1 /* 2 * Copyright (c) 2001, 2013, 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/symbolTable.hpp" 27 #include "gc_implementation/g1/concurrentMark.inline.hpp" 28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 30 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 31 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 32 #include "gc_implementation/g1/g1Log.hpp" 33 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 34 #include "gc_implementation/g1/g1RemSet.hpp" 35 #include "gc_implementation/g1/heapRegion.inline.hpp" 36 #include "gc_implementation/g1/heapRegionRemSet.hpp" 37 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 38 #include "gc_implementation/shared/vmGCOperations.hpp" 39 #include "gc_implementation/shared/gcTimer.hpp" 40 #include "gc_implementation/shared/gcTrace.hpp" 41 #include "gc_implementation/shared/gcTraceTime.hpp" 42 #include "memory/genOopClosures.inline.hpp" 43 #include "memory/referencePolicy.hpp" 44 #include "memory/resourceArea.hpp" 45 #include "oops/oop.inline.hpp" 46 #include "runtime/handles.inline.hpp" 47 #include "runtime/java.hpp" 48 #include "services/memTracker.hpp" 49 50 // Concurrent marking bit map wrapper 51 52 CMBitMapRO::CMBitMapRO(ReservedSpace rs, int shifter) : 53 _bm((uintptr_t*)NULL,0), 54 _shifter(shifter) { 55 _bmStartWord = (HeapWord*)(rs.base()); 56 _bmWordSize = rs.size()/HeapWordSize; // rs.size() is in bytes 57 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 58 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 59 60 MemTracker::record_virtual_memory_type((address)brs.base(), mtGC); 61 62 guarantee(brs.is_reserved(), "couldn't allocate concurrent marking bit map"); 63 // For now we'll just commit all of the bit map up fromt. 64 // Later on we'll try to be more parsimonious with swap. 65 guarantee(_virtual_space.initialize(brs, brs.size()), 66 "couldn't reseve backing store for concurrent marking bit map"); 67 assert(_virtual_space.committed_size() == brs.size(), 68 "didn't reserve backing store for all of concurrent marking bit map?"); 69 _bm.set_map((uintptr_t*)_virtual_space.low()); 70 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 71 _bmWordSize, "inconsistency in bit map sizing"); 72 _bm.set_size(_bmWordSize >> _shifter); 73 } 74 75 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr, 76 HeapWord* limit) const { 77 // First we must round addr *up* to a possible object boundary. 78 addr = (HeapWord*)align_size_up((intptr_t)addr, 79 HeapWordSize << _shifter); 80 size_t addrOffset = heapWordToOffset(addr); 81 if (limit == NULL) { 82 limit = _bmStartWord + _bmWordSize; 83 } 84 size_t limitOffset = heapWordToOffset(limit); 85 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); 86 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 87 assert(nextAddr >= addr, "get_next_one postcondition"); 88 assert(nextAddr == limit || isMarked(nextAddr), 89 "get_next_one postcondition"); 90 return nextAddr; 91 } 92 93 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr, 94 HeapWord* limit) const { 95 size_t addrOffset = heapWordToOffset(addr); 96 if (limit == NULL) { 97 limit = _bmStartWord + _bmWordSize; 98 } 99 size_t limitOffset = heapWordToOffset(limit); 100 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset); 101 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 102 assert(nextAddr >= addr, "get_next_one postcondition"); 103 assert(nextAddr == limit || !isMarked(nextAddr), 104 "get_next_one postcondition"); 105 return nextAddr; 106 } 107 108 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const { 109 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check"); 110 return (int) (diff >> _shifter); 111 } 112 113 #ifndef PRODUCT 114 bool CMBitMapRO::covers(ReservedSpace rs) const { 115 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 116 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize, 117 "size inconsistency"); 118 return _bmStartWord == (HeapWord*)(rs.base()) && 119 _bmWordSize == rs.size()>>LogHeapWordSize; 120 } 121 #endif 122 123 void CMBitMap::clearAll() { 124 _bm.clear(); 125 return; 126 } 127 128 void CMBitMap::markRange(MemRegion mr) { 129 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 130 assert(!mr.is_empty(), "unexpected empty region"); 131 assert((offsetToHeapWord(heapWordToOffset(mr.end())) == 132 ((HeapWord *) mr.end())), 133 "markRange memory region end is not card aligned"); 134 // convert address range into offset range 135 _bm.at_put_range(heapWordToOffset(mr.start()), 136 heapWordToOffset(mr.end()), true); 137 } 138 139 void CMBitMap::clearRange(MemRegion mr) { 140 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 141 assert(!mr.is_empty(), "unexpected empty region"); 142 // convert address range into offset range 143 _bm.at_put_range(heapWordToOffset(mr.start()), 144 heapWordToOffset(mr.end()), false); 145 } 146 147 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr, 148 HeapWord* end_addr) { 149 HeapWord* start = getNextMarkedWordAddress(addr); 150 start = MIN2(start, end_addr); 151 HeapWord* end = getNextUnmarkedWordAddress(start); 152 end = MIN2(end, end_addr); 153 assert(start <= end, "Consistency check"); 154 MemRegion mr(start, end); 155 if (!mr.is_empty()) { 156 clearRange(mr); 157 } 158 return mr; 159 } 160 161 CMMarkStack::CMMarkStack(ConcurrentMark* cm) : 162 _base(NULL), _cm(cm) 163 #ifdef ASSERT 164 , _drain_in_progress(false) 165 , _drain_in_progress_yields(false) 166 #endif 167 {} 168 169 void CMMarkStack::allocate(size_t size) { 170 _base = NEW_C_HEAP_ARRAY(oop, size, mtGC); 171 if (_base == NULL) { 172 vm_exit_during_initialization("Failed to allocate CM region mark stack"); 173 } 174 _index = 0; 175 _capacity = (jint) size; 176 _saved_index = -1; 177 NOT_PRODUCT(_max_depth = 0); 178 } 179 180 CMMarkStack::~CMMarkStack() { 181 if (_base != NULL) { 182 FREE_C_HEAP_ARRAY(oop, _base, mtGC); 183 } 184 } 185 186 void CMMarkStack::par_push(oop ptr) { 187 while (true) { 188 if (isFull()) { 189 _overflow = true; 190 return; 191 } 192 // Otherwise... 193 jint index = _index; 194 jint next_index = index+1; 195 jint res = Atomic::cmpxchg(next_index, &_index, index); 196 if (res == index) { 197 _base[index] = ptr; 198 // Note that we don't maintain this atomically. We could, but it 199 // doesn't seem necessary. 200 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 201 return; 202 } 203 // Otherwise, we need to try again. 204 } 205 } 206 207 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) { 208 while (true) { 209 if (isFull()) { 210 _overflow = true; 211 return; 212 } 213 // Otherwise... 214 jint index = _index; 215 jint next_index = index + n; 216 if (next_index > _capacity) { 217 _overflow = true; 218 return; 219 } 220 jint res = Atomic::cmpxchg(next_index, &_index, index); 221 if (res == index) { 222 for (int i = 0; i < n; i++) { 223 int ind = index + i; 224 assert(ind < _capacity, "By overflow test above."); 225 _base[ind] = ptr_arr[i]; 226 } 227 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 228 return; 229 } 230 // Otherwise, we need to try again. 231 } 232 } 233 234 235 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { 236 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 237 jint start = _index; 238 jint next_index = start + n; 239 if (next_index > _capacity) { 240 _overflow = true; 241 return; 242 } 243 // Otherwise. 244 _index = next_index; 245 for (int i = 0; i < n; i++) { 246 int ind = start + i; 247 assert(ind < _capacity, "By overflow test above."); 248 _base[ind] = ptr_arr[i]; 249 } 250 } 251 252 253 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { 254 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 255 jint index = _index; 256 if (index == 0) { 257 *n = 0; 258 return false; 259 } else { 260 int k = MIN2(max, index); 261 jint new_ind = index - k; 262 for (int j = 0; j < k; j++) { 263 ptr_arr[j] = _base[new_ind + j]; 264 } 265 _index = new_ind; 266 *n = k; 267 return true; 268 } 269 } 270 271 template<class OopClosureClass> 272 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) { 273 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after 274 || SafepointSynchronize::is_at_safepoint(), 275 "Drain recursion must be yield-safe."); 276 bool res = true; 277 debug_only(_drain_in_progress = true); 278 debug_only(_drain_in_progress_yields = yield_after); 279 while (!isEmpty()) { 280 oop newOop = pop(); 281 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop"); 282 assert(newOop->is_oop(), "Expected an oop"); 283 assert(bm == NULL || bm->isMarked((HeapWord*)newOop), 284 "only grey objects on this stack"); 285 newOop->oop_iterate(cl); 286 if (yield_after && _cm->do_yield_check()) { 287 res = false; 288 break; 289 } 290 } 291 debug_only(_drain_in_progress = false); 292 return res; 293 } 294 295 void CMMarkStack::note_start_of_gc() { 296 assert(_saved_index == -1, 297 "note_start_of_gc()/end_of_gc() bracketed incorrectly"); 298 _saved_index = _index; 299 } 300 301 void CMMarkStack::note_end_of_gc() { 302 // This is intentionally a guarantee, instead of an assert. If we 303 // accidentally add something to the mark stack during GC, it 304 // will be a correctness issue so it's better if we crash. we'll 305 // only check this once per GC anyway, so it won't be a performance 306 // issue in any way. 307 guarantee(_saved_index == _index, 308 err_msg("saved index: %d index: %d", _saved_index, _index)); 309 _saved_index = -1; 310 } 311 312 void CMMarkStack::oops_do(OopClosure* f) { 313 assert(_saved_index == _index, 314 err_msg("saved index: %d index: %d", _saved_index, _index)); 315 for (int i = 0; i < _index; i += 1) { 316 f->do_oop(&_base[i]); 317 } 318 } 319 320 bool ConcurrentMark::not_yet_marked(oop obj) const { 321 return (_g1h->is_obj_ill(obj) 322 || (_g1h->is_in_permanent(obj) 323 && !nextMarkBitMap()->isMarked((HeapWord*)obj))); 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 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 405 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 406 #endif // _MSC_VER 407 408 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 409 return MAX2((n_par_threads + 2) / 4, 1U); 410 } 411 412 ConcurrentMark::ConcurrentMark(ReservedSpace rs, uint max_regions) : 413 _markBitMap1(rs, MinObjAlignment - 1), 414 _markBitMap2(rs, MinObjAlignment - 1), 415 416 _parallel_marking_threads(0), 417 _max_parallel_marking_threads(0), 418 _sleep_factor(0.0), 419 _marking_task_overhead(1.0), 420 _cleanup_sleep_factor(0.0), 421 _cleanup_task_overhead(1.0), 422 _cleanup_list("Cleanup List"), 423 _region_bm((BitMap::idx_t) max_regions, false /* in_resource_area*/), 424 _card_bm((rs.size() + CardTableModRefBS::card_size - 1) >> 425 CardTableModRefBS::card_shift, 426 false /* in_resource_area*/), 427 428 _prevMarkBitMap(&_markBitMap1), 429 _nextMarkBitMap(&_markBitMap2), 430 431 _markStack(this), 432 // _finger set in set_non_marking_state 433 434 _max_task_num(MAX2((uint)ParallelGCThreads, 1U)), 435 // _active_tasks set in set_non_marking_state 436 // _tasks set inside the constructor 437 _task_queues(new CMTaskQueueSet((int) _max_task_num)), 438 _terminator(ParallelTaskTerminator((int) _max_task_num, _task_queues)), 439 440 _has_overflown(false), 441 _concurrent(false), 442 _has_aborted(false), 443 _restart_for_overflow(false), 444 _concurrent_marking_in_progress(false), 445 446 // _verbose_level set below 447 448 _init_times(), 449 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 450 _cleanup_times(), 451 _total_counting_time(0.0), 452 _total_rs_scrub_time(0.0), 453 454 _parallel_workers(NULL), 455 456 _count_card_bitmaps(NULL), 457 _count_marked_bytes(NULL) { 458 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel; 459 if (verbose_level < no_verbose) { 460 verbose_level = no_verbose; 461 } 462 if (verbose_level > high_verbose) { 463 verbose_level = high_verbose; 464 } 465 _verbose_level = verbose_level; 466 467 if (verbose_low()) { 468 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", " 469 "heap end = "PTR_FORMAT, _heap_start, _heap_end); 470 } 471 472 _markStack.allocate(MarkStackSize); 473 474 // Create & start a ConcurrentMark thread. 