMark::~G1ConcurrentMark() { 598 // The G1ConcurrentMark instance is never freed. 599 ShouldNotReachHere(); 600 } 601 602 class G1ClearBitMapTask : public AbstractGangTask { 603 public: 604 static size_t chunk_size() { return M; } 605 606 private: 607 // Heap region closure used for clearing the given mark bitmap. 608 class G1ClearBitmapHRClosure : public HeapRegionClosure { 609 private: 610 G1CMBitMap* _bitmap; 611 G1ConcurrentMark* _cm; 612 public: 613 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) { 614 } 615 616 virtual bool doHeapRegion(HeapRegion* r) { 617 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize; 618 619 HeapWord* cur = r->bottom(); 620 HeapWord* const end = r->end(); 621 622 while (cur < end) { 623 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end)); 624 _bitmap->clear_range(mr); 625 626 cur += chunk_size_in_words; 627 628 // Abort iteration if after yielding the marking has been aborted. 629 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) { 630 return true; 631 } 632 // Repeat the asserts from before the start of the closure. We will do them 633 // as asserts here to minimize their overhead on the product. However, we 634 // will have them as guarantees at the beginning / end of the bitmap 635 // clearing to get some checking in the product. 636 assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant"); 637 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant"); 638 } 639 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index()); 640 641 return false; 642 } 643 }; 644 645 G1ClearBitmapHRClosure _cl; 646 HeapRegionClaimer _hr_claimer; 647 bool _suspendible; // If the task is suspendible, workers must join the STS. 648 649 public: 650 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) : 651 AbstractGangTask("G1 Clear Bitmap"), 652 _cl(bitmap, suspendible ? cm : NULL), 653 _hr_claimer(n_workers), 654 _suspendible(suspendible) 655 { } 656 657 void work(uint worker_id) { 658 SuspendibleThreadSetJoiner sts_join(_suspendible); 659 G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer, true); 660 } 661 662 bool is_complete() { 663 return _cl.complete(); 664 } 665 }; 666 667 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) { 668 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint."); 669 670 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor(); 671 size_t const num_chunks = align_size_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size(); 672 673 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers()); 674 675 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield); 676 677 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks); 678 workers->run_task(&cl, num_workers); 679 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding."); 680 } 681 682 void G1ConcurrentMark::cleanup_for_next_mark() { 683 // Make sure that the concurrent mark thread looks to still be in 684 // the current cycle. 685 guarantee(cmThread()->during_cycle(), "invariant"); 686 687 // We are finishing up the current cycle by clearing the next 688 // marking bitmap and getting it ready for the next cycle. During 689 // this time no other cycle can start. So, let's make sure that this 690 // is the case. 691 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant"); 692 693 clear_bitmap(_nextMarkBitMap, _parallel_workers, true); 694 695 // Clear the live count data. If the marking has been aborted, the abort() 696 // call already did that. 697 if (!has_aborted()) { 698 clear_live_data(_parallel_workers); 699 DEBUG_ONLY(verify_live_data_clear()); 700 } 701 702 // Repeat the asserts from above. 703 guarantee(cmThread()->during_cycle(), "invariant"); 704 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant"); 705 } 706 707 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) { 708 assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint."); 709 clear_bitmap((G1CMBitMap*)_prevMarkBitMap, workers, false); 710 } 711 712 class CheckBitmapClearHRClosure : public HeapRegionClosure { 713 G1CMBitMap* _bitmap; 714 bool _error; 715 public: 716 CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) { 717 } 718 719 virtual bool doHeapRegion(HeapRegion* r) { 720 // This closure can be called concurrently to the mutator, so we must make sure 721 // that the result of the getNextMarkedWordAddress() call is compared to the 722 // value passed to it as limit to detect any found bits. 723 // end never changes in G1. 724 HeapWord* end = r->end(); 725 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end; 726 } 727 }; 728 729 bool G1ConcurrentMark::nextMarkBitmapIsClear() { 730 CheckBitmapClearHRClosure cl(_nextMarkBitMap); 731 _g1h->heap_region_iterate(&cl); 732 return cl.complete(); 733 } 734 735 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 736 public: 737 bool doHeapRegion(HeapRegion* r) { 738 r->note_start_of_marking(); 739 return false; 740 } 741 }; 742 743 void G1ConcurrentMark::checkpointRootsInitialPre() { 744 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 745 G1Policy* g1p = g1h->g1_policy(); 746 747 _has_aborted = false; 748 749 // Initialize marking structures. This has to be done in a STW phase. 750 reset(); 751 752 // For each region note start of marking. 753 NoteStartOfMarkHRClosure startcl; 754 g1h->heap_region_iterate(&startcl); 755 } 756 757 758 void G1ConcurrentMark::checkpointRootsInitialPost() { 759 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 760 761 // Start Concurrent Marking weak-reference discovery. 762 ReferenceProcessor* rp = g1h->ref_processor_cm(); 763 // enable ("weak") refs discovery 764 rp->enable_discovery(); 765 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 766 767 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 768 // This is the start of the marking cycle, we're expected all 769 // threads to have SATB queues with active set to false. 770 satb_mq_set.set_active_all_threads(true, /* new active value */ 771 false /* expected_active */); 772 773 _root_regions.prepare_for_scan(); 774 775 // update_g1_committed() will be called at the end of an evac pause 776 // when marking is on. So, it's also called at the end of the 777 // initial-mark pause to update the heap end, if the heap expands 778 // during it. No need to call it here. 779 } 780 781 /* 782 * Notice that in the next two methods, we actually leave the STS 783 * during the barrier sync and join it immediately afterwards. If we 784 * do not do this, the following deadlock can occur: one thread could 785 * be in the barrier sync code, waiting for the other thread to also 786 * sync up, whereas another one could be trying to yield, while also 787 * waiting for the other threads to sync up too. 788 * 789 * Note, however, that this code is also used during remark and in 790 * this case we should not attempt to leave / enter the STS, otherwise 791 * we'll either hit an assert (debug / fastdebug) or deadlock 792 * (product). So we should only leave / enter the STS if we are 793 * operating concurrently. 794 * 795 * Because the thread that does the sync barrier has left the STS, it 796 * is possible to be suspended for a Full GC or an evacuation pause 797 * could occur. This is actually safe, since the entering the sync 798 * barrier is one of the last things do_marking_step() does, and it 799 * doesn't manipulate any data structures afterwards. 800 */ 801 802 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 803 bool barrier_aborted; 804 { 805 SuspendibleThreadSetLeaver sts_leave(concurrent()); 806 barrier_aborted = !_first_overflow_barrier_sync.enter(); 807 } 808 809 // at this point everyone should have synced up and not be doing any 810 // more work 811 812 if (barrier_aborted) { 813 // If the barrier aborted we ignore the overflow condition and 814 // just abort the whole marking phase as quickly as possible. 815 return; 816 } 817 818 // If we're executing the concurrent phase of marking, reset the marking 819 // state; otherwise the marking state is reset after reference processing, 820 // during the remark pause. 821 // If we reset here as a result of an overflow during the remark we will 822 // see assertion failures from any subsequent set_concurrency_and_phase() 823 // calls. 824 if (concurrent()) { 825 // let the task associated with with worker 0 do this 826 if (worker_id == 0) { 827 // task 0 is responsible for clearing the global data structures 828 // We should be here because of an overflow. During STW we should 829 // not clear the overflow flag since we rely on it being true when 830 // we exit this method to abort the pause and restart concurrent 831 // marking. 832 reset_marking_state(true /* clear_overflow */); 833 834 log_info(gc, marking)("Concurrent Mark reset for overflow"); 835 } 836 } 837 838 // after this, each task should reset its own data structures then 839 // then go into the second barrier 840 } 841 842 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 843 SuspendibleThreadSetLeaver sts_leave(concurrent()); 844 _second_overflow_barrier_sync.enter(); 845 846 // at this point everything should be re-initialized and ready to go 847 } 848 849 class G1CMConcurrentMarkingTask: public AbstractGangTask { 850 private: 851 G1ConcurrentMark* _cm; 852 ConcurrentMarkThread* _cmt; 853 854 public: 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 ResourceMark rm; 859 860 double start_vtime = os::elapsedVTime(); 861 862 { 863 SuspendibleThreadSetJoiner sts_join; 864 865 assert(worker_id < _cm->active_tasks(), "invariant"); 866 G1CMTask* the_task = _cm->task(worker_id); 867 the_task->record_start_time(); 868 if (!_cm->has_aborted()) { 869 do { 870 double start_vtime_sec = os::elapsedVTime(); 871 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 872 873 the_task->do_marking_step(mark_step_duration_ms, 874 true /* do_termination */, 875 false /* is_serial*/); 876 877 double end_vtime_sec = os::elapsedVTime(); 878 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 879 _cm->clear_has_overflown(); 880 881 _cm->do_yield_check(); 882 883 jlong sleep_time_ms; 884 if (!_cm->has_aborted() && the_task->has_aborted()) { 885 sleep_time_ms = 886 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 887 { 888 SuspendibleThreadSetLeaver sts_leave; 889 os::sleep(Thread::current(), sleep_time_ms, false); 890 } 891 } 892 } while (!_cm->has_aborted() && the_task->has_aborted()); 893 } 894 the_task->record_end_time(); 895 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 896 } 897 898 double end_vtime = os::elapsedVTime(); 899 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 900 } 901 902 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm, 903 ConcurrentMarkThread* cmt) : 904 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 905 906 ~G1CMConcurrentMarkingTask() { } 907 }; 908 909 // Calculates the number of active workers for a concurrent 910 // phase. 911 uint G1ConcurrentMark::calc_parallel_marking_threads() { 912 uint n_conc_workers = 0; 913 if (!UseDynamicNumberOfGCThreads || 914 (!FLAG_IS_DEFAULT(ConcGCThreads) && 915 !ForceDynamicNumberOfGCThreads)) { 916 n_conc_workers = max_parallel_marking_threads(); 917 } else { 918 n_conc_workers = 919 AdaptiveSizePolicy::calc_default_active_workers(max_parallel_marking_threads(), 920 1, /* Minimum workers */ 921 parallel_marking_threads(), 922 Threads::number_of_non_daemon_threads()); 923 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 924 // that scaling has already gone into "_max_parallel_marking_threads". 925 } 926 assert(n_conc_workers > 0 && n_conc_workers <= max_parallel_marking_threads(), 927 "Calculated number of workers must be larger than zero and at most the maximum %u, but is %u", 928 max_parallel_marking_threads(), n_conc_workers); 929 return n_conc_workers; 930 } 931 932 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr) { 933 // Currently, only survivors can be root regions. 