1 /* 2 * Copyright (c) 2001, 2015, 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/concurrentMark.inline.hpp" 30 #include "gc/g1/concurrentMarkThread.inline.hpp" 31 #include "gc/g1/g1CollectedHeap.inline.hpp" 32 #include "gc/g1/g1CollectorPolicy.hpp" 33 #include "gc/g1/g1CollectorState.hpp" 34 #include "gc/g1/g1OopClosures.inline.hpp" 35 #include "gc/g1/g1RemSet.hpp" 36 #include "gc/g1/g1StringDedup.hpp" 37 #include "gc/g1/heapRegion.inline.hpp" 38 #include "gc/g1/heapRegionManager.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 61 // Concurrent marking bit map wrapper 62 63 CMBitMapRO::CMBitMapRO(int shifter) : 64 _bm(), 65 _shifter(shifter) { 66 _bmStartWord = 0; 67 _bmWordSize = 0; 68 } 69 70 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr, 71 const HeapWord* limit) const { 72 // First we must round addr *up* to a possible object boundary. 73 addr = (HeapWord*)align_size_up((intptr_t)addr, 74 HeapWordSize << _shifter); 75 size_t addrOffset = heapWordToOffset(addr); 76 assert(limit != NULL, "limit must not be NULL"); 77 size_t limitOffset = heapWordToOffset(limit); 78 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); 79 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 80 assert(nextAddr >= addr, "get_next_one postcondition"); 81 assert(nextAddr == limit || isMarked(nextAddr), 82 "get_next_one postcondition"); 83 return nextAddr; 84 } 85 86 #ifndef PRODUCT 87 bool CMBitMapRO::covers(MemRegion heap_rs) const { 88 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 89 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize, 90 "size inconsistency"); 91 return _bmStartWord == (HeapWord*)(heap_rs.start()) && 92 _bmWordSize == heap_rs.word_size(); 93 } 94 #endif 95 96 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const { 97 _bm.print_on_error(st, prefix); 98 } 99 100 size_t CMBitMap::compute_size(size_t heap_size) { 101 return ReservedSpace::allocation_align_size_up(heap_size / mark_distance()); 102 } 103 104 size_t CMBitMap::mark_distance() { 105 return MinObjAlignmentInBytes * BitsPerByte; 106 } 107 108 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) { 109 _bmStartWord = heap.start(); 110 _bmWordSize = heap.word_size(); 111 112 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start()); 113 _bm.set_size(_bmWordSize >> _shifter); 114 115 storage->set_mapping_changed_listener(&_listener); 116 } 117 118 void CMBitMapMappingChangedListener::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->clearRange(mr); 125 } 126 127 // Closure used for clearing the given mark bitmap. 128 class ClearBitmapHRClosure : public HeapRegionClosure { 129 private: 130 ConcurrentMark* _cm; 131 CMBitMap* _bitmap; 132 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration. 133 public: 134 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) { 135 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield."); 136 } 137 138 virtual bool doHeapRegion(HeapRegion* r) { 139 size_t const chunk_size_in_words = M / HeapWordSize; 140 141 HeapWord* cur = r->bottom(); 142 HeapWord* const end = r->end(); 143 144 while (cur < end) { 145 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end)); 146 _bitmap->clearRange(mr); 147 148 cur += chunk_size_in_words; 149 150 // Abort iteration if after yielding the marking has been aborted. 151 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) { 152 return true; 153 } 154 // Repeat the asserts from before the start of the closure. We will do them 155 // as asserts here to minimize their overhead on the product. However, we 156 // will have them as guarantees at the beginning / end of the bitmap 157 // clearing to get some checking in the product. 158 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant"); 159 assert(!_may_yield || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant"); 160 } 161 162 return false; 163 } 164 }; 165 166 class ParClearNextMarkBitmapTask : public AbstractGangTask { 167 ClearBitmapHRClosure* _cl; 168 HeapRegionClaimer _hrclaimer; 169 bool _suspendible; // If the task is suspendible, workers must join the STS. 170 171 public: 172 ParClearNextMarkBitmapTask(ClearBitmapHRClosure *cl, uint n_workers, bool suspendible) : 173 _cl(cl), _suspendible(suspendible), AbstractGangTask("Parallel Clear Bitmap Task"), _hrclaimer(n_workers) {} 174 175 void work(uint worker_id) { 176 SuspendibleThreadSetJoiner sts_join(_suspendible); 177 G1CollectedHeap::heap()->heap_region_par_iterate(_cl, worker_id, &_hrclaimer, true); 178 } 179 }; 180 181 void CMBitMap::clearAll() { 182 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 183 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */); 184 uint n_workers = g1h->workers()->active_workers(); 185 ParClearNextMarkBitmapTask task(&cl, n_workers, false); 186 g1h->workers()->run_task(&task); 187 guarantee(cl.complete(), "Must have completed iteration."); 188 return; 189 } 190 191 void CMBitMap::clearRange(MemRegion mr) { 192 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 193 assert(!mr.is_empty(), "unexpected empty region"); 194 // convert address range into offset range 195 _bm.at_put_range(heapWordToOffset(mr.start()), 196 heapWordToOffset(mr.end()), false); 197 } 198 199 CMMarkStack::CMMarkStack(ConcurrentMark* cm) : 200 _base(NULL), _cm(cm) 201 {} 202 203 bool CMMarkStack::allocate(size_t capacity) { 204 // allocate a stack of the requisite depth 205 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop))); 206 if (!rs.is_reserved()) { 207 warning("ConcurrentMark MarkStack allocation failure"); 208 return false; 209 } 210 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 211 if (!_virtual_space.initialize(rs, rs.size())) { 212 warning("ConcurrentMark MarkStack backing store failure"); 213 // Release the virtual memory reserved for the marking stack 214 rs.release(); 215 return false; 216 } 217 assert(_virtual_space.committed_size() == rs.size(), 218 "Didn't reserve backing store for all of ConcurrentMark stack?"); 219 _base = (oop*) _virtual_space.low(); 220 setEmpty(); 221 _capacity = (jint) capacity; 222 _saved_index = -1; 223 _should_expand = false; 224 return true; 225 } 226 227 void CMMarkStack::expand() { 228 // Called, during remark, if we've overflown the marking stack during marking. 229 assert(isEmpty(), "stack should been emptied while handling overflow"); 230 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted"); 231 // Clear expansion flag 232 _should_expand = false; 233 if (_capacity == (jint) MarkStackSizeMax) { 234 log_trace(gc)("(benign) Can't expand marking stack capacity, at max size limit"); 235 return; 236 } 237 // Double capacity if possible 238 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax); 239 // Do not give up existing stack until we have managed to 240 // get the double capacity that we desired. 241 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity * 242 sizeof(oop))); 243 if (rs.is_reserved()) { 244 // Release the backing store associated with old stack 245 _virtual_space.release(); 246 // Reinitialize virtual space for new stack 247 if (!_virtual_space.initialize(rs, rs.size())) { 248 fatal("Not enough swap for expanded marking stack capacity"); 249 } 250 _base = (oop*)(_virtual_space.low()); 251 _index = 0; 252 _capacity = new_capacity; 253 } else { 254 // Failed to double capacity, continue; 255 log_trace(gc)("(benign) Failed to expand marking stack capacity from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 256 _capacity / K, new_capacity / K); 257 } 258 } 259 260 void CMMarkStack::set_should_expand() { 261 // If we're resetting the marking state because of an 262 // marking stack overflow, record that we should, if 263 // possible, expand the stack. 264 _should_expand = _cm->has_overflown(); 265 } 266 267 CMMarkStack::~CMMarkStack() { 268 if (_base != NULL) { 269 _base = NULL; 270 _virtual_space.release(); 271 } 272 } 273 274 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { 275 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 276 jint start = _index; 277 jint next_index = start + n; 278 if (next_index > _capacity) { 279 _overflow = true; 280 return; 281 } 282 // Otherwise. 283 _index = next_index; 284 for (int i = 0; i < n; i++) { 285 int ind = start + i; 286 assert(ind < _capacity, "By overflow test above."); 287 _base[ind] = ptr_arr[i]; 288 } 289 } 290 291 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { 292 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 293 jint index = _index; 294 if (index == 0) { 295 *n = 0; 296 return false; 297 } else { 298 int k = MIN2(max, index); 299 jint new_ind = index - k; 300 for (int j = 0; j < k; j++) { 301 ptr_arr[j] = _base[new_ind + j]; 302 } 303 _index = new_ind; 304 *n = k; 305 return true; 306 } 307 } 308 309 void CMMarkStack::note_start_of_gc() { 310 assert(_saved_index == -1, 311 "note_start_of_gc()/end_of_gc() bracketed incorrectly"); 312 _saved_index = _index; 313 } 314 315 void CMMarkStack::note_end_of_gc() { 316 // This is intentionally a guarantee, instead of an assert. If we 317 // accidentally add something to the mark stack during GC, it 318 // will be a correctness issue so it's better if we crash. we'll 319 // only check this once per GC anyway, so it won't be a performance 320 // issue in any way. 321 guarantee(_saved_index == _index, 322 "saved index: %d index: %d", _saved_index, _index); 323 _saved_index = -1; 324 } 325 326 CMRootRegions::CMRootRegions() : 327 _young_list(NULL), _cm(NULL), _scan_in_progress(false), 328 _should_abort(false), _next_survivor(NULL) { } 329 330 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) { 331 _young_list = g1h->young_list(); 332 _cm = cm; 333 } 334 335 void CMRootRegions::prepare_for_scan() { 336 assert(!scan_in_progress(), "pre-condition"); 337 338 // Currently, only survivors can be root regions. 339 assert(_next_survivor == NULL, "pre-condition"); 340 _next_survivor = _young_list->first_survivor_region(); 341 _scan_in_progress = (_next_survivor != NULL); 342 _should_abort = false; 343 } 344 345 HeapRegion* CMRootRegions::claim_next() { 346 if (_should_abort) { 347 // If someone has set the should_abort flag, we return NULL to 348 // force the caller to bail out of their loop. 349 return NULL; 350 } 351 352 // Currently, only survivors can be root regions. 353 HeapRegion* res = _next_survivor; 354 if (res != NULL) { 355 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 356 // Read it again in case it changed while we were waiting for the lock. 357 res = _next_survivor; 358 if (res != NULL) { 359 if (res == _young_list->last_survivor_region()) { 360 // We just claimed the last survivor so store NULL to indicate 361 // that we're done. 362 _next_survivor = NULL; 363 } else { 364 _next_survivor = res->get_next_young_region(); 365 } 366 } else { 367 // Someone else claimed the last survivor while we were trying 368 // to take the lock so nothing else to do. 369 } 370 } 371 assert(res == NULL || res->is_survivor(), "post-condition"); 372 373 return res; 374 } 375 376 void CMRootRegions::scan_finished() { 377 assert(scan_in_progress(), "pre-condition"); 378 379 // Currently, only survivors can be root regions. 380 if (!_should_abort) { 381 assert(_next_survivor == NULL, "we should have claimed all survivors"); 382 } 383 _next_survivor = NULL; 384 385 { 386 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 387 _scan_in_progress = false; 388 RootRegionScan_lock->notify_all(); 389 } 390 } 391 392 bool CMRootRegions::wait_until_scan_finished() { 393 if (!scan_in_progress()) return false; 394 395 { 396 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 397 while (scan_in_progress()) { 398 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 399 } 400 } 401 return true; 402 } 403 404 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 405 return MAX2((n_par_threads + 2) / 4, 1U); 406 } 407 408 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) : 409 _g1h(g1h), 410 _markBitMap1(), 411 _markBitMap2(), 412 _parallel_marking_threads(0), 413 _max_parallel_marking_threads(0), 414 _sleep_factor(0.0), 415 _marking_task_overhead(1.0), 416 _cleanup_list("Cleanup List"), 417 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/), 418 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >> 419 CardTableModRefBS::card_shift, 420 false /* in_resource_area*/), 421 422 _prevMarkBitMap(&_markBitMap1), 423 _nextMarkBitMap(&_markBitMap2), 424 425 _markStack(this), 426 // _finger set in set_non_marking_state 427 428 _max_worker_id(ParallelGCThreads), 429 // _active_tasks set in set_non_marking_state 430 // _tasks set inside the constructor 431 _task_queues(new CMTaskQueueSet((int) _max_worker_id)), 432 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)), 433 434 _has_overflown(false), 435 _concurrent(false), 436 _has_aborted(false), 437 _restart_for_overflow(false), 438 _concurrent_marking_in_progress(false), 439 _concurrent_phase_started(false), 440 441 // _verbose_level set below 442 443 _init_times(), 444 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 445 _cleanup_times(), 446 _total_counting_time(0.0), 447 _total_rs_scrub_time(0.0), 448 449 _parallel_workers(NULL), 450 451 _count_card_bitmaps(NULL), 452 _count_marked_bytes(NULL), 453 _completed_initialization(false) { 454 455 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage); 456 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage); 457 458 // Create & start a ConcurrentMark thread. 459 _cmThread = new ConcurrentMarkThread(this); 460 assert(cmThread() != NULL, "CM Thread should have been created"); 461 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 462 if (_cmThread->osthread() == NULL) { 463 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread"); 464 } 465 466 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 467 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency"); 468 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency"); 469 470 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 471 satb_qs.set_buffer_size(G1SATBBufferSize); 472 473 _root_regions.init(_g1h, this); 474 475 if (ConcGCThreads > ParallelGCThreads) { 476 warning("Can't have more ConcGCThreads (%u) " 477 "than ParallelGCThreads (%u).", 478 ConcGCThreads, ParallelGCThreads); 479 return; 480 } 481 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 482 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 483 // if both are set 484 _sleep_factor = 0.0; 485 _marking_task_overhead = 1.0; 486 } else if (G1MarkingOverheadPercent > 0) { 487 // We will calculate the number of parallel marking threads based 488 // on a target overhead with respect to the soft real-time goal 489 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 490 double overall_cm_overhead = 491 (double) MaxGCPauseMillis * marking_overhead / 492 (double) GCPauseIntervalMillis; 493 double cpu_ratio = 1.0 / (double) os::processor_count(); 494 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 495 double marking_task_overhead = 496 overall_cm_overhead / marking_thread_num * 497 (double) os::processor_count(); 498 double sleep_factor = 499 (1.0 - marking_task_overhead) / marking_task_overhead; 500 501 FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num); 502 _sleep_factor = sleep_factor; 503 _marking_task_overhead = marking_task_overhead; 504 } else { 505 // Calculate the number of parallel marking threads by scaling 506 // the number of parallel GC threads. 507 uint marking_thread_num = scale_parallel_threads(ParallelGCThreads); 508 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num); 509 _sleep_factor = 0.0; 510 _marking_task_overhead = 1.0; 511 } 512 513 assert(ConcGCThreads > 0, "Should have been set"); 514 _parallel_marking_threads = ConcGCThreads; 515 _max_parallel_marking_threads = _parallel_marking_threads; 516 517 _parallel_workers = new WorkGang("G1 Marker", 518 _max_parallel_marking_threads, false, true); 519 if (_parallel_workers == NULL) { 520 vm_exit_during_initialization("Failed necessary allocation."); 521 } else { 522 _parallel_workers->initialize_workers(); 523 } 524 525 if (FLAG_IS_DEFAULT(MarkStackSize)) { 526 size_t mark_stack_size = 527 MIN2(MarkStackSizeMax, 528 MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE))); 529 // Verify that the calculated value for MarkStackSize is in range. 530 // It would be nice to use the private utility routine from Arguments. 531 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 532 warning("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): " 533 "must be between 1 and " SIZE_FORMAT, 534 mark_stack_size, MarkStackSizeMax); 535 return; 536 } 537 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size); 538 } else { 539 // Verify MarkStackSize is in range. 