475 _cmThread = new ConcurrentMarkThread(this); 476 assert(cmThread() != NULL, "CM Thread should have been created"); 477 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 478 479 _g1h = G1CollectedHeap::heap(); 480 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 481 assert(_markBitMap1.covers(rs), "_markBitMap1 inconsistency"); 482 assert(_markBitMap2.covers(rs), "_markBitMap2 inconsistency"); 483 484 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 485 satb_qs.set_buffer_size(G1SATBBufferSize); 486 487 _root_regions.init(_g1h, this); 488 489 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_task_num, mtGC); 490 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_task_num, mtGC); 491 492 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_task_num, mtGC); 493 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_task_num, mtGC); 494 495 BitMap::idx_t card_bm_size = _card_bm.size(); 496 497 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 498 _active_tasks = _max_task_num; 499 for (int i = 0; i < (int) _max_task_num; ++i) { 500 CMTaskQueue* task_queue = new CMTaskQueue(); 501 task_queue->initialize(); 502 _task_queues->register_queue(i, task_queue); 503 504 _count_card_bitmaps[i] = BitMap(card_bm_size, false); 505 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, (size_t) max_regions, mtGC); 506 507 _tasks[i] = new CMTask(i, this, 508 _count_marked_bytes[i], 509 &_count_card_bitmaps[i], 510 task_queue, _task_queues); 511 512 _accum_task_vtime[i] = 0.0; 513 } 514 515 // Calculate the card number for the bottom of the heap. Used 516 // in biasing indexes into the accounting card bitmaps. 517 _heap_bottom_card_num = 518 intptr_t(uintptr_t(_g1h->reserved_region().start()) >> 519 CardTableModRefBS::card_shift); 520 521 // Clear all the liveness counting data 522 clear_all_count_data(); 523 524 if (ConcGCThreads > ParallelGCThreads) { 525 vm_exit_during_initialization("Can't have more ConcGCThreads " 526 "than ParallelGCThreads."); 527 } 528 if (ParallelGCThreads == 0) { 529 // if we are not running with any parallel GC threads we will not 530 // spawn any marking threads either 531 _parallel_marking_threads = 0; 532 _max_parallel_marking_threads = 0; 533 _sleep_factor = 0.0; 534 _marking_task_overhead = 1.0; 535 } else { 536 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 537 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 538 // if both are set 539 _sleep_factor = 0.0; 540 _marking_task_overhead = 1.0; 541 } else if (G1MarkingOverheadPercent > 0) { 542 // We will calculate the number of parallel marking threads based 543 // on a target overhead with respect to the soft real-time goal 544 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 545 double overall_cm_overhead = 546 (double) MaxGCPauseMillis * marking_overhead / 547 (double) GCPauseIntervalMillis; 548 double cpu_ratio = 1.0 / (double) os::processor_count(); 549 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 550 double marking_task_overhead = 551 overall_cm_overhead / marking_thread_num * 552 (double) os::processor_count(); 553 double sleep_factor = 554 (1.0 - marking_task_overhead) / marking_task_overhead; 555 556 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num); 557 _sleep_factor = sleep_factor; 558 _marking_task_overhead = marking_task_overhead; 559 } else { 560 // Calculate the number of parallel marking threads by scaling 561 // the number of parallel GC threads. 562 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads); 563 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num); 564 _sleep_factor = 0.0; 565 _marking_task_overhead = 1.0; 566 } 567 568 assert(ConcGCThreads > 0, "Should have been set"); 569 _parallel_marking_threads = (uint) ConcGCThreads; 570 _max_parallel_marking_threads = _parallel_marking_threads; 571 572 if (parallel_marking_threads() > 1) { 573 _cleanup_task_overhead = 1.0; 574 } else { 575 _cleanup_task_overhead = marking_task_overhead(); 576 } 577 _cleanup_sleep_factor = 578 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead(); 579 580 #if 0 581 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads()); 582 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead()); 583 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor()); 584 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead()); 585 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor()); 586 #endif 587 588 guarantee(parallel_marking_threads() > 0, "peace of mind"); 589 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads", 590 _max_parallel_marking_threads, false, true); 591 if (_parallel_workers == NULL) { 592 vm_exit_during_initialization("Failed necessary allocation."); 593 } else { 594 _parallel_workers->initialize_workers(); 595 } 596 } 597 598 // so that the call below can read a sensible value 599 _heap_start = (HeapWord*) rs.base(); 600 set_non_marking_state(); 601 } 602 603 void ConcurrentMark::update_g1_committed(bool force) { 604 // If concurrent marking is not in progress, then we do not need to 605 // update _heap_end. 606 if (!concurrent_marking_in_progress() && !force) return; 607 608 MemRegion committed = _g1h->g1_committed(); 609 assert(committed.start() == _heap_start, "start shouldn't change"); 610 HeapWord* new_end = committed.end(); 611 if (new_end > _heap_end) { 612 // The heap has been expanded. 613 614 _heap_end = new_end; 615 } 616 // Notice that the heap can also shrink. However, this only happens 617 // during a Full GC (at least currently) and the entire marking 618 // phase will bail out and the task will not be restarted. So, let's 619 // do nothing. 620 } 621 622 void ConcurrentMark::reset() { 623 // Starting values for these two. This should be called in a STW 624 // phase. CM will be notified of any future g1_committed expansions 625 // will be at the end of evacuation pauses, when tasks are 626 // inactive. 627 MemRegion committed = _g1h->g1_committed(); 628 _heap_start = committed.start(); 629 _heap_end = committed.end(); 630 631 // Separated the asserts so that we know which one fires. 632 assert(_heap_start != NULL, "heap bounds should look ok"); 633 assert(_heap_end != NULL, "heap bounds should look ok"); 634 assert(_heap_start < _heap_end, "heap bounds should look ok"); 635 636 // Reset all the marking data structures and any necessary flags 637 reset_marking_state(); 638 639 if (verbose_low()) { 640 gclog_or_tty->print_cr("[global] resetting"); 641 } 642 643 // We do reset all of them, since different phases will use 644 // different number of active threads. So, it's easiest to have all 645 // of them ready. 646 for (int i = 0; i < (int) _max_task_num; ++i) { 647 _tasks[i]->reset(_nextMarkBitMap); 648 } 649 650 // we need this to make sure that the flag is on during the evac 651 // pause with initial mark piggy-backed 652 set_concurrent_marking_in_progress(); 653 } 654 655 656 void ConcurrentMark::reset_marking_state(bool clear_overflow) { 657 _markStack.setEmpty(); 658 _markStack.clear_overflow(); 659 if (clear_overflow) { 660 clear_has_overflown(); 661 } else { 662 assert(has_overflown(), "pre-condition"); 663 } 664 _finger = _heap_start; 665 666 for (uint i = 0; i < _max_task_num; ++i) { 667 CMTaskQueue* queue = _task_queues->queue(i); 668 queue->set_empty(); 669 } 670 } 671 672 void ConcurrentMark::set_phase(uint active_tasks, bool concurrent) { 673 assert(active_tasks <= _max_task_num, "we should not have more"); 674 675 _active_tasks = active_tasks; 676 // Need to update the three data structures below according to the 677 // number of active threads for this phase. 678 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 679 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 680 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 681 682 _concurrent = concurrent; 683 // We propagate this to all tasks, not just the active ones. 684 for (int i = 0; i < (int) _max_task_num; ++i) 685 _tasks[i]->set_concurrent(concurrent); 686 687 if (concurrent) { 688 set_concurrent_marking_in_progress(); 689 } else { 690 // We currently assume that the concurrent flag has been set to 691 // false before we start remark. At this point we should also be 692 // in a STW phase. 693 assert(!concurrent_marking_in_progress(), "invariant"); 694 assert(_finger == _heap_end, "only way to get here"); 695 update_g1_committed(true); 696 } 697 } 698 699 void ConcurrentMark::set_non_marking_state() { 700 // We set the global marking state to some default values when we're 701 // not doing marking. 702 reset_marking_state(); 703 _active_tasks = 0; 704 clear_concurrent_marking_in_progress(); 705 } 706 707 ConcurrentMark::~ConcurrentMark() { 708 // The ConcurrentMark instance is never freed. 709 ShouldNotReachHere(); 710 } 711 712 void ConcurrentMark::clearNextBitmap() { 713 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 714 G1CollectorPolicy* g1p = g1h->g1_policy(); 715 716 // Make sure that the concurrent mark thread looks to still be in 717 // the current cycle. 718 guarantee(cmThread()->during_cycle(), "invariant"); 719 720 // We are finishing up the current cycle by clearing the next 721 // marking bitmap and getting it ready for the next cycle. During 722 // this time no other cycle can start. So, let's make sure that this 723 // is the case. 724 guarantee(!g1h->mark_in_progress(), "invariant"); 725 726 // clear the mark bitmap (no grey objects to start with). 727 // We need to do this in chunks and offer to yield in between 728 // each chunk. 729 HeapWord* start = _nextMarkBitMap->startWord(); 730 HeapWord* end = _nextMarkBitMap->endWord(); 731 HeapWord* cur = start; 732 size_t chunkSize = M; 733 while (cur < end) { 734 HeapWord* next = cur + chunkSize; 735 if (next > end) { 736 next = end; 737 } 738 MemRegion mr(cur,next); 739 _nextMarkBitMap->clearRange(mr); 740 cur = next; 741 do_yield_check(); 742 743 // Repeat the asserts from above. We'll do them as asserts here to 744 // minimize their overhead on the product. However, we'll have 745 // them as guarantees at the beginning / end of the bitmap 746 // clearing to get some checking in the product. 747 assert(cmThread()->during_cycle(), "invariant"); 748 assert(!g1h->mark_in_progress(), "invariant"); 749 } 750 751 // Clear the liveness counting data 752 clear_all_count_data(); 753 754 // Repeat the asserts from above. 755 guarantee(cmThread()->during_cycle(), "invariant"); 756 guarantee(!g1h->mark_in_progress(), "invariant"); 757 } 758 759 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 760 public: 761 bool doHeapRegion(HeapRegion* r) { 762 if (!r->continuesHumongous()) { 763 r->note_start_of_marking(); 764 } 765 return false; 766 } 767 }; 768 769 void ConcurrentMark::checkpointRootsInitialPre() { 770 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 771 G1CollectorPolicy* g1p = g1h->g1_policy(); 772 773 _has_aborted = false; 774 775 #ifndef PRODUCT 776 if (G1PrintReachableAtInitialMark) { 777 print_reachable("at-cycle-start", 778 VerifyOption_G1UsePrevMarking, true /* all */); 779 } 780 #endif 781 782 // Initialise marking structures. This has to be done in a STW phase. 783 reset(); 784 785 // For each region note start of marking. 786 NoteStartOfMarkHRClosure startcl; 787 g1h->heap_region_iterate(&startcl); 788 } 789 790 791 void ConcurrentMark::checkpointRootsInitialPost() { 792 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 793 794 // If we force an overflow during remark, the remark operation will 795 // actually abort and we'll restart concurrent marking. If we always 796 // force an oveflow during remark we'll never actually complete the 797 // marking phase. So, we initilize this here, at the start of the 798 // cycle, so that at the remaining overflow number will decrease at 799 // every remark and we'll eventually not need to cause one. 800 force_overflow_stw()->init(); 801 802 // Start Concurrent Marking weak-reference discovery. 803 ReferenceProcessor* rp = g1h->ref_processor_cm(); 804 // enable ("weak") refs discovery 805 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 806 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 807 808 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 809 // This is the start of the marking cycle, we're expected all 810 // threads to have SATB queues with active set to false. 811 satb_mq_set.set_active_all_threads(true, /* new active value */ 812 false /* expected_active */); 813 814 _root_regions.prepare_for_scan(); 815 816 // update_g1_committed() will be called at the end of an evac pause 817 // when marking is on. So, it's also called at the end of the 818 // initial-mark pause to update the heap end, if the heap expands 819 // during it. No need to call it here. 820 } 821 822 /* 823 * Notice that in the next two methods, we actually leave the STS 824 * during the barrier sync and join it immediately afterwards. If we 825 * do not do this, the following deadlock can occur: one thread could 826 * be in the barrier sync code, waiting for the other thread to also 827 * sync up, whereas another one could be trying to yield, while also 828 * waiting for the other threads to sync up too. 829 * 830 * Note, however, that this code is also used during remark and in 831 * this case we should not attempt to leave / enter the STS, otherwise 832 * we'll either hit an asseert (debug / fastdebug) or deadlock 833 * (product). So we should only leave / enter the STS if we are 834 * operating concurrently. 835 * 836 * Because the thread that does the sync barrier has left the STS, it 837 * is possible to be suspended for a Full GC or an evacuation pause 838 * could occur. This is actually safe, since the entering the sync 839 * barrier is one of the last things do_marking_step() does, and it 840 * doesn't manipulate any data structures afterwards. 