934 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 935 G1RootRegionScanClosure cl(_g1h, this); 936 937 const uintx interval = PrefetchScanIntervalInBytes; 938 HeapWord* curr = hr->bottom(); 939 const HeapWord* end = hr->top(); 940 while (curr < end) { 941 Prefetch::read(curr, interval); 942 oop obj = oop(curr); 943 int size = obj->oop_iterate_size(&cl); 944 assert(size == obj->size(), "sanity"); 945 curr += size; 946 } 947 } 948 949 class G1CMRootRegionScanTask : public AbstractGangTask { 950 private: 951 G1ConcurrentMark* _cm; 952 953 public: 954 G1CMRootRegionScanTask(G1ConcurrentMark* cm) : 955 AbstractGangTask("G1 Root Region Scan"), _cm(cm) { } 956 957 void work(uint worker_id) { 958 assert(Thread::current()->is_ConcurrentGC_thread(), 959 "this should only be done by a conc GC thread"); 960 961 G1CMRootRegions* root_regions = _cm->root_regions(); 962 HeapRegion* hr = root_regions->claim_next(); 963 while (hr != NULL) { 964 _cm->scanRootRegion(hr); 965 hr = root_regions->claim_next(); 966 } 967 } 968 }; 969 970 void G1ConcurrentMark::scan_root_regions() { 971 // scan_in_progress() will have been set to true only if there was 972 // at least one root region to scan. So, if it's false, we 973 // should not attempt to do any further work. 974 if (root_regions()->scan_in_progress()) { 975 assert(!has_aborted(), "Aborting before root region scanning is finished not supported."); 976 977 _parallel_marking_threads = MIN2(calc_parallel_marking_threads(), 978 // We distribute work on a per-region basis, so starting 979 // more threads than that is useless. 980 root_regions()->num_root_regions()); 981 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 982 "Maximum number of marking threads exceeded"); 983 984 G1CMRootRegionScanTask task(this); 985 log_debug(gc, ergo)("Running %s using %u workers for %u work units.", 986 task.name(), _parallel_marking_threads, root_regions()->num_root_regions()); 987 _parallel_workers->run_task(&task, _parallel_marking_threads); 988 989 // It's possible that has_aborted() is true here without actually 990 // aborting the survivor scan earlier. This is OK as it's 991 // mainly used for sanity checking. 992 root_regions()->scan_finished(); 993 } 994 } 995 996 void G1ConcurrentMark::concurrent_cycle_start() { 997 _gc_timer_cm->register_gc_start(); 998 999 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start()); 1000 1001 _g1h->trace_heap_before_gc(_gc_tracer_cm); 1002 } 1003 1004 void G1ConcurrentMark::concurrent_cycle_end() { 1005 _g1h->trace_heap_after_gc(_gc_tracer_cm); 1006 1007 if (has_aborted()) { 1008 _gc_tracer_cm->report_concurrent_mode_failure(); 1009 } 1010 1011 _gc_timer_cm->register_gc_end(); 1012 1013 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1014 } 1015 1016 void G1ConcurrentMark::mark_from_roots() { 1017 // we might be tempted to assert that: 1018 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1019 // "inconsistent argument?"); 1020 // However that wouldn't be right, because it's possible that 1021 // a safepoint is indeed in progress as a younger generation 1022 // stop-the-world GC happens even as we mark in this generation. 1023 1024 _restart_for_overflow = false; 1025 1026 // _g1h has _n_par_threads 1027 _parallel_marking_threads = calc_parallel_marking_threads(); 1028 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1029 "Maximum number of marking threads exceeded"); 1030 1031 uint active_workers = MAX2(1U, parallel_marking_threads()); 1032 assert(active_workers > 0, "Should have been set"); 1033 1034 // Parallel task terminator is set in "set_concurrency_and_phase()" 1035 set_concurrency_and_phase(active_workers, true /* concurrent */); 1036 1037 G1CMConcurrentMarkingTask markingTask(this, cmThread()); 1038 _parallel_workers->set_active_workers(active_workers); 1039 _parallel_workers->run_task(&markingTask); 1040 print_stats(); 1041 } 1042 1043 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1044 // world is stopped at this checkpoint 1045 assert(SafepointSynchronize::is_at_safepoint(), 1046 "world should be stopped"); 1047 1048 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1049 1050 // If a full collection has happened, we shouldn't do this. 1051 if (has_aborted()) { 1052 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1053 return; 1054 } 1055 1056 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1057 1058 if (VerifyDuringGC) { 1059 HandleMark hm; // handle scope 1060 g1h->prepare_for_verify(); 1061 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1062 } 1063 g1h->verifier()->check_bitmaps("Remark Start"); 1064 1065 G1Policy* g1p = g1h->g1_policy(); 1066 g1p->record_concurrent_mark_remark_start(); 1067 1068 double start = os::elapsedTime(); 1069 1070 checkpointRootsFinalWork(); 1071 1072 double mark_work_end = os::elapsedTime(); 1073 1074 weakRefsWork(clear_all_soft_refs); 1075 1076 if (has_overflown()) { 1077 // We overflowed. Restart concurrent marking. 1078 _restart_for_overflow = true; 1079 1080 // Verify the heap w.r.t. the previous marking bitmap. 1081 if (VerifyDuringGC) { 1082 HandleMark hm; // handle scope 1083 g1h->prepare_for_verify(); 1084 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)"); 1085 } 1086 1087 // Clear the marking state because we will be restarting 1088 // marking due to overflowing the global mark stack. 1089 reset_marking_state(); 1090 } else { 1091 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1092 // We're done with marking. 1093 // This is the end of the marking cycle, we're expected all 1094 // threads to have SATB queues with active set to true. 1095 satb_mq_set.set_active_all_threads(false, /* new active value */ 1096 true /* expected_active */); 1097 1098 if (VerifyDuringGC) { 1099 HandleMark hm; // handle scope 1100 g1h->prepare_for_verify(); 1101 Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)"); 1102 } 1103 g1h->verifier()->check_bitmaps("Remark End"); 1104 assert(!restart_for_overflow(), "sanity"); 1105 // Completely reset the marking state since marking completed 1106 set_non_marking_state(); 1107 } 1108 1109 // Expand the marking stack, if we have to and if we can. 1110 if (_global_mark_stack.should_expand()) { 1111 _global_mark_stack.expand(); 1112 } 1113 1114 // Statistics 1115 double now = os::elapsedTime(); 1116 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1117 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1118 _remark_times.add((now - start) * 1000.0); 1119 1120 g1p->record_concurrent_mark_remark_end(); 1121 1122 G1CMIsAliveClosure is_alive(g1h); 1123 _gc_tracer_cm->report_object_count_after_gc(&is_alive); 1124 } 1125 1126 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1127 G1CollectedHeap* _g1; 1128 size_t _freed_bytes; 1129 FreeRegionList* _local_cleanup_list; 1130 uint _old_regions_removed; 1131 uint _humongous_regions_removed; 1132 HRRSCleanupTask* _hrrs_cleanup_task; 1133 1134 public: 1135 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1136 FreeRegionList* local_cleanup_list, 1137 HRRSCleanupTask* hrrs_cleanup_task) : 1138 _g1(g1), 1139 _freed_bytes(0), 1140 _local_cleanup_list(local_cleanup_list), 1141 _old_regions_removed(0), 1142 _humongous_regions_removed(0), 1143 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1144 1145 size_t freed_bytes() { return _freed_bytes; } 1146 const uint old_regions_removed() { return _old_regions_removed; } 1147 const uint humongous_regions_removed() { return _humongous_regions_removed; } 1148 1149 bool doHeapRegion(HeapRegion *hr) { 1150 if (hr->is_archive()) { 1151 return false; 1152 } 1153 _g1->reset_gc_time_stamps(hr); 1154 hr->note_end_of_marking(); 1155 1156 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 1157 _freed_bytes += hr->used(); 1158 hr->set_containing_set(NULL); 1159 if (hr->is_humongous()) { 1160 _humongous_regions_removed++; 1161 _g1->free_humongous_region(hr, _local_cleanup_list, true /* skip_remset */); 1162 } else { 1163 _old_regions_removed++; 1164 _g1->free_region(hr, _local_cleanup_list, true /* skip_remset */); 1165 } 1166 } else { 1167 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task); 1168 } 1169 1170 return false; 1171 } 1172 }; 1173 1174 class G1ParNoteEndTask: public AbstractGangTask { 1175 friend class G1NoteEndOfConcMarkClosure; 1176 1177 protected: 1178 G1CollectedHeap* _g1h; 1179 FreeRegionList* _cleanup_list; 1180 HeapRegionClaimer _hrclaimer; 1181 1182 public: 1183 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) : 1184 AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) { 1185 } 1186 1187 void work(uint worker_id) { 1188 FreeRegionList local_cleanup_list("Local Cleanup List"); 1189 HRRSCleanupTask hrrs_cleanup_task; 1190 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list, 1191 &hrrs_cleanup_task); 1192 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer); 1193 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1194 1195 // Now update the lists 1196 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed()); 1197 { 1198 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1199 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes()); 1200 1201 // If we iterate over the global cleanup list at the end of 1202 // cleanup to do this printing we will not guarantee to only 1203 // generate output for the newly-reclaimed regions (the list 1204 // might not be empty at the beginning of cleanup; we might 1205 // still be working on its previous contents). So we do the 1206 // printing here, before we append the new regions to the global 1207 // cleanup list. 1208 1209 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1210 if (hr_printer->is_active()) { 1211 FreeRegionListIterator iter(&local_cleanup_list); 1212 while (iter.more_available()) { 1213 HeapRegion* hr = iter.get_next(); 1214 hr_printer->cleanup(hr); 1215 } 1216 } 1217 1218 _cleanup_list->add_ordered(&local_cleanup_list); 1219 assert(local_cleanup_list.is_empty(), "post-condition"); 1220 1221 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1222 } 1223 } 1224 }; 1225 1226 void G1ConcurrentMark::cleanup() { 1227 // world is stopped at this checkpoint 1228 assert(SafepointSynchronize::is_at_safepoint(), 1229 "world should be stopped"); 1230 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1231 1232 // If a full collection has happened, we shouldn't do this. 1233 if (has_aborted()) { 1234 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1235 return; 1236 } 1237 1238 g1h->verifier()->verify_region_sets_optional(); 1239 1240 if (VerifyDuringGC) { 1241 HandleMark hm; // handle scope 1242 g1h->prepare_for_verify(); 1243 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1244 } 1245 g1h->verifier()->check_bitmaps("Cleanup Start"); 1246 1247 G1Policy* g1p = g1h->g1_policy(); 1248 g1p->record_concurrent_mark_cleanup_start(); 1249 1250 double start = os::elapsedTime(); 1251 1252 HeapRegionRemSet::reset_for_cleanup_tasks(); 1253 1254 { 1255 GCTraceTime(Debug, gc)("Finalize Live Data"); 1256 finalize_live_data(); 1257 } 1258 1259 if (VerifyDuringGC) { 1260 GCTraceTime(Debug, gc)("Verify Live Data"); 1261 verify_live_data(); 1262 } 1263 1264 g1h->collector_state()->set_mark_in_progress(false); 1265 1266 double count_end = os::elapsedTime(); 1267 double this_final_counting_time = (count_end - start); 1268 _total_counting_time += this_final_counting_time; 1269 1270 if (log_is_enabled(Trace, gc, liveness)) { 1271 G1PrintRegionLivenessInfoClosure cl("Post-Marking"); 1272 _g1h->heap_region_iterate(&cl); 1273 } 1274 1275 // Install newly created mark bitMap as "prev". 