540 if (FLAG_IS_CMDLINE(MarkStackSize)) { 541 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 542 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 543 warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): " 544 "must be between 1 and " SIZE_FORMAT, 545 MarkStackSize, MarkStackSizeMax); 546 return; 547 } 548 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 549 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 550 warning("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")" 551 " or for MarkStackSizeMax (" SIZE_FORMAT ")", 552 MarkStackSize, MarkStackSizeMax); 553 return; 554 } 555 } 556 } 557 } 558 559 if (!_markStack.allocate(MarkStackSize)) { 560 warning("Failed to allocate CM marking stack"); 561 return; 562 } 563 564 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC); 565 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC); 566 567 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC); 568 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC); 569 570 BitMap::idx_t card_bm_size = _card_bm.size(); 571 572 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 573 _active_tasks = _max_worker_id; 574 575 uint max_regions = _g1h->max_regions(); 576 for (uint i = 0; i < _max_worker_id; ++i) { 577 CMTaskQueue* task_queue = new CMTaskQueue(); 578 task_queue->initialize(); 579 _task_queues->register_queue(i, task_queue); 580 581 _count_card_bitmaps[i] = BitMap(card_bm_size, false); 582 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC); 583 584 _tasks[i] = new CMTask(i, this, 585 _count_marked_bytes[i], 586 &_count_card_bitmaps[i], 587 task_queue, _task_queues); 588 589 _accum_task_vtime[i] = 0.0; 590 } 591 592 // Calculate the card number for the bottom of the heap. Used 593 // in biasing indexes into the accounting card bitmaps. 594 _heap_bottom_card_num = 595 intptr_t(uintptr_t(_g1h->reserved_region().start()) >> 596 CardTableModRefBS::card_shift); 597 598 // Clear all the liveness counting data 599 clear_all_count_data(); 600 601 // so that the call below can read a sensible value 602 _heap_start = g1h->reserved_region().start(); 603 set_non_marking_state(); 604 _completed_initialization = true; 605 } 606 607 void ConcurrentMark::reset() { 608 // Starting values for these two. This should be called in a STW 609 // phase. 610 MemRegion reserved = _g1h->g1_reserved(); 611 _heap_start = reserved.start(); 612 _heap_end = reserved.end(); 613 614 // Separated the asserts so that we know which one fires. 615 assert(_heap_start != NULL, "heap bounds should look ok"); 616 assert(_heap_end != NULL, "heap bounds should look ok"); 617 assert(_heap_start < _heap_end, "heap bounds should look ok"); 618 619 // Reset all the marking data structures and any necessary flags 620 reset_marking_state(); 621 622 // We do reset all of them, since different phases will use 623 // different number of active threads. So, it's easiest to have all 624 // of them ready. 625 for (uint i = 0; i < _max_worker_id; ++i) { 626 _tasks[i]->reset(_nextMarkBitMap); 627 } 628 629 // we need this to make sure that the flag is on during the evac 630 // pause with initial mark piggy-backed 631 set_concurrent_marking_in_progress(); 632 } 633 634 635 void ConcurrentMark::reset_marking_state(bool clear_overflow) { 636 _markStack.set_should_expand(); 637 _markStack.setEmpty(); // Also clears the _markStack overflow flag 638 if (clear_overflow) { 639 clear_has_overflown(); 640 } else { 641 assert(has_overflown(), "pre-condition"); 642 } 643 _finger = _heap_start; 644 645 for (uint i = 0; i < _max_worker_id; ++i) { 646 CMTaskQueue* queue = _task_queues->queue(i); 647 queue->set_empty(); 648 } 649 } 650 651 void ConcurrentMark::set_concurrency(uint active_tasks) { 652 assert(active_tasks <= _max_worker_id, "we should not have more"); 653 654 _active_tasks = active_tasks; 655 // Need to update the three data structures below according to the 656 // number of active threads for this phase. 657 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 658 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 659 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 660 } 661 662 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 663 set_concurrency(active_tasks); 664 665 _concurrent = concurrent; 666 // We propagate this to all tasks, not just the active ones. 667 for (uint i = 0; i < _max_worker_id; ++i) 668 _tasks[i]->set_concurrent(concurrent); 669 670 if (concurrent) { 671 set_concurrent_marking_in_progress(); 672 } else { 673 // We currently assume that the concurrent flag has been set to 674 // false before we start remark. At this point we should also be 675 // in a STW phase. 676 assert(!concurrent_marking_in_progress(), "invariant"); 677 assert(out_of_regions(), 678 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT, 679 p2i(_finger), p2i(_heap_end)); 680 } 681 } 682 683 void ConcurrentMark::set_non_marking_state() { 684 // We set the global marking state to some default values when we're 685 // not doing marking. 686 reset_marking_state(); 687 _active_tasks = 0; 688 clear_concurrent_marking_in_progress(); 689 } 690 691 ConcurrentMark::~ConcurrentMark() { 692 // The ConcurrentMark instance is never freed. 693 ShouldNotReachHere(); 694 } 695 696 void ConcurrentMark::clearNextBitmap() { 697 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 698 699 // Make sure that the concurrent mark thread looks to still be in 700 // the current cycle. 701 guarantee(cmThread()->during_cycle(), "invariant"); 702 703 // We are finishing up the current cycle by clearing the next 704 // marking bitmap and getting it ready for the next cycle. During 705 // this time no other cycle can start. So, let's make sure that this 706 // is the case. 707 guarantee(!g1h->collector_state()->mark_in_progress(), "invariant"); 708 709 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */); 710 ParClearNextMarkBitmapTask task(&cl, parallel_marking_threads(), true); 711 _parallel_workers->run_task(&task); 712 713 // Clear the liveness counting data. If the marking has been aborted, the abort() 714 // call already did that. 715 if (cl.complete()) { 716 clear_all_count_data(); 717 } 718 719 // Repeat the asserts from above. 720 guarantee(cmThread()->during_cycle(), "invariant"); 721 guarantee(!g1h->collector_state()->mark_in_progress(), "invariant"); 722 } 723 724 class CheckBitmapClearHRClosure : public HeapRegionClosure { 725 CMBitMap* _bitmap; 726 bool _error; 727 public: 728 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) { 729 } 730 731 virtual bool doHeapRegion(HeapRegion* r) { 732 // This closure can be called concurrently to the mutator, so we must make sure 733 // that the result of the getNextMarkedWordAddress() call is compared to the 734 // value passed to it as limit to detect any found bits. 735 // end never changes in G1. 736 HeapWord* end = r->end(); 737 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end; 738 } 739 }; 740 741 bool ConcurrentMark::nextMarkBitmapIsClear() { 742 CheckBitmapClearHRClosure cl(_nextMarkBitMap); 743 _g1h->heap_region_iterate(&cl); 744 return cl.complete(); 745 } 746 747 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 748 public: 749 bool doHeapRegion(HeapRegion* r) { 750 r->note_start_of_marking(); 751 return false; 752 } 753 }; 754 755 void ConcurrentMark::checkpointRootsInitialPre() { 756 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 757 G1CollectorPolicy* g1p = g1h->g1_policy(); 758 759 _has_aborted = false; 760 761 // Initialize marking structures. This has to be done in a STW phase. 762 reset(); 763 764 // For each region note start of marking. 765 NoteStartOfMarkHRClosure startcl; 766 g1h->heap_region_iterate(&startcl); 767 } 768 769 770 void ConcurrentMark::checkpointRootsInitialPost() { 771 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 772 773 // Start Concurrent Marking weak-reference discovery. 774 ReferenceProcessor* rp = g1h->ref_processor_cm(); 775 // enable ("weak") refs discovery 776 rp->enable_discovery(); 777 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 778 779 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 780 // This is the start of the marking cycle, we're expected all 781 // threads to have SATB queues with active set to false. 782 satb_mq_set.set_active_all_threads(true, /* new active value */ 783 false /* expected_active */); 784 785 _root_regions.prepare_for_scan(); 786 787 // update_g1_committed() will be called at the end of an evac pause 788 // when marking is on. So, it's also called at the end of the 789 // initial-mark pause to update the heap end, if the heap expands 790 // during it. No need to call it here. 791 } 792 793 /* 794 * Notice that in the next two methods, we actually leave the STS 795 * during the barrier sync and join it immediately afterwards. If we 796 * do not do this, the following deadlock can occur: one thread could 797 * be in the barrier sync code, waiting for the other thread to also 798 * sync up, whereas another one could be trying to yield, while also 799 * waiting for the other threads to sync up too. 800 * 801 * Note, however, that this code is also used during remark and in 802 * this case we should not attempt to leave / enter the STS, otherwise 803 * we'll either hit an assert (debug / fastdebug) or deadlock 804 * (product). So we should only leave / enter the STS if we are 805 * operating concurrently. 806 * 807 * Because the thread that does the sync barrier has left the STS, it 808 * is possible to be suspended for a Full GC or an evacuation pause 809 * could occur. This is actually safe, since the entering the sync 810 * barrier is one of the last things do_marking_step() does, and it 811 * doesn't manipulate any data structures afterwards. 812 */ 813 814 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 815 bool barrier_aborted; 816 { 817 SuspendibleThreadSetLeaver sts_leave(concurrent()); 818 barrier_aborted = !_first_overflow_barrier_sync.enter(); 819 } 820 821 // at this point everyone should have synced up and not be doing any 822 // more work 823 824 if (barrier_aborted) { 825 // If the barrier aborted we ignore the overflow condition and 826 // just abort the whole marking phase as quickly as possible. 827 return; 828 } 829 830 // If we're executing the concurrent phase of marking, reset the marking 831 // state; otherwise the marking state is reset after reference processing, 832 // during the remark pause. 833 // If we reset here as a result of an overflow during the remark we will 834 // see assertion failures from any subsequent set_concurrency_and_phase() 835 // calls. 836 if (concurrent()) { 837 // let the task associated with with worker 0 do this 838 if (worker_id == 0) { 839 // task 0 is responsible for clearing the global data structures 840 // We should be here because of an overflow. During STW we should 841 // not clear the overflow flag since we rely on it being true when 842 // we exit this method to abort the pause and restart concurrent 843 // marking. 844 reset_marking_state(true /* clear_overflow */); 845 846 log_info(gc)("Concurrent Mark reset for overflow"); 847 } 848 } 849 850 // after this, each task should reset its own data structures then 851 // then go into the second barrier 852 } 853 854 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 855 SuspendibleThreadSetLeaver sts_leave(concurrent()); 856 _second_overflow_barrier_sync.enter(); 857 858 // at this point everything should be re-initialized and ready to go 859 } 860 861 class CMConcurrentMarkingTask: public AbstractGangTask { 862 private: 863 ConcurrentMark* _cm; 864 ConcurrentMarkThread* _cmt; 865 866 public: 867 void work(uint worker_id) { 868 assert(Thread::current()->is_ConcurrentGC_thread(), 869 "this should only be done by a conc GC thread"); 870 ResourceMark rm; 871 872 double start_vtime = os::elapsedVTime(); 873 874 { 875 SuspendibleThreadSetJoiner sts_join; 876 877 assert(worker_id < _cm->active_tasks(), "invariant"); 878 CMTask* the_task = _cm->task(worker_id); 879 the_task->record_start_time(); 880 if (!_cm->has_aborted()) { 881 do { 882 double start_vtime_sec = os::elapsedVTime(); 883 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 884 885 the_task->do_marking_step(mark_step_duration_ms, 886 true /* do_termination */, 887 false /* is_serial*/); 888 889 double end_vtime_sec = os::elapsedVTime(); 890 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 891 _cm->clear_has_overflown(); 892 893 _cm->do_yield_check(worker_id); 894 895 jlong sleep_time_ms; 896 if (!_cm->has_aborted() && the_task->has_aborted()) { 897 sleep_time_ms = 898 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 899 { 900 SuspendibleThreadSetLeaver sts_leave; 901 os::sleep(Thread::current(), sleep_time_ms, false); 902 } 903 } 904 } while (!_cm->has_aborted() && the_task->has_aborted()); 905 } 906 the_task->record_end_time(); 907 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 908 } 909 910 double end_vtime = os::elapsedVTime(); 911 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 912 } 913 914 CMConcurrentMarkingTask(ConcurrentMark* cm, 915 ConcurrentMarkThread* cmt) : 916 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 917 918 ~CMConcurrentMarkingTask() { } 919 }; 920 921 // Calculates the number of active workers for a concurrent 922 // phase. 923 uint ConcurrentMark::calc_parallel_marking_threads() { 924 uint n_conc_workers = 0; 925 if (!UseDynamicNumberOfGCThreads || 926 (!FLAG_IS_DEFAULT(ConcGCThreads) && 927 !ForceDynamicNumberOfGCThreads)) { 928 n_conc_workers = max_parallel_marking_threads(); 929 } else { 930 n_conc_workers = 931 AdaptiveSizePolicy::calc_default_active_workers( 932 max_parallel_marking_threads(), 933 1, /* Minimum workers */ 934 parallel_marking_threads(), 935 Threads::number_of_non_daemon_threads()); 936 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 937 // that scaling has already gone into "_max_parallel_marking_threads". 938 } 939 assert(n_conc_workers > 0, "Always need at least 1"); 940 return n_conc_workers; 941 } 942 943 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) { 944 // Currently, only survivors can be root regions. 945 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 946 G1RootRegionScanClosure cl(_g1h, this, worker_id); 947 948 const uintx interval = PrefetchScanIntervalInBytes; 949 HeapWord* curr = hr->bottom(); 950 const HeapWord* end = hr->top(); 951 while (curr < end) { 952 Prefetch::read(curr, interval); 953 oop obj = oop(curr); 954 int size = obj->oop_iterate_size(&cl); 955 assert(size == obj->size(), "sanity"); 956 curr += size; 957 } 958 } 959 960 class CMRootRegionScanTask : public AbstractGangTask { 961 private: 962 ConcurrentMark* _cm; 963 964 public: 965 CMRootRegionScanTask(ConcurrentMark* cm) : 966 AbstractGangTask("Root Region Scan"), _cm(cm) { } 967 968 void work(uint worker_id) { 969 assert(Thread::current()->is_ConcurrentGC_thread(), 970 "this should only be done by a conc GC thread"); 971 972 CMRootRegions* root_regions = _cm->root_regions(); 973 HeapRegion* hr = root_regions->claim_next(); 974 while (hr != NULL) { 975 _cm->scanRootRegion(hr, worker_id); 976 hr = root_regions->claim_next(); 977 } 978 } 979 }; 980 981 void ConcurrentMark::scanRootRegions() { 982 // Start of concurrent marking. 983 ClassLoaderDataGraph::clear_claimed_marks(); 984 985 // scan_in_progress() will have been set to true only if there was 986 // at least one root region to scan. So, if it's false, we 987 // should not attempt to do any further work. 988 if (root_regions()->scan_in_progress()) { 989 GCTraceConcTime(Info, gc) tt("Concurrent Root Region Scan"); 990 991 _parallel_marking_threads = calc_parallel_marking_threads(); 992 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 993 "Maximum number of marking threads exceeded"); 994 uint active_workers = MAX2(1U, parallel_marking_threads()); 995 996 CMRootRegionScanTask task(this); 997 _parallel_workers->set_active_workers(active_workers); 998 _parallel_workers->run_task(&task); 999 1000 // It's possible that has_aborted() is true here without actually 1001 // aborting the survivor scan earlier. This is OK as it's 1002 // mainly used for sanity checking. 1003 root_regions()->scan_finished(); 1004 } 1005 } 1006 1007 void ConcurrentMark::register_concurrent_phase_start(const char* title) { 1008 assert(!_concurrent_phase_started, "Sanity"); 1009 _concurrent_phase_started = true; 1010 _g1h->gc_timer_cm()->register_gc_concurrent_start(title); 1011 } 1012 1013 void ConcurrentMark::register_concurrent_phase_end() { 1014 if (_concurrent_phase_started) { 1015 _concurrent_phase_started = false; 1016 _g1h->gc_timer_cm()->register_gc_concurrent_end(); 1017 } 1018 } 1019 1020 void ConcurrentMark::markFromRoots() { 1021 // we might be tempted to assert that: 1022 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1023 // "inconsistent argument?"); 1024 // However that wouldn't be right, because it's possible that 1025 // a safepoint is indeed in progress as a younger generation 1026 // stop-the-world GC happens even as we mark in this generation. 