841 */ 842 843 void ConcurrentMark::enter_first_sync_barrier(int task_num) { 844 if (verbose_low()) { 845 gclog_or_tty->print_cr("[%d] entering first barrier", task_num); 846 } 847 848 if (concurrent()) { 849 ConcurrentGCThread::stsLeave(); 850 } 851 _first_overflow_barrier_sync.enter(); 852 if (concurrent()) { 853 ConcurrentGCThread::stsJoin(); 854 } 855 // at this point everyone should have synced up and not be doing any 856 // more work 857 858 if (verbose_low()) { 859 gclog_or_tty->print_cr("[%d] leaving first barrier", task_num); 860 } 861 862 // let task 0 do this 863 if (task_num == 0) { 864 // task 0 is responsible for clearing the global data structures 865 // We should be here because of an overflow. During STW we should 866 // not clear the overflow flag since we rely on it being true when 867 // we exit this method to abort the pause and restart concurent 868 // marking. 869 reset_marking_state(concurrent() /* clear_overflow */); 870 force_overflow()->update(); 871 872 if (G1Log::fine()) { 873 gclog_or_tty->date_stamp(PrintGCDateStamps); 874 gclog_or_tty->stamp(PrintGCTimeStamps); 875 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]"); 876 } 877 } 878 879 // after this, each task should reset its own data structures then 880 // then go into the second barrier 881 } 882 883 void ConcurrentMark::enter_second_sync_barrier(int task_num) { 884 if (verbose_low()) { 885 gclog_or_tty->print_cr("[%d] entering second barrier", task_num); 886 } 887 888 if (concurrent()) { 889 ConcurrentGCThread::stsLeave(); 890 } 891 _second_overflow_barrier_sync.enter(); 892 if (concurrent()) { 893 ConcurrentGCThread::stsJoin(); 894 } 895 // at this point everything should be re-initialised and ready to go 896 897 if (verbose_low()) { 898 gclog_or_tty->print_cr("[%d] leaving second barrier", task_num); 899 } 900 } 901 902 #ifndef PRODUCT 903 void ForceOverflowSettings::init() { 904 _num_remaining = G1ConcMarkForceOverflow; 905 _force = false; 906 update(); 907 } 908 909 void ForceOverflowSettings::update() { 910 if (_num_remaining > 0) { 911 _num_remaining -= 1; 912 _force = true; 913 } else { 914 _force = false; 915 } 916 } 917 918 bool ForceOverflowSettings::should_force() { 919 if (_force) { 920 _force = false; 921 return true; 922 } else { 923 return false; 924 } 925 } 926 #endif // !PRODUCT 927 928 class CMConcurrentMarkingTask: public AbstractGangTask { 929 private: 930 ConcurrentMark* _cm; 931 ConcurrentMarkThread* _cmt; 932 933 public: 934 void work(uint worker_id) { 935 assert(Thread::current()->is_ConcurrentGC_thread(), 936 "this should only be done by a conc GC thread"); 937 ResourceMark rm; 938 939 double start_vtime = os::elapsedVTime(); 940 941 ConcurrentGCThread::stsJoin(); 942 943 assert(worker_id < _cm->active_tasks(), "invariant"); 944 CMTask* the_task = _cm->task(worker_id); 945 the_task->record_start_time(); 946 if (!_cm->has_aborted()) { 947 do { 948 double start_vtime_sec = os::elapsedVTime(); 949 double start_time_sec = os::elapsedTime(); 950 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 951 952 the_task->do_marking_step(mark_step_duration_ms, 953 true /* do_stealing */, 954 true /* do_termination */); 955 956 double end_time_sec = os::elapsedTime(); 957 double end_vtime_sec = os::elapsedVTime(); 958 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 959 double elapsed_time_sec = end_time_sec - start_time_sec; 960 _cm->clear_has_overflown(); 961 962 bool ret = _cm->do_yield_check(worker_id); 963 964 jlong sleep_time_ms; 965 if (!_cm->has_aborted() && the_task->has_aborted()) { 966 sleep_time_ms = 967 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 968 ConcurrentGCThread::stsLeave(); 969 os::sleep(Thread::current(), sleep_time_ms, false); 970 ConcurrentGCThread::stsJoin(); 971 } 972 double end_time2_sec = os::elapsedTime(); 973 double elapsed_time2_sec = end_time2_sec - start_time_sec; 974 975 #if 0 976 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, " 977 "overhead %1.4lf", 978 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms, 979 the_task->conc_overhead(os::elapsedTime()) * 8.0); 980 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms", 981 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0); 982 #endif 983 } while (!_cm->has_aborted() && the_task->has_aborted()); 984 } 985 the_task->record_end_time(); 986 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 987 988 ConcurrentGCThread::stsLeave(); 989 990 double end_vtime = os::elapsedVTime(); 991 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 992 } 993 994 CMConcurrentMarkingTask(ConcurrentMark* cm, 995 ConcurrentMarkThread* cmt) : 996 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 997 998 ~CMConcurrentMarkingTask() { } 999 }; 1000 1001 // Calculates the number of active workers for a concurrent 1002 // phase. 1003 uint ConcurrentMark::calc_parallel_marking_threads() { 1004 if (G1CollectedHeap::use_parallel_gc_threads()) { 1005 uint n_conc_workers = 0; 1006 if (!UseDynamicNumberOfGCThreads || 1007 (!FLAG_IS_DEFAULT(ConcGCThreads) && 1008 !ForceDynamicNumberOfGCThreads)) { 1009 n_conc_workers = max_parallel_marking_threads(); 1010 } else { 1011 n_conc_workers = 1012 AdaptiveSizePolicy::calc_default_active_workers( 1013 max_parallel_marking_threads(), 1014 1, /* Minimum workers */ 1015 parallel_marking_threads(), 1016 Threads::number_of_non_daemon_threads()); 1017 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 1018 // that scaling has already gone into "_max_parallel_marking_threads". 1019 } 1020 assert(n_conc_workers > 0, "Always need at least 1"); 1021 return n_conc_workers; 1022 } 1023 // If we are not running with any parallel GC threads we will not 1024 // have spawned any marking threads either. Hence the number of 1025 // concurrent workers should be 0. 1026 return 0; 1027 } 1028 1029 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) { 1030 // Currently, only survivors can be root regions. 1031 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 1032 G1RootRegionScanClosure cl(_g1h, this, worker_id); 1033 1034 const uintx interval = PrefetchScanIntervalInBytes; 1035 HeapWord* curr = hr->bottom(); 1036 const HeapWord* end = hr->top(); 1037 while (curr < end) { 1038 Prefetch::read(curr, interval); 1039 oop obj = oop(curr); 1040 int size = obj->oop_iterate(&cl); 1041 assert(size == obj->size(), "sanity"); 1042 curr += size; 1043 } 1044 } 1045 1046 class CMRootRegionScanTask : public AbstractGangTask { 1047 private: 1048 ConcurrentMark* _cm; 1049 1050 public: 1051 CMRootRegionScanTask(ConcurrentMark* cm) : 1052 AbstractGangTask("Root Region Scan"), _cm(cm) { } 1053 1054 void work(uint worker_id) { 1055 assert(Thread::current()->is_ConcurrentGC_thread(), 1056 "this should only be done by a conc GC thread"); 1057 1058 CMRootRegions* root_regions = _cm->root_regions(); 1059 HeapRegion* hr = root_regions->claim_next(); 1060 while (hr != NULL) { 1061 _cm->scanRootRegion(hr, worker_id); 1062 hr = root_regions->claim_next(); 1063 } 1064 } 1065 }; 1066 1067 void ConcurrentMark::scanRootRegions() { 1068 // scan_in_progress() will have been set to true only if there was 1069 // at least one root region to scan. So, if it's false, we 1070 // should not attempt to do any further work. 1071 if (root_regions()->scan_in_progress()) { 1072 _parallel_marking_threads = calc_parallel_marking_threads(); 1073 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1074 "Maximum number of marking threads exceeded"); 1075 uint active_workers = MAX2(1U, parallel_marking_threads()); 1076 1077 CMRootRegionScanTask task(this); 1078 if (use_parallel_marking_threads()) { 1079 _parallel_workers->set_active_workers((int) active_workers); 1080 _parallel_workers->run_task(&task); 1081 } else { 1082 task.work(0); 1083 } 1084 1085 // It's possible that has_aborted() is true here without actually 1086 // aborting the survivor scan earlier. This is OK as it's 1087 // mainly used for sanity checking. 1088 root_regions()->scan_finished(); 1089 } 1090 } 1091 1092 void ConcurrentMark::markFromRoots() { 1093 // we might be tempted to assert that: 1094 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1095 // "inconsistent argument?"); 1096 // However that wouldn't be right, because it's possible that 1097 // a safepoint is indeed in progress as a younger generation 1098 // stop-the-world GC happens even as we mark in this generation. 1099 1100 _restart_for_overflow = false; 1101 force_overflow_conc()->init(); 1102 1103 // _g1h has _n_par_threads 1104 _parallel_marking_threads = calc_parallel_marking_threads(); 1105 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1106 "Maximum number of marking threads exceeded"); 1107 1108 uint active_workers = MAX2(1U, parallel_marking_threads()); 1109 1110 // Parallel task terminator is set in "set_phase()" 1111 set_phase(active_workers, true /* concurrent */); 1112 1113 CMConcurrentMarkingTask markingTask(this, cmThread()); 1114 if (use_parallel_marking_threads()) { 1115 _parallel_workers->set_active_workers((int)active_workers); 1116 // Don't set _n_par_threads because it affects MT in proceess_strong_roots() 1117 // and the decisions on that MT processing is made elsewhere. 1118 assert(_parallel_workers->active_workers() > 0, "Should have been set"); 1119 _parallel_workers->run_task(&markingTask); 1120 } else { 1121 markingTask.work(0); 1122 } 1123 print_stats(); 1124 } 1125 1126 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1127 // world is stopped at this checkpoint 1128 assert(SafepointSynchronize::is_at_safepoint(), 1129 "world should be stopped"); 1130 1131 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1132 1133 // If a full collection has happened, we shouldn't do this. 1134 if (has_aborted()) { 1135 g1h->set_marking_complete(); // So bitmap clearing isn't confused 1136 return; 1137 } 1138 1139 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1140 1141 if (VerifyDuringGC) { 1142 HandleMark hm; // handle scope 1143 gclog_or_tty->print(" VerifyDuringGC:(before)"); 1144 Universe::heap()->prepare_for_verify(); 1145 Universe::verify(/* silent */ false, 1146 /* option */ VerifyOption_G1UsePrevMarking); 1147 } 1148 1149 G1CollectorPolicy* g1p = g1h->g1_policy(); 1150 g1p->record_concurrent_mark_remark_start(); 1151 1152 double start = os::elapsedTime(); 1153 1154 checkpointRootsFinalWork(); 1155 1156 double mark_work_end = os::elapsedTime(); 1157 1158 weakRefsWork(clear_all_soft_refs); 1159 1160 if (has_overflown()) { 1161 // Oops. We overflowed. Restart concurrent marking. 1162 _restart_for_overflow = true; 1163 // Clear the marking state because we will be restarting 1164 // marking due to overflowing the global mark stack. 1165 reset_marking_state(); 1166 if (G1TraceMarkStackOverflow) { 1167 gclog_or_tty->print_cr("\nRemark led to restart for overflow."); 1168 } 1169 } else { 1170 // Aggregate the per-task counting data that we have accumulated 1171 // while marking. 1172 aggregate_count_data(); 1173 1174 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1175 // We're done with marking. 1176 // This is the end of the marking cycle, we're expected all 1177 // threads to have SATB queues with active set to true. 1178 satb_mq_set.set_active_all_threads(false, /* new active value */ 1179 true /* expected_active */); 1180 1181 if (VerifyDuringGC) { 1182 HandleMark hm; // handle scope 1183 gclog_or_tty->print(" VerifyDuringGC:(after)"); 1184 Universe::heap()->prepare_for_verify(); 1185 Universe::verify(/* silent */ false, 1186 /* option */ VerifyOption_G1UseNextMarking); 1187 } 1188 assert(!restart_for_overflow(), "sanity"); 1189 // Completely reset the marking state since marking completed 1190 set_non_marking_state(); 1191 } 1192 1193 #if VERIFY_OBJS_PROCESSED 1194 _scan_obj_cl.objs_processed = 0; 1195 ThreadLocalObjQueue::objs_enqueued = 0; 1196 #endif 1197 1198 // Statistics 1199 double now = os::elapsedTime(); 1200 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1201 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1202 _remark_times.add((now - start) * 1000.0); 1203 1204 g1p->record_concurrent_mark_remark_end(); 1205 1206 G1CMIsAliveClosure is_alive(g1h); 1207 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive); 1208 } 1209 1210 // Base class of the closures that finalize and verify the 1211 // liveness counting data. 1212 class CMCountDataClosureBase: public HeapRegionClosure { 1213 protected: 1214 G1CollectedHeap* _g1h; 1215 ConcurrentMark* _cm; 1216 CardTableModRefBS* _ct_bs; 1217 1218 BitMap* _region_bm; 1219 BitMap* _card_bm; 1220 1221 // Takes a region that's not empty (i.e., it has at least one 1222 // live object in it and sets its corresponding bit on the region 1223 // bitmap to 1. If the region is "starts humongous" it will also set 1224 // to 1 the bits on the region bitmap that correspond to its 1225 // associated "continues humongous" regions. 