1276 swapMarkBitMaps(); 1277 1278 g1h->reset_gc_time_stamp(); 1279 1280 uint n_workers = _g1h->workers()->active_workers(); 1281 1282 // Note end of marking in all heap regions. 1283 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers); 1284 g1h->workers()->run_task(&g1_par_note_end_task); 1285 g1h->check_gc_time_stamps(); 1286 1287 if (!cleanup_list_is_empty()) { 1288 // The cleanup list is not empty, so we'll have to process it 1289 // concurrently. Notify anyone else that might be wanting free 1290 // regions that there will be more free regions coming soon. 1291 g1h->set_free_regions_coming(); 1292 } 1293 1294 // call below, since it affects the metric by which we sort the heap 1295 // regions. 1296 if (G1ScrubRemSets) { 1297 double rs_scrub_start = os::elapsedTime(); 1298 g1h->scrub_rem_set(); 1299 _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start); 1300 } 1301 1302 // this will also free any regions totally full of garbage objects, 1303 // and sort the regions. 1304 g1h->g1_policy()->record_concurrent_mark_cleanup_end(); 1305 1306 // Statistics. 1307 double end = os::elapsedTime(); 1308 _cleanup_times.add((end - start) * 1000.0); 1309 1310 // Clean up will have freed any regions completely full of garbage. 1311 // Update the soft reference policy with the new heap occupancy. 1312 Universe::update_heap_info_at_gc(); 1313 1314 if (VerifyDuringGC) { 1315 HandleMark hm; // handle scope 1316 g1h->prepare_for_verify(); 1317 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)"); 1318 } 1319 1320 g1h->verifier()->check_bitmaps("Cleanup End"); 1321 1322 g1h->verifier()->verify_region_sets_optional(); 1323 1324 // We need to make this be a "collection" so any collection pause that 1325 // races with it goes around and waits for completeCleanup to finish. 1326 g1h->increment_total_collections(); 1327 1328 // Clean out dead classes and update Metaspace sizes. 1329 if (ClassUnloadingWithConcurrentMark) { 1330 ClassLoaderDataGraph::purge(); 1331 } 1332 MetaspaceGC::compute_new_size(); 1333 1334 // We reclaimed old regions so we should calculate the sizes to make 1335 // sure we update the old gen/space data. 1336 g1h->g1mm()->update_sizes(); 1337 g1h->allocation_context_stats().update_after_mark(); 1338 } 1339 1340 void G1ConcurrentMark::complete_cleanup() { 1341 if (has_aborted()) return; 1342 1343 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1344 1345 _cleanup_list.verify_optional(); 1346 FreeRegionList tmp_free_list("Tmp Free List"); 1347 1348 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1349 "cleanup list has %u entries", 1350 _cleanup_list.length()); 1351 1352 // No one else should be accessing the _cleanup_list at this point, 1353 // so it is not necessary to take any locks 1354 while (!_cleanup_list.is_empty()) { 1355 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */); 1356 assert(hr != NULL, "Got NULL from a non-empty list"); 1357 hr->par_clear(); 1358 tmp_free_list.add_ordered(hr); 1359 1360 // Instead of adding one region at a time to the secondary_free_list, 1361 // we accumulate them in the local list and move them a few at a 1362 // time. This also cuts down on the number of notify_all() calls 1363 // we do during this process. We'll also append the local list when 1364 // _cleanup_list is empty (which means we just removed the last 1365 // region from the _cleanup_list). 1366 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 1367 _cleanup_list.is_empty()) { 1368 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1369 "appending %u entries to the secondary_free_list, " 1370 "cleanup list still has %u entries", 1371 tmp_free_list.length(), 1372 _cleanup_list.length()); 1373 1374 { 1375 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 1376 g1h->secondary_free_list_add(&tmp_free_list); 1377 SecondaryFreeList_lock->notify_all(); 1378 } 1379 #ifndef PRODUCT 1380 if (G1StressConcRegionFreeing) { 1381 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 1382 os::sleep(Thread::current(), (jlong) 1, false); 1383 } 1384 } 1385 #endif 1386 } 1387 } 1388 assert(tmp_free_list.is_empty(), "post-condition"); 1389 } 1390 1391 // Supporting Object and Oop closures for reference discovery 1392 // and processing in during marking 1393 1394 bool G1CMIsAliveClosure::do_object_b(oop obj) { 1395 HeapWord* addr = (HeapWord*)obj; 1396 return addr != NULL && 1397 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 1398 } 1399 1400 // 'Keep Alive' oop closure used by both serial parallel reference processing. 1401 // Uses the G1CMTask associated with a worker thread (for serial reference 1402 // processing the G1CMTask for worker 0 is used) to preserve (mark) and 1403 // trace referent objects. 1404 // 1405 // Using the G1CMTask and embedded local queues avoids having the worker 1406 // threads operating on the global mark stack. This reduces the risk 1407 // of overflowing the stack - which we would rather avoid at this late 1408 // state. Also using the tasks' local queues removes the potential 1409 // of the workers interfering with each other that could occur if 1410 // operating on the global stack. 1411 1412 class G1CMKeepAliveAndDrainClosure: public OopClosure { 1413 G1ConcurrentMark* _cm; 1414 G1CMTask* _task; 1415 int _ref_counter_limit; 1416 int _ref_counter; 1417 bool _is_serial; 1418 public: 1419 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1420 _cm(cm), _task(task), _is_serial(is_serial), 1421 _ref_counter_limit(G1RefProcDrainInterval) { 1422 assert(_ref_counter_limit > 0, "sanity"); 1423 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1424 _ref_counter = _ref_counter_limit; 1425 } 1426 1427 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 1428 virtual void do_oop( oop* p) { do_oop_work(p); } 1429 1430 template <class T> void do_oop_work(T* p) { 1431 if (!_cm->has_overflown()) { 1432 oop obj = oopDesc::load_decode_heap_oop(p); 1433 _task->deal_with_reference(obj); 1434 _ref_counter--; 1435 1436 if (_ref_counter == 0) { 1437 // We have dealt with _ref_counter_limit references, pushing them 1438 // and objects reachable from them on to the local stack (and 1439 // possibly the global stack). Call G1CMTask::do_marking_step() to 1440 // process these entries. 1441 // 1442 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if 1443 // there's nothing more to do (i.e. we're done with the entries that 1444 // were pushed as a result of the G1CMTask::deal_with_reference() calls 1445 // above) or we overflow. 1446 // 1447 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1448 // flag while there may still be some work to do. (See the comment at 1449 // the beginning of G1CMTask::do_marking_step() for those conditions - 1450 // one of which is reaching the specified time target.) It is only 1451 // when G1CMTask::do_marking_step() returns without setting the 1452 // has_aborted() flag that the marking step has completed. 1453 do { 1454 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1455 _task->do_marking_step(mark_step_duration_ms, 1456 false /* do_termination */, 1457 _is_serial); 1458 } while (_task->has_aborted() && !_cm->has_overflown()); 1459 _ref_counter = _ref_counter_limit; 1460 } 1461 } 1462 } 1463 }; 1464 1465 // 'Drain' oop closure used by both serial and parallel reference processing. 1466 // Uses the G1CMTask associated with a given worker thread (for serial 1467 // reference processing the G1CMtask for worker 0 is used). Calls the 1468 // do_marking_step routine, with an unbelievably large timeout value, 1469 // to drain the marking data structures of the remaining entries 1470 // added by the 'keep alive' oop closure above. 1471 1472 class G1CMDrainMarkingStackClosure: public VoidClosure { 1473 G1ConcurrentMark* _cm; 1474 G1CMTask* _task; 1475 bool _is_serial; 1476 public: 1477 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1478 _cm(cm), _task(task), _is_serial(is_serial) { 1479 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1480 } 1481 1482 void do_void() { 1483 do { 1484 // We call G1CMTask::do_marking_step() to completely drain the local 1485 // and global marking stacks of entries pushed by the 'keep alive' 1486 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 1487 // 1488 // G1CMTask::do_marking_step() is called in a loop, which we'll exit 1489 // if there's nothing more to do (i.e. we've completely drained the 1490 // entries that were pushed as a a result of applying the 'keep alive' 1491 // closure to the entries on the discovered ref lists) or we overflow 1492 // the global marking stack. 1493 // 1494 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1495 // flag while there may still be some work to do. (See the comment at 1496 // the beginning of G1CMTask::do_marking_step() for those conditions - 1497 // one of which is reaching the specified time target.) It is only 1498 // when G1CMTask::do_marking_step() returns without setting the 1499 // has_aborted() flag that the marking step has completed. 1500 1501 _task->do_marking_step(1000000000.0 /* something very large */, 1502 true /* do_termination */, 1503 _is_serial); 1504 } while (_task->has_aborted() && !_cm->has_overflown()); 1505 } 1506 }; 1507 1508 // Implementation of AbstractRefProcTaskExecutor for parallel 1509 // reference processing at the end of G1 concurrent marking 1510 1511 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 1512 private: 1513 G1CollectedHeap* _g1h; 1514 G1ConcurrentMark* _cm; 1515 WorkGang* _workers; 1516 uint _active_workers; 1517 1518 public: 1519 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 1520 G1ConcurrentMark* cm, 1521 WorkGang* workers, 1522 uint n_workers) : 1523 _g1h(g1h), _cm(cm), 1524 _workers(workers), _active_workers(n_workers) { } 1525 1526 // Executes the given task using concurrent marking worker threads. 1527 virtual void execute(ProcessTask& task); 1528 virtual void execute(EnqueueTask& task); 1529 }; 1530 1531 class G1CMRefProcTaskProxy: public AbstractGangTask { 1532 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 1533 ProcessTask& _proc_task; 1534 G1CollectedHeap* _g1h; 1535 G1ConcurrentMark* _cm; 1536 1537 public: 1538 G1CMRefProcTaskProxy(ProcessTask& proc_task, 1539 G1CollectedHeap* g1h, 1540 G1ConcurrentMark* cm) : 1541 AbstractGangTask("Process reference objects in parallel"), 1542 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 1543 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1544 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 1545 } 1546 1547 virtual void work(uint worker_id) { 1548 ResourceMark rm; 1549 HandleMark hm; 1550 G1CMTask* task = _cm->task(worker_id); 1551 G1CMIsAliveClosure g1_is_alive(_g1h); 1552 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 1553 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 1554 1555 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 1556 } 1557 }; 1558 1559 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 1560 assert(_workers != NULL, "Need parallel worker threads."); 1561 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1562 1563 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 1564 1565 // We need to reset the concurrency level before each 1566 // proxy task execution, so that the termination protocol 1567 // and overflow handling in G1CMTask::do_marking_step() knows 1568 // how many workers to wait for. 1569 _cm->set_concurrency(_active_workers); 1570 _workers->run_task(&proc_task_proxy); 1571 } 1572 1573 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 1574 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 1575 EnqueueTask& _enq_task; 1576 1577 public: 1578 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 1579 AbstractGangTask("Enqueue reference objects in parallel"), 1580 _enq_task(enq_task) { } 1581 1582 virtual void work(uint worker_id) { 1583 _enq_task.work(worker_id); 1584 } 1585 }; 1586 1587 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 1588 assert(_workers != NULL, "Need parallel worker threads."); 1589 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1590 1591 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 1592 1593 // Not strictly necessary but... 1594 // 1595 // We need to reset the concurrency level before each 1596 // proxy task execution, so that the termination protocol 1597 // and overflow handling in G1CMTask::do_marking_step() knows 1598 // how many workers to wait for. 1599 _cm->set_concurrency(_active_workers); 1600 _workers->run_task(&enq_task_proxy); 1601 } 1602 1603 void G1ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) { 1604 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes); 1605 } 1606 1607 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 1608 if (has_overflown()) { 1609 // Skip processing the discovered references if we have 1610 // overflown the global marking stack. Reference objects 1611 // only get discovered once so it is OK to not 1612 // de-populate the discovered reference lists. We could have, 1613 // but the only benefit would be that, when marking restarts, 1614 // less reference objects are discovered. 