1027 1028 _restart_for_overflow = false; 1029 1030 // _g1h has _n_par_threads 1031 _parallel_marking_threads = calc_parallel_marking_threads(); 1032 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1033 "Maximum number of marking threads exceeded"); 1034 1035 uint active_workers = MAX2(1U, parallel_marking_threads()); 1036 assert(active_workers > 0, "Should have been set"); 1037 1038 // Parallel task terminator is set in "set_concurrency_and_phase()" 1039 set_concurrency_and_phase(active_workers, true /* concurrent */); 1040 1041 CMConcurrentMarkingTask markingTask(this, cmThread()); 1042 _parallel_workers->set_active_workers(active_workers); 1043 _parallel_workers->run_task(&markingTask); 1044 print_stats(); 1045 } 1046 1047 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1048 // world is stopped at this checkpoint 1049 assert(SafepointSynchronize::is_at_safepoint(), 1050 "world should be stopped"); 1051 1052 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1053 1054 // If a full collection has happened, we shouldn't do this. 1055 if (has_aborted()) { 1056 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1057 return; 1058 } 1059 1060 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1061 1062 if (VerifyDuringGC) { 1063 HandleMark hm; // handle scope 1064 g1h->prepare_for_verify(); 1065 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1066 } 1067 g1h->check_bitmaps("Remark Start"); 1068 1069 G1CollectorPolicy* g1p = g1h->g1_policy(); 1070 g1p->record_concurrent_mark_remark_start(); 1071 1072 double start = os::elapsedTime(); 1073 1074 checkpointRootsFinalWork(); 1075 1076 double mark_work_end = os::elapsedTime(); 1077 1078 weakRefsWork(clear_all_soft_refs); 1079 1080 if (has_overflown()) { 1081 // Oops. We overflowed. Restart concurrent marking. 1082 _restart_for_overflow = true; 1083 log_develop_trace(gc)("Remark led to restart for overflow."); 1084 1085 // Verify the heap w.r.t. the previous marking bitmap. 1086 if (VerifyDuringGC) { 1087 HandleMark hm; // handle scope 1088 g1h->prepare_for_verify(); 1089 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)"); 1090 } 1091 1092 // Clear the marking state because we will be restarting 1093 // marking due to overflowing the global mark stack. 1094 reset_marking_state(); 1095 } else { 1096 { 1097 GCTraceTime(Debug, gc) trace("GC Aggregate Data", g1h->gc_timer_cm()); 1098 1099 // Aggregate the per-task counting data that we have accumulated 1100 // while marking. 1101 aggregate_count_data(); 1102 } 1103 1104 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1105 // We're done with marking. 1106 // This is the end of the marking cycle, we're expected all 1107 // threads to have SATB queues with active set to true. 1108 satb_mq_set.set_active_all_threads(false, /* new active value */ 1109 true /* expected_active */); 1110 1111 if (VerifyDuringGC) { 1112 HandleMark hm; // handle scope 1113 g1h->prepare_for_verify(); 1114 Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)"); 1115 } 1116 g1h->check_bitmaps("Remark End"); 1117 assert(!restart_for_overflow(), "sanity"); 1118 // Completely reset the marking state since marking completed 1119 set_non_marking_state(); 1120 } 1121 1122 // Expand the marking stack, if we have to and if we can. 1123 if (_markStack.should_expand()) { 1124 _markStack.expand(); 1125 } 1126 1127 // Statistics 1128 double now = os::elapsedTime(); 1129 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1130 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1131 _remark_times.add((now - start) * 1000.0); 1132 1133 g1p->record_concurrent_mark_remark_end(); 1134 1135 G1CMIsAliveClosure is_alive(g1h); 1136 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive); 1137 } 1138 1139 // Base class of the closures that finalize and verify the 1140 // liveness counting data. 1141 class CMCountDataClosureBase: public HeapRegionClosure { 1142 protected: 1143 G1CollectedHeap* _g1h; 1144 ConcurrentMark* _cm; 1145 CardTableModRefBS* _ct_bs; 1146 1147 BitMap* _region_bm; 1148 BitMap* _card_bm; 1149 1150 // Takes a region that's not empty (i.e., it has at least one 1151 // live object in it and sets its corresponding bit on the region 1152 // bitmap to 1. 1153 void set_bit_for_region(HeapRegion* hr) { 1154 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index(); 1155 _region_bm->par_at_put(index, true); 1156 } 1157 1158 public: 1159 CMCountDataClosureBase(G1CollectedHeap* g1h, 1160 BitMap* region_bm, BitMap* card_bm): 1161 _g1h(g1h), _cm(g1h->concurrent_mark()), 1162 _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())), 1163 _region_bm(region_bm), _card_bm(card_bm) { } 1164 }; 1165 1166 // Closure that calculates the # live objects per region. Used 1167 // for verification purposes during the cleanup pause. 1168 class CalcLiveObjectsClosure: public CMCountDataClosureBase { 1169 CMBitMapRO* _bm; 1170 size_t _region_marked_bytes; 1171 1172 public: 1173 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h, 1174 BitMap* region_bm, BitMap* card_bm) : 1175 CMCountDataClosureBase(g1h, region_bm, card_bm), 1176 _bm(bm), _region_marked_bytes(0) { } 1177 1178 bool doHeapRegion(HeapRegion* hr) { 1179 HeapWord* ntams = hr->next_top_at_mark_start(); 1180 HeapWord* start = hr->bottom(); 1181 1182 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(), 1183 "Preconditions not met - " 1184 "start: " PTR_FORMAT ", ntams: " PTR_FORMAT ", end: " PTR_FORMAT, 1185 p2i(start), p2i(ntams), p2i(hr->end())); 1186 1187 // Find the first marked object at or after "start". 1188 start = _bm->getNextMarkedWordAddress(start, ntams); 1189 1190 size_t marked_bytes = 0; 1191 1192 while (start < ntams) { 1193 oop obj = oop(start); 1194 int obj_sz = obj->size(); 1195 HeapWord* obj_end = start + obj_sz; 1196 1197 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 1198 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end); 1199 1200 // Note: if we're looking at the last region in heap - obj_end 1201 // could be actually just beyond the end of the heap; end_idx 1202 // will then correspond to a (non-existent) card that is also 1203 // just beyond the heap. 1204 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) { 1205 // end of object is not card aligned - increment to cover 1206 // all the cards spanned by the object 1207 end_idx += 1; 1208 } 1209 1210 // Set the bits in the card BM for the cards spanned by this object. 1211 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1212 1213 // Add the size of this object to the number of marked bytes. 1214 marked_bytes += (size_t)obj_sz * HeapWordSize; 1215 1216 // This will happen if we are handling a humongous object that spans 1217 // several heap regions. 1218 if (obj_end > hr->end()) { 1219 break; 1220 } 1221 // Find the next marked object after this one. 1222 start = _bm->getNextMarkedWordAddress(obj_end, ntams); 1223 } 1224 1225 // Mark the allocated-since-marking portion... 1226 HeapWord* top = hr->top(); 1227 if (ntams < top) { 1228 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1229 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1230 1231 // Note: if we're looking at the last region in heap - top 1232 // could be actually just beyond the end of the heap; end_idx 1233 // will then correspond to a (non-existent) card that is also 1234 // just beyond the heap. 1235 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1236 // end of object is not card aligned - increment to cover 1237 // all the cards spanned by the object 1238 end_idx += 1; 1239 } 1240 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1241 1242 // This definitely means the region has live objects. 1243 set_bit_for_region(hr); 1244 } 1245 1246 // Update the live region bitmap. 1247 if (marked_bytes > 0) { 1248 set_bit_for_region(hr); 1249 } 1250 1251 // Set the marked bytes for the current region so that 1252 // it can be queried by a calling verification routine 1253 _region_marked_bytes = marked_bytes; 1254 1255 return false; 1256 } 1257 1258 size_t region_marked_bytes() const { return _region_marked_bytes; } 1259 }; 1260 1261 // Heap region closure used for verifying the counting data 1262 // that was accumulated concurrently and aggregated during 1263 // the remark pause. This closure is applied to the heap 1264 // regions during the STW cleanup pause. 1265 1266 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure { 1267 G1CollectedHeap* _g1h; 1268 ConcurrentMark* _cm; 1269 CalcLiveObjectsClosure _calc_cl; 1270 BitMap* _region_bm; // Region BM to be verified 1271 BitMap* _card_bm; // Card BM to be verified 1272 1273 BitMap* _exp_region_bm; // Expected Region BM values 1274 BitMap* _exp_card_bm; // Expected card BM values 1275 1276 int _failures; 1277 1278 public: 1279 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h, 1280 BitMap* region_bm, 1281 BitMap* card_bm, 1282 BitMap* exp_region_bm, 1283 BitMap* exp_card_bm) : 1284 _g1h(g1h), _cm(g1h->concurrent_mark()), 1285 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm), 1286 _region_bm(region_bm), _card_bm(card_bm), 1287 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm), 1288 _failures(0) { } 1289 1290 int failures() const { return _failures; } 1291 1292 bool doHeapRegion(HeapRegion* hr) { 1293 int failures = 0; 1294 1295 // Call the CalcLiveObjectsClosure to walk the marking bitmap for 1296 // this region and set the corresponding bits in the expected region 1297 // and card bitmaps. 1298 bool res = _calc_cl.doHeapRegion(hr); 1299 assert(res == false, "should be continuing"); 1300 1301 // Verify the marked bytes for this region. 1302 size_t exp_marked_bytes = _calc_cl.region_marked_bytes(); 1303 size_t act_marked_bytes = hr->next_marked_bytes(); 1304 1305 if (exp_marked_bytes > act_marked_bytes) { 1306 if (hr->is_starts_humongous()) { 1307 // For start_humongous regions, the size of the whole object will be 1308 // in exp_marked_bytes. 1309 HeapRegion* region = hr; 1310 int num_regions; 1311 for (num_regions = 0; region != NULL; num_regions++) { 1312 region = _g1h->next_region_in_humongous(region); 1313 } 1314 if ((num_regions-1) * HeapRegion::GrainBytes >= exp_marked_bytes) { 1315 failures += 1; 1316 } else if (num_regions * HeapRegion::GrainBytes < exp_marked_bytes) { 1317 failures += 1; 1318 } 1319 } else { 1320 // We're not OK if expected marked bytes > actual marked bytes. It means 1321 // we have missed accounting some objects during the actual marking. 1322 failures += 1; 1323 } 1324 } 1325 1326 // Verify the bit, for this region, in the actual and expected 1327 // (which was just calculated) region bit maps. 1328 // We're not OK if the bit in the calculated expected region 1329 // bitmap is set and the bit in the actual region bitmap is not. 1330 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index(); 1331 1332 bool expected = _exp_region_bm->at(index); 1333 bool actual = _region_bm->at(index); 1334 if (expected && !actual) { 1335 failures += 1; 1336 } 1337 1338 // Verify that the card bit maps for the cards spanned by the current 1339 // region match. We have an error if we have a set bit in the expected 1340 // bit map and the corresponding bit in the actual bitmap is not set. 1341 1342 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom()); 1343 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top()); 1344 1345 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) { 1346 expected = _exp_card_bm->at(i); 1347 actual = _card_bm->at(i); 1348 1349 if (expected && !actual) { 1350 failures += 1; 1351 } 1352 } 1353 1354 _failures += failures; 1355 1356 // We could stop iteration over the heap when we 1357 // find the first violating region by returning true. 1358 return false; 1359 } 1360 }; 1361 1362 class G1ParVerifyFinalCountTask: public AbstractGangTask { 1363 protected: 1364 G1CollectedHeap* _g1h; 1365 ConcurrentMark* _cm; 1366 BitMap* _actual_region_bm; 1367 BitMap* _actual_card_bm; 1368 1369 uint _n_workers; 1370 1371 BitMap* _expected_region_bm; 1372 BitMap* _expected_card_bm; 1373 1374 int _failures; 1375 1376 HeapRegionClaimer _hrclaimer; 1377 1378 public: 1379 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h, 1380 BitMap* region_bm, BitMap* card_bm, 1381 BitMap* expected_region_bm, BitMap* expected_card_bm) 1382 : AbstractGangTask("G1 verify final counting"), 1383 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1384 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1385 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm), 1386 _failures(0), 1387 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) { 1388 assert(VerifyDuringGC, "don't call this otherwise"); 1389 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity"); 1390 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity"); 1391 } 1392 1393 void work(uint worker_id) { 1394 assert(worker_id < _n_workers, "invariant"); 1395 1396 VerifyLiveObjectDataHRClosure verify_cl(_g1h, 1397 _actual_region_bm, _actual_card_bm, 1398 _expected_region_bm, 1399 _expected_card_bm); 1400 1401 _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer); 1402 1403 Atomic::add(verify_cl.failures(), &_failures); 1404 } 1405 1406 int failures() const { return _failures; } 1407 }; 1408 1409 // Closure that finalizes the liveness counting data. 1410 // Used during the cleanup pause. 1411 // Sets the bits corresponding to the interval [NTAMS, top] 1412 // (which contains the implicitly live objects) in the 1413 // card liveness bitmap. Also sets the bit for each region, 1414 // containing live data, in the region liveness bitmap. 1415 1416 class FinalCountDataUpdateClosure: public CMCountDataClosureBase { 1417 public: 1418 FinalCountDataUpdateClosure(G1CollectedHeap* g1h, 1419 BitMap* region_bm, 1420 BitMap* card_bm) : 1421 CMCountDataClosureBase(g1h, region_bm, card_bm) { } 1422 1423 bool doHeapRegion(HeapRegion* hr) { 1424 HeapWord* ntams = hr->next_top_at_mark_start(); 1425 HeapWord* top = hr->top(); 1426 1427 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions."); 1428 1429 // Mark the allocated-since-marking portion... 1430 if (ntams < top) { 1431 // This definitely means the region has live objects. 