1226 void set_bit_for_region(HeapRegion* hr) { 1227 assert(!hr->continuesHumongous(), "should have filtered those out"); 1228 1229 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index(); 1230 if (!hr->startsHumongous()) { 1231 // Normal (non-humongous) case: just set the bit. 1232 _region_bm->par_at_put(index, true); 1233 } else { 1234 // Starts humongous case: calculate how many regions are part of 1235 // this humongous region and then set the bit range. 1236 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index(); 1237 _region_bm->par_at_put_range(index, end_index, true); 1238 } 1239 } 1240 1241 public: 1242 CMCountDataClosureBase(G1CollectedHeap* g1h, 1243 BitMap* region_bm, BitMap* card_bm): 1244 _g1h(g1h), _cm(g1h->concurrent_mark()), 1245 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 1246 _region_bm(region_bm), _card_bm(card_bm) { } 1247 }; 1248 1249 // Closure that calculates the # live objects per region. Used 1250 // for verification purposes during the cleanup pause. 1251 class CalcLiveObjectsClosure: public CMCountDataClosureBase { 1252 CMBitMapRO* _bm; 1253 size_t _region_marked_bytes; 1254 1255 public: 1256 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h, 1257 BitMap* region_bm, BitMap* card_bm) : 1258 CMCountDataClosureBase(g1h, region_bm, card_bm), 1259 _bm(bm), _region_marked_bytes(0) { } 1260 1261 bool doHeapRegion(HeapRegion* hr) { 1262 1263 if (hr->continuesHumongous()) { 1264 // We will ignore these here and process them when their 1265 // associated "starts humongous" region is processed (see 1266 // set_bit_for_heap_region()). Note that we cannot rely on their 1267 // associated "starts humongous" region to have their bit set to 1268 // 1 since, due to the region chunking in the parallel region 1269 // iteration, a "continues humongous" region might be visited 1270 // before its associated "starts humongous". 1271 return false; 1272 } 1273 1274 HeapWord* ntams = hr->next_top_at_mark_start(); 1275 HeapWord* start = hr->bottom(); 1276 1277 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(), 1278 err_msg("Preconditions not met - " 1279 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT, 1280 start, ntams, hr->end())); 1281 1282 // Find the first marked object at or after "start". 1283 start = _bm->getNextMarkedWordAddress(start, ntams); 1284 1285 size_t marked_bytes = 0; 1286 1287 while (start < ntams) { 1288 oop obj = oop(start); 1289 int obj_sz = obj->size(); 1290 HeapWord* obj_end = start + obj_sz; 1291 1292 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 1293 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end); 1294 1295 // Note: if we're looking at the last region in heap - obj_end 1296 // could be actually just beyond the end of the heap; end_idx 1297 // will then correspond to a (non-existent) card that is also 1298 // just beyond the heap. 1299 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) { 1300 // end of object is not card aligned - increment to cover 1301 // all the cards spanned by the object 1302 end_idx += 1; 1303 } 1304 1305 // Set the bits in the card BM for the cards spanned by this object. 1306 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1307 1308 // Add the size of this object to the number of marked bytes. 1309 marked_bytes += (size_t)obj_sz * HeapWordSize; 1310 1311 // Find the next marked object after this one. 1312 start = _bm->getNextMarkedWordAddress(obj_end, ntams); 1313 } 1314 1315 // Mark the allocated-since-marking portion... 1316 HeapWord* top = hr->top(); 1317 if (ntams < top) { 1318 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1319 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1320 1321 // Note: if we're looking at the last region in heap - top 1322 // could be actually just beyond the end of the heap; end_idx 1323 // will then correspond to a (non-existent) card that is also 1324 // just beyond the heap. 1325 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1326 // end of object is not card aligned - increment to cover 1327 // all the cards spanned by the object 1328 end_idx += 1; 1329 } 1330 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1331 1332 // This definitely means the region has live objects. 1333 set_bit_for_region(hr); 1334 } 1335 1336 // Update the live region bitmap. 1337 if (marked_bytes > 0) { 1338 set_bit_for_region(hr); 1339 } 1340 1341 // Set the marked bytes for the current region so that 1342 // it can be queried by a calling verificiation routine 1343 _region_marked_bytes = marked_bytes; 1344 1345 return false; 1346 } 1347 1348 size_t region_marked_bytes() const { return _region_marked_bytes; } 1349 }; 1350 1351 // Heap region closure used for verifying the counting data 1352 // that was accumulated concurrently and aggregated during 1353 // the remark pause. This closure is applied to the heap 1354 // regions during the STW cleanup pause. 1355 1356 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure { 1357 G1CollectedHeap* _g1h; 1358 ConcurrentMark* _cm; 1359 CalcLiveObjectsClosure _calc_cl; 1360 BitMap* _region_bm; // Region BM to be verified 1361 BitMap* _card_bm; // Card BM to be verified 1362 bool _verbose; // verbose output? 1363 1364 BitMap* _exp_region_bm; // Expected Region BM values 1365 BitMap* _exp_card_bm; // Expected card BM values 1366 1367 int _failures; 1368 1369 public: 1370 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h, 1371 BitMap* region_bm, 1372 BitMap* card_bm, 1373 BitMap* exp_region_bm, 1374 BitMap* exp_card_bm, 1375 bool verbose) : 1376 _g1h(g1h), _cm(g1h->concurrent_mark()), 1377 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm), 1378 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose), 1379 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm), 1380 _failures(0) { } 1381 1382 int failures() const { return _failures; } 1383 1384 bool doHeapRegion(HeapRegion* hr) { 1385 if (hr->continuesHumongous()) { 1386 // We will ignore these here and process them when their 1387 // associated "starts humongous" region is processed (see 1388 // set_bit_for_heap_region()). Note that we cannot rely on their 1389 // associated "starts humongous" region to have their bit set to 1390 // 1 since, due to the region chunking in the parallel region 1391 // iteration, a "continues humongous" region might be visited 1392 // before its associated "starts humongous". 1393 return false; 1394 } 1395 1396 int failures = 0; 1397 1398 // Call the CalcLiveObjectsClosure to walk the marking bitmap for 1399 // this region and set the corresponding bits in the expected region 1400 // and card bitmaps. 1401 bool res = _calc_cl.doHeapRegion(hr); 1402 assert(res == false, "should be continuing"); 1403 1404 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL), 1405 Mutex::_no_safepoint_check_flag); 1406 1407 // Verify the marked bytes for this region. 1408 size_t exp_marked_bytes = _calc_cl.region_marked_bytes(); 1409 size_t act_marked_bytes = hr->next_marked_bytes(); 1410 1411 // We're not OK if expected marked bytes > actual marked bytes. It means 1412 // we have missed accounting some objects during the actual marking. 1413 if (exp_marked_bytes > act_marked_bytes) { 1414 if (_verbose) { 1415 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: " 1416 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT, 1417 hr->hrs_index(), exp_marked_bytes, act_marked_bytes); 1418 } 1419 failures += 1; 1420 } 1421 1422 // Verify the bit, for this region, in the actual and expected 1423 // (which was just calculated) region bit maps. 1424 // We're not OK if the bit in the calculated expected region 1425 // bitmap is set and the bit in the actual region bitmap is not. 1426 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index(); 1427 1428 bool expected = _exp_region_bm->at(index); 1429 bool actual = _region_bm->at(index); 1430 if (expected && !actual) { 1431 if (_verbose) { 1432 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: " 1433 "expected: %s, actual: %s", 1434 hr->hrs_index(), 1435 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1436 } 1437 failures += 1; 1438 } 1439 1440 // Verify that the card bit maps for the cards spanned by the current 1441 // region match. We have an error if we have a set bit in the expected 1442 // bit map and the corresponding bit in the actual bitmap is not set. 1443 1444 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom()); 1445 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top()); 1446 1447 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) { 1448 expected = _exp_card_bm->at(i); 1449 actual = _card_bm->at(i); 1450 1451 if (expected && !actual) { 1452 if (_verbose) { 1453 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": " 1454 "expected: %s, actual: %s", 1455 hr->hrs_index(), i, 1456 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1457 } 1458 failures += 1; 1459 } 1460 } 1461 1462 if (failures > 0 && _verbose) { 1463 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", " 1464 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT, 1465 HR_FORMAT_PARAMS(hr), hr->next_top_at_mark_start(), 1466 _calc_cl.region_marked_bytes(), hr->next_marked_bytes()); 1467 } 1468 1469 _failures += failures; 1470 1471 // We could stop iteration over the heap when we 1472 // find the first violating region by returning true. 1473 return false; 1474 } 1475 }; 1476 1477 1478 class G1ParVerifyFinalCountTask: public AbstractGangTask { 1479 protected: 1480 G1CollectedHeap* _g1h; 1481 ConcurrentMark* _cm; 1482 BitMap* _actual_region_bm; 1483 BitMap* _actual_card_bm; 1484 1485 uint _n_workers; 1486 1487 BitMap* _expected_region_bm; 1488 BitMap* _expected_card_bm; 1489 1490 int _failures; 1491 bool _verbose; 1492 1493 public: 1494 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h, 1495 BitMap* region_bm, BitMap* card_bm, 1496 BitMap* expected_region_bm, BitMap* expected_card_bm) 1497 : AbstractGangTask("G1 verify final counting"), 1498 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1499 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1500 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm), 1501 _failures(0), _verbose(false), 1502 _n_workers(0) { 1503 assert(VerifyDuringGC, "don't call this otherwise"); 1504 1505 // Use the value already set as the number of active threads 1506 // in the call to run_task(). 1507 if (G1CollectedHeap::use_parallel_gc_threads()) { 1508 assert( _g1h->workers()->active_workers() > 0, 1509 "Should have been previously set"); 1510 _n_workers = _g1h->workers()->active_workers(); 1511 } else { 1512 _n_workers = 1; 1513 } 1514 1515 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity"); 1516 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity"); 1517 1518 _verbose = _cm->verbose_medium(); 1519 } 1520 1521 void work(uint worker_id) { 1522 assert(worker_id < _n_workers, "invariant"); 1523 1524 VerifyLiveObjectDataHRClosure verify_cl(_g1h, 1525 _actual_region_bm, _actual_card_bm, 1526 _expected_region_bm, 1527 _expected_card_bm, 1528 _verbose); 1529 1530 if (G1CollectedHeap::use_parallel_gc_threads()) { 1531 _g1h->heap_region_par_iterate_chunked(&verify_cl, 1532 worker_id, 1533 _n_workers, 1534 HeapRegion::VerifyCountClaimValue); 1535 } else { 1536 _g1h->heap_region_iterate(&verify_cl); 1537 } 1538 1539 Atomic::add(verify_cl.failures(), &_failures); 1540 } 1541 1542 int failures() const { return _failures; } 1543 }; 1544 1545 // Closure that finalizes the liveness counting data. 1546 // Used during the cleanup pause. 1547 // Sets the bits corresponding to the interval [NTAMS, top] 1548 // (which contains the implicitly live objects) in the 1549 // card liveness bitmap. Also sets the bit for each region, 1550 // containing live data, in the region liveness bitmap. 1551 1552 class FinalCountDataUpdateClosure: public CMCountDataClosureBase { 1553 public: 1554 FinalCountDataUpdateClosure(G1CollectedHeap* g1h, 1555 BitMap* region_bm, 1556 BitMap* card_bm) : 1557 CMCountDataClosureBase(g1h, region_bm, card_bm) { } 1558 1559 bool doHeapRegion(HeapRegion* hr) { 1560 1561 if (hr->continuesHumongous()) { 1562 // We will ignore these here and process them when their 1563 // associated "starts humongous" region is processed (see 1564 // set_bit_for_heap_region()). Note that we cannot rely on their 1565 // associated "starts humongous" region to have their bit set to 1566 // 1 since, due to the region chunking in the parallel region 1567 // iteration, a "continues humongous" region might be visited 1568 // before its associated "starts humongous". 1569 return false; 1570 } 1571 1572 HeapWord* ntams = hr->next_top_at_mark_start(); 1573 HeapWord* top = hr->top(); 1574 1575 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions."); 1576 1577 // Mark the allocated-since-marking portion... 