1615 return; 1616 } 1617 1618 ResourceMark rm; 1619 HandleMark hm; 1620 1621 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1622 1623 // Is alive closure. 1624 G1CMIsAliveClosure g1_is_alive(g1h); 1625 1626 // Inner scope to exclude the cleaning of the string and symbol 1627 // tables from the displayed time. 1628 { 1629 GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm); 1630 1631 ReferenceProcessor* rp = g1h->ref_processor_cm(); 1632 1633 // See the comment in G1CollectedHeap::ref_processing_init() 1634 // about how reference processing currently works in G1. 1635 1636 // Set the soft reference policy 1637 rp->setup_policy(clear_all_soft_refs); 1638 assert(_global_mark_stack.is_empty(), "mark stack should be empty"); 1639 1640 // Instances of the 'Keep Alive' and 'Complete GC' closures used 1641 // in serial reference processing. Note these closures are also 1642 // used for serially processing (by the the current thread) the 1643 // JNI references during parallel reference processing. 1644 // 1645 // These closures do not need to synchronize with the worker 1646 // threads involved in parallel reference processing as these 1647 // instances are executed serially by the current thread (e.g. 1648 // reference processing is not multi-threaded and is thus 1649 // performed by the current thread instead of a gang worker). 1650 // 1651 // The gang tasks involved in parallel reference processing create 1652 // their own instances of these closures, which do their own 1653 // synchronization among themselves. 1654 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 1655 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 1656 1657 // We need at least one active thread. If reference processing 1658 // is not multi-threaded we use the current (VMThread) thread, 1659 // otherwise we use the work gang from the G1CollectedHeap and 1660 // we utilize all the worker threads we can. 1661 bool processing_is_mt = rp->processing_is_mt(); 1662 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U); 1663 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U); 1664 1665 // Parallel processing task executor. 1666 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 1667 g1h->workers(), active_workers); 1668 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 1669 1670 // Set the concurrency level. The phase was already set prior to 1671 // executing the remark task. 1672 set_concurrency(active_workers); 1673 1674 // Set the degree of MT processing here. If the discovery was done MT, 1675 // the number of threads involved during discovery could differ from 1676 // the number of active workers. This is OK as long as the discovered 1677 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 1678 rp->set_active_mt_degree(active_workers); 1679 1680 // Process the weak references. 1681 const ReferenceProcessorStats& stats = 1682 rp->process_discovered_references(&g1_is_alive, 1683 &g1_keep_alive, 1684 &g1_drain_mark_stack, 1685 executor, 1686 _gc_timer_cm); 1687 _gc_tracer_cm->report_gc_reference_stats(stats); 1688 1689 // The do_oop work routines of the keep_alive and drain_marking_stack 1690 // oop closures will set the has_overflown flag if we overflow the 1691 // global marking stack. 1692 1693 assert(_global_mark_stack.overflow() || _global_mark_stack.is_empty(), 1694 "mark stack should be empty (unless it overflowed)"); 1695 1696 if (_global_mark_stack.overflow()) { 1697 // This should have been done already when we tried to push an 1698 // entry on to the global mark stack. But let's do it again. 1699 set_has_overflown(); 1700 } 1701 1702 assert(rp->num_q() == active_workers, "why not"); 1703 1704 rp->enqueue_discovered_references(executor); 1705 1706 rp->verify_no_references_recorded(); 1707 assert(!rp->discovery_enabled(), "Post condition"); 1708 } 1709 1710 if (has_overflown()) { 1711 // We can not trust g1_is_alive if the marking stack overflowed 1712 return; 1713 } 1714 1715 assert(_global_mark_stack.is_empty(), "Marking should have completed"); 1716 1717 // Unload Klasses, String, Symbols, Code Cache, etc. 1718 if (ClassUnloadingWithConcurrentMark) { 1719 bool purged_classes; 1720 1721 { 1722 GCTraceTime(Debug, gc, phases) trace("System Dictionary Unloading", _gc_timer_cm); 1723 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */); 1724 } 1725 1726 { 1727 GCTraceTime(Debug, gc, phases) trace("Parallel Unloading", _gc_timer_cm); 1728 weakRefsWorkParallelPart(&g1_is_alive, purged_classes); 1729 } 1730 } 1731 1732 if (G1StringDedup::is_enabled()) { 1733 GCTraceTime(Debug, gc, phases) trace("String Deduplication Unlink", _gc_timer_cm); 1734 G1StringDedup::unlink(&g1_is_alive); 1735 } 1736 } 1737 1738 void G1ConcurrentMark::swapMarkBitMaps() { 1739 G1CMBitMapRO* temp = _prevMarkBitMap; 1740 _prevMarkBitMap = (G1CMBitMapRO*)_nextMarkBitMap; 1741 _nextMarkBitMap = (G1CMBitMap*) temp; 1742 } 1743 1744 // Closure for marking entries in SATB buffers. 1745 class G1CMSATBBufferClosure : public SATBBufferClosure { 1746 private: 1747 G1CMTask* _task; 1748 G1CollectedHeap* _g1h; 1749 1750 // This is very similar to G1CMTask::deal_with_reference, but with 1751 // more relaxed requirements for the argument, so this must be more 1752 // circumspect about treating the argument as an object. 1753 void do_entry(void* entry) const { 1754 _task->increment_refs_reached(); 1755 HeapRegion* hr = _g1h->heap_region_containing(entry); 1756 if (entry < hr->next_top_at_mark_start()) { 1757 // Until we get here, we don't know whether entry refers to a valid 1758 // object; it could instead have been a stale reference. 1759 oop obj = static_cast<oop>(entry); 1760 assert(obj->is_oop(true /* ignore mark word */), 1761 "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj)); 1762 _task->make_reference_grey(obj); 1763 } 1764 } 1765 1766 public: 1767 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h) 1768 : _task(task), _g1h(g1h) { } 1769 1770 virtual void do_buffer(void** buffer, size_t size) { 1771 for (size_t i = 0; i < size; ++i) { 1772 do_entry(buffer[i]); 1773 } 1774 } 1775 }; 1776 1777 class G1RemarkThreadsClosure : public ThreadClosure { 1778 G1CMSATBBufferClosure _cm_satb_cl; 1779 G1CMOopClosure _cm_cl; 1780 MarkingCodeBlobClosure _code_cl; 1781 int _thread_parity; 1782 1783 public: 1784 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) : 1785 _cm_satb_cl(task, g1h), 1786 _cm_cl(g1h, g1h->concurrent_mark(), task), 1787 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations), 1788 _thread_parity(Threads::thread_claim_parity()) {} 1789 1790 void do_thread(Thread* thread) { 1791 if (thread->is_Java_thread()) { 1792 if (thread->claim_oops_do(true, _thread_parity)) { 1793 JavaThread* jt = (JavaThread*)thread; 1794 1795 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking 1796 // however the liveness of oops reachable from nmethods have very complex lifecycles: 1797 // * Alive if on the stack of an executing method 1798 // * Weakly reachable otherwise 1799 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be 1800 // live by the SATB invariant but other oops recorded in nmethods may behave differently. 1801 jt->nmethods_do(&_code_cl); 1802 1803 jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl); 1804 } 1805 } else if (thread->is_VM_thread()) { 1806 if (thread->claim_oops_do(true, _thread_parity)) { 1807 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl); 1808 } 1809 } 1810 } 1811 }; 1812 1813 class G1CMRemarkTask: public AbstractGangTask { 1814 private: 1815 G1ConcurrentMark* _cm; 1816 public: 1817 void work(uint worker_id) { 1818 // Since all available tasks are actually started, we should 1819 // only proceed if we're supposed to be active. 1820 if (worker_id < _cm->active_tasks()) { 1821 G1CMTask* task = _cm->task(worker_id); 1822 task->record_start_time(); 1823 { 1824 ResourceMark rm; 1825 HandleMark hm; 1826 1827 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task); 1828 Threads::threads_do(&threads_f); 1829 } 1830 1831 do { 1832 task->do_marking_step(1000000000.0 /* something very large */, 1833 true /* do_termination */, 1834 false /* is_serial */); 1835 } while (task->has_aborted() && !_cm->has_overflown()); 1836 // If we overflow, then we do not want to restart. We instead 1837 // want to abort remark and do concurrent marking again. 1838 task->record_end_time(); 1839 } 1840 } 1841 1842 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) : 1843 AbstractGangTask("Par Remark"), _cm(cm) { 1844 _cm->terminator()->reset_for_reuse(active_workers); 1845 } 1846 }; 1847 1848 void G1ConcurrentMark::checkpointRootsFinalWork() { 1849 ResourceMark rm; 1850 HandleMark hm; 1851 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1852 1853 GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm); 1854 1855 g1h->ensure_parsability(false); 1856 1857 // this is remark, so we'll use up all active threads 1858 uint active_workers = g1h->workers()->active_workers(); 1859 set_concurrency_and_phase(active_workers, false /* concurrent */); 1860 // Leave _parallel_marking_threads at it's 1861 // value originally calculated in the G1ConcurrentMark 1862 // constructor and pass values of the active workers 1863 // through the gang in the task. 1864 1865 { 1866 StrongRootsScope srs(active_workers); 1867 1868 G1CMRemarkTask remarkTask(this, active_workers); 1869 // We will start all available threads, even if we decide that the 1870 // active_workers will be fewer. The extra ones will just bail out 1871 // immediately. 1872 g1h->workers()->run_task(&remarkTask); 1873 } 1874 1875 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1876 guarantee(has_overflown() || 1877 satb_mq_set.completed_buffers_num() == 0, 1878 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT, 1879 BOOL_TO_STR(has_overflown()), 1880 satb_mq_set.completed_buffers_num()); 1881 1882 print_stats(); 1883 } 1884 1885 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 1886 // Note we are overriding the read-only view of the prev map here, via 1887 // the cast. 1888 ((G1CMBitMap*)_prevMarkBitMap)->clear_range(mr); 1889 } 1890 1891 HeapRegion* 1892 G1ConcurrentMark::claim_region(uint worker_id) { 1893 // "checkpoint" the finger 1894 HeapWord* finger = _finger; 1895 1896 // _heap_end will not change underneath our feet; it only changes at 1897 // yield points. 