1432 set_bit_for_region(hr); 1433 1434 // Now set the bits in the card bitmap for [ntams, top) 1435 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1436 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1437 1438 // Note: if we're looking at the last region in heap - top 1439 // could be actually just beyond the end of the heap; end_idx 1440 // will then correspond to a (non-existent) card that is also 1441 // just beyond the heap. 1442 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1443 // end of object is not card aligned - increment to cover 1444 // all the cards spanned by the object 1445 end_idx += 1; 1446 } 1447 1448 assert(end_idx <= _card_bm->size(), 1449 "oob: end_idx= " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT, 1450 end_idx, _card_bm->size()); 1451 assert(start_idx < _card_bm->size(), 1452 "oob: start_idx= " SIZE_FORMAT ", bitmap size= " SIZE_FORMAT, 1453 start_idx, _card_bm->size()); 1454 1455 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1456 } 1457 1458 // Set the bit for the region if it contains live data 1459 if (hr->next_marked_bytes() > 0) { 1460 set_bit_for_region(hr); 1461 } 1462 1463 return false; 1464 } 1465 }; 1466 1467 class G1ParFinalCountTask: public AbstractGangTask { 1468 protected: 1469 G1CollectedHeap* _g1h; 1470 ConcurrentMark* _cm; 1471 BitMap* _actual_region_bm; 1472 BitMap* _actual_card_bm; 1473 1474 uint _n_workers; 1475 HeapRegionClaimer _hrclaimer; 1476 1477 public: 1478 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm) 1479 : AbstractGangTask("G1 final counting"), 1480 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1481 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1482 _n_workers(_g1h->workers()->active_workers()), _hrclaimer(_n_workers) { 1483 } 1484 1485 void work(uint worker_id) { 1486 assert(worker_id < _n_workers, "invariant"); 1487 1488 FinalCountDataUpdateClosure final_update_cl(_g1h, 1489 _actual_region_bm, 1490 _actual_card_bm); 1491 1492 _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer); 1493 } 1494 }; 1495 1496 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1497 G1CollectedHeap* _g1; 1498 size_t _freed_bytes; 1499 FreeRegionList* _local_cleanup_list; 1500 uint _old_regions_removed; 1501 uint _humongous_regions_removed; 1502 HRRSCleanupTask* _hrrs_cleanup_task; 1503 1504 public: 1505 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1506 FreeRegionList* local_cleanup_list, 1507 HRRSCleanupTask* hrrs_cleanup_task) : 1508 _g1(g1), 1509 _freed_bytes(0), 1510 _local_cleanup_list(local_cleanup_list), 1511 _old_regions_removed(0), 1512 _humongous_regions_removed(0), 1513 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1514 1515 size_t freed_bytes() { return _freed_bytes; } 1516 const uint old_regions_removed() { return _old_regions_removed; } 1517 const uint humongous_regions_removed() { return _humongous_regions_removed; } 1518 1519 bool doHeapRegion(HeapRegion *hr) { 1520 if (hr->is_archive()) { 1521 return false; 1522 } 1523 // We use a claim value of zero here because all regions 1524 // were claimed with value 1 in the FinalCount task. 1525 _g1->reset_gc_time_stamps(hr); 1526 hr->note_end_of_marking(); 1527 1528 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 1529 _freed_bytes += hr->used(); 1530 hr->set_containing_set(NULL); 1531 if (hr->is_humongous()) { 1532 _humongous_regions_removed++; 1533 _g1->free_humongous_region(hr, _local_cleanup_list, true); 1534 } else { 1535 _old_regions_removed++; 1536 _g1->free_region(hr, _local_cleanup_list, true); 1537 } 1538 } else { 1539 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task); 1540 } 1541 1542 return false; 1543 } 1544 }; 1545 1546 class G1ParNoteEndTask: public AbstractGangTask { 1547 friend class G1NoteEndOfConcMarkClosure; 1548 1549 protected: 1550 G1CollectedHeap* _g1h; 1551 FreeRegionList* _cleanup_list; 1552 HeapRegionClaimer _hrclaimer; 1553 1554 public: 1555 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) : 1556 AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) { 1557 } 1558 1559 void work(uint worker_id) { 1560 FreeRegionList local_cleanup_list("Local Cleanup List"); 1561 HRRSCleanupTask hrrs_cleanup_task; 1562 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list, 1563 &hrrs_cleanup_task); 1564 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer); 1565 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1566 1567 // Now update the lists 1568 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed()); 1569 { 1570 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1571 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes()); 1572 1573 // If we iterate over the global cleanup list at the end of 1574 // cleanup to do this printing we will not guarantee to only 1575 // generate output for the newly-reclaimed regions (the list 1576 // might not be empty at the beginning of cleanup; we might 1577 // still be working on its previous contents). So we do the 1578 // printing here, before we append the new regions to the global 1579 // cleanup list. 1580 1581 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1582 if (hr_printer->is_active()) { 1583 FreeRegionListIterator iter(&local_cleanup_list); 1584 while (iter.more_available()) { 1585 HeapRegion* hr = iter.get_next(); 1586 hr_printer->cleanup(hr); 1587 } 1588 } 1589 1590 _cleanup_list->add_ordered(&local_cleanup_list); 1591 assert(local_cleanup_list.is_empty(), "post-condition"); 1592 1593 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1594 } 1595 } 1596 }; 1597 1598 class G1ParScrubRemSetTask: public AbstractGangTask { 1599 protected: 1600 G1RemSet* _g1rs; 1601 BitMap* _region_bm; 1602 BitMap* _card_bm; 1603 HeapRegionClaimer _hrclaimer; 1604 1605 public: 1606 G1ParScrubRemSetTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm, uint n_workers) : 1607 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), _region_bm(region_bm), _card_bm(card_bm), _hrclaimer(n_workers) { 1608 } 1609 1610 void work(uint worker_id) { 1611 _g1rs->scrub(_region_bm, _card_bm, worker_id, &_hrclaimer); 1612 } 1613 1614 }; 1615 1616 void ConcurrentMark::cleanup() { 1617 // world is stopped at this checkpoint 1618 assert(SafepointSynchronize::is_at_safepoint(), 1619 "world should be stopped"); 1620 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1621 1622 // If a full collection has happened, we shouldn't do this. 1623 if (has_aborted()) { 1624 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1625 return; 1626 } 1627 1628 g1h->verify_region_sets_optional(); 1629 1630 if (VerifyDuringGC) { 1631 HandleMark hm; // handle scope 1632 g1h->prepare_for_verify(); 1633 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1634 } 1635 g1h->check_bitmaps("Cleanup Start"); 1636 1637 G1CollectorPolicy* g1p = g1h->g1_policy(); 1638 g1p->record_concurrent_mark_cleanup_start(); 1639 1640 double start = os::elapsedTime(); 1641 1642 HeapRegionRemSet::reset_for_cleanup_tasks(); 1643 1644 // Do counting once more with the world stopped for good measure. 1645 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm); 1646 1647 g1h->workers()->run_task(&g1_par_count_task); 1648 1649 if (VerifyDuringGC) { 1650 // Verify that the counting data accumulated during marking matches 1651 // that calculated by walking the marking bitmap. 1652 1653 // Bitmaps to hold expected values 1654 BitMap expected_region_bm(_region_bm.size(), true); 1655 BitMap expected_card_bm(_card_bm.size(), true); 1656 1657 G1ParVerifyFinalCountTask g1_par_verify_task(g1h, 1658 &_region_bm, 1659 &_card_bm, 1660 &expected_region_bm, 1661 &expected_card_bm); 1662 1663 g1h->workers()->run_task(&g1_par_verify_task); 1664 1665 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures"); 1666 } 1667 1668 size_t start_used_bytes = g1h->used(); 1669 g1h->collector_state()->set_mark_in_progress(false); 1670 1671 double count_end = os::elapsedTime(); 1672 double this_final_counting_time = (count_end - start); 1673 _total_counting_time += this_final_counting_time; 1674 1675 if (log_is_enabled(Trace, gc, liveness)) { 1676 G1PrintRegionLivenessInfoClosure cl("Post-Marking"); 1677 _g1h->heap_region_iterate(&cl); 1678 } 1679 1680 // Install newly created mark bitMap as "prev". 1681 swapMarkBitMaps(); 1682 1683 g1h->reset_gc_time_stamp(); 1684 1685 uint n_workers = _g1h->workers()->active_workers(); 1686 1687 // Note end of marking in all heap regions. 1688 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers); 1689 g1h->workers()->run_task(&g1_par_note_end_task); 1690 g1h->check_gc_time_stamps(); 1691 1692 if (!cleanup_list_is_empty()) { 1693 // The cleanup list is not empty, so we'll have to process it 1694 // concurrently. Notify anyone else that might be wanting free 1695 // regions that there will be more free regions coming soon. 1696 g1h->set_free_regions_coming(); 1697 } 1698 1699 // call below, since it affects the metric by which we sort the heap 1700 // regions. 1701 if (G1ScrubRemSets) { 1702 double rs_scrub_start = os::elapsedTime(); 1703 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm, n_workers); 1704 g1h->workers()->run_task(&g1_par_scrub_rs_task); 1705 1706 double rs_scrub_end = os::elapsedTime(); 1707 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); 1708 _total_rs_scrub_time += this_rs_scrub_time; 1709 } 1710 1711 // this will also free any regions totally full of garbage objects, 1712 // and sort the regions. 1713 g1h->g1_policy()->record_concurrent_mark_cleanup_end(); 1714 1715 // Statistics. 1716 double end = os::elapsedTime(); 1717 _cleanup_times.add((end - start) * 1000.0); 1718 1719 // Clean up will have freed any regions completely full of garbage. 1720 // Update the soft reference policy with the new heap occupancy. 1721 Universe::update_heap_info_at_gc(); 1722 1723 if (VerifyDuringGC) { 1724 HandleMark hm; // handle scope 1725 g1h->prepare_for_verify(); 1726 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)"); 1727 } 1728 1729 g1h->check_bitmaps("Cleanup End"); 1730 1731 g1h->verify_region_sets_optional(); 1732 1733 // We need to make this be a "collection" so any collection pause that 1734 // races with it goes around and waits for completeCleanup to finish. 1735 g1h->increment_total_collections(); 1736 1737 // Clean out dead classes and update Metaspace sizes. 1738 if (ClassUnloadingWithConcurrentMark) { 1739 ClassLoaderDataGraph::purge(); 1740 } 1741 MetaspaceGC::compute_new_size(); 1742 1743 // We reclaimed old regions so we should calculate the sizes to make 1744 // sure we update the old gen/space data. 1745 g1h->g1mm()->update_sizes(); 1746 g1h->allocation_context_stats().update_after_mark(); 1747 1748 g1h->trace_heap_after_concurrent_cycle(); 1749 } 1750 1751 void ConcurrentMark::completeCleanup() { 1752 if (has_aborted()) return; 1753 1754 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1755 1756 _cleanup_list.verify_optional(); 1757 FreeRegionList tmp_free_list("Tmp Free List"); 1758 1759 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1760 "cleanup list has %u entries", 1761 _cleanup_list.length()); 1762 1763 // No one else should be accessing the _cleanup_list at this point, 1764 // so it is not necessary to take any locks 1765 while (!_cleanup_list.is_empty()) { 1766 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */); 1767 assert(hr != NULL, "Got NULL from a non-empty list"); 1768 hr->par_clear(); 1769 tmp_free_list.add_ordered(hr); 1770 1771 // Instead of adding one region at a time to the secondary_free_list, 1772 // we accumulate them in the local list and move them a few at a 1773 // time. This also cuts down on the number of notify_all() calls 1774 // we do during this process. We'll also append the local list when 1775 // _cleanup_list is empty (which means we just removed the last 1776 // region from the _cleanup_list). 1777 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 1778 _cleanup_list.is_empty()) { 1779 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1780 "appending %u entries to the secondary_free_list, " 1781 "cleanup list still has %u entries", 1782 tmp_free_list.length(), 1783 _cleanup_list.length()); 1784 1785 { 1786 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 1787 g1h->secondary_free_list_add(&tmp_free_list); 1788 SecondaryFreeList_lock->notify_all(); 1789 } 1790 #ifndef PRODUCT 1791 if (G1StressConcRegionFreeing) { 1792 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 1793 os::sleep(Thread::current(), (jlong) 1, false); 1794 } 1795 } 1796 #endif 1797 } 1798 } 1799 assert(tmp_free_list.is_empty(), "post-condition"); 1800 } 1801 1802 // Supporting Object and Oop closures for reference discovery 1803 // and processing in during marking 1804 1805 bool G1CMIsAliveClosure::do_object_b(oop obj) { 1806 HeapWord* addr = (HeapWord*)obj; 1807 return addr != NULL && 1808 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 1809 } 1810 1811 // 'Keep Alive' oop closure used by both serial parallel reference processing. 1812 // Uses the CMTask associated with a worker thread (for serial reference 1813 // processing the CMTask for worker 0 is used) to preserve (mark) and 1814 // trace referent objects. 1815 // 1816 // Using the CMTask and embedded local queues avoids having the worker 1817 // threads operating on the global mark stack. This reduces the risk 1818 // of overflowing the stack - which we would rather avoid at this late 1819 // state. Also using the tasks' local queues removes the potential 1820 // of the workers interfering with each other that could occur if 1821 // operating on the global stack. 1822 1823 class G1CMKeepAliveAndDrainClosure: public OopClosure { 1824 ConcurrentMark* _cm; 1825 CMTask* _task; 1826 int _ref_counter_limit; 1827 int _ref_counter; 1828 bool _is_serial; 1829 public: 1830 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) : 1831 _cm(cm), _task(task), _is_serial(is_serial), 1832 _ref_counter_limit(G1RefProcDrainInterval) { 1833 assert(_ref_counter_limit > 0, "sanity"); 1834 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1835 _ref_counter = _ref_counter_limit; 1836 } 1837 1838 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 1839 virtual void do_oop( oop* p) { do_oop_work(p); } 1840 1841 template <class T> void do_oop_work(T* p) { 1842 if (!_cm->has_overflown()) { 1843 oop obj = oopDesc::load_decode_heap_oop(p); 1844 _task->deal_with_reference(obj); 1845 _ref_counter--; 1846 1847 if (_ref_counter == 0) { 1848 // We have dealt with _ref_counter_limit references, pushing them 1849 // and objects reachable from them on to the local stack (and 1850 // possibly the global stack). Call CMTask::do_marking_step() to 1851 // process these entries. 1852 // 1853 // We call CMTask::do_marking_step() in a loop, which we'll exit if 1854 // there's nothing more to do (i.e. we're done with the entries that 1855 // were pushed as a result of the CMTask::deal_with_reference() calls 1856 // above) or we overflow. 1857 // 1858 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 1859 // flag while there may still be some work to do. (See the comment at 1860 // the beginning of CMTask::do_marking_step() for those conditions - 1861 // one of which is reaching the specified time target.) It is only 1862 // when CMTask::do_marking_step() returns without setting the 1863 // has_aborted() flag that the marking step has completed. 1864 do { 1865 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1866 _task->do_marking_step(mark_step_duration_ms, 1867 false /* do_termination */, 1868 _is_serial); 1869 } while (_task->has_aborted() && !_cm->has_overflown()); 1870 _ref_counter = _ref_counter_limit; 1871 } 1872 } 1873 } 1874 }; 1875 1876 // 'Drain' oop closure used by both serial and parallel reference processing. 1877 // Uses the CMTask associated with a given worker thread (for serial 1878 // reference processing the CMtask for worker 0 is used). Calls the 1879 // do_marking_step routine, with an unbelievably large timeout value, 1880 // to drain the marking data structures of the remaining entries 1881 // added by the 'keep alive' oop closure above. 1882 1883 class G1CMDrainMarkingStackClosure: public VoidClosure { 1884 ConcurrentMark* _cm; 1885 CMTask* _task; 1886 bool _is_serial; 1887 public: 1888 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) : 1889 _cm(cm), _task(task), _is_serial(is_serial) { 1890 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1891 } 1892 1893 void do_void() { 1894 do { 1895 // We call CMTask::do_marking_step() to completely drain the local 1896 // and global marking stacks of entries pushed by the 'keep alive' 1897 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 1898 // 1899 // CMTask::do_marking_step() is called in a loop, which we'll exit 1900 // if there's nothing more to do (i.e. we've completely drained the 1901 // entries that were pushed as a a result of applying the 'keep alive' 1902 // closure to the entries on the discovered ref lists) or we overflow 1903 // the global marking stack. 1904 // 1905 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 1906 // flag while there may still be some work to do. (See the comment at 1907 // the beginning of CMTask::do_marking_step() for those conditions - 1908 // one of which is reaching the specified time target.) It is only 1909 // when CMTask::do_marking_step() returns without setting the 1910 // has_aborted() flag that the marking step has completed. 1911 1912 _task->do_marking_step(1000000000.0 /* something very large */, 1913 true /* do_termination */, 1914 _is_serial); 1915 } while (_task->has_aborted() && !_cm->has_overflown()); 1916 } 1917 }; 1918 1919 // Implementation of AbstractRefProcTaskExecutor for parallel 1920 // reference processing at the end of G1 concurrent marking 1921 1922 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 1923 private: 1924 G1CollectedHeap* _g1h; 1925 ConcurrentMark* _cm; 1926 WorkGang* _workers; 1927 uint _active_workers; 1928 1929 public: 1930 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 1931 ConcurrentMark* cm, 1932 WorkGang* workers, 1933 uint n_workers) : 1934 _g1h(g1h), _cm(cm), 1935 _workers(workers), _active_workers(n_workers) { } 1936 1937 // Executes the given task using concurrent marking worker threads. 