1578 if (ntams < top) { 1579 // This definitely means the region has live objects. 1580 set_bit_for_region(hr); 1581 1582 // Now set the bits in the card bitmap for [ntams, top) 1583 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1584 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1585 1586 // Note: if we're looking at the last region in heap - top 1587 // could be actually just beyond the end of the heap; end_idx 1588 // will then correspond to a (non-existent) card that is also 1589 // just beyond the heap. 1590 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1591 // end of object is not card aligned - increment to cover 1592 // all the cards spanned by the object 1593 end_idx += 1; 1594 } 1595 1596 assert(end_idx <= _card_bm->size(), 1597 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1598 end_idx, _card_bm->size())); 1599 assert(start_idx < _card_bm->size(), 1600 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1601 start_idx, _card_bm->size())); 1602 1603 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1604 } 1605 1606 // Set the bit for the region if it contains live data 1607 if (hr->next_marked_bytes() > 0) { 1608 set_bit_for_region(hr); 1609 } 1610 1611 return false; 1612 } 1613 }; 1614 1615 class G1ParFinalCountTask: public AbstractGangTask { 1616 protected: 1617 G1CollectedHeap* _g1h; 1618 ConcurrentMark* _cm; 1619 BitMap* _actual_region_bm; 1620 BitMap* _actual_card_bm; 1621 1622 uint _n_workers; 1623 1624 public: 1625 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm) 1626 : AbstractGangTask("G1 final counting"), 1627 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1628 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1629 _n_workers(0) { 1630 // Use the value already set as the number of active threads 1631 // in the call to run_task(). 1632 if (G1CollectedHeap::use_parallel_gc_threads()) { 1633 assert( _g1h->workers()->active_workers() > 0, 1634 "Should have been previously set"); 1635 _n_workers = _g1h->workers()->active_workers(); 1636 } else { 1637 _n_workers = 1; 1638 } 1639 } 1640 1641 void work(uint worker_id) { 1642 assert(worker_id < _n_workers, "invariant"); 1643 1644 FinalCountDataUpdateClosure final_update_cl(_g1h, 1645 _actual_region_bm, 1646 _actual_card_bm); 1647 1648 if (G1CollectedHeap::use_parallel_gc_threads()) { 1649 _g1h->heap_region_par_iterate_chunked(&final_update_cl, 1650 worker_id, 1651 _n_workers, 1652 HeapRegion::FinalCountClaimValue); 1653 } else { 1654 _g1h->heap_region_iterate(&final_update_cl); 1655 } 1656 } 1657 }; 1658 1659 class G1ParNoteEndTask; 1660 1661 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1662 G1CollectedHeap* _g1; 1663 int _worker_num; 1664 size_t _max_live_bytes; 1665 uint _regions_claimed; 1666 size_t _freed_bytes; 1667 FreeRegionList* _local_cleanup_list; 1668 OldRegionSet* _old_proxy_set; 1669 HumongousRegionSet* _humongous_proxy_set; 1670 HRRSCleanupTask* _hrrs_cleanup_task; 1671 double _claimed_region_time; 1672 double _max_region_time; 1673 1674 public: 1675 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1676 int worker_num, 1677 FreeRegionList* local_cleanup_list, 1678 OldRegionSet* old_proxy_set, 1679 HumongousRegionSet* humongous_proxy_set, 1680 HRRSCleanupTask* hrrs_cleanup_task) : 1681 _g1(g1), _worker_num(worker_num), 1682 _max_live_bytes(0), _regions_claimed(0), 1683 _freed_bytes(0), 1684 _claimed_region_time(0.0), _max_region_time(0.0), 1685 _local_cleanup_list(local_cleanup_list), 1686 _old_proxy_set(old_proxy_set), 1687 _humongous_proxy_set(humongous_proxy_set), 1688 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1689 1690 size_t freed_bytes() { return _freed_bytes; } 1691 1692 bool doHeapRegion(HeapRegion *hr) { 1693 if (hr->continuesHumongous()) { 1694 return false; 1695 } 1696 // We use a claim value of zero here because all regions 1697 // were claimed with value 1 in the FinalCount task. 1698 _g1->reset_gc_time_stamps(hr); 1699 double start = os::elapsedTime(); 1700 _regions_claimed++; 1701 hr->note_end_of_marking(); 1702 _max_live_bytes += hr->max_live_bytes(); 1703 _g1->free_region_if_empty(hr, 1704 &_freed_bytes, 1705 _local_cleanup_list, 1706 _old_proxy_set, 1707 _humongous_proxy_set, 1708 _hrrs_cleanup_task, 1709 true /* par */); 1710 double region_time = (os::elapsedTime() - start); 1711 _claimed_region_time += region_time; 1712 if (region_time > _max_region_time) { 1713 _max_region_time = region_time; 1714 } 1715 return false; 1716 } 1717 1718 size_t max_live_bytes() { return _max_live_bytes; } 1719 uint regions_claimed() { return _regions_claimed; } 1720 double claimed_region_time_sec() { return _claimed_region_time; } 1721 double max_region_time_sec() { return _max_region_time; } 1722 }; 1723 1724 class G1ParNoteEndTask: public AbstractGangTask { 1725 friend class G1NoteEndOfConcMarkClosure; 1726 1727 protected: 1728 G1CollectedHeap* _g1h; 1729 size_t _max_live_bytes; 1730 size_t _freed_bytes; 1731 FreeRegionList* _cleanup_list; 1732 1733 public: 1734 G1ParNoteEndTask(G1CollectedHeap* g1h, 1735 FreeRegionList* cleanup_list) : 1736 AbstractGangTask("G1 note end"), _g1h(g1h), 1737 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { } 1738 1739 void work(uint worker_id) { 1740 double start = os::elapsedTime(); 1741 FreeRegionList local_cleanup_list("Local Cleanup List"); 1742 OldRegionSet old_proxy_set("Local Cleanup Old Proxy Set"); 1743 HumongousRegionSet humongous_proxy_set("Local Cleanup Humongous Proxy Set"); 1744 HRRSCleanupTask hrrs_cleanup_task; 1745 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, worker_id, &local_cleanup_list, 1746 &old_proxy_set, 1747 &humongous_proxy_set, 1748 &hrrs_cleanup_task); 1749 if (G1CollectedHeap::use_parallel_gc_threads()) { 1750 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id, 1751 _g1h->workers()->active_workers(), 1752 HeapRegion::NoteEndClaimValue); 1753 } else { 1754 _g1h->heap_region_iterate(&g1_note_end); 1755 } 1756 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1757 1758 // Now update the lists 1759 _g1h->update_sets_after_freeing_regions(g1_note_end.freed_bytes(), 1760 NULL /* free_list */, 1761 &old_proxy_set, 1762 &humongous_proxy_set, 1763 true /* par */); 1764 { 1765 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1766 _max_live_bytes += g1_note_end.max_live_bytes(); 1767 _freed_bytes += g1_note_end.freed_bytes(); 1768 1769 // If we iterate over the global cleanup list at the end of 1770 // cleanup to do this printing we will not guarantee to only 1771 // generate output for the newly-reclaimed regions (the list 1772 // might not be empty at the beginning of cleanup; we might 1773 // still be working on its previous contents). So we do the 1774 // printing here, before we append the new regions to the global 1775 // cleanup list. 1776 1777 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1778 if (hr_printer->is_active()) { 1779 HeapRegionLinkedListIterator iter(&local_cleanup_list); 1780 while (iter.more_available()) { 1781 HeapRegion* hr = iter.get_next(); 1782 hr_printer->cleanup(hr); 1783 } 1784 } 1785 1786 _cleanup_list->add_as_tail(&local_cleanup_list); 1787 assert(local_cleanup_list.is_empty(), "post-condition"); 1788 1789 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1790 } 1791 } 1792 size_t max_live_bytes() { return _max_live_bytes; } 1793 size_t freed_bytes() { return _freed_bytes; } 1794 }; 1795 1796 class G1ParScrubRemSetTask: public AbstractGangTask { 1797 protected: 1798 G1RemSet* _g1rs; 1799 BitMap* _region_bm; 1800 BitMap* _card_bm; 1801 public: 1802 G1ParScrubRemSetTask(G1CollectedHeap* g1h, 1803 BitMap* region_bm, BitMap* card_bm) : 1804 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), 1805 _region_bm(region_bm), _card_bm(card_bm) { } 1806 1807 void work(uint worker_id) { 1808 if (G1CollectedHeap::use_parallel_gc_threads()) { 1809 _g1rs->scrub_par(_region_bm, _card_bm, worker_id, 1810 HeapRegion::ScrubRemSetClaimValue); 1811 } else { 1812 _g1rs->scrub(_region_bm, _card_bm); 1813 } 1814 } 1815 1816 }; 1817 1818 void ConcurrentMark::cleanup() { 1819 // world is stopped at this checkpoint 1820 assert(SafepointSynchronize::is_at_safepoint(), 1821 "world should be stopped"); 1822 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1823 1824 // If a full collection has happened, we shouldn't do this. 1825 if (has_aborted()) { 1826 g1h->set_marking_complete(); // So bitmap clearing isn't confused 1827 return; 1828 } 1829 1830 HRSPhaseSetter x(HRSPhaseCleanup); 1831 g1h->verify_region_sets_optional(); 1832 1833 if (VerifyDuringGC) { 1834 HandleMark hm; // handle scope 1835 gclog_or_tty->print(" VerifyDuringGC:(before)"); 1836 Universe::heap()->prepare_for_verify(); 1837 Universe::verify(/* silent */ false, 1838 /* option */ VerifyOption_G1UsePrevMarking); 1839 } 1840 1841 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); 1842 g1p->record_concurrent_mark_cleanup_start(); 1843 1844 double start = os::elapsedTime(); 1845 1846 HeapRegionRemSet::reset_for_cleanup_tasks(); 1847 1848 uint n_workers; 1849 1850 // Do counting once more with the world stopped for good measure. 1851 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm); 1852 1853 if (G1CollectedHeap::use_parallel_gc_threads()) { 1854 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 1855 "sanity check"); 1856 1857 g1h->set_par_threads(); 1858 n_workers = g1h->n_par_threads(); 1859 assert(g1h->n_par_threads() == n_workers, 1860 "Should not have been reset"); 1861 g1h->workers()->run_task(&g1_par_count_task); 1862 // Done with the parallel phase so reset to 0. 1863 g1h->set_par_threads(0); 1864 1865 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue), 1866 "sanity check"); 1867 } else { 1868 n_workers = 1; 1869 g1_par_count_task.work(0); 1870 } 1871 1872 if (VerifyDuringGC) { 1873 // Verify that the counting data accumulated during marking matches 1874 // that calculated by walking the marking bitmap. 1875 1876 // Bitmaps to hold expected values 1877 BitMap expected_region_bm(_region_bm.size(), false); 1878 BitMap expected_card_bm(_card_bm.size(), false); 1879 1880 G1ParVerifyFinalCountTask g1_par_verify_task(g1h, 1881 &_region_bm, 1882 &_card_bm, 1883 &expected_region_bm, 1884 &expected_card_bm); 1885 1886 if (G1CollectedHeap::use_parallel_gc_threads()) { 1887 g1h->set_par_threads((int)n_workers); 1888 g1h->workers()->run_task(&g1_par_verify_task); 1889 // Done with the parallel phase so reset to 0. 1890 g1h->set_par_threads(0); 1891 1892 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue), 1893 "sanity check"); 1894 } else { 1895 g1_par_verify_task.work(0); 1896 } 1897 1898 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures"); 1899 } 1900 1901 size_t start_used_bytes = g1h->used(); 1902 g1h->set_marking_complete(); 1903 1904 double count_end = os::elapsedTime(); 1905 double this_final_counting_time = (count_end - start); 1906 _total_counting_time += this_final_counting_time; 1907 1908 if (G1PrintRegionLivenessInfo) { 1909 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking"); 1910 _g1h->heap_region_iterate(&cl); 1911 } 1912 1913 // Install newly created mark bitMap as "prev". 1914 swapMarkBitMaps(); 1915 1916 g1h->reset_gc_time_stamp(); 1917 1918 // Note end of marking in all heap regions. 1919 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list); 1920 if (G1CollectedHeap::use_parallel_gc_threads()) { 1921 g1h->set_par_threads((int)n_workers); 1922 g1h->workers()->run_task(&g1_par_note_end_task); 1923 g1h->set_par_threads(0); 1924 1925 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue), 1926 "sanity check"); 1927 } else { 1928 g1_par_note_end_task.work(0); 1929 } 1930 g1h->check_gc_time_stamps(); 1931 1932 if (!cleanup_list_is_empty()) { 1933 // The cleanup list is not empty, so we'll have to process it 1934 // concurrently. Notify anyone else that might be wanting free 1935 // regions that there will be more free regions coming soon. 1936 g1h->set_free_regions_coming(); 1937 } 1938 1939 // call below, since it affects the metric by which we sort the heap 1940 // regions. 1941 if (G1ScrubRemSets) { 1942 double rs_scrub_start = os::elapsedTime(); 1943 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm); 1944 if (G1CollectedHeap::use_parallel_gc_threads()) { 1945 g1h->set_par_threads((int)n_workers); 1946 g1h->workers()->run_task(&g1_par_scrub_rs_task); 1947 g1h->set_par_threads(0); 1948 1949 assert(g1h->check_heap_region_claim_values( 1950 HeapRegion::ScrubRemSetClaimValue), 1951 "sanity check"); 1952 } else { 1953 g1_par_scrub_rs_task.work(0); 1954 } 1955 1956 double rs_scrub_end = os::elapsedTime(); 1957 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); 1958 _total_rs_scrub_time += this_rs_scrub_time; 1959 } 1960 1961 // this will also free any regions totally full of garbage objects, 1962 // and sort the regions. 