1898 while (finger < _heap_end) { 1899 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 1900 1901 HeapRegion* curr_region = _g1h->heap_region_containing(finger); 1902 1903 // Above heap_region_containing may return NULL as we always scan claim 1904 // until the end of the heap. In this case, just jump to the next region. 1905 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords; 1906 1907 // Is the gap between reading the finger and doing the CAS too long? 1908 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 1909 if (res == finger && curr_region != NULL) { 1910 // we succeeded 1911 HeapWord* bottom = curr_region->bottom(); 1912 HeapWord* limit = curr_region->next_top_at_mark_start(); 1913 1914 // notice that _finger == end cannot be guaranteed here since, 1915 // someone else might have moved the finger even further 1916 assert(_finger >= end, "the finger should have moved forward"); 1917 1918 if (limit > bottom) { 1919 return curr_region; 1920 } else { 1921 assert(limit == bottom, 1922 "the region limit should be at bottom"); 1923 // we return NULL and the caller should try calling 1924 // claim_region() again. 1925 return NULL; 1926 } 1927 } else { 1928 assert(_finger > finger, "the finger should have moved forward"); 1929 // read it again 1930 finger = _finger; 1931 } 1932 } 1933 1934 return NULL; 1935 } 1936 1937 #ifndef PRODUCT 1938 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC { 1939 private: 1940 G1CollectedHeap* _g1h; 1941 const char* _phase; 1942 int _info; 1943 1944 public: 1945 VerifyNoCSetOops(const char* phase, int info = -1) : 1946 _g1h(G1CollectedHeap::heap()), 1947 _phase(phase), 1948 _info(info) 1949 { } 1950 1951 void operator()(oop obj) const { 1952 guarantee(obj->is_oop(), 1953 "Non-oop " PTR_FORMAT ", phase: %s, info: %d", 1954 p2i(obj), _phase, _info); 1955 guarantee(!_g1h->obj_in_cs(obj), 1956 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d", 1957 p2i(obj), _phase, _info); 1958 } 1959 }; 1960 1961 void G1ConcurrentMark::verify_no_cset_oops() { 1962 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1963 if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) { 1964 return; 1965 } 1966 1967 // Verify entries on the global mark stack 1968 _global_mark_stack.iterate(VerifyNoCSetOops("Stack")); 1969 1970 // Verify entries on the task queues 1971 for (uint i = 0; i < _max_worker_id; ++i) { 1972 G1CMTaskQueue* queue = _task_queues->queue(i); 1973 queue->iterate(VerifyNoCSetOops("Queue", i)); 1974 } 1975 1976 // Verify the global finger 1977 HeapWord* global_finger = finger(); 1978 if (global_finger != NULL && global_finger < _heap_end) { 1979 // Since we always iterate over all regions, we might get a NULL HeapRegion 1980 // here. 1981 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger); 1982 guarantee(global_hr == NULL || global_finger == global_hr->bottom(), 1983 "global finger: " PTR_FORMAT " region: " HR_FORMAT, 1984 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)); 1985 } 1986 1987 // Verify the task fingers 1988 assert(parallel_marking_threads() <= _max_worker_id, "sanity"); 1989 for (uint i = 0; i < parallel_marking_threads(); ++i) { 1990 G1CMTask* task = _tasks[i]; 1991 HeapWord* task_finger = task->finger(); 1992 if (task_finger != NULL && task_finger < _heap_end) { 1993 // See above note on the global finger verification. 1994 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger); 1995 guarantee(task_hr == NULL || task_finger == task_hr->bottom() || 1996 !task_hr->in_collection_set(), 1997 "task finger: " PTR_FORMAT " region: " HR_FORMAT, 1998 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)); 1999 } 2000 } 2001 } 2002 #endif // PRODUCT 2003 void G1ConcurrentMark::create_live_data() { 2004 _g1h->g1_rem_set()->create_card_live_data(_parallel_workers, _nextMarkBitMap); 2005 } 2006 2007 void G1ConcurrentMark::finalize_live_data() { 2008 _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _nextMarkBitMap); 2009 } 2010 2011 void G1ConcurrentMark::verify_live_data() { 2012 _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _nextMarkBitMap); 2013 } 2014 2015 void G1ConcurrentMark::clear_live_data(WorkGang* workers) { 2016 _g1h->g1_rem_set()->clear_card_live_data(workers); 2017 } 2018 2019 #ifdef ASSERT 2020 void G1ConcurrentMark::verify_live_data_clear() { 2021 _g1h->g1_rem_set()->verify_card_live_data_is_clear(); 2022 } 2023 #endif 2024 2025 void G1ConcurrentMark::print_stats() { 2026 if (!log_is_enabled(Debug, gc, stats)) { 2027 return; 2028 } 2029 log_debug(gc, stats)("---------------------------------------------------------------------"); 2030 for (size_t i = 0; i < _active_tasks; ++i) { 2031 _tasks[i]->print_stats(); 2032 log_debug(gc, stats)("---------------------------------------------------------------------"); 2033 } 2034 } 2035 2036 void G1ConcurrentMark::abort() { 2037 if (!cmThread()->during_cycle() || _has_aborted) { 2038 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything. 2039 return; 2040 } 2041 2042 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next 2043 // concurrent bitmap clearing. 2044 { 2045 GCTraceTime(Debug, gc)("Clear Next Bitmap"); 2046 clear_bitmap(_nextMarkBitMap, _g1h->workers(), false); 2047 } 2048 // Note we cannot clear the previous marking bitmap here 2049 // since VerifyDuringGC verifies the objects marked during 2050 // a full GC against the previous bitmap. 2051 2052 { 2053 GCTraceTime(Debug, gc)("Clear Live Data"); 2054 clear_live_data(_g1h->workers()); 2055 } 2056 DEBUG_ONLY({ 2057 GCTraceTime(Debug, gc)("Verify Live Data Clear"); 2058 verify_live_data_clear(); 2059 }) 2060 // Empty mark stack 2061 reset_marking_state(); 2062 for (uint i = 0; i < _max_worker_id; ++i) { 2063 _tasks[i]->clear_region_fields(); 2064 } 2065 _first_overflow_barrier_sync.abort(); 2066 _second_overflow_barrier_sync.abort(); 2067 _has_aborted = true; 2068 2069 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2070 satb_mq_set.abandon_partial_marking(); 2071 // This can be called either during or outside marking, we'll read 2072 // the expected_active value from the SATB queue set. 2073 satb_mq_set.set_active_all_threads( 2074 false, /* new active value */ 2075 satb_mq_set.is_active() /* expected_active */); 2076 } 2077 2078 static void print_ms_time_info(const char* prefix, const char* name, 2079 NumberSeq& ns) { 2080 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 2081 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 2082 if (ns.num() > 0) { 2083 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]", 2084 prefix, ns.sd(), ns.maximum()); 2085 } 2086 } 2087 2088 void G1ConcurrentMark::print_summary_info() { 2089 Log(gc, marking) log; 2090 if (!log.is_trace()) { 2091 return; 2092 } 2093 2094 log.trace(" Concurrent marking:"); 2095 print_ms_time_info(" ", "init marks", _init_times); 2096 print_ms_time_info(" ", "remarks", _remark_times); 2097 { 2098 print_ms_time_info(" ", "final marks", _remark_mark_times); 2099 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 2100 2101 } 2102 print_ms_time_info(" ", "cleanups", _cleanup_times); 2103 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).", 2104 _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2105 if (G1ScrubRemSets) { 2106 log.trace(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 2107 _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2108 } 2109 log.trace(" Total stop_world time = %8.2f s.", 2110 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0); 2111 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).", 2112 cmThread()->vtime_accum(), cmThread()->vtime_mark_accum()); 2113 } 2114 2115 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const { 2116 _parallel_workers->print_worker_threads_on(st); 2117 } 2118 2119 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const { 2120 _parallel_workers->threads_do(tc); 2121 } 2122 2123 void G1ConcurrentMark::print_on_error(outputStream* st) const { 2124 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT, 2125 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap)); 2126 _prevMarkBitMap->print_on_error(st, " Prev Bits: "); 2127 _nextMarkBitMap->print_on_error(st, " Next Bits: "); 2128 } 2129 2130 // Closure for iteration over bitmaps 2131 class G1CMBitMapClosure : public BitMapClosure { 2132 private: 2133 // the bitmap that is being iterated over 2134 G1CMBitMap* _nextMarkBitMap; 2135 G1ConcurrentMark* _cm; 2136 G1CMTask* _task; 2137 2138 public: 2139 G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm, G1CMBitMap* nextMarkBitMap) : 2140 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } 2141 2142 bool do_bit(size_t offset) { 2143 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); 2144 assert(_nextMarkBitMap->isMarked(addr), "invariant"); 2145 assert( addr < _cm->finger(), "invariant"); 2146 assert(addr >= _task->finger(), "invariant"); 2147 2148 // We move that task's local finger along. 2149 _task->move_finger_to(addr); 2150 2151 _task->scan_object(oop(addr)); 2152 // we only partially drain the local queue and global stack 2153 _task->drain_local_queue(true); 2154 _task->drain_global_stack(true); 2155 2156 // if the has_aborted flag has been raised, we need to bail out of 2157 // the iteration 2158 return !_task->has_aborted(); 2159 } 2160 }; 2161 2162 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) { 2163 ReferenceProcessor* result = g1h->ref_processor_cm(); 2164 assert(result != NULL, "CM reference processor should not be NULL"); 2165 return result; 2166 } 2167 2168 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 2169 G1ConcurrentMark* cm, 2170 G1CMTask* task) 2171 : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)), 2172 _g1h(g1h), _cm(cm), _task(task) 2173 { } 2174 2175 void G1CMTask::setup_for_region(HeapRegion* hr) { 2176 assert(hr != NULL, 2177 "claim_region() should have filtered out NULL regions"); 2178 _curr_region = hr; 2179 _finger = hr->bottom(); 2180 update_region_limit(); 2181 } 2182 2183 void G1CMTask::update_region_limit() { 2184 HeapRegion* hr = _curr_region; 2185 HeapWord* bottom = hr->bottom(); 2186 HeapWord* limit = hr->next_top_at_mark_start(); 2187 2188 if (limit == bottom) { 2189 // The region was collected underneath our feet. 2190 // We set the finger to bottom to ensure that the bitmap 2191 // iteration that will follow this will not do anything. 2192 // (this is not a condition that holds when we set the region up, 2193 // as the region is not supposed to be empty in the first place) 2194 _finger = bottom; 2195 } else if (limit >= _region_limit) { 2196 assert(limit >= _finger, "peace of mind"); 2197 } else { 2198 assert(limit < _region_limit, "only way to get here"); 2199 // This can happen under some pretty unusual circumstances. An 2200 // evacuation pause empties the region underneath our feet (NTAMS 2201 // at bottom). We then do some allocation in the region (NTAMS 2202 // stays at bottom), followed by the region being used as a GC 2203 // alloc region (NTAMS will move to top() and the objects 2204 // originally below it will be grayed). All objects now marked in 2205 // the region are explicitly grayed, if below the global finger, 2206 // and we do not need in fact to scan anything else. So, we simply 2207 // set _finger to be limit to ensure that the bitmap iteration 2208 // doesn't do anything. 2209 _finger = limit; 2210 } 2211 2212 _region_limit = limit; 2213 } 2214 2215 void G1CMTask::giveup_current_region() { 2216 assert(_curr_region != NULL, "invariant"); 2217 clear_region_fields(); 2218 } 2219 2220 void G1CMTask::clear_region_fields() { 2221 // Values for these three fields that indicate that we're not 2222 // holding on to a region. 