1938 virtual void execute(ProcessTask& task); 1939 virtual void execute(EnqueueTask& task); 1940 }; 1941 1942 class G1CMRefProcTaskProxy: public AbstractGangTask { 1943 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 1944 ProcessTask& _proc_task; 1945 G1CollectedHeap* _g1h; 1946 ConcurrentMark* _cm; 1947 1948 public: 1949 G1CMRefProcTaskProxy(ProcessTask& proc_task, 1950 G1CollectedHeap* g1h, 1951 ConcurrentMark* cm) : 1952 AbstractGangTask("Process reference objects in parallel"), 1953 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 1954 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1955 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 1956 } 1957 1958 virtual void work(uint worker_id) { 1959 ResourceMark rm; 1960 HandleMark hm; 1961 CMTask* task = _cm->task(worker_id); 1962 G1CMIsAliveClosure g1_is_alive(_g1h); 1963 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 1964 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 1965 1966 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 1967 } 1968 }; 1969 1970 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 1971 assert(_workers != NULL, "Need parallel worker threads."); 1972 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1973 1974 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 1975 1976 // We need to reset the concurrency level before each 1977 // proxy task execution, so that the termination protocol 1978 // and overflow handling in CMTask::do_marking_step() knows 1979 // how many workers to wait for. 1980 _cm->set_concurrency(_active_workers); 1981 _workers->run_task(&proc_task_proxy); 1982 } 1983 1984 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 1985 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 1986 EnqueueTask& _enq_task; 1987 1988 public: 1989 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 1990 AbstractGangTask("Enqueue reference objects in parallel"), 1991 _enq_task(enq_task) { } 1992 1993 virtual void work(uint worker_id) { 1994 _enq_task.work(worker_id); 1995 } 1996 }; 1997 1998 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 1999 assert(_workers != NULL, "Need parallel worker threads."); 2000 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2001 2002 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 2003 2004 // Not strictly necessary but... 2005 // 2006 // We need to reset the concurrency level before each 2007 // proxy task execution, so that the termination protocol 2008 // and overflow handling in CMTask::do_marking_step() knows 2009 // how many workers to wait for. 2010 _cm->set_concurrency(_active_workers); 2011 _workers->run_task(&enq_task_proxy); 2012 } 2013 2014 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) { 2015 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes); 2016 } 2017 2018 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 2019 if (has_overflown()) { 2020 // Skip processing the discovered references if we have 2021 // overflown the global marking stack. Reference objects 2022 // only get discovered once so it is OK to not 2023 // de-populate the discovered reference lists. We could have, 2024 // but the only benefit would be that, when marking restarts, 2025 // less reference objects are discovered. 2026 return; 2027 } 2028 2029 ResourceMark rm; 2030 HandleMark hm; 2031 2032 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2033 2034 // Is alive closure. 2035 G1CMIsAliveClosure g1_is_alive(g1h); 2036 2037 // Inner scope to exclude the cleaning of the string and symbol 2038 // tables from the displayed time. 2039 { 2040 GCTraceTime(Debug, gc) trace("GC Ref Proc", g1h->gc_timer_cm()); 2041 2042 ReferenceProcessor* rp = g1h->ref_processor_cm(); 2043 2044 // See the comment in G1CollectedHeap::ref_processing_init() 2045 // about how reference processing currently works in G1. 2046 2047 // Set the soft reference policy 2048 rp->setup_policy(clear_all_soft_refs); 2049 assert(_markStack.isEmpty(), "mark stack should be empty"); 2050 2051 // Instances of the 'Keep Alive' and 'Complete GC' closures used 2052 // in serial reference processing. Note these closures are also 2053 // used for serially processing (by the the current thread) the 2054 // JNI references during parallel reference processing. 2055 // 2056 // These closures do not need to synchronize with the worker 2057 // threads involved in parallel reference processing as these 2058 // instances are executed serially by the current thread (e.g. 2059 // reference processing is not multi-threaded and is thus 2060 // performed by the current thread instead of a gang worker). 2061 // 2062 // The gang tasks involved in parallel reference processing create 2063 // their own instances of these closures, which do their own 2064 // synchronization among themselves. 2065 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 2066 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 2067 2068 // We need at least one active thread. If reference processing 2069 // is not multi-threaded we use the current (VMThread) thread, 2070 // otherwise we use the work gang from the G1CollectedHeap and 2071 // we utilize all the worker threads we can. 2072 bool processing_is_mt = rp->processing_is_mt(); 2073 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U); 2074 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U); 2075 2076 // Parallel processing task executor. 2077 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 2078 g1h->workers(), active_workers); 2079 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 2080 2081 // Set the concurrency level. The phase was already set prior to 2082 // executing the remark task. 2083 set_concurrency(active_workers); 2084 2085 // Set the degree of MT processing here. If the discovery was done MT, 2086 // the number of threads involved during discovery could differ from 2087 // the number of active workers. This is OK as long as the discovered 2088 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 2089 rp->set_active_mt_degree(active_workers); 2090 2091 // Process the weak references. 2092 const ReferenceProcessorStats& stats = 2093 rp->process_discovered_references(&g1_is_alive, 2094 &g1_keep_alive, 2095 &g1_drain_mark_stack, 2096 executor, 2097 g1h->gc_timer_cm()); 2098 g1h->gc_tracer_cm()->report_gc_reference_stats(stats); 2099 2100 // The do_oop work routines of the keep_alive and drain_marking_stack 2101 // oop closures will set the has_overflown flag if we overflow the 2102 // global marking stack. 2103 2104 assert(_markStack.overflow() || _markStack.isEmpty(), 2105 "mark stack should be empty (unless it overflowed)"); 2106 2107 if (_markStack.overflow()) { 2108 // This should have been done already when we tried to push an 2109 // entry on to the global mark stack. But let's do it again. 2110 set_has_overflown(); 2111 } 2112 2113 assert(rp->num_q() == active_workers, "why not"); 2114 2115 rp->enqueue_discovered_references(executor); 2116 2117 rp->verify_no_references_recorded(); 2118 assert(!rp->discovery_enabled(), "Post condition"); 2119 } 2120 2121 if (has_overflown()) { 2122 // We can not trust g1_is_alive if the marking stack overflowed 2123 return; 2124 } 2125 2126 assert(_markStack.isEmpty(), "Marking should have completed"); 2127 2128 // Unload Klasses, String, Symbols, Code Cache, etc. 2129 { 2130 GCTraceTime(Debug, gc) trace("Unloading", g1h->gc_timer_cm()); 2131 2132 if (ClassUnloadingWithConcurrentMark) { 2133 bool purged_classes; 2134 2135 { 2136 GCTraceTime(Trace, gc) trace("System Dictionary Unloading", g1h->gc_timer_cm()); 2137 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */); 2138 } 2139 2140 { 2141 GCTraceTime(Trace, gc) trace("Parallel Unloading", g1h->gc_timer_cm()); 2142 weakRefsWorkParallelPart(&g1_is_alive, purged_classes); 2143 } 2144 } 2145 2146 if (G1StringDedup::is_enabled()) { 2147 GCTraceTime(Trace, gc) trace("String Deduplication Unlink", g1h->gc_timer_cm()); 2148 G1StringDedup::unlink(&g1_is_alive); 2149 } 2150 } 2151 } 2152 2153 void ConcurrentMark::swapMarkBitMaps() { 2154 CMBitMapRO* temp = _prevMarkBitMap; 2155 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; 2156 _nextMarkBitMap = (CMBitMap*) temp; 2157 } 2158 2159 // Closure for marking entries in SATB buffers. 2160 class CMSATBBufferClosure : public SATBBufferClosure { 2161 private: 2162 CMTask* _task; 2163 G1CollectedHeap* _g1h; 2164 2165 // This is very similar to CMTask::deal_with_reference, but with 2166 // more relaxed requirements for the argument, so this must be more 2167 // circumspect about treating the argument as an object. 2168 void do_entry(void* entry) const { 2169 _task->increment_refs_reached(); 2170 HeapRegion* hr = _g1h->heap_region_containing(entry); 2171 if (entry < hr->next_top_at_mark_start()) { 2172 // Until we get here, we don't know whether entry refers to a valid 2173 // object; it could instead have been a stale reference. 2174 oop obj = static_cast<oop>(entry); 2175 assert(obj->is_oop(true /* ignore mark word */), 2176 "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj)); 2177 _task->make_reference_grey(obj, hr); 2178 } 2179 } 2180 2181 public: 2182 CMSATBBufferClosure(CMTask* task, G1CollectedHeap* g1h) 2183 : _task(task), _g1h(g1h) { } 2184 2185 virtual void do_buffer(void** buffer, size_t size) { 2186 for (size_t i = 0; i < size; ++i) { 2187 do_entry(buffer[i]); 2188 } 2189 } 2190 }; 2191 2192 class G1RemarkThreadsClosure : public ThreadClosure { 2193 CMSATBBufferClosure _cm_satb_cl; 2194 G1CMOopClosure _cm_cl; 2195 MarkingCodeBlobClosure _code_cl; 2196 int _thread_parity; 2197 2198 public: 2199 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task) : 2200 _cm_satb_cl(task, g1h), 2201 _cm_cl(g1h, g1h->concurrent_mark(), task), 2202 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations), 2203 _thread_parity(Threads::thread_claim_parity()) {} 2204 2205 void do_thread(Thread* thread) { 2206 if (thread->is_Java_thread()) { 2207 if (thread->claim_oops_do(true, _thread_parity)) { 2208 JavaThread* jt = (JavaThread*)thread; 2209 2210 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking 2211 // however the liveness of oops reachable from nmethods have very complex lifecycles: 2212 // * Alive if on the stack of an executing method 2213 // * Weakly reachable otherwise 2214 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be 2215 // live by the SATB invariant but other oops recorded in nmethods may behave differently. 2216 jt->nmethods_do(&_code_cl); 2217 2218 jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl); 2219 } 2220 } else if (thread->is_VM_thread()) { 2221 if (thread->claim_oops_do(true, _thread_parity)) { 2222 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl); 2223 } 2224 } 2225 } 2226 }; 2227 2228 class CMRemarkTask: public AbstractGangTask { 2229 private: 2230 ConcurrentMark* _cm; 2231 public: 2232 void work(uint worker_id) { 2233 // Since all available tasks are actually started, we should 2234 // only proceed if we're supposed to be active. 2235 if (worker_id < _cm->active_tasks()) { 2236 CMTask* task = _cm->task(worker_id); 2237 task->record_start_time(); 2238 { 2239 ResourceMark rm; 2240 HandleMark hm; 2241 2242 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task); 2243 Threads::threads_do(&threads_f); 2244 } 2245 2246 do { 2247 task->do_marking_step(1000000000.0 /* something very large */, 2248 true /* do_termination */, 2249 false /* is_serial */); 2250 } while (task->has_aborted() && !_cm->has_overflown()); 2251 // If we overflow, then we do not want to restart. We instead 2252 // want to abort remark and do concurrent marking again. 2253 task->record_end_time(); 2254 } 2255 } 2256 2257 CMRemarkTask(ConcurrentMark* cm, uint active_workers) : 2258 AbstractGangTask("Par Remark"), _cm(cm) { 2259 _cm->terminator()->reset_for_reuse(active_workers); 2260 } 2261 }; 2262 2263 void ConcurrentMark::checkpointRootsFinalWork() { 2264 ResourceMark rm; 2265 HandleMark hm; 2266 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2267 2268 GCTraceTime(Debug, gc) trace("Finalize Marking", g1h->gc_timer_cm()); 2269 2270 g1h->ensure_parsability(false); 2271 2272 // this is remark, so we'll use up all active threads 2273 uint active_workers = g1h->workers()->active_workers(); 2274 set_concurrency_and_phase(active_workers, false /* concurrent */); 2275 // Leave _parallel_marking_threads at it's 2276 // value originally calculated in the ConcurrentMark 2277 // constructor and pass values of the active workers 2278 // through the gang in the task. 2279 2280 { 2281 StrongRootsScope srs(active_workers); 2282 2283 CMRemarkTask remarkTask(this, active_workers); 2284 // We will start all available threads, even if we decide that the 2285 // active_workers will be fewer. The extra ones will just bail out 2286 // immediately. 2287 g1h->workers()->run_task(&remarkTask); 2288 } 2289 2290 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2291 guarantee(has_overflown() || 2292 satb_mq_set.completed_buffers_num() == 0, 2293 "Invariant: has_overflown = %s, num buffers = %d", 2294 BOOL_TO_STR(has_overflown()), 2295 satb_mq_set.completed_buffers_num()); 2296 2297 print_stats(); 2298 } 2299 2300 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 2301 // Note we are overriding the read-only view of the prev map here, via 2302 // the cast. 2303 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); 2304 } 2305 2306 HeapRegion* 2307 ConcurrentMark::claim_region(uint worker_id) { 2308 // "checkpoint" the finger 2309 HeapWord* finger = _finger; 2310 2311 // _heap_end will not change underneath our feet; it only changes at 2312 // yield points. 2313 while (finger < _heap_end) { 2314 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 2315 2316 HeapRegion* curr_region = _g1h->heap_region_containing(finger); 2317 2318 // Above heap_region_containing may return NULL as we always scan claim 2319 // until the end of the heap. In this case, just jump to the next region. 2320 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords; 2321 2322 // Is the gap between reading the finger and doing the CAS too long? 2323 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 2324 if (res == finger && curr_region != NULL) { 2325 // we succeeded 2326 HeapWord* bottom = curr_region->bottom(); 2327 HeapWord* limit = curr_region->next_top_at_mark_start(); 2328 2329 // notice that _finger == end cannot be guaranteed here since, 2330 // someone else might have moved the finger even further 2331 assert(_finger >= end, "the finger should have moved forward"); 2332 2333 if (limit > bottom) { 2334 return curr_region; 2335 } else { 2336 assert(limit == bottom, 2337 "the region limit should be at bottom"); 2338 // we return NULL and the caller should try calling 2339 // claim_region() again. 2340 return NULL; 2341 } 2342 } else { 2343 assert(_finger > finger, "the finger should have moved forward"); 2344 // read it again 2345 finger = _finger; 2346 } 2347 } 2348 2349 return NULL; 2350 } 2351 2352 #ifndef PRODUCT 2353 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC { 2354 private: 2355 G1CollectedHeap* _g1h; 2356 const char* _phase; 2357 int _info; 2358 2359 public: 2360 VerifyNoCSetOops(const char* phase, int info = -1) : 2361 _g1h(G1CollectedHeap::heap()), 2362 _phase(phase), 2363 _info(info) 2364 { } 2365 2366 void operator()(oop obj) const { 2367 guarantee(obj->is_oop(), 2368 "Non-oop " PTR_FORMAT ", phase: %s, info: %d", 2369 p2i(obj), _phase, _info); 2370 guarantee(!_g1h->obj_in_cs(obj), 2371 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d", 2372 p2i(obj), _phase, _info); 2373 } 2374 }; 2375 2376 void ConcurrentMark::verify_no_cset_oops() { 2377 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 2378 if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) { 2379 return; 2380 } 2381 2382 // Verify entries on the global mark stack 2383 _markStack.iterate(VerifyNoCSetOops("Stack")); 2384 2385 // Verify entries on the task queues 2386 for (uint i = 0; i < _max_worker_id; ++i) { 2387 CMTaskQueue* queue = _task_queues->queue(i); 2388 queue->iterate(VerifyNoCSetOops("Queue", i)); 2389 } 2390 2391 // Verify the global finger 2392 HeapWord* global_finger = finger(); 2393 if (global_finger != NULL && global_finger < _heap_end) { 2394 // Since we always iterate over all regions, we might get a NULL HeapRegion 2395 // here. 2396 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger); 2397 guarantee(global_hr == NULL || global_finger == global_hr->bottom(), 2398 "global finger: " PTR_FORMAT " region: " HR_FORMAT, 2399 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)); 2400 } 2401 2402 // Verify the task fingers 2403 assert(parallel_marking_threads() <= _max_worker_id, "sanity"); 2404 for (uint i = 0; i < parallel_marking_threads(); ++i) { 2405 CMTask* task = _tasks[i]; 2406 HeapWord* task_finger = task->finger(); 2407 if (task_finger != NULL && task_finger < _heap_end) { 2408 // See above note on the global finger verification. 2409 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger); 2410 guarantee(task_hr == NULL || task_finger == task_hr->bottom() || 2411 !task_hr->in_collection_set(), 2412 "task finger: " PTR_FORMAT " region: " HR_FORMAT, 2413 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)); 2414 } 2415 } 2416 } 2417 #endif // PRODUCT 2418 2419 // Aggregate the counting data that was constructed concurrently 2420 // with marking. 