1963 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers); 1964 1965 // Statistics. 1966 double end = os::elapsedTime(); 1967 _cleanup_times.add((end - start) * 1000.0); 1968 1969 if (G1Log::fine()) { 1970 g1h->print_size_transition(gclog_or_tty, 1971 start_used_bytes, 1972 g1h->used(), 1973 g1h->capacity()); 1974 } 1975 1976 // Clean up will have freed any regions completely full of garbage. 1977 // Update the soft reference policy with the new heap occupancy. 1978 Universe::update_heap_info_at_gc(); 1979 1980 // We need to make this be a "collection" so any collection pause that 1981 // races with it goes around and waits for completeCleanup to finish. 1982 g1h->increment_total_collections(); 1983 1984 // We reclaimed old regions so we should calculate the sizes to make 1985 // sure we update the old gen/space data. 1986 g1h->g1mm()->update_sizes(); 1987 1988 if (VerifyDuringGC) { 1989 HandleMark hm; // handle scope 1990 gclog_or_tty->print(" VerifyDuringGC:(after)"); 1991 Universe::heap()->prepare_for_verify(); 1992 Universe::verify(/* silent */ false, 1993 /* option */ VerifyOption_G1UsePrevMarking); 1994 } 1995 1996 g1h->verify_region_sets_optional(); 1997 g1h->trace_heap_after_concurrent_cycle(); 1998 } 1999 2000 void ConcurrentMark::completeCleanup() { 2001 if (has_aborted()) return; 2002 2003 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2004 2005 _cleanup_list.verify_optional(); 2006 FreeRegionList tmp_free_list("Tmp Free List"); 2007 2008 if (G1ConcRegionFreeingVerbose) { 2009 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2010 "cleanup list has %u entries", 2011 _cleanup_list.length()); 2012 } 2013 2014 // Noone else should be accessing the _cleanup_list at this point, 2015 // so it's not necessary to take any locks 2016 while (!_cleanup_list.is_empty()) { 2017 HeapRegion* hr = _cleanup_list.remove_head(); 2018 assert(hr != NULL, "the list was not empty"); 2019 hr->par_clear(); 2020 tmp_free_list.add_as_tail(hr); 2021 2022 // Instead of adding one region at a time to the secondary_free_list, 2023 // we accumulate them in the local list and move them a few at a 2024 // time. This also cuts down on the number of notify_all() calls 2025 // we do during this process. We'll also append the local list when 2026 // _cleanup_list is empty (which means we just removed the last 2027 // region from the _cleanup_list). 2028 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 2029 _cleanup_list.is_empty()) { 2030 if (G1ConcRegionFreeingVerbose) { 2031 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2032 "appending %u entries to the secondary_free_list, " 2033 "cleanup list still has %u entries", 2034 tmp_free_list.length(), 2035 _cleanup_list.length()); 2036 } 2037 2038 { 2039 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 2040 g1h->secondary_free_list_add_as_tail(&tmp_free_list); 2041 SecondaryFreeList_lock->notify_all(); 2042 } 2043 2044 if (G1StressConcRegionFreeing) { 2045 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 2046 os::sleep(Thread::current(), (jlong) 1, false); 2047 } 2048 } 2049 } 2050 } 2051 assert(tmp_free_list.is_empty(), "post-condition"); 2052 } 2053 2054 // Supporting Object and Oop closures for reference discovery 2055 // and processing in during marking 2056 2057 bool G1CMIsAliveClosure::do_object_b(oop obj) { 2058 HeapWord* addr = (HeapWord*)obj; 2059 return addr != NULL && 2060 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 2061 } 2062 2063 // 'Keep Alive' oop closure used by both serial parallel reference processing. 2064 // Uses the CMTask associated with a worker thread (for serial reference 2065 // processing the CMTask for worker 0 is used) to preserve (mark) and 2066 // trace referent objects. 2067 // 2068 // Using the CMTask and embedded local queues avoids having the worker 2069 // threads operating on the global mark stack. This reduces the risk 2070 // of overflowing the stack - which we would rather avoid at this late 2071 // state. Also using the tasks' local queues removes the potential 2072 // of the workers interfering with each other that could occur if 2073 // operating on the global stack. 2074 2075 class G1CMKeepAliveAndDrainClosure: public OopClosure { 2076 ConcurrentMark* _cm; 2077 CMTask* _task; 2078 int _ref_counter_limit; 2079 int _ref_counter; 2080 public: 2081 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task) : 2082 _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval) { 2083 assert(_ref_counter_limit > 0, "sanity"); 2084 _ref_counter = _ref_counter_limit; 2085 } 2086 2087 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 2088 virtual void do_oop( oop* p) { do_oop_work(p); } 2089 2090 template <class T> void do_oop_work(T* p) { 2091 if (!_cm->has_overflown()) { 2092 oop obj = oopDesc::load_decode_heap_oop(p); 2093 if (_cm->verbose_high()) { 2094 gclog_or_tty->print_cr("\t[%d] we're looking at location " 2095 "*"PTR_FORMAT" = "PTR_FORMAT, 2096 _task->task_id(), p, (void*) obj); 2097 } 2098 2099 _task->deal_with_reference(obj); 2100 _ref_counter--; 2101 2102 if (_ref_counter == 0) { 2103 // We have dealt with _ref_counter_limit references, pushing them 2104 // and objects reachable from them on to the local stack (and 2105 // possibly the global stack). Call CMTask::do_marking_step() to 2106 // process these entries. 2107 // 2108 // We call CMTask::do_marking_step() in a loop, which we'll exit if 2109 // there's nothing more to do (i.e. we're done with the entries that 2110 // were pushed as a result of the CMTask::deal_with_reference() calls 2111 // above) or we overflow. 2112 // 2113 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2114 // flag while there may still be some work to do. (See the comment at 2115 // the beginning of CMTask::do_marking_step() for those conditions - 2116 // one of which is reaching the specified time target.) It is only 2117 // when CMTask::do_marking_step() returns without setting the 2118 // has_aborted() flag that the marking step has completed. 2119 do { 2120 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 2121 _task->do_marking_step(mark_step_duration_ms, 2122 false /* do_stealing */, 2123 false /* do_termination */); 2124 } while (_task->has_aborted() && !_cm->has_overflown()); 2125 _ref_counter = _ref_counter_limit; 2126 } 2127 } else { 2128 if (_cm->verbose_high()) { 2129 gclog_or_tty->print_cr("\t[%d] CM Overflow", _task->task_id()); 2130 } 2131 } 2132 } 2133 }; 2134 2135 // 'Drain' oop closure used by both serial and parallel reference processing. 2136 // Uses the CMTask associated with a given worker thread (for serial 2137 // reference processing the CMtask for worker 0 is used). Calls the 2138 // do_marking_step routine, with an unbelievably large timeout value, 2139 // to drain the marking data structures of the remaining entries 2140 // added by the 'keep alive' oop closure above. 2141 2142 class G1CMDrainMarkingStackClosure: public VoidClosure { 2143 ConcurrentMark* _cm; 2144 CMTask* _task; 2145 bool _do_stealing; 2146 bool _do_termination; 2147 public: 2148 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_par) : 2149 _cm(cm), _task(task) { 2150 assert(is_par || _task->task_id() == 0, 2151 "Only task for worker 0 should be used if ref processing is single threaded"); 2152 // We only allow stealing and only enter the termination protocol 2153 // in CMTask::do_marking_step() if this closure is being instantiated 2154 // for parallel reference processing. 2155 _do_stealing = _do_termination = is_par; 2156 } 2157 2158 void do_void() { 2159 do { 2160 if (_cm->verbose_high()) { 2161 gclog_or_tty->print_cr("\t[%d] Drain: Calling do_marking_step - " 2162 "stealing: %s, termination: %s", 2163 _task->task_id(), 2164 BOOL_TO_STR(_do_stealing), 2165 BOOL_TO_STR(_do_termination)); 2166 } 2167 2168 // We call CMTask::do_marking_step() to completely drain the local 2169 // and global marking stacks of entries pushed by the 'keep alive' 2170 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 2171 // 2172 // CMTask::do_marking_step() is called in a loop, which we'll exit 2173 // if there's nothing more to do (i.e. we'completely drained the 2174 // entries that were pushed as a a result of applying the 'keep alive' 2175 // closure to the entries on the discovered ref lists) or we overflow 2176 // the global marking stack. 2177 // 2178 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2179 // flag while there may still be some work to do. (See the comment at 2180 // the beginning of CMTask::do_marking_step() for those conditions - 2181 // one of which is reaching the specified time target.) It is only 2182 // when CMTask::do_marking_step() returns without setting the 2183 // has_aborted() flag that the marking step has completed. 2184 2185 _task->do_marking_step(1000000000.0 /* something very large */, 2186 _do_stealing, 2187 _do_termination); 2188 } while (_task->has_aborted() && !_cm->has_overflown()); 2189 } 2190 }; 2191 2192 // Implementation of AbstractRefProcTaskExecutor for parallel 2193 // reference processing at the end of G1 concurrent marking 2194 2195 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2196 private: 2197 G1CollectedHeap* _g1h; 2198 ConcurrentMark* _cm; 2199 WorkGang* _workers; 2200 int _active_workers; 2201 2202 public: 2203 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 2204 ConcurrentMark* cm, 2205 WorkGang* workers, 2206 int n_workers) : 2207 _g1h(g1h), _cm(cm), 2208 _workers(workers), _active_workers(n_workers) { } 2209 2210 // Executes the given task using concurrent marking worker threads. 2211 virtual void execute(ProcessTask& task); 2212 virtual void execute(EnqueueTask& task); 2213 }; 2214 2215 class G1CMRefProcTaskProxy: public AbstractGangTask { 2216 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2217 ProcessTask& _proc_task; 2218 G1CollectedHeap* _g1h; 2219 ConcurrentMark* _cm; 2220 bool _processing_is_mt; 2221 2222 public: 2223 G1CMRefProcTaskProxy(ProcessTask& proc_task, 2224 G1CollectedHeap* g1h, 2225 ConcurrentMark* cm) : 2226 AbstractGangTask("Process reference objects in parallel"), 2227 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 2228 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 2229 _processing_is_mt = rp->processing_is_mt(); 2230 } 2231 2232 virtual void work(uint worker_id) { 2233 CMTask* marking_task = _cm->task(worker_id); 2234 G1CMIsAliveClosure g1_is_alive(_g1h); 2235 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, marking_task); 2236 G1CMDrainMarkingStackClosure g1_par_drain(_cm, marking_task, _processing_is_mt); 2237 2238 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 2239 } 2240 }; 2241 2242 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 2243 assert(_workers != NULL, "Need parallel worker threads."); 2244 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2245 2246 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 2247 2248 // We need to reset the phase for each task execution so that 2249 // the termination protocol of CMTask::do_marking_step works. 2250 _cm->set_phase(_active_workers, false /* concurrent */); 2251 _g1h->set_par_threads(_active_workers); 2252 _workers->run_task(&proc_task_proxy); 2253 _g1h->set_par_threads(0); 2254 } 2255 2256 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 2257 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 2258 EnqueueTask& _enq_task; 2259 2260 public: 2261 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 2262 AbstractGangTask("Enqueue reference objects in parallel"), 2263 _enq_task(enq_task) { } 2264 2265 virtual void work(uint worker_id) { 2266 _enq_task.work(worker_id); 2267 } 2268 }; 2269 2270 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 2271 assert(_workers != NULL, "Need parallel worker threads."); 2272 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2273 2274 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 2275 2276 _g1h->set_par_threads(_active_workers); 2277 _workers->run_task(&enq_task_proxy); 2278 _g1h->set_par_threads(0); 2279 } 2280 2281 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 2282 ResourceMark rm; 2283 HandleMark hm; 2284 2285 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2286 2287 // Is alive closure. 2288 G1CMIsAliveClosure g1_is_alive(g1h); 2289 2290 // Inner scope to exclude the cleaning of the string and symbol 2291 // tables from the displayed time. 2292 { 2293 if (G1Log::finer()) { 2294 gclog_or_tty->put(' '); 2295 } 2296 GCTraceTime t("GC ref-proc", G1Log::finer(), false, g1h->gc_timer_cm()); 2297 2298 ReferenceProcessor* rp = g1h->ref_processor_cm(); 2299 2300 // See the comment in G1CollectedHeap::ref_processing_init() 2301 // about how reference processing currently works in G1. 2302 2303 // Set the soft reference policy 2304 rp->setup_policy(clear_all_soft_refs); 2305 assert(_markStack.isEmpty(), "mark stack should be empty"); 2306 2307 // Non-MT instances 'Keep Alive' and 'Complete GC' oop closures. 