2223 _curr_region = NULL; 2224 _finger = NULL; 2225 _region_limit = NULL; 2226 } 2227 2228 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 2229 if (cm_oop_closure == NULL) { 2230 assert(_cm_oop_closure != NULL, "invariant"); 2231 } else { 2232 assert(_cm_oop_closure == NULL, "invariant"); 2233 } 2234 _cm_oop_closure = cm_oop_closure; 2235 } 2236 2237 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) { 2238 guarantee(nextMarkBitMap != NULL, "invariant"); 2239 _nextMarkBitMap = nextMarkBitMap; 2240 clear_region_fields(); 2241 2242 _calls = 0; 2243 _elapsed_time_ms = 0.0; 2244 _termination_time_ms = 0.0; 2245 _termination_start_time_ms = 0.0; 2246 } 2247 2248 bool G1CMTask::should_exit_termination() { 2249 regular_clock_call(); 2250 // This is called when we are in the termination protocol. We should 2251 // quit if, for some reason, this task wants to abort or the global 2252 // stack is not empty (this means that we can get work from it). 2253 return !_cm->mark_stack_empty() || has_aborted(); 2254 } 2255 2256 void G1CMTask::reached_limit() { 2257 assert(_words_scanned >= _words_scanned_limit || 2258 _refs_reached >= _refs_reached_limit , 2259 "shouldn't have been called otherwise"); 2260 regular_clock_call(); 2261 } 2262 2263 void G1CMTask::regular_clock_call() { 2264 if (has_aborted()) return; 2265 2266 // First, we need to recalculate the words scanned and refs reached 2267 // limits for the next clock call. 2268 recalculate_limits(); 2269 2270 // During the regular clock call we do the following 2271 2272 // (1) If an overflow has been flagged, then we abort. 2273 if (_cm->has_overflown()) { 2274 set_has_aborted(); 2275 return; 2276 } 2277 2278 // If we are not concurrent (i.e. we're doing remark) we don't need 2279 // to check anything else. The other steps are only needed during 2280 // the concurrent marking phase. 2281 if (!concurrent()) return; 2282 2283 // (2) If marking has been aborted for Full GC, then we also abort. 2284 if (_cm->has_aborted()) { 2285 set_has_aborted(); 2286 return; 2287 } 2288 2289 double curr_time_ms = os::elapsedVTime() * 1000.0; 2290 2291 // (4) We check whether we should yield. If we have to, then we abort. 2292 if (SuspendibleThreadSet::should_yield()) { 2293 // We should yield. To do this we abort the task. The caller is 2294 // responsible for yielding. 2295 set_has_aborted(); 2296 return; 2297 } 2298 2299 // (5) We check whether we've reached our time quota. If we have, 2300 // then we abort. 2301 double elapsed_time_ms = curr_time_ms - _start_time_ms; 2302 if (elapsed_time_ms > _time_target_ms) { 2303 set_has_aborted(); 2304 _has_timed_out = true; 2305 return; 2306 } 2307 2308 // (6) Finally, we check whether there are enough completed STAB 2309 // buffers available for processing. If there are, we abort. 2310 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2311 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 2312 // we do need to process SATB buffers, we'll abort and restart 2313 // the marking task to do so 2314 set_has_aborted(); 2315 return; 2316 } 2317 } 2318 2319 void G1CMTask::recalculate_limits() { 2320 _real_words_scanned_limit = _words_scanned + words_scanned_period; 2321 _words_scanned_limit = _real_words_scanned_limit; 2322 2323 _real_refs_reached_limit = _refs_reached + refs_reached_period; 2324 _refs_reached_limit = _real_refs_reached_limit; 2325 } 2326 2327 void G1CMTask::decrease_limits() { 2328 // This is called when we believe that we're going to do an infrequent 2329 // operation which will increase the per byte scanned cost (i.e. move 2330 // entries to/from the global stack). It basically tries to decrease the 2331 // scanning limit so that the clock is called earlier. 2332 2333 _words_scanned_limit = _real_words_scanned_limit - 2334 3 * words_scanned_period / 4; 2335 _refs_reached_limit = _real_refs_reached_limit - 2336 3 * refs_reached_period / 4; 2337 } 2338 2339 void G1CMTask::move_entries_to_global_stack() { 2340 // local array where we'll store the entries that will be popped 2341 // from the local queue 2342 oop buffer[global_stack_transfer_size]; 2343 2344 int n = 0; 2345 oop obj; 2346 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { 2347 buffer[n] = obj; 2348 ++n; 2349 } 2350 2351 if (n > 0) { 2352 // we popped at least one entry from the local queue 2353 2354 if (!_cm->mark_stack_push(buffer, n)) { 2355 set_has_aborted(); 2356 } 2357 } 2358 2359 // this operation was quite expensive, so decrease the limits 2360 decrease_limits(); 2361 } 2362 2363 void G1CMTask::get_entries_from_global_stack() { 2364 // local array where we'll store the entries that will be popped 2365 // from the global stack. 2366 oop buffer[global_stack_transfer_size]; 2367 size_t n; 2368 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); 2369 assert(n <= global_stack_transfer_size, 2370 "we should not pop more than the given limit"); 2371 if (n > 0) { 2372 // yes, we did actually pop at least one entry 2373 for (size_t i = 0; i < n; ++i) { 2374 bool success = _task_queue->push(buffer[i]); 2375 // We only call this when the local queue is empty or under a 2376 // given target limit. So, we do not expect this push to fail. 2377 assert(success, "invariant"); 2378 } 2379 } 2380 2381 // this operation was quite expensive, so decrease the limits 2382 decrease_limits(); 2383 } 2384 2385 void G1CMTask::drain_local_queue(bool partially) { 2386 if (has_aborted()) return; 2387 2388 // Decide what the target size is, depending whether we're going to 2389 // drain it partially (so that other tasks can steal if they run out 2390 // of things to do) or totally (at the very end). 2391 size_t target_size; 2392 if (partially) { 2393 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 2394 } else { 2395 target_size = 0; 2396 } 2397 2398 if (_task_queue->size() > target_size) { 2399 oop obj; 2400 bool ret = _task_queue->pop_local(obj); 2401 while (ret) { 2402 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); 2403 assert(!_g1h->is_on_master_free_list( 2404 _g1h->heap_region_containing((HeapWord*) obj)), "invariant"); 2405 2406 scan_object(obj); 2407 2408 if (_task_queue->size() <= target_size || has_aborted()) { 2409 ret = false; 2410 } else { 2411 ret = _task_queue->pop_local(obj); 2412 } 2413 } 2414 } 2415 } 2416 2417 void G1CMTask::drain_global_stack(bool partially) { 2418 if (has_aborted()) return; 2419 2420 // We have a policy to drain the local queue before we attempt to 2421 // drain the global stack. 2422 assert(partially || _task_queue->size() == 0, "invariant"); 2423 2424 // Decide what the target size is, depending whether we're going to 2425 // drain it partially (so that other tasks can steal if they run out 2426 // of things to do) or totally (at the very end). Notice that, 2427 // because we move entries from the global stack in chunks or 2428 // because another task might be doing the same, we might in fact 2429 // drop below the target. But, this is not a problem. 2430 size_t target_size; 2431 if (partially) { 2432 target_size = _cm->partial_mark_stack_size_target(); 2433 } else { 2434 target_size = 0; 2435 } 2436 2437 if (_cm->mark_stack_size() > target_size) { 2438 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 2439 get_entries_from_global_stack(); 2440 drain_local_queue(partially); 2441 } 2442 } 2443 } 2444 2445 // SATB Queue has several assumptions on whether to call the par or 2446 // non-par versions of the methods. this is why some of the code is 2447 // replicated. We should really get rid of the single-threaded version 2448 // of the code to simplify things. 2449 void G1CMTask::drain_satb_buffers() { 2450 if (has_aborted()) return; 2451 2452 // We set this so that the regular clock knows that we're in the 2453 // middle of draining buffers and doesn't set the abort flag when it 2454 // notices that SATB buffers are available for draining. It'd be 2455 // very counter productive if it did that. :-) 2456 _draining_satb_buffers = true; 2457 2458 G1CMSATBBufferClosure satb_cl(this, _g1h); 2459 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2460 2461 // This keeps claiming and applying the closure to completed buffers 2462 // until we run out of buffers or we need to abort. 2463 while (!has_aborted() && 2464 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) { 2465 regular_clock_call(); 2466 } 2467 2468 _draining_satb_buffers = false; 2469 2470 assert(has_aborted() || 2471 concurrent() || 2472 satb_mq_set.completed_buffers_num() == 0, "invariant"); 2473 2474 // again, this was a potentially expensive operation, decrease the 2475 // limits to get the regular clock call early 2476 decrease_limits(); 2477 } 2478 2479 void G1CMTask::print_stats() { 2480 log_debug(gc, stats)("Marking Stats, task = %u, calls = %d", 2481 _worker_id, _calls); 2482 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 2483 _elapsed_time_ms, _termination_time_ms); 2484 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 2485 _step_times_ms.num(), _step_times_ms.avg(), 2486 _step_times_ms.sd()); 2487 log_debug(gc, stats)(" max = %1.2lfms, total = %1.2lfms", 2488 _step_times_ms.maximum(), _step_times_ms.sum()); 2489 } 2490 2491 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) { 2492 return _task_queues->steal(worker_id, hash_seed, obj); 2493 } 2494 2495 /***************************************************************************** 2496 2497 The do_marking_step(time_target_ms, ...) method is the building 2498 block of the parallel marking framework. It can be called in parallel 2499 with other invocations of do_marking_step() on different tasks 2500 (but only one per task, obviously) and concurrently with the 2501 mutator threads, or during remark, hence it eliminates the need 2502 for two versions of the code. When called during remark, it will 2503 pick up from where the task left off during the concurrent marking 2504 phase. Interestingly, tasks are also claimable during evacuation 2505 pauses too, since do_marking_step() ensures that it aborts before 2506 it needs to yield. 2507 2508 The data structures that it uses to do marking work are the 2509 following: 2510 2511 (1) Marking Bitmap. If there are gray objects that appear only 2512 on the bitmap (this happens either when dealing with an overflow 2513 or when the initial marking phase has simply marked the roots 2514 and didn't push them on the stack), then tasks claim heap 2515 regions whose bitmap they then scan to find gray objects. A 2516 global finger indicates where the end of the last claimed region 2517 is. A local finger indicates how far into the region a task has 2518 scanned. The two fingers are used to determine how to gray an 2519 object (i.e. whether simply marking it is OK, as it will be 2520 visited by a task in the future, or whether it needs to be also 2521 pushed on a stack). 2522 2523 (2) Local Queue. The local queue of the task which is accessed 2524 reasonably efficiently by the task. Other tasks can steal from 2525 it when they run out of work. Throughout the marking phase, a 2526 task attempts to keep its local queue short but not totally 2527 empty, so that entries are available for stealing by other 2528 tasks. Only when there is no more work, a task will totally 2529 drain its local queue. 