2421 class AggregateCountDataHRClosure: public HeapRegionClosure { 2422 G1CollectedHeap* _g1h; 2423 ConcurrentMark* _cm; 2424 CardTableModRefBS* _ct_bs; 2425 BitMap* _cm_card_bm; 2426 uint _max_worker_id; 2427 2428 public: 2429 AggregateCountDataHRClosure(G1CollectedHeap* g1h, 2430 BitMap* cm_card_bm, 2431 uint max_worker_id) : 2432 _g1h(g1h), _cm(g1h->concurrent_mark()), 2433 _ct_bs(barrier_set_cast<CardTableModRefBS>(g1h->barrier_set())), 2434 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { } 2435 2436 bool doHeapRegion(HeapRegion* hr) { 2437 HeapWord* start = hr->bottom(); 2438 HeapWord* limit = hr->next_top_at_mark_start(); 2439 HeapWord* end = hr->end(); 2440 2441 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(), 2442 "Preconditions not met - " 2443 "start: " PTR_FORMAT ", limit: " PTR_FORMAT ", " 2444 "top: " PTR_FORMAT ", end: " PTR_FORMAT, 2445 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end())); 2446 2447 assert(hr->next_marked_bytes() == 0, "Precondition"); 2448 2449 if (start == limit) { 2450 // NTAMS of this region has not been set so nothing to do. 2451 return false; 2452 } 2453 2454 // 'start' should be in the heap. 2455 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity"); 2456 // 'end' *may* be just beyond the end of the heap (if hr is the last region) 2457 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity"); 2458 2459 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 2460 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit); 2461 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end); 2462 2463 // If ntams is not card aligned then we bump card bitmap index 2464 // for limit so that we get the all the cards spanned by 2465 // the object ending at ntams. 2466 // Note: if this is the last region in the heap then ntams 2467 // could be actually just beyond the end of the the heap; 2468 // limit_idx will then correspond to a (non-existent) card 2469 // that is also outside the heap. 2470 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) { 2471 limit_idx += 1; 2472 } 2473 2474 assert(limit_idx <= end_idx, "or else use atomics"); 2475 2476 // Aggregate the "stripe" in the count data associated with hr. 2477 uint hrm_index = hr->hrm_index(); 2478 size_t marked_bytes = 0; 2479 2480 for (uint i = 0; i < _max_worker_id; i += 1) { 2481 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i); 2482 BitMap* task_card_bm = _cm->count_card_bitmap_for(i); 2483 2484 // Fetch the marked_bytes in this region for task i and 2485 // add it to the running total for this region. 2486 marked_bytes += marked_bytes_array[hrm_index]; 2487 2488 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx) 2489 // into the global card bitmap. 2490 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx); 2491 2492 while (scan_idx < limit_idx) { 2493 assert(task_card_bm->at(scan_idx) == true, "should be"); 2494 _cm_card_bm->set_bit(scan_idx); 2495 assert(_cm_card_bm->at(scan_idx) == true, "should be"); 2496 2497 // BitMap::get_next_one_offset() can handle the case when 2498 // its left_offset parameter is greater than its right_offset 2499 // parameter. It does, however, have an early exit if 2500 // left_offset == right_offset. So let's limit the value 2501 // passed in for left offset here. 2502 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx); 2503 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx); 2504 } 2505 } 2506 2507 // Update the marked bytes for this region. 2508 hr->add_to_marked_bytes(marked_bytes); 2509 2510 // Next heap region 2511 return false; 2512 } 2513 }; 2514 2515 class G1AggregateCountDataTask: public AbstractGangTask { 2516 protected: 2517 G1CollectedHeap* _g1h; 2518 ConcurrentMark* _cm; 2519 BitMap* _cm_card_bm; 2520 uint _max_worker_id; 2521 uint _active_workers; 2522 HeapRegionClaimer _hrclaimer; 2523 2524 public: 2525 G1AggregateCountDataTask(G1CollectedHeap* g1h, 2526 ConcurrentMark* cm, 2527 BitMap* cm_card_bm, 2528 uint max_worker_id, 2529 uint n_workers) : 2530 AbstractGangTask("Count Aggregation"), 2531 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm), 2532 _max_worker_id(max_worker_id), 2533 _active_workers(n_workers), 2534 _hrclaimer(_active_workers) { 2535 } 2536 2537 void work(uint worker_id) { 2538 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id); 2539 2540 _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer); 2541 } 2542 }; 2543 2544 2545 void ConcurrentMark::aggregate_count_data() { 2546 uint n_workers = _g1h->workers()->active_workers(); 2547 2548 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm, 2549 _max_worker_id, n_workers); 2550 2551 _g1h->workers()->run_task(&g1_par_agg_task); 2552 } 2553 2554 // Clear the per-worker arrays used to store the per-region counting data 2555 void ConcurrentMark::clear_all_count_data() { 2556 // Clear the global card bitmap - it will be filled during 2557 // liveness count aggregation (during remark) and the 2558 // final counting task. 2559 _card_bm.clear(); 2560 2561 // Clear the global region bitmap - it will be filled as part 2562 // of the final counting task. 2563 _region_bm.clear(); 2564 2565 uint max_regions = _g1h->max_regions(); 2566 assert(_max_worker_id > 0, "uninitialized"); 2567 2568 for (uint i = 0; i < _max_worker_id; i += 1) { 2569 BitMap* task_card_bm = count_card_bitmap_for(i); 2570 size_t* marked_bytes_array = count_marked_bytes_array_for(i); 2571 2572 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 2573 assert(marked_bytes_array != NULL, "uninitialized"); 2574 2575 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t)); 2576 task_card_bm->clear(); 2577 } 2578 } 2579 2580 void ConcurrentMark::print_stats() { 2581 if (!log_is_enabled(Debug, gc, stats)) { 2582 return; 2583 } 2584 log_debug(gc, stats)("---------------------------------------------------------------------"); 2585 for (size_t i = 0; i < _active_tasks; ++i) { 2586 _tasks[i]->print_stats(); 2587 log_debug(gc, stats)("---------------------------------------------------------------------"); 2588 } 2589 } 2590 2591 // abandon current marking iteration due to a Full GC 2592 void ConcurrentMark::abort() { 2593 if (!cmThread()->during_cycle() || _has_aborted) { 2594 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything. 2595 return; 2596 } 2597 2598 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next 2599 // concurrent bitmap clearing. 2600 _nextMarkBitMap->clearAll(); 2601 2602 // Note we cannot clear the previous marking bitmap here 2603 // since VerifyDuringGC verifies the objects marked during 2604 // a full GC against the previous bitmap. 2605 2606 // Clear the liveness counting data 2607 clear_all_count_data(); 2608 // Empty mark stack 2609 reset_marking_state(); 2610 for (uint i = 0; i < _max_worker_id; ++i) { 2611 _tasks[i]->clear_region_fields(); 2612 } 2613 _first_overflow_barrier_sync.abort(); 2614 _second_overflow_barrier_sync.abort(); 2615 _has_aborted = true; 2616 2617 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2618 satb_mq_set.abandon_partial_marking(); 2619 // This can be called either during or outside marking, we'll read 2620 // the expected_active value from the SATB queue set. 2621 satb_mq_set.set_active_all_threads( 2622 false, /* new active value */ 2623 satb_mq_set.is_active() /* expected_active */); 2624 2625 _g1h->trace_heap_after_concurrent_cycle(); 2626 2627 // Close any open concurrent phase timing 2628 register_concurrent_phase_end(); 2629 2630 _g1h->register_concurrent_cycle_end(); 2631 } 2632 2633 static void print_ms_time_info(const char* prefix, const char* name, 2634 NumberSeq& ns) { 2635 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 2636 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 2637 if (ns.num() > 0) { 2638 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]", 2639 prefix, ns.sd(), ns.maximum()); 2640 } 2641 } 2642 2643 void ConcurrentMark::print_summary_info() { 2644 LogHandle(gc, marking) log; 2645 if (!log.is_trace()) { 2646 return; 2647 } 2648 2649 log.trace(" Concurrent marking:"); 2650 print_ms_time_info(" ", "init marks", _init_times); 2651 print_ms_time_info(" ", "remarks", _remark_times); 2652 { 2653 print_ms_time_info(" ", "final marks", _remark_mark_times); 2654 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 2655 2656 } 2657 print_ms_time_info(" ", "cleanups", _cleanup_times); 2658 log.trace(" Final counting total time = %8.2f s (avg = %8.2f ms).", 2659 _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2660 if (G1ScrubRemSets) { 2661 log.trace(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 2662 _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2663 } 2664 log.trace(" Total stop_world time = %8.2f s.", 2665 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0); 2666 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).", 2667 cmThread()->vtime_accum(), cmThread()->vtime_mark_accum()); 2668 } 2669 2670 void ConcurrentMark::print_worker_threads_on(outputStream* st) const { 2671 _parallel_workers->print_worker_threads_on(st); 2672 } 2673 2674 void ConcurrentMark::print_on_error(outputStream* st) const { 2675 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT, 2676 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap)); 2677 _prevMarkBitMap->print_on_error(st, " Prev Bits: "); 2678 _nextMarkBitMap->print_on_error(st, " Next Bits: "); 2679 } 2680 2681 // We take a break if someone is trying to stop the world. 2682 bool ConcurrentMark::do_yield_check(uint worker_id) { 2683 if (SuspendibleThreadSet::should_yield()) { 2684 if (worker_id == 0) { 2685 _g1h->g1_policy()->record_concurrent_pause(); 2686 } 2687 SuspendibleThreadSet::yield(); 2688 return true; 2689 } else { 2690 return false; 2691 } 2692 } 2693 2694 // Closure for iteration over bitmaps 2695 class CMBitMapClosure : public BitMapClosure { 2696 private: 2697 // the bitmap that is being iterated over 2698 CMBitMap* _nextMarkBitMap; 2699 ConcurrentMark* _cm; 2700 CMTask* _task; 2701 2702 public: 2703 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) : 2704 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } 2705 2706 bool do_bit(size_t offset) { 2707 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); 2708 assert(_nextMarkBitMap->isMarked(addr), "invariant"); 2709 assert( addr < _cm->finger(), "invariant"); 2710 assert(addr >= _task->finger(), "invariant"); 2711 2712 // We move that task's local finger along. 2713 _task->move_finger_to(addr); 2714 2715 _task->scan_object(oop(addr)); 2716 // we only partially drain the local queue and global stack 2717 _task->drain_local_queue(true); 2718 _task->drain_global_stack(true); 2719 2720 // if the has_aborted flag has been raised, we need to bail out of 2721 // the iteration 2722 return !_task->has_aborted(); 2723 } 2724 }; 2725 2726 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) { 2727 ReferenceProcessor* result = NULL; 2728 if (G1UseConcMarkReferenceProcessing) { 2729 result = g1h->ref_processor_cm(); 2730 assert(result != NULL, "should not be NULL"); 2731 } 2732 return result; 2733 } 2734 2735 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 2736 ConcurrentMark* cm, 2737 CMTask* task) 2738 : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)), 2739 _g1h(g1h), _cm(cm), _task(task) 2740 { } 2741 2742 void CMTask::setup_for_region(HeapRegion* hr) { 2743 assert(hr != NULL, 2744 "claim_region() should have filtered out NULL regions"); 2745 _curr_region = hr; 2746 _finger = hr->bottom(); 2747 update_region_limit(); 2748 } 2749 2750 void CMTask::update_region_limit() { 2751 HeapRegion* hr = _curr_region; 2752 HeapWord* bottom = hr->bottom(); 2753 HeapWord* limit = hr->next_top_at_mark_start(); 2754 2755 if (limit == bottom) { 2756 // The region was collected underneath our feet. 2757 // We set the finger to bottom to ensure that the bitmap 2758 // iteration that will follow this will not do anything. 2759 // (this is not a condition that holds when we set the region up, 2760 // as the region is not supposed to be empty in the first place) 2761 _finger = bottom; 2762 } else if (limit >= _region_limit) { 2763 assert(limit >= _finger, "peace of mind"); 2764 } else { 2765 assert(limit < _region_limit, "only way to get here"); 2766 // This can happen under some pretty unusual circumstances. An 2767 // evacuation pause empties the region underneath our feet (NTAMS 2768 // at bottom). We then do some allocation in the region (NTAMS 2769 // stays at bottom), followed by the region being used as a GC 2770 // alloc region (NTAMS will move to top() and the objects 2771 // originally below it will be grayed). All objects now marked in 2772 // the region are explicitly grayed, if below the global finger, 2773 // and we do not need in fact to scan anything else. So, we simply 2774 // set _finger to be limit to ensure that the bitmap iteration 2775 // doesn't do anything. 2776 _finger = limit; 2777 } 2778 2779 _region_limit = limit; 2780 } 2781 2782 void CMTask::giveup_current_region() { 2783 assert(_curr_region != NULL, "invariant"); 2784 clear_region_fields(); 2785 } 2786 2787 void CMTask::clear_region_fields() { 2788 // Values for these three fields that indicate that we're not 2789 // holding on to a region. 2790 _curr_region = NULL; 2791 _finger = NULL; 2792 _region_limit = NULL; 2793 } 2794 2795 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 2796 if (cm_oop_closure == NULL) { 2797 assert(_cm_oop_closure != NULL, "invariant"); 2798 } else { 2799 assert(_cm_oop_closure == NULL, "invariant"); 2800 } 2801 _cm_oop_closure = cm_oop_closure; 2802 } 2803 2804 void CMTask::reset(CMBitMap* nextMarkBitMap) { 2805 guarantee(nextMarkBitMap != NULL, "invariant"); 2806 _nextMarkBitMap = nextMarkBitMap; 2807 clear_region_fields(); 2808 2809 _calls = 0; 2810 _elapsed_time_ms = 0.0; 2811 _termination_time_ms = 0.0; 2812 _termination_start_time_ms = 0.0; 2813 } 2814 2815 bool CMTask::should_exit_termination() { 2816 regular_clock_call(); 2817 // This is called when we are in the termination protocol. We should 2818 // quit if, for some reason, this task wants to abort or the global 2819 // stack is not empty (this means that we can get work from it). 2820 return !_cm->mark_stack_empty() || has_aborted(); 2821 } 2822 2823 void CMTask::reached_limit() { 2824 assert(_words_scanned >= _words_scanned_limit || 2825 _refs_reached >= _refs_reached_limit , 2826 "shouldn't have been called otherwise"); 2827 regular_clock_call(); 2828 } 2829 2830 void CMTask::regular_clock_call() { 2831 if (has_aborted()) return; 2832 2833 // First, we need to recalculate the words scanned and refs reached 2834 // limits for the next clock call. 2835 recalculate_limits(); 2836 2837 // During the regular clock call we do the following 2838 2839 // (1) If an overflow has been flagged, then we abort. 2840 if (_cm->has_overflown()) { 2841 set_has_aborted(); 2842 return; 2843 } 2844 2845 // If we are not concurrent (i.e. we're doing remark) we don't need 2846 // to check anything else. The other steps are only needed during 2847 // the concurrent marking phase. 2848 if (!concurrent()) return; 2849 2850 // (2) If marking has been aborted for Full GC, then we also abort. 2851 if (_cm->has_aborted()) { 2852 set_has_aborted(); 2853 return; 2854 } 2855 2856 double curr_time_ms = os::elapsedVTime() * 1000.0; 2857 2858 // (4) We check whether we should yield. If we have to, then we abort. 2859 if (SuspendibleThreadSet::should_yield()) { 2860 // We should yield. To do this we abort the task. The caller is 2861 // responsible for yielding. 2862 set_has_aborted(); 2863 return; 2864 } 2865 2866 // (5) We check whether we've reached our time quota. If we have, 2867 // then we abort. 2868 double elapsed_time_ms = curr_time_ms - _start_time_ms; 2869 if (elapsed_time_ms > _time_target_ms) { 2870 set_has_aborted(); 2871 _has_timed_out = true; 2872 return; 2873 } 2874 2875 // (6) Finally, we check whether there are enough completed STAB 2876 // buffers available for processing. If there are, we abort. 2877 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2878 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 2879 // we do need to process SATB buffers, we'll abort and restart 2880 // the marking task to do so 2881 set_has_aborted(); 2882 return; 2883 } 2884 } 2885 2886 void CMTask::recalculate_limits() { 2887 _real_words_scanned_limit = _words_scanned + words_scanned_period; 2888 _words_scanned_limit = _real_words_scanned_limit; 2889 2890 _real_refs_reached_limit = _refs_reached + refs_reached_period; 2891 _refs_reached_limit = _real_refs_reached_limit; 2892 } 2893 2894 void CMTask::decrease_limits() { 2895 // This is called when we believe that we're going to do an infrequent 2896 // operation which will increase the per byte scanned cost (i.e. move 2897 // entries to/from the global stack). It basically tries to decrease the 2898 // scanning limit so that the clock is called earlier. 