2308 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0)); 2309 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), false); 2310 2311 // We need at least one active thread. If reference processing is 2312 // not multi-threaded we use the current (ConcurrentMarkThread) thread, 2313 // otherwise we use the work gang from the G1CollectedHeap and we 2314 // utilize all the worker threads we can. 2315 uint active_workers = (rp->processing_is_mt() && g1h->workers() != NULL 2316 ? g1h->workers()->active_workers() 2317 : 1U); 2318 2319 active_workers = MAX2(MIN2(active_workers, _max_task_num), 1U); 2320 2321 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 2322 g1h->workers(), active_workers); 2323 2324 AbstractRefProcTaskExecutor* executor = (rp->processing_is_mt() 2325 ? &par_task_executor 2326 : NULL); 2327 2328 ReferenceProcessorStats stats; 2329 2330 // Set the degree of MT processing here. If the discovery was done MT, 2331 // the number of threads involved during discovery could differ from 2332 // the number of active workers. This is OK as long as the discovered 2333 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 2334 rp->set_active_mt_degree(active_workers); 2335 2336 // Process the weak references. 2337 stats = rp->process_discovered_references(&g1_is_alive, 2338 &g1_keep_alive, 2339 &g1_drain_mark_stack, 2340 executor, 2341 g1h->gc_timer_cm()); 2342 2343 // The do_oop work routines of the keep_alive and drain_marking_stack 2344 // oop closures will set the has_overflown flag if we overflow the 2345 // global marking stack. 2346 2347 g1h->gc_tracer_cm()->report_gc_reference_stats(stats); 2348 2349 assert(_markStack.overflow() || _markStack.isEmpty(), 2350 "mark stack should be empty (unless it overflowed)"); 2351 if (_markStack.overflow()) { 2352 // This should have been done already when we tried to push an 2353 // entry on to the global mark stack. But let's do it again. 2354 set_has_overflown(); 2355 } 2356 2357 assert(rp->num_q() == active_workers, "why not"); 2358 2359 rp->enqueue_discovered_references(executor); 2360 2361 rp->verify_no_references_recorded(); 2362 assert(!rp->discovery_enabled(), "Post condition"); 2363 } 2364 2365 // Now clean up stale oops in StringTable 2366 StringTable::unlink(&g1_is_alive); 2367 // Clean up unreferenced symbols in symbol table. 2368 SymbolTable::unlink(); 2369 } 2370 2371 void ConcurrentMark::swapMarkBitMaps() { 2372 CMBitMapRO* temp = _prevMarkBitMap; 2373 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; 2374 _nextMarkBitMap = (CMBitMap*) temp; 2375 } 2376 2377 class CMRemarkTask: public AbstractGangTask { 2378 private: 2379 ConcurrentMark *_cm; 2380 2381 public: 2382 void work(uint worker_id) { 2383 // Since all available tasks are actually started, we should 2384 // only proceed if we're supposed to be actived. 2385 if (worker_id < _cm->active_tasks()) { 2386 CMTask* task = _cm->task(worker_id); 2387 task->record_start_time(); 2388 do { 2389 task->do_marking_step(1000000000.0 /* something very large */, 2390 true /* do_stealing */, 2391 true /* do_termination */); 2392 } while (task->has_aborted() && !_cm->has_overflown()); 2393 // If we overflow, then we do not want to restart. We instead 2394 // want to abort remark and do concurrent marking again. 2395 task->record_end_time(); 2396 } 2397 } 2398 2399 CMRemarkTask(ConcurrentMark* cm, int active_workers) : 2400 AbstractGangTask("Par Remark"), _cm(cm) { 2401 _cm->terminator()->reset_for_reuse(active_workers); 2402 } 2403 }; 2404 2405 void ConcurrentMark::checkpointRootsFinalWork() { 2406 ResourceMark rm; 2407 HandleMark hm; 2408 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2409 2410 g1h->ensure_parsability(false); 2411 2412 if (G1CollectedHeap::use_parallel_gc_threads()) { 2413 G1CollectedHeap::StrongRootsScope srs(g1h); 2414 // this is remark, so we'll use up all active threads 2415 uint active_workers = g1h->workers()->active_workers(); 2416 if (active_workers == 0) { 2417 assert(active_workers > 0, "Should have been set earlier"); 2418 active_workers = (uint) ParallelGCThreads; 2419 g1h->workers()->set_active_workers(active_workers); 2420 } 2421 set_phase(active_workers, false /* concurrent */); 2422 // Leave _parallel_marking_threads at it's 2423 // value originally calculated in the ConcurrentMark 2424 // constructor and pass values of the active workers 2425 // through the gang in the task. 2426 2427 CMRemarkTask remarkTask(this, active_workers); 2428 g1h->set_par_threads(active_workers); 2429 g1h->workers()->run_task(&remarkTask); 2430 g1h->set_par_threads(0); 2431 } else { 2432 G1CollectedHeap::StrongRootsScope srs(g1h); 2433 // this is remark, so we'll use up all available threads 2434 uint active_workers = 1; 2435 set_phase(active_workers, false /* concurrent */); 2436 2437 CMRemarkTask remarkTask(this, active_workers); 2438 // We will start all available threads, even if we decide that the 2439 // active_workers will be fewer. The extra ones will just bail out 2440 // immediately. 2441 remarkTask.work(0); 2442 } 2443 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2444 guarantee(satb_mq_set.completed_buffers_num() == 0, "invariant"); 2445 2446 print_stats(); 2447 2448 #if VERIFY_OBJS_PROCESSED 2449 if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) { 2450 gclog_or_tty->print_cr("Processed = %d, enqueued = %d.", 2451 _scan_obj_cl.objs_processed, 2452 ThreadLocalObjQueue::objs_enqueued); 2453 guarantee(_scan_obj_cl.objs_processed == 2454 ThreadLocalObjQueue::objs_enqueued, 2455 "Different number of objs processed and enqueued."); 2456 } 2457 #endif 2458 } 2459 2460 #ifndef PRODUCT 2461 2462 class PrintReachableOopClosure: public OopClosure { 2463 private: 2464 G1CollectedHeap* _g1h; 2465 outputStream* _out; 2466 VerifyOption _vo; 2467 bool _all; 2468 2469 public: 2470 PrintReachableOopClosure(outputStream* out, 2471 VerifyOption vo, 2472 bool all) : 2473 _g1h(G1CollectedHeap::heap()), 2474 _out(out), _vo(vo), _all(all) { } 2475 2476 void do_oop(narrowOop* p) { do_oop_work(p); } 2477 void do_oop( oop* p) { do_oop_work(p); } 2478 2479 template <class T> void do_oop_work(T* p) { 2480 oop obj = oopDesc::load_decode_heap_oop(p); 2481 const char* str = NULL; 2482 const char* str2 = ""; 2483 2484 if (obj == NULL) { 2485 str = ""; 2486 } else if (!_g1h->is_in_g1_reserved(obj)) { 2487 str = " O"; 2488 } else { 2489 HeapRegion* hr = _g1h->heap_region_containing(obj); 2490 guarantee(hr != NULL, "invariant"); 2491 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo); 2492 bool marked = _g1h->is_marked(obj, _vo); 2493 2494 if (over_tams) { 2495 str = " >"; 2496 if (marked) { 2497 str2 = " AND MARKED"; 2498 } 2499 } else if (marked) { 2500 str = " M"; 2501 } else { 2502 str = " NOT"; 2503 } 2504 } 2505 2506 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s", 2507 p, (void*) obj, str, str2); 2508 } 2509 }; 2510 2511 class PrintReachableObjectClosure : public ObjectClosure { 2512 private: 2513 G1CollectedHeap* _g1h; 2514 outputStream* _out; 2515 VerifyOption _vo; 2516 bool _all; 2517 HeapRegion* _hr; 2518 2519 public: 2520 PrintReachableObjectClosure(outputStream* out, 2521 VerifyOption vo, 2522 bool all, 2523 HeapRegion* hr) : 2524 _g1h(G1CollectedHeap::heap()), 2525 _out(out), _vo(vo), _all(all), _hr(hr) { } 2526 2527 void do_object(oop o) { 2528 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo); 2529 bool marked = _g1h->is_marked(o, _vo); 2530 bool print_it = _all || over_tams || marked; 2531 2532 if (print_it) { 2533 _out->print_cr(" "PTR_FORMAT"%s", 2534 o, (over_tams) ? " >" : (marked) ? " M" : ""); 2535 PrintReachableOopClosure oopCl(_out, _vo, _all); 2536 o->oop_iterate(&oopCl); 2537 } 2538 } 2539 }; 2540 2541 class PrintReachableRegionClosure : public HeapRegionClosure { 2542 private: 2543 G1CollectedHeap* _g1h; 2544 outputStream* _out; 2545 VerifyOption _vo; 2546 bool _all; 2547 2548 public: 2549 bool doHeapRegion(HeapRegion* hr) { 2550 HeapWord* b = hr->bottom(); 2551 HeapWord* e = hr->end(); 2552 HeapWord* t = hr->top(); 2553 HeapWord* p = _g1h->top_at_mark_start(hr, _vo); 2554 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" " 2555 "TAMS: "PTR_FORMAT, b, e, t, p); 2556 _out->cr(); 2557 2558 HeapWord* from = b; 2559 HeapWord* to = t; 2560 2561 if (to > from) { 2562 _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to); 2563 _out->cr(); 2564 PrintReachableObjectClosure ocl(_out, _vo, _all, hr); 2565 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl); 2566 _out->cr(); 2567 } 2568 2569 return false; 2570 } 2571 2572 PrintReachableRegionClosure(outputStream* out, 2573 VerifyOption vo, 2574 bool all) : 2575 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { } 2576 }; 2577 2578 void ConcurrentMark::print_reachable(const char* str, 2579 VerifyOption vo, 2580 bool all) { 2581 gclog_or_tty->cr(); 2582 gclog_or_tty->print_cr("== Doing heap dump... "); 2583 2584 if (G1PrintReachableBaseFile == NULL) { 2585 gclog_or_tty->print_cr(" #### error: no base file defined"); 2586 return; 2587 } 2588 2589 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) > 2590 (JVM_MAXPATHLEN - 1)) { 2591 gclog_or_tty->print_cr(" #### error: file name too long"); 2592 return; 2593 } 2594 2595 char file_name[JVM_MAXPATHLEN]; 2596 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str); 2597 gclog_or_tty->print_cr(" dumping to file %s", file_name); 2598 2599 fileStream fout(file_name); 2600 if (!fout.is_open()) { 2601 gclog_or_tty->print_cr(" #### error: could not open file"); 2602 return; 2603 } 2604 2605 outputStream* out = &fout; 2606 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo)); 2607 out->cr(); 2608 2609 out->print_cr("--- ITERATING OVER REGIONS"); 2610 out->cr(); 2611 PrintReachableRegionClosure rcl(out, vo, all); 2612 _g1h->heap_region_iterate(&rcl); 2613 out->cr(); 2614 2615 gclog_or_tty->print_cr(" done"); 2616 gclog_or_tty->flush(); 2617 } 2618 2619 #endif // PRODUCT 2620 2621 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 2622 // Note we are overriding the read-only view of the prev map here, via 2623 // the cast. 2624 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); 2625 } 2626 2627 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) { 2628 _nextMarkBitMap->clearRange(mr); 2629 } 2630 2631 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) { 2632 clearRangePrevBitmap(mr); 2633 clearRangeNextBitmap(mr); 2634 } 2635 2636 HeapRegion* 2637 ConcurrentMark::claim_region(int task_num) { 2638 // "checkpoint" the finger 2639 HeapWord* finger = _finger; 2640 2641 // _heap_end will not change underneath our feet; it only changes at 2642 // yield points. 2643 while (finger < _heap_end) { 2644 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 2645 2646 // Note on how this code handles humongous regions. In the 2647 // normal case the finger will reach the start of a "starts 2648 // humongous" (SH) region. Its end will either be the end of the 2649 // last "continues humongous" (CH) region in the sequence, or the 2650 // standard end of the SH region (if the SH is the only region in 2651 // the sequence). That way claim_region() will skip over the CH 2652 // regions. However, there is a subtle race between a CM thread 2653 // executing this method and a mutator thread doing a humongous 2654 // object allocation. The two are not mutually exclusive as the CM 2655 // thread does not need to hold the Heap_lock when it gets 2656 // here. So there is a chance that claim_region() will come across 2657 // a free region that's in the progress of becoming a SH or a CH 2658 // region. In the former case, it will either 2659 // a) Miss the update to the region's end, in which case it will 2660 // visit every subsequent CH region, will find their bitmaps 2661 // empty, and do nothing, or 2662 // b) Will observe the update of the region's end (in which case 2663 // it will skip the subsequent CH regions). 2664 // If it comes across a region that suddenly becomes CH, the 2665 // scenario will be similar to b). So, the race between 2666 // claim_region() and a humongous object allocation might force us 2667 // to do a bit of unnecessary work (due to some unnecessary bitmap 2668 // iterations) but it should not introduce and correctness issues. 