2530 2531 (3) Global Mark Stack. This handles local queue overflow. During 2532 marking only sets of entries are moved between it and the local 2533 queues, as access to it requires a mutex and more fine-grain 2534 interaction with it which might cause contention. If it 2535 overflows, then the marking phase should restart and iterate 2536 over the bitmap to identify gray objects. Throughout the marking 2537 phase, tasks attempt to keep the global mark stack at a small 2538 length but not totally empty, so that entries are available for 2539 popping by other tasks. Only when there is no more work, tasks 2540 will totally drain the global mark stack. 2541 2542 (4) SATB Buffer Queue. This is where completed SATB buffers are 2543 made available. Buffers are regularly removed from this queue 2544 and scanned for roots, so that the queue doesn't get too 2545 long. During remark, all completed buffers are processed, as 2546 well as the filled in parts of any uncompleted buffers. 2547 2548 The do_marking_step() method tries to abort when the time target 2549 has been reached. There are a few other cases when the 2550 do_marking_step() method also aborts: 2551 2552 (1) When the marking phase has been aborted (after a Full GC). 2553 2554 (2) When a global overflow (on the global stack) has been 2555 triggered. Before the task aborts, it will actually sync up with 2556 the other tasks to ensure that all the marking data structures 2557 (local queues, stacks, fingers etc.) are re-initialized so that 2558 when do_marking_step() completes, the marking phase can 2559 immediately restart. 2560 2561 (3) When enough completed SATB buffers are available. The 2562 do_marking_step() method only tries to drain SATB buffers right 2563 at the beginning. So, if enough buffers are available, the 2564 marking step aborts and the SATB buffers are processed at 2565 the beginning of the next invocation. 2566 2567 (4) To yield. when we have to yield then we abort and yield 2568 right at the end of do_marking_step(). This saves us from a lot 2569 of hassle as, by yielding we might allow a Full GC. If this 2570 happens then objects will be compacted underneath our feet, the 2571 heap might shrink, etc. We save checking for this by just 2572 aborting and doing the yield right at the end. 2573 2574 From the above it follows that the do_marking_step() method should 2575 be called in a loop (or, otherwise, regularly) until it completes. 2576 2577 If a marking step completes without its has_aborted() flag being 2578 true, it means it has completed the current marking phase (and 2579 also all other marking tasks have done so and have all synced up). 2580 2581 A method called regular_clock_call() is invoked "regularly" (in 2582 sub ms intervals) throughout marking. It is this clock method that 2583 checks all the abort conditions which were mentioned above and 2584 decides when the task should abort. A work-based scheme is used to 2585 trigger this clock method: when the number of object words the 2586 marking phase has scanned or the number of references the marking 2587 phase has visited reach a given limit. Additional invocations to 2588 the method clock have been planted in a few other strategic places 2589 too. The initial reason for the clock method was to avoid calling 2590 vtime too regularly, as it is quite expensive. So, once it was in 2591 place, it was natural to piggy-back all the other conditions on it 2592 too and not constantly check them throughout the code. 2593 2594 If do_termination is true then do_marking_step will enter its 2595 termination protocol. 2596 2597 The value of is_serial must be true when do_marking_step is being 2598 called serially (i.e. by the VMThread) and do_marking_step should 2599 skip any synchronization in the termination and overflow code. 2600 Examples include the serial remark code and the serial reference 2601 processing closures. 2602 2603 The value of is_serial must be false when do_marking_step is 2604 being called by any of the worker threads in a work gang. 2605 Examples include the concurrent marking code (CMMarkingTask), 2606 the MT remark code, and the MT reference processing closures. 2607 2608 *****************************************************************************/ 2609 2610 void G1CMTask::do_marking_step(double time_target_ms, 2611 bool do_termination, 2612 bool is_serial) { 2613 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 2614 assert(concurrent() == _cm->concurrent(), "they should be the same"); 2615 2616 G1Policy* g1_policy = _g1h->g1_policy(); 2617 assert(_task_queues != NULL, "invariant"); 2618 assert(_task_queue != NULL, "invariant"); 2619 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant"); 2620 2621 assert(!_claimed, 2622 "only one thread should claim this task at any one time"); 2623 2624 // OK, this doesn't safeguard again all possible scenarios, as it is 2625 // possible for two threads to set the _claimed flag at the same 2626 // time. But it is only for debugging purposes anyway and it will 2627 // catch most problems. 2628 _claimed = true; 2629 2630 _start_time_ms = os::elapsedVTime() * 1000.0; 2631 2632 // If do_stealing is true then do_marking_step will attempt to 2633 // steal work from the other G1CMTasks. It only makes sense to 2634 // enable stealing when the termination protocol is enabled 2635 // and do_marking_step() is not being called serially. 2636 bool do_stealing = do_termination && !is_serial; 2637 2638 double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms); 2639 _time_target_ms = time_target_ms - diff_prediction_ms; 2640 2641 // set up the variables that are used in the work-based scheme to 2642 // call the regular clock method 2643 _words_scanned = 0; 2644 _refs_reached = 0; 2645 recalculate_limits(); 2646 2647 // clear all flags 2648 clear_has_aborted(); 2649 _has_timed_out = false; 2650 _draining_satb_buffers = false; 2651 2652 ++_calls; 2653 2654 // Set up the bitmap and oop closures. Anything that uses them is 2655 // eventually called from this method, so it is OK to allocate these 2656 // statically. 2657 G1CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); 2658 G1CMOopClosure cm_oop_closure(_g1h, _cm, this); 2659 set_cm_oop_closure(&cm_oop_closure); 2660 2661 if (_cm->has_overflown()) { 2662 // This can happen if the mark stack overflows during a GC pause 2663 // and this task, after a yield point, restarts. We have to abort 2664 // as we need to get into the overflow protocol which happens 2665 // right at the end of this task. 2666 set_has_aborted(); 2667 } 2668 2669 // First drain any available SATB buffers. After this, we will not 2670 // look at SATB buffers before the next invocation of this method. 2671 // If enough completed SATB buffers are queued up, the regular clock 2672 // will abort this task so that it restarts. 2673 drain_satb_buffers(); 2674 // ...then partially drain the local queue and the global stack 2675 drain_local_queue(true); 2676 drain_global_stack(true); 2677 2678 do { 2679 if (!has_aborted() && _curr_region != NULL) { 2680 // This means that we're already holding on to a region. 2681 assert(_finger != NULL, "if region is not NULL, then the finger " 2682 "should not be NULL either"); 2683 2684 // We might have restarted this task after an evacuation pause 2685 // which might have evacuated the region we're holding on to 2686 // underneath our feet. Let's read its limit again to make sure 2687 // that we do not iterate over a region of the heap that 2688 // contains garbage (update_region_limit() will also move 2689 // _finger to the start of the region if it is found empty). 2690 update_region_limit(); 2691 // We will start from _finger not from the start of the region, 2692 // as we might be restarting this task after aborting half-way 2693 // through scanning this region. In this case, _finger points to 2694 // the address where we last found a marked object. If this is a 2695 // fresh region, _finger points to start(). 2696 MemRegion mr = MemRegion(_finger, _region_limit); 2697 2698 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(), 2699 "humongous regions should go around loop once only"); 2700 2701 // Some special cases: 2702 // If the memory region is empty, we can just give up the region. 2703 // If the current region is humongous then we only need to check 2704 // the bitmap for the bit associated with the start of the object, 2705 // scan the object if it's live, and give up the region. 2706 // Otherwise, let's iterate over the bitmap of the part of the region 2707 // that is left. 2708 // If the iteration is successful, give up the region. 2709 if (mr.is_empty()) { 2710 giveup_current_region(); 2711 regular_clock_call(); 2712 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) { 2713 if (_nextMarkBitMap->isMarked(mr.start())) { 2714 // The object is marked - apply the closure 2715 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start()); 2716 bitmap_closure.do_bit(offset); 2717 } 2718 // Even if this task aborted while scanning the humongous object 2719 // we can (and should) give up the current region. 2720 giveup_current_region(); 2721 regular_clock_call(); 2722 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) { 2723 giveup_current_region(); 2724 regular_clock_call(); 2725 } else { 2726 assert(has_aborted(), "currently the only way to do so"); 2727 // The only way to abort the bitmap iteration is to return 2728 // false from the do_bit() method. However, inside the 2729 // do_bit() method we move the _finger to point to the 2730 // object currently being looked at. So, if we bail out, we 2731 // have definitely set _finger to something non-null. 2732 assert(_finger != NULL, "invariant"); 2733 2734 // Region iteration was actually aborted. So now _finger 2735 // points to the address of the object we last scanned. If we 2736 // leave it there, when we restart this task, we will rescan 2737 // the object. It is easy to avoid this. We move the finger by 2738 // enough to point to the next possible object header (the 2739 // bitmap knows by how much we need to move it as it knows its 2740 // granularity). 2741 assert(_finger < _region_limit, "invariant"); 2742 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger); 2743 // Check if bitmap iteration was aborted while scanning the last object 2744 if (new_finger >= _region_limit) { 2745 giveup_current_region(); 2746 } else { 2747 move_finger_to(new_finger); 2748 } 2749 } 2750 } 2751 // At this point we have either completed iterating over the 2752 // region we were holding on to, or we have aborted. 2753 2754 // We then partially drain the local queue and the global stack. 2755 // (Do we really need this?) 2756 drain_local_queue(true); 2757 drain_global_stack(true); 2758 2759 // Read the note on the claim_region() method on why it might 2760 // return NULL with potentially more regions available for 2761 // claiming and why we have to check out_of_regions() to determine 2762 // whether we're done or not. 2763 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 2764 // We are going to try to claim a new region. We should have 2765 // given up on the previous one. 2766 // Separated the asserts so that we know which one fires. 2767 assert(_curr_region == NULL, "invariant"); 2768 assert(_finger == NULL, "invariant"); 2769 assert(_region_limit == NULL, "invariant"); 2770 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 2771 if (claimed_region != NULL) { 2772 // Yes, we managed to claim one 2773 setup_for_region(claimed_region); 2774 assert(_curr_region == claimed_region, "invariant"); 2775 } 2776 // It is important to call the regular clock here. It might take 2777 // a while to claim a region if, for example, we hit a large 2778 // block of empty regions. So we need to call the regular clock 2779 // method once round the loop to make sure it's called 2780 // frequently enough. 