2899 2900 _words_scanned_limit = _real_words_scanned_limit - 2901 3 * words_scanned_period / 4; 2902 _refs_reached_limit = _real_refs_reached_limit - 2903 3 * refs_reached_period / 4; 2904 } 2905 2906 void CMTask::move_entries_to_global_stack() { 2907 // local array where we'll store the entries that will be popped 2908 // from the local queue 2909 oop buffer[global_stack_transfer_size]; 2910 2911 int n = 0; 2912 oop obj; 2913 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { 2914 buffer[n] = obj; 2915 ++n; 2916 } 2917 2918 if (n > 0) { 2919 // we popped at least one entry from the local queue 2920 2921 if (!_cm->mark_stack_push(buffer, n)) { 2922 set_has_aborted(); 2923 } 2924 } 2925 2926 // this operation was quite expensive, so decrease the limits 2927 decrease_limits(); 2928 } 2929 2930 void CMTask::get_entries_from_global_stack() { 2931 // local array where we'll store the entries that will be popped 2932 // from the global stack. 2933 oop buffer[global_stack_transfer_size]; 2934 int n; 2935 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); 2936 assert(n <= global_stack_transfer_size, 2937 "we should not pop more than the given limit"); 2938 if (n > 0) { 2939 // yes, we did actually pop at least one entry 2940 for (int i = 0; i < n; ++i) { 2941 bool success = _task_queue->push(buffer[i]); 2942 // We only call this when the local queue is empty or under a 2943 // given target limit. So, we do not expect this push to fail. 2944 assert(success, "invariant"); 2945 } 2946 } 2947 2948 // this operation was quite expensive, so decrease the limits 2949 decrease_limits(); 2950 } 2951 2952 void CMTask::drain_local_queue(bool partially) { 2953 if (has_aborted()) return; 2954 2955 // Decide what the target size is, depending whether we're going to 2956 // drain it partially (so that other tasks can steal if they run out 2957 // of things to do) or totally (at the very end). 2958 size_t target_size; 2959 if (partially) { 2960 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 2961 } else { 2962 target_size = 0; 2963 } 2964 2965 if (_task_queue->size() > target_size) { 2966 oop obj; 2967 bool ret = _task_queue->pop_local(obj); 2968 while (ret) { 2969 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); 2970 assert(!_g1h->is_on_master_free_list( 2971 _g1h->heap_region_containing((HeapWord*) obj)), "invariant"); 2972 2973 scan_object(obj); 2974 2975 if (_task_queue->size() <= target_size || has_aborted()) { 2976 ret = false; 2977 } else { 2978 ret = _task_queue->pop_local(obj); 2979 } 2980 } 2981 } 2982 } 2983 2984 void CMTask::drain_global_stack(bool partially) { 2985 if (has_aborted()) return; 2986 2987 // We have a policy to drain the local queue before we attempt to 2988 // drain the global stack. 2989 assert(partially || _task_queue->size() == 0, "invariant"); 2990 2991 // Decide what the target size is, depending whether we're going to 2992 // drain it partially (so that other tasks can steal if they run out 2993 // of things to do) or totally (at the very end). Notice that, 2994 // because we move entries from the global stack in chunks or 2995 // because another task might be doing the same, we might in fact 2996 // drop below the target. But, this is not a problem. 2997 size_t target_size; 2998 if (partially) { 2999 target_size = _cm->partial_mark_stack_size_target(); 3000 } else { 3001 target_size = 0; 3002 } 3003 3004 if (_cm->mark_stack_size() > target_size) { 3005 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 3006 get_entries_from_global_stack(); 3007 drain_local_queue(partially); 3008 } 3009 } 3010 } 3011 3012 // SATB Queue has several assumptions on whether to call the par or 3013 // non-par versions of the methods. this is why some of the code is 3014 // replicated. We should really get rid of the single-threaded version 3015 // of the code to simplify things. 3016 void CMTask::drain_satb_buffers() { 3017 if (has_aborted()) return; 3018 3019 // We set this so that the regular clock knows that we're in the 3020 // middle of draining buffers and doesn't set the abort flag when it 3021 // notices that SATB buffers are available for draining. It'd be 3022 // very counter productive if it did that. :-) 3023 _draining_satb_buffers = true; 3024 3025 CMSATBBufferClosure satb_cl(this, _g1h); 3026 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3027 3028 // This keeps claiming and applying the closure to completed buffers 3029 // until we run out of buffers or we need to abort. 3030 while (!has_aborted() && 3031 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) { 3032 regular_clock_call(); 3033 } 3034 3035 _draining_satb_buffers = false; 3036 3037 assert(has_aborted() || 3038 concurrent() || 3039 satb_mq_set.completed_buffers_num() == 0, "invariant"); 3040 3041 // again, this was a potentially expensive operation, decrease the 3042 // limits to get the regular clock call early 3043 decrease_limits(); 3044 } 3045 3046 void CMTask::print_stats() { 3047 log_debug(gc, stats)("Marking Stats, task = %u, calls = %d", 3048 _worker_id, _calls); 3049 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 3050 _elapsed_time_ms, _termination_time_ms); 3051 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 3052 _step_times_ms.num(), _step_times_ms.avg(), 3053 _step_times_ms.sd()); 3054 log_debug(gc, stats)(" max = %1.2lfms, total = %1.2lfms", 3055 _step_times_ms.maximum(), _step_times_ms.sum()); 3056 } 3057 3058 bool ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) { 3059 return _task_queues->steal(worker_id, hash_seed, obj); 3060 } 3061 3062 /***************************************************************************** 3063 3064 The do_marking_step(time_target_ms, ...) method is the building 3065 block of the parallel marking framework. It can be called in parallel 3066 with other invocations of do_marking_step() on different tasks 3067 (but only one per task, obviously) and concurrently with the 3068 mutator threads, or during remark, hence it eliminates the need 3069 for two versions of the code. When called during remark, it will 3070 pick up from where the task left off during the concurrent marking 3071 phase. Interestingly, tasks are also claimable during evacuation 3072 pauses too, since do_marking_step() ensures that it aborts before 3073 it needs to yield. 3074 3075 The data structures that it uses to do marking work are the 3076 following: 3077 3078 (1) Marking Bitmap. If there are gray objects that appear only 3079 on the bitmap (this happens either when dealing with an overflow 3080 or when the initial marking phase has simply marked the roots 3081 and didn't push them on the stack), then tasks claim heap 3082 regions whose bitmap they then scan to find gray objects. A 3083 global finger indicates where the end of the last claimed region 3084 is. A local finger indicates how far into the region a task has 3085 scanned. The two fingers are used to determine how to gray an 3086 object (i.e. whether simply marking it is OK, as it will be 3087 visited by a task in the future, or whether it needs to be also 3088 pushed on a stack). 3089 3090 (2) Local Queue. The local queue of the task which is accessed 3091 reasonably efficiently by the task. Other tasks can steal from 3092 it when they run out of work. Throughout the marking phase, a 3093 task attempts to keep its local queue short but not totally 3094 empty, so that entries are available for stealing by other 3095 tasks. Only when there is no more work, a task will totally 3096 drain its local queue. 3097 3098 (3) Global Mark Stack. This handles local queue overflow. During 3099 marking only sets of entries are moved between it and the local 3100 queues, as access to it requires a mutex and more fine-grain 3101 interaction with it which might cause contention. If it 3102 overflows, then the marking phase should restart and iterate 3103 over the bitmap to identify gray objects. Throughout the marking 3104 phase, tasks attempt to keep the global mark stack at a small 3105 length but not totally empty, so that entries are available for 3106 popping by other tasks. Only when there is no more work, tasks 3107 will totally drain the global mark stack. 3108 3109 (4) SATB Buffer Queue. This is where completed SATB buffers are 3110 made available. Buffers are regularly removed from this queue 3111 and scanned for roots, so that the queue doesn't get too 3112 long. During remark, all completed buffers are processed, as 3113 well as the filled in parts of any uncompleted buffers. 3114 3115 The do_marking_step() method tries to abort when the time target 3116 has been reached. There are a few other cases when the 3117 do_marking_step() method also aborts: 3118 3119 (1) When the marking phase has been aborted (after a Full GC). 3120 3121 (2) When a global overflow (on the global stack) has been 3122 triggered. Before the task aborts, it will actually sync up with 3123 the other tasks to ensure that all the marking data structures 3124 (local queues, stacks, fingers etc.) are re-initialized so that 3125 when do_marking_step() completes, the marking phase can 3126 immediately restart. 3127 3128 (3) When enough completed SATB buffers are available. The 3129 do_marking_step() method only tries to drain SATB buffers right 3130 at the beginning. So, if enough buffers are available, the 3131 marking step aborts and the SATB buffers are processed at 3132 the beginning of the next invocation. 3133 3134 (4) To yield. when we have to yield then we abort and yield 3135 right at the end of do_marking_step(). This saves us from a lot 3136 of hassle as, by yielding we might allow a Full GC. If this 3137 happens then objects will be compacted underneath our feet, the 3138 heap might shrink, etc. We save checking for this by just 3139 aborting and doing the yield right at the end. 3140 3141 From the above it follows that the do_marking_step() method should 3142 be called in a loop (or, otherwise, regularly) until it completes. 3143 3144 If a marking step completes without its has_aborted() flag being 3145 true, it means it has completed the current marking phase (and 3146 also all other marking tasks have done so and have all synced up). 3147 3148 A method called regular_clock_call() is invoked "regularly" (in 3149 sub ms intervals) throughout marking. It is this clock method that 3150 checks all the abort conditions which were mentioned above and 3151 decides when the task should abort. A work-based scheme is used to 3152 trigger this clock method: when the number of object words the 3153 marking phase has scanned or the number of references the marking 3154 phase has visited reach a given limit. Additional invocations to 3155 the method clock have been planted in a few other strategic places 3156 too. The initial reason for the clock method was to avoid calling 3157 vtime too regularly, as it is quite expensive. So, once it was in 3158 place, it was natural to piggy-back all the other conditions on it 3159 too and not constantly check them throughout the code. 3160 3161 If do_termination is true then do_marking_step will enter its 3162 termination protocol. 3163 3164 The value of is_serial must be true when do_marking_step is being 3165 called serially (i.e. by the VMThread) and do_marking_step should 3166 skip any synchronization in the termination and overflow code. 3167 Examples include the serial remark code and the serial reference 3168 processing closures. 3169 3170 The value of is_serial must be false when do_marking_step is 3171 being called by any of the worker threads in a work gang. 3172 Examples include the concurrent marking code (CMMarkingTask), 3173 the MT remark code, and the MT reference processing closures. 3174 3175 *****************************************************************************/ 3176 3177 void CMTask::do_marking_step(double time_target_ms, 3178 bool do_termination, 3179 bool is_serial) { 3180 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 3181 assert(concurrent() == _cm->concurrent(), "they should be the same"); 3182 3183 G1CollectorPolicy* g1_policy = _g1h->g1_policy(); 3184 assert(_task_queues != NULL, "invariant"); 3185 assert(_task_queue != NULL, "invariant"); 3186 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant"); 3187 3188 assert(!_claimed, 3189 "only one thread should claim this task at any one time"); 3190 3191 // OK, this doesn't safeguard again all possible scenarios, as it is 3192 // possible for two threads to set the _claimed flag at the same 3193 // time. But it is only for debugging purposes anyway and it will 3194 // catch most problems. 3195 _claimed = true; 3196 3197 _start_time_ms = os::elapsedVTime() * 1000.0; 3198 3199 // If do_stealing is true then do_marking_step will attempt to 3200 // steal work from the other CMTasks. It only makes sense to 3201 // enable stealing when the termination protocol is enabled 3202 // and do_marking_step() is not being called serially. 3203 bool do_stealing = do_termination && !is_serial; 3204 3205 double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms); 3206 _time_target_ms = time_target_ms - diff_prediction_ms; 3207 3208 // set up the variables that are used in the work-based scheme to 3209 // call the regular clock method 3210 _words_scanned = 0; 3211 _refs_reached = 0; 3212 recalculate_limits(); 3213 3214 // clear all flags 3215 clear_has_aborted(); 3216 _has_timed_out = false; 3217 _draining_satb_buffers = false; 3218 3219 ++_calls; 3220 3221 // Set up the bitmap and oop closures. Anything that uses them is 3222 // eventually called from this method, so it is OK to allocate these 3223 // statically. 3224 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); 3225 G1CMOopClosure cm_oop_closure(_g1h, _cm, this); 3226 set_cm_oop_closure(&cm_oop_closure); 3227 3228 if (_cm->has_overflown()) { 3229 // This can happen if the mark stack overflows during a GC pause 3230 // and this task, after a yield point, restarts. We have to abort 3231 // as we need to get into the overflow protocol which happens 3232 // right at the end of this task. 3233 set_has_aborted(); 3234 } 3235 3236 // First drain any available SATB buffers. After this, we will not 3237 // look at SATB buffers before the next invocation of this method. 3238 // If enough completed SATB buffers are queued up, the regular clock 3239 // will abort this task so that it restarts. 3240 drain_satb_buffers(); 3241 // ...then partially drain the local queue and the global stack 3242 drain_local_queue(true); 3243 drain_global_stack(true); 3244 3245 do { 3246 if (!has_aborted() && _curr_region != NULL) { 3247 // This means that we're already holding on to a region. 3248 assert(_finger != NULL, "if region is not NULL, then the finger " 3249 "should not be NULL either"); 3250 3251 // We might have restarted this task after an evacuation pause 3252 // which might have evacuated the region we're holding on to 3253 // underneath our feet. Let's read its limit again to make sure 3254 // that we do not iterate over a region of the heap that 3255 // contains garbage (update_region_limit() will also move 3256 // _finger to the start of the region if it is found empty). 3257 update_region_limit(); 3258 // We will start from _finger not from the start of the region, 3259 // as we might be restarting this task after aborting half-way 3260 // through scanning this region. In this case, _finger points to 3261 // the address where we last found a marked object. If this is a 3262 // fresh region, _finger points to start(). 3263 MemRegion mr = MemRegion(_finger, _region_limit); 3264 3265 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(), 3266 "humongous regions should go around loop once only"); 3267 3268 // Some special cases: 3269 // If the memory region is empty, we can just give up the region. 3270 // If the current region is humongous then we only need to check 3271 // the bitmap for the bit associated with the start of the object, 3272 // scan the object if it's live, and give up the region. 3273 // Otherwise, let's iterate over the bitmap of the part of the region 3274 // that is left. 3275 // If the iteration is successful, give up the region. 3276 if (mr.is_empty()) { 3277 giveup_current_region(); 3278 regular_clock_call(); 3279 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) { 3280 if (_nextMarkBitMap->isMarked(mr.start())) { 3281 // The object is marked - apply the closure 3282 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start()); 3283 bitmap_closure.do_bit(offset); 3284 } 3285 // Even if this task aborted while scanning the humongous object 3286 // we can (and should) give up the current region. 3287 giveup_current_region(); 3288 regular_clock_call(); 3289 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) { 3290 giveup_current_region(); 3291 regular_clock_call(); 3292 } else { 3293 assert(has_aborted(), "currently the only way to do so"); 3294 // The only way to abort the bitmap iteration is to return 3295 // false from the do_bit() method. However, inside the 3296 // do_bit() method we move the _finger to point to the 3297 // object currently being looked at. So, if we bail out, we 3298 // have definitely set _finger to something non-null. 3299 assert(_finger != NULL, "invariant"); 3300 3301 // Region iteration was actually aborted. So now _finger 3302 // points to the address of the object we last scanned. If we 3303 // leave it there, when we restart this task, we will rescan 3304 // the object. It is easy to avoid this. We move the finger by 3305 // enough to point to the next possible object header (the 3306 // bitmap knows by how much we need to move it as it knows its 3307 // granularity). 3308 assert(_finger < _region_limit, "invariant"); 3309 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger); 3310 // Check if bitmap iteration was aborted while scanning the last object 3311 if (new_finger >= _region_limit) { 3312 giveup_current_region(); 3313 } else { 3314 move_finger_to(new_finger); 3315 } 3316 } 3317 } 3318 // At this point we have either completed iterating over the 3319 // region we were holding on to, or we have aborted. 3320 3321 // We then partially drain the local queue and the global stack. 3322 // (Do we really need this?) 3323 drain_local_queue(true); 3324 drain_global_stack(true); 3325 3326 // Read the note on the claim_region() method on why it might 3327 // return NULL with potentially more regions available for 3328 // claiming and why we have to check out_of_regions() to determine 3329 // whether we're done or not. 3330 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 3331 // We are going to try to claim a new region. We should have 3332 // given up on the previous one. 3333 // Separated the asserts so that we know which one fires. 3334 assert(_curr_region == NULL, "invariant"); 3335 assert(_finger == NULL, "invariant"); 3336 assert(_region_limit == NULL, "invariant"); 3337 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 3338 if (claimed_region != NULL) { 3339 // Yes, we managed to claim one 3340 setup_for_region(claimed_region); 3341 assert(_curr_region == claimed_region, "invariant"); 3342 } 3343 // It is important to call the regular clock here. It might take 3344 // a while to claim a region if, for example, we hit a large 3345 // block of empty regions. So we need to call the regular clock 3346 // method once round the loop to make sure it's called 3347 // frequently enough. 3348 regular_clock_call(); 3349 } 3350 3351 if (!has_aborted() && _curr_region == NULL) { 3352 assert(_cm->out_of_regions(), 3353 "at this point we should be out of regions"); 3354 } 3355 } while ( _curr_region != NULL && !has_aborted()); 3356 3357 if (!has_aborted()) { 3358 // We cannot check whether the global stack is empty, since other 3359 // tasks might be pushing objects to it concurrently. 3360 assert(_cm->out_of_regions(), 3361 "at this point we should be out of regions"); 3362 // Try to reduce the number of available SATB buffers so that 3363 // remark has less work to do. 3364 drain_satb_buffers(); 3365 } 3366 3367 // Since we've done everything else, we can now totally drain the 3368 // local queue and global stack. 3369 drain_local_queue(false); 3370 drain_global_stack(false); 3371 3372 // Attempt at work stealing from other task's queues. 3373 if (do_stealing && !has_aborted()) { 3374 // We have not aborted. This means that we have finished all that 3375 // we could. Let's try to do some stealing... 3376 3377 // We cannot check whether the global stack is empty, since other 3378 // tasks might be pushing objects to it concurrently. 3379 assert(_cm->out_of_regions() && _task_queue->size() == 0, 3380 "only way to reach here"); 3381 while (!has_aborted()) { 3382 oop obj; 3383 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) { 3384 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), 3385 "any stolen object should be marked"); 3386 scan_object(obj); 3387 3388 // And since we're towards the end, let's totally drain the 3389 // local queue and global stack. 3390 drain_local_queue(false); 3391 drain_global_stack(false); 3392 } else { 3393 break; 3394 } 3395 } 3396 } 3397 3398 // We still haven't aborted. Now, let's try to get into the 3399 // termination protocol. 3400 if (do_termination && !has_aborted()) { 3401 // We cannot check whether the global stack is empty, since other 3402 // tasks might be concurrently pushing objects on it. 3403 // Separated the asserts so that we know which one fires. 3404 assert(_cm->out_of_regions(), "only way to reach here"); 3405 assert(_task_queue->size() == 0, "only way to reach here"); 3406 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 3407 3408 // The CMTask class also extends the TerminatorTerminator class, 3409 // hence its should_exit_termination() method will also decide 3410 // whether to exit the termination protocol or not. 3411 bool finished = (is_serial || 3412 _cm->terminator()->offer_termination(this)); 3413 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 3414 _termination_time_ms += 3415 termination_end_time_ms - _termination_start_time_ms; 3416 3417 if (finished) { 3418 // We're all done. 3419 3420 if (_worker_id == 0) { 3421 // let's allow task 0 to do this 3422 if (concurrent()) { 3423 assert(_cm->concurrent_marking_in_progress(), "invariant"); 3424 // we need to set this to false before the next 3425 // safepoint. This way we ensure that the marking phase 3426 // doesn't observe any more heap expansions. 3427 _cm->clear_concurrent_marking_in_progress(); 3428 } 3429 } 3430 3431 // We can now guarantee that the global stack is empty, since 3432 // all other tasks have finished. We separated the guarantees so 3433 // that, if a condition is false, we can immediately find out 3434 // which one. 3435 guarantee(_cm->out_of_regions(), "only way to reach here"); 3436 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 3437 guarantee(_task_queue->size() == 0, "only way to reach here"); 3438 guarantee(!_cm->has_overflown(), "only way to reach here"); 3439 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); 3440 } else { 3441 // Apparently there's more work to do. Let's abort this task. It 3442 // will restart it and we can hopefully find more things to do. 3443 set_has_aborted(); 3444 } 3445 } 3446 3447 // Mainly for debugging purposes to make sure that a pointer to the 3448 // closure which was statically allocated in this frame doesn't 3449 // escape it by accident. 3450 set_cm_oop_closure(NULL); 3451 double end_time_ms = os::elapsedVTime() * 1000.0; 3452 double elapsed_time_ms = end_time_ms - _start_time_ms; 3453 // Update the step history. 3454 _step_times_ms.add(elapsed_time_ms); 3455 3456 if (has_aborted()) { 3457 // The task was aborted for some reason. 3458 if (_has_timed_out) { 3459 double diff_ms = elapsed_time_ms - _time_target_ms; 3460 // Keep statistics of how well we did with respect to hitting 3461 // our target only if we actually timed out (if we aborted for 3462 // other reasons, then the results might get skewed). 3463 _marking_step_diffs_ms.add(diff_ms); 3464 } 3465 3466 if (_cm->has_overflown()) { 3467 // This is the interesting one. We aborted because a global 3468 // overflow was raised. This means we have to restart the 3469 // marking phase and start iterating over regions. However, in 3470 // order to do this we have to make sure that all tasks stop 3471 // what they are doing and re-initialize in a safe manner. We 3472 // will achieve this with the use of two barrier sync points. 3473 3474 if (!is_serial) { 3475 // We only need to enter the sync barrier if being called 3476 // from a parallel context 3477 _cm->enter_first_sync_barrier(_worker_id); 3478 3479 // When we exit this sync barrier we know that all tasks have 3480 // stopped doing marking work. So, it's now safe to 3481 // re-initialize our data structures. At the end of this method, 3482 // task 0 will clear the global data structures. 3483 } 3484 3485 // We clear the local state of this task... 3486 clear_region_fields(); 3487 3488 if (!is_serial) { 3489 // ...and enter the second barrier. 3490 _cm->enter_second_sync_barrier(_worker_id); 3491 } 3492 // At this point, if we're during the concurrent phase of 3493 // marking, everything has been re-initialized and we're 3494 // ready to restart. 3495 } 3496 } 3497 3498 _claimed = false; 3499 } 3500 3501 CMTask::CMTask(uint worker_id, 3502 ConcurrentMark* cm, 3503 size_t* marked_bytes, 3504 BitMap* card_bm, 3505 CMTaskQueue* task_queue, 3506 CMTaskQueueSet* task_queues) 3507 : _g1h(G1CollectedHeap::heap()), 3508 _worker_id(worker_id), _cm(cm), 3509 _claimed(false), 3510 _nextMarkBitMap(NULL), _hash_seed(17), 3511 _task_queue(task_queue), 3512 _task_queues(task_queues), 3513 _cm_oop_closure(NULL), 3514 _marked_bytes_array(marked_bytes), 3515 _card_bm(card_bm) { 3516 guarantee(task_queue != NULL, "invariant"); 3517 guarantee(task_queues != NULL, "invariant"); 3518 3519 _marking_step_diffs_ms.add(0.5); 3520 } 3521 3522 // These are formatting macros that are used below to ensure 3523 // consistent formatting. The *_H_* versions are used to format the 3524 // header for a particular value and they should be kept consistent 3525 // with the corresponding macro. Also note that most of the macros add 3526 // the necessary white space (as a prefix) which makes them a bit 3527 // easier to compose. 3528 3529 // All the output lines are prefixed with this string to be able to 3530 // identify them easily in a large log file. 3531 #define G1PPRL_LINE_PREFIX "###" 3532 3533 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT 3534 #ifdef _LP64 3535 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 3536 #else // _LP64 3537 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 3538 #endif // _LP64 3539 3540 // For per-region info 3541 #define G1PPRL_TYPE_FORMAT " %-4s" 3542 #define G1PPRL_TYPE_H_FORMAT " %4s" 3543 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9) 3544 #define G1PPRL_BYTE_H_FORMAT " %9s" 3545 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 3546 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 3547 3548 // For summary info 3549 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT 3550 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT 3551 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB" 3552 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%" 3553 3554 G1PrintRegionLivenessInfoClosure:: 3555 G1PrintRegionLivenessInfoClosure(const char* phase_name) 3556 : _total_used_bytes(0), _total_capacity_bytes(0), 3557 _total_prev_live_bytes(0), _total_next_live_bytes(0), 3558 _hum_used_bytes(0), _hum_capacity_bytes(0), 3559 _hum_prev_live_bytes(0), _hum_next_live_bytes(0), 3560 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) { 3561 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 3562 MemRegion g1_reserved = g1h->g1_reserved(); 3563 double now = os::elapsedTime(); 3564 3565 // Print the header of the output. 3566 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 3567 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP" 3568 G1PPRL_SUM_ADDR_FORMAT("reserved") 3569 G1PPRL_SUM_BYTE_FORMAT("region-size"), 3570 p2i(g1_reserved.start()), p2i(g1_reserved.end()), 3571 HeapRegion::GrainBytes); 3572 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 3573 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3574 G1PPRL_TYPE_H_FORMAT 3575 G1PPRL_ADDR_BASE_H_FORMAT 3576 G1PPRL_BYTE_H_FORMAT 3577 G1PPRL_BYTE_H_FORMAT 3578 G1PPRL_BYTE_H_FORMAT 3579 G1PPRL_DOUBLE_H_FORMAT 3580 G1PPRL_BYTE_H_FORMAT 3581 G1PPRL_BYTE_H_FORMAT, 3582 "type", "address-range", 3583 "used", "prev-live", "next-live", "gc-eff", 3584 "remset", "code-roots"); 3585 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3586 G1PPRL_TYPE_H_FORMAT 3587 G1PPRL_ADDR_BASE_H_FORMAT 3588 G1PPRL_BYTE_H_FORMAT 3589 G1PPRL_BYTE_H_FORMAT 3590 G1PPRL_BYTE_H_FORMAT 3591 G1PPRL_DOUBLE_H_FORMAT 3592 G1PPRL_BYTE_H_FORMAT 3593 G1PPRL_BYTE_H_FORMAT, 3594 "", "", 3595 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", 3596 "(bytes)", "(bytes)"); 3597 } 3598 3599 // It takes as a parameter a reference to one of the _hum_* fields, it 3600 // deduces the corresponding value for a region in a humongous region 3601 // series (either the region size, or what's left if the _hum_* field 3602 // is < the region size), and updates the _hum_* field accordingly. 3603 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) { 3604 size_t bytes = 0; 3605 // The > 0 check is to deal with the prev and next live bytes which 3606 // could be 0. 3607 if (*hum_bytes > 0) { 3608 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes); 3609 *hum_bytes -= bytes; 3610 } 3611 return bytes; 3612 } 3613 3614 // It deduces the values for a region in a humongous region series 3615 // from the _hum_* fields and updates those accordingly. It assumes 3616 // that that _hum_* fields have already been set up from the "starts 3617 // humongous" region and we visit the regions in address order. 3618 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes, 3619 size_t* capacity_bytes, 3620 size_t* prev_live_bytes, 3621 size_t* next_live_bytes) { 3622 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition"); 3623 *used_bytes = get_hum_bytes(&_hum_used_bytes); 3624 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes); 3625 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes); 3626 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes); 3627 } 3628 3629 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) { 3630 const char* type = r->get_type_str(); 3631 HeapWord* bottom = r->bottom(); 3632 HeapWord* end = r->end(); 3633 size_t capacity_bytes = r->capacity(); 3634 size_t used_bytes = r->used(); 3635 size_t prev_live_bytes = r->live_bytes(); 3636 size_t next_live_bytes = r->next_live_bytes(); 3637 double gc_eff = r->gc_efficiency(); 3638 size_t remset_bytes = r->rem_set()->mem_size(); 3639 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size(); 3640 3641 if (r->is_starts_humongous()) { 3642 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 && 3643 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0, 3644 "they should have been zeroed after the last time we used them"); 3645 // Set up the _hum_* fields. 3646 _hum_capacity_bytes = capacity_bytes; 3647 _hum_used_bytes = used_bytes; 3648 _hum_prev_live_bytes = prev_live_bytes; 3649 _hum_next_live_bytes = next_live_bytes; 3650 get_hum_bytes(&used_bytes, &capacity_bytes, 3651 &prev_live_bytes, &next_live_bytes); 3652 end = bottom + HeapRegion::GrainWords; 3653 } else if (r->is_continues_humongous()) { 3654 get_hum_bytes(&used_bytes, &capacity_bytes, 3655 &prev_live_bytes, &next_live_bytes); 3656 assert(end == bottom + HeapRegion::GrainWords, "invariant"); 3657 } 3658 3659 _total_used_bytes += used_bytes; 3660 _total_capacity_bytes += capacity_bytes; 3661 _total_prev_live_bytes += prev_live_bytes; 3662 _total_next_live_bytes += next_live_bytes; 3663 _total_remset_bytes += remset_bytes; 3664 _total_strong_code_roots_bytes += strong_code_roots_bytes; 3665 3666 // Print a line for this particular region. 3667 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3668 G1PPRL_TYPE_FORMAT 3669 G1PPRL_ADDR_BASE_FORMAT 3670 G1PPRL_BYTE_FORMAT 3671 G1PPRL_BYTE_FORMAT 3672 G1PPRL_BYTE_FORMAT 3673 G1PPRL_DOUBLE_FORMAT 3674 G1PPRL_BYTE_FORMAT 3675 G1PPRL_BYTE_FORMAT, 3676 type, p2i(bottom), p2i(end), 3677 used_bytes, prev_live_bytes, next_live_bytes, gc_eff, 3678 remset_bytes, strong_code_roots_bytes); 3679 3680 return false; 3681 } 3682 3683 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 3684 // add static memory usages to remembered set sizes 3685 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size(); 3686 // Print the footer of the output. 3687 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 3688 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3689 " SUMMARY" 3690 G1PPRL_SUM_MB_FORMAT("capacity") 3691 G1PPRL_SUM_MB_PERC_FORMAT("used") 3692 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 3693 G1PPRL_SUM_MB_PERC_FORMAT("next-live") 3694 G1PPRL_SUM_MB_FORMAT("remset") 3695 G1PPRL_SUM_MB_FORMAT("code-roots"), 3696 bytes_to_mb(_total_capacity_bytes), 3697 bytes_to_mb(_total_used_bytes), 3698 perc(_total_used_bytes, _total_capacity_bytes), 3699 bytes_to_mb(_total_prev_live_bytes), 3700 perc(_total_prev_live_bytes, _total_capacity_bytes), 3701 bytes_to_mb(_total_next_live_bytes), 3702 perc(_total_next_live_bytes, _total_capacity_bytes), 3703 bytes_to_mb(_total_remset_bytes), 3704 bytes_to_mb(_total_strong_code_roots_bytes)); 3705 }