2781 regular_clock_call(); 2782 } 2783 2784 if (!has_aborted() && _curr_region == NULL) { 2785 assert(_cm->out_of_regions(), 2786 "at this point we should be out of regions"); 2787 } 2788 } while ( _curr_region != NULL && !has_aborted()); 2789 2790 if (!has_aborted()) { 2791 // We cannot check whether the global stack is empty, since other 2792 // tasks might be pushing objects to it concurrently. 2793 assert(_cm->out_of_regions(), 2794 "at this point we should be out of regions"); 2795 // Try to reduce the number of available SATB buffers so that 2796 // remark has less work to do. 2797 drain_satb_buffers(); 2798 } 2799 2800 // Since we've done everything else, we can now totally drain the 2801 // local queue and global stack. 2802 drain_local_queue(false); 2803 drain_global_stack(false); 2804 2805 // Attempt at work stealing from other task's queues. 2806 if (do_stealing && !has_aborted()) { 2807 // We have not aborted. This means that we have finished all that 2808 // we could. Let's try to do some stealing... 2809 2810 // We cannot check whether the global stack is empty, since other 2811 // tasks might be pushing objects to it concurrently. 2812 assert(_cm->out_of_regions() && _task_queue->size() == 0, 2813 "only way to reach here"); 2814 while (!has_aborted()) { 2815 oop obj; 2816 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) { 2817 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), 2818 "any stolen object should be marked"); 2819 scan_object(obj); 2820 2821 // And since we're towards the end, let's totally drain the 2822 // local queue and global stack. 2823 drain_local_queue(false); 2824 drain_global_stack(false); 2825 } else { 2826 break; 2827 } 2828 } 2829 } 2830 2831 // We still haven't aborted. Now, let's try to get into the 2832 // termination protocol. 2833 if (do_termination && !has_aborted()) { 2834 // We cannot check whether the global stack is empty, since other 2835 // tasks might be concurrently pushing objects on it. 2836 // Separated the asserts so that we know which one fires. 2837 assert(_cm->out_of_regions(), "only way to reach here"); 2838 assert(_task_queue->size() == 0, "only way to reach here"); 2839 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 2840 2841 // The G1CMTask class also extends the TerminatorTerminator class, 2842 // hence its should_exit_termination() method will also decide 2843 // whether to exit the termination protocol or not. 2844 bool finished = (is_serial || 2845 _cm->terminator()->offer_termination(this)); 2846 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 2847 _termination_time_ms += 2848 termination_end_time_ms - _termination_start_time_ms; 2849 2850 if (finished) { 2851 // We're all done. 2852 2853 if (_worker_id == 0) { 2854 // let's allow task 0 to do this 2855 if (concurrent()) { 2856 assert(_cm->concurrent_marking_in_progress(), "invariant"); 2857 // we need to set this to false before the next 2858 // safepoint. This way we ensure that the marking phase 2859 // doesn't observe any more heap expansions. 2860 _cm->clear_concurrent_marking_in_progress(); 2861 } 2862 } 2863 2864 // We can now guarantee that the global stack is empty, since 2865 // all other tasks have finished. We separated the guarantees so 2866 // that, if a condition is false, we can immediately find out 2867 // which one. 2868 guarantee(_cm->out_of_regions(), "only way to reach here"); 2869 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 2870 guarantee(_task_queue->size() == 0, "only way to reach here"); 2871 guarantee(!_cm->has_overflown(), "only way to reach here"); 2872 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); 2873 } else { 2874 // Apparently there's more work to do. Let's abort this task. It 2875 // will restart it and we can hopefully find more things to do. 2876 set_has_aborted(); 2877 } 2878 } 2879 2880 // Mainly for debugging purposes to make sure that a pointer to the 2881 // closure which was statically allocated in this frame doesn't 2882 // escape it by accident. 2883 set_cm_oop_closure(NULL); 2884 double end_time_ms = os::elapsedVTime() * 1000.0; 2885 double elapsed_time_ms = end_time_ms - _start_time_ms; 2886 // Update the step history. 2887 _step_times_ms.add(elapsed_time_ms); 2888 2889 if (has_aborted()) { 2890 // The task was aborted for some reason. 2891 if (_has_timed_out) { 2892 double diff_ms = elapsed_time_ms - _time_target_ms; 2893 // Keep statistics of how well we did with respect to hitting 2894 // our target only if we actually timed out (if we aborted for 2895 // other reasons, then the results might get skewed). 2896 _marking_step_diffs_ms.add(diff_ms); 2897 } 2898 2899 if (_cm->has_overflown()) { 2900 // This is the interesting one. We aborted because a global 2901 // overflow was raised. This means we have to restart the 2902 // marking phase and start iterating over regions. However, in 2903 // order to do this we have to make sure that all tasks stop 2904 // what they are doing and re-initialize in a safe manner. We 2905 // will achieve this with the use of two barrier sync points. 2906 2907 if (!is_serial) { 2908 // We only need to enter the sync barrier if being called 2909 // from a parallel context 2910 _cm->enter_first_sync_barrier(_worker_id); 2911 2912 // When we exit this sync barrier we know that all tasks have 2913 // stopped doing marking work. So, it's now safe to 2914 // re-initialize our data structures. At the end of this method, 2915 // task 0 will clear the global data structures. 2916 } 2917 2918 // We clear the local state of this task... 2919 clear_region_fields(); 2920 2921 if (!is_serial) { 2922 // ...and enter the second barrier. 2923 _cm->enter_second_sync_barrier(_worker_id); 2924 } 2925 // At this point, if we're during the concurrent phase of 2926 // marking, everything has been re-initialized and we're 2927 // ready to restart. 2928 } 2929 } 2930 2931 _claimed = false; 2932 } 2933 2934 G1CMTask::G1CMTask(uint worker_id, 2935 G1ConcurrentMark* cm, 2936 G1CMTaskQueue* task_queue, 2937 G1CMTaskQueueSet* task_queues) 2938 : _g1h(G1CollectedHeap::heap()), 2939 _worker_id(worker_id), _cm(cm), 2940 _claimed(false), 2941 _nextMarkBitMap(NULL), _hash_seed(17), 2942 _task_queue(task_queue), 2943 _task_queues(task_queues), 2944 _cm_oop_closure(NULL) { 2945 guarantee(task_queue != NULL, "invariant"); 2946 guarantee(task_queues != NULL, "invariant"); 2947 2948 _marking_step_diffs_ms.add(0.5); 2949 } 2950 2951 // These are formatting macros that are used below to ensure 2952 // consistent formatting. The *_H_* versions are used to format the 2953 // header for a particular value and they should be kept consistent 2954 // with the corresponding macro. Also note that most of the macros add 2955 // the necessary white space (as a prefix) which makes them a bit 2956 // easier to compose. 2957 2958 // All the output lines are prefixed with this string to be able to 2959 // identify them easily in a large log file. 2960 #define G1PPRL_LINE_PREFIX "###" 2961 2962 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT 2963 #ifdef _LP64 2964 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 2965 #else // _LP64 2966 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 2967 #endif // _LP64 2968 2969 // For per-region info 2970 #define G1PPRL_TYPE_FORMAT " %-4s" 2971 #define G1PPRL_TYPE_H_FORMAT " %4s" 2972 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9) 2973 #define G1PPRL_BYTE_H_FORMAT " %9s" 2974 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 2975 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 2976 2977 // For summary info 2978 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT 2979 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT 2980 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB" 2981 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%" 2982 2983 G1PrintRegionLivenessInfoClosure:: 2984 G1PrintRegionLivenessInfoClosure(const char* phase_name) 2985 : _total_used_bytes(0), _total_capacity_bytes(0), 2986 _total_prev_live_bytes(0), _total_next_live_bytes(0), 2987 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) { 2988 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2989 MemRegion g1_reserved = g1h->g1_reserved(); 2990 double now = os::elapsedTime(); 2991 2992 // Print the header of the output. 2993 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 2994 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP" 2995 G1PPRL_SUM_ADDR_FORMAT("reserved") 2996 G1PPRL_SUM_BYTE_FORMAT("region-size"), 2997 p2i(g1_reserved.start()), p2i(g1_reserved.end()), 2998 HeapRegion::GrainBytes); 2999 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 3000 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3001 G1PPRL_TYPE_H_FORMAT 3002 G1PPRL_ADDR_BASE_H_FORMAT 3003 G1PPRL_BYTE_H_FORMAT 3004 G1PPRL_BYTE_H_FORMAT 3005 G1PPRL_BYTE_H_FORMAT 3006 G1PPRL_DOUBLE_H_FORMAT 3007 G1PPRL_BYTE_H_FORMAT 3008 G1PPRL_BYTE_H_FORMAT, 3009 "type", "address-range", 3010 "used", "prev-live", "next-live", "gc-eff", 3011 "remset", "code-roots"); 3012 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3013 G1PPRL_TYPE_H_FORMAT 3014 G1PPRL_ADDR_BASE_H_FORMAT 3015 G1PPRL_BYTE_H_FORMAT 3016 G1PPRL_BYTE_H_FORMAT 3017 G1PPRL_BYTE_H_FORMAT 3018 G1PPRL_DOUBLE_H_FORMAT 3019 G1PPRL_BYTE_H_FORMAT 3020 G1PPRL_BYTE_H_FORMAT, 3021 "", "", 3022 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", 3023 "(bytes)", "(bytes)"); 3024 } 3025 3026 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) { 3027 const char* type = r->get_type_str(); 3028 HeapWord* bottom = r->bottom(); 3029 HeapWord* end = r->end(); 3030 size_t capacity_bytes = r->capacity(); 3031 size_t used_bytes = r->used(); 3032 size_t prev_live_bytes = r->live_bytes(); 3033 size_t next_live_bytes = r->next_live_bytes(); 3034 double gc_eff = r->gc_efficiency(); 3035 size_t remset_bytes = r->rem_set()->mem_size(); 3036 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size(); 3037 3038 _total_used_bytes += used_bytes; 3039 _total_capacity_bytes += capacity_bytes; 3040 _total_prev_live_bytes += prev_live_bytes; 3041 _total_next_live_bytes += next_live_bytes; 3042 _total_remset_bytes += remset_bytes; 3043 _total_strong_code_roots_bytes += strong_code_roots_bytes; 3044 3045 // Print a line for this particular region. 3046 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3047 G1PPRL_TYPE_FORMAT 3048 G1PPRL_ADDR_BASE_FORMAT 3049 G1PPRL_BYTE_FORMAT 3050 G1PPRL_BYTE_FORMAT 3051 G1PPRL_BYTE_FORMAT 3052 G1PPRL_DOUBLE_FORMAT 3053 G1PPRL_BYTE_FORMAT 3054 G1PPRL_BYTE_FORMAT, 3055 type, p2i(bottom), p2i(end), 3056 used_bytes, prev_live_bytes, next_live_bytes, gc_eff, 3057 remset_bytes, strong_code_roots_bytes); 3058 3059 return false; 3060 } 3061 3062 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 3063 // add static memory usages to remembered set sizes 3064 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size(); 3065 // Print the footer of the output. 3066 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 3067 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3068 " SUMMARY" 3069 G1PPRL_SUM_MB_FORMAT("capacity") 3070 G1PPRL_SUM_MB_PERC_FORMAT("used") 3071 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 3072 G1PPRL_SUM_MB_PERC_FORMAT("next-live") 3073 G1PPRL_SUM_MB_FORMAT("remset") 3074 G1PPRL_SUM_MB_FORMAT("code-roots"), 3075 bytes_to_mb(_total_capacity_bytes), 3076 bytes_to_mb(_total_used_bytes), 3077 perc(_total_used_bytes, _total_capacity_bytes), 3078 bytes_to_mb(_total_prev_live_bytes), 3079 perc(_total_prev_live_bytes, _total_capacity_bytes), 3080 bytes_to_mb(_total_next_live_bytes), 3081 perc(_total_next_live_bytes, _total_capacity_bytes), 3082 bytes_to_mb(_total_remset_bytes), 3083 bytes_to_mb(_total_strong_code_roots_bytes)); 3084 }