2669 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger); 2670 HeapWord* bottom = curr_region->bottom(); 2671 HeapWord* end = curr_region->end(); 2672 HeapWord* limit = curr_region->next_top_at_mark_start(); 2673 2674 if (verbose_low()) { 2675 gclog_or_tty->print_cr("[%d] curr_region = "PTR_FORMAT" " 2676 "["PTR_FORMAT", "PTR_FORMAT"), " 2677 "limit = "PTR_FORMAT, 2678 task_num, curr_region, bottom, end, limit); 2679 } 2680 2681 // Is the gap between reading the finger and doing the CAS too long? 2682 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 2683 if (res == finger) { 2684 // we succeeded 2685 2686 // notice that _finger == end cannot be guaranteed here since, 2687 // someone else might have moved the finger even further 2688 assert(_finger >= end, "the finger should have moved forward"); 2689 2690 if (verbose_low()) { 2691 gclog_or_tty->print_cr("[%d] we were successful with region = " 2692 PTR_FORMAT, task_num, curr_region); 2693 } 2694 2695 if (limit > bottom) { 2696 if (verbose_low()) { 2697 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is not empty, " 2698 "returning it ", task_num, curr_region); 2699 } 2700 return curr_region; 2701 } else { 2702 assert(limit == bottom, 2703 "the region limit should be at bottom"); 2704 if (verbose_low()) { 2705 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is empty, " 2706 "returning NULL", task_num, curr_region); 2707 } 2708 // we return NULL and the caller should try calling 2709 // claim_region() again. 2710 return NULL; 2711 } 2712 } else { 2713 assert(_finger > finger, "the finger should have moved forward"); 2714 if (verbose_low()) { 2715 gclog_or_tty->print_cr("[%d] somebody else moved the finger, " 2716 "global finger = "PTR_FORMAT", " 2717 "our finger = "PTR_FORMAT, 2718 task_num, _finger, finger); 2719 } 2720 2721 // read it again 2722 finger = _finger; 2723 } 2724 } 2725 2726 return NULL; 2727 } 2728 2729 #ifndef PRODUCT 2730 enum VerifyNoCSetOopsPhase { 2731 VerifyNoCSetOopsStack, 2732 VerifyNoCSetOopsQueues, 2733 VerifyNoCSetOopsSATBCompleted, 2734 VerifyNoCSetOopsSATBThread 2735 }; 2736 2737 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure { 2738 private: 2739 G1CollectedHeap* _g1h; 2740 VerifyNoCSetOopsPhase _phase; 2741 int _info; 2742 2743 const char* phase_str() { 2744 switch (_phase) { 2745 case VerifyNoCSetOopsStack: return "Stack"; 2746 case VerifyNoCSetOopsQueues: return "Queue"; 2747 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers"; 2748 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers"; 2749 default: ShouldNotReachHere(); 2750 } 2751 return NULL; 2752 } 2753 2754 void do_object_work(oop obj) { 2755 guarantee(!_g1h->obj_in_cs(obj), 2756 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d", 2757 (void*) obj, phase_str(), _info)); 2758 } 2759 2760 public: 2761 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { } 2762 2763 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) { 2764 _phase = phase; 2765 _info = info; 2766 } 2767 2768 virtual void do_oop(oop* p) { 2769 oop obj = oopDesc::load_decode_heap_oop(p); 2770 do_object_work(obj); 2771 } 2772 2773 virtual void do_oop(narrowOop* p) { 2774 // We should not come across narrow oops while scanning marking 2775 // stacks and SATB buffers. 2776 ShouldNotReachHere(); 2777 } 2778 2779 virtual void do_object(oop obj) { 2780 do_object_work(obj); 2781 } 2782 }; 2783 2784 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks, 2785 bool verify_enqueued_buffers, 2786 bool verify_thread_buffers, 2787 bool verify_fingers) { 2788 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 2789 if (!G1CollectedHeap::heap()->mark_in_progress()) { 2790 return; 2791 } 2792 2793 VerifyNoCSetOopsClosure cl; 2794 2795 if (verify_stacks) { 2796 // Verify entries on the global mark stack 2797 cl.set_phase(VerifyNoCSetOopsStack); 2798 _markStack.oops_do(&cl); 2799 2800 // Verify entries on the task queues 2801 for (int i = 0; i < (int) _max_task_num; i += 1) { 2802 cl.set_phase(VerifyNoCSetOopsQueues, i); 2803 OopTaskQueue* queue = _task_queues->queue(i); 2804 queue->oops_do(&cl); 2805 } 2806 } 2807 2808 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 2809 2810 // Verify entries on the enqueued SATB buffers 2811 if (verify_enqueued_buffers) { 2812 cl.set_phase(VerifyNoCSetOopsSATBCompleted); 2813 satb_qs.iterate_completed_buffers_read_only(&cl); 2814 } 2815 2816 // Verify entries on the per-thread SATB buffers 2817 if (verify_thread_buffers) { 2818 cl.set_phase(VerifyNoCSetOopsSATBThread); 2819 satb_qs.iterate_thread_buffers_read_only(&cl); 2820 } 2821 2822 if (verify_fingers) { 2823 // Verify the global finger 2824 HeapWord* global_finger = finger(); 2825 if (global_finger != NULL && global_finger < _heap_end) { 2826 // The global finger always points to a heap region boundary. We 2827 // use heap_region_containing_raw() to get the containing region 2828 // given that the global finger could be pointing to a free region 2829 // which subsequently becomes continues humongous. If that 2830 // happens, heap_region_containing() will return the bottom of the 2831 // corresponding starts humongous region and the check below will 2832 // not hold any more. 2833 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger); 2834 guarantee(global_finger == global_hr->bottom(), 2835 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT, 2836 global_finger, HR_FORMAT_PARAMS(global_hr))); 2837 } 2838 2839 // Verify the task fingers 2840 assert(parallel_marking_threads() <= _max_task_num, "sanity"); 2841 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) { 2842 CMTask* task = _tasks[i]; 2843 HeapWord* task_finger = task->finger(); 2844 if (task_finger != NULL && task_finger < _heap_end) { 2845 // See above note on the global finger verification. 2846 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger); 2847 guarantee(task_finger == task_hr->bottom() || 2848 !task_hr->in_collection_set(), 2849 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT, 2850 task_finger, HR_FORMAT_PARAMS(task_hr))); 2851 } 2852 } 2853 } 2854 } 2855 #endif // PRODUCT 2856 2857 // Aggregate the counting data that was constructed concurrently 2858 // with marking. 2859 class AggregateCountDataHRClosure: public HeapRegionClosure { 2860 G1CollectedHeap* _g1h; 2861 ConcurrentMark* _cm; 2862 CardTableModRefBS* _ct_bs; 2863 BitMap* _cm_card_bm; 2864 size_t _max_task_num; 2865 2866 public: 2867 AggregateCountDataHRClosure(G1CollectedHeap* g1h, 2868 BitMap* cm_card_bm, 2869 size_t max_task_num) : 2870 _g1h(g1h), _cm(g1h->concurrent_mark()), 2871 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 2872 _cm_card_bm(cm_card_bm), _max_task_num(max_task_num) { } 2873 2874 bool doHeapRegion(HeapRegion* hr) { 2875 if (hr->continuesHumongous()) { 2876 // We will ignore these here and process them when their 2877 // associated "starts humongous" region is processed. 2878 // Note that we cannot rely on their associated 2879 // "starts humongous" region to have their bit set to 1 2880 // since, due to the region chunking in the parallel region 2881 // iteration, a "continues humongous" region might be visited 2882 // before its associated "starts humongous". 2883 return false; 2884 } 2885 2886 HeapWord* start = hr->bottom(); 2887 HeapWord* limit = hr->next_top_at_mark_start(); 2888 HeapWord* end = hr->end(); 2889 2890 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(), 2891 err_msg("Preconditions not met - " 2892 "start: "PTR_FORMAT", limit: "PTR_FORMAT", " 2893 "top: "PTR_FORMAT", end: "PTR_FORMAT, 2894 start, limit, hr->top(), hr->end())); 2895 2896 assert(hr->next_marked_bytes() == 0, "Precondition"); 2897 2898 if (start == limit) { 2899 // NTAMS of this region has not been set so nothing to do. 2900 return false; 2901 } 2902 2903 // 'start' should be in the heap. 2904 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity"); 2905 // 'end' *may* be just beyone the end of the heap (if hr is the last region) 2906 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity"); 2907 2908 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 2909 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit); 2910 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end); 2911 2912 // If ntams is not card aligned then we bump card bitmap index 2913 // for limit so that we get the all the cards spanned by 2914 // the object ending at ntams. 2915 // Note: if this is the last region in the heap then ntams 2916 // could be actually just beyond the end of the the heap; 2917 // limit_idx will then correspond to a (non-existent) card 2918 // that is also outside the heap. 2919 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) { 2920 limit_idx += 1; 2921 } 2922 2923 assert(limit_idx <= end_idx, "or else use atomics"); 2924 2925 // Aggregate the "stripe" in the count data associated with hr. 2926 uint hrs_index = hr->hrs_index(); 2927 size_t marked_bytes = 0; 2928 2929 for (int i = 0; (size_t)i < _max_task_num; i += 1) { 2930 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i); 2931 BitMap* task_card_bm = _cm->count_card_bitmap_for(i); 2932 2933 // Fetch the marked_bytes in this region for task i and 2934 // add it to the running total for this region. 2935 marked_bytes += marked_bytes_array[hrs_index]; 2936 2937 // Now union the bitmaps[0,max_task_num)[start_idx..limit_idx) 2938 // into the global card bitmap. 2939 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx); 2940 2941 while (scan_idx < limit_idx) { 2942 assert(task_card_bm->at(scan_idx) == true, "should be"); 2943 _cm_card_bm->set_bit(scan_idx); 2944 assert(_cm_card_bm->at(scan_idx) == true, "should be"); 2945 2946 // BitMap::get_next_one_offset() can handle the case when 2947 // its left_offset parameter is greater than its right_offset 2948 // parameter. It does, however, have an early exit if 2949 // left_offset == right_offset. So let's limit the value 2950 // passed in for left offset here. 2951 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx); 2952 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx); 2953 } 2954 } 2955 2956 // Update the marked bytes for this region. 2957 hr->add_to_marked_bytes(marked_bytes); 2958 2959 // Next heap region 2960 return false; 2961 } 2962 }; 2963 2964 class G1AggregateCountDataTask: public AbstractGangTask { 2965 protected: 2966 G1CollectedHeap* _g1h; 2967 ConcurrentMark* _cm; 2968 BitMap* _cm_card_bm; 2969 size_t _max_task_num; 2970 int _active_workers; 2971 2972 public: 2973 G1AggregateCountDataTask(G1CollectedHeap* g1h, 2974 ConcurrentMark* cm, 2975 BitMap* cm_card_bm, 2976 size_t max_task_num, 2977 int n_workers) : 2978 AbstractGangTask("Count Aggregation"), 2979 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm), 2980 _max_task_num(max_task_num), 2981 _active_workers(n_workers) { } 2982 2983 void work(uint worker_id) { 2984 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_task_num); 2985 2986 if (G1CollectedHeap::use_parallel_gc_threads()) { 2987 _g1h->heap_region_par_iterate_chunked(&cl, worker_id, 2988 _active_workers, 2989 HeapRegion::AggregateCountClaimValue); 2990 } else { 2991 _g1h->heap_region_iterate(&cl); 2992 } 2993 } 2994 }; 2995 2996 2997 void ConcurrentMark::aggregate_count_data() { 2998 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2999 _g1h->workers()->active_workers() : 3000 1); 3001 3002 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm, 3003 _max_task_num, n_workers); 3004 3005 if (G1CollectedHeap::use_parallel_gc_threads()) { 3006 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3007 "sanity check"); 3008 _g1h->set_par_threads(n_workers); 3009 _g1h->workers()->run_task(&g1_par_agg_task); 3010 _g1h->set_par_threads(0); 3011 3012 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue), 3013 "sanity check"); 3014 _g1h->reset_heap_region_claim_values(); 3015 } else { 3016 g1_par_agg_task.work(0); 3017 } 3018 } 3019 3020 // Clear the per-worker arrays used to store the per-region counting data 3021 void ConcurrentMark::clear_all_count_data() { 3022 // Clear the global card bitmap - it will be filled during 3023 // liveness count aggregation (during remark) and the 3024 // final counting task. 3025 _card_bm.clear(); 3026 3027 // Clear the global region bitmap - it will be filled as part 3028 // of the final counting task. 3029 _region_bm.clear(); 3030 3031 uint max_regions = _g1h->max_regions(); 3032 assert(_max_task_num != 0, "unitialized"); 3033 3034 for (int i = 0; (size_t) i < _max_task_num; i += 1) { 3035 BitMap* task_card_bm = count_card_bitmap_for(i); 3036 size_t* marked_bytes_array = count_marked_bytes_array_for(i); 3037 3038 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 3039 assert(marked_bytes_array != NULL, "uninitialized"); 3040 3041 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t)); 3042 task_card_bm->clear(); 3043 } 3044 } 3045 3046 void ConcurrentMark::print_stats() { 3047 if (verbose_stats()) { 3048 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3049 for (size_t i = 0; i < _active_tasks; ++i) { 3050 _tasks[i]->print_stats(); 3051 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3052 } 3053 } 3054 } 3055 3056 // abandon current marking iteration due to a Full GC 3057 void ConcurrentMark::abort() { 3058 // Clear all marks to force marking thread to do nothing 3059 _nextMarkBitMap->clearAll(); 3060 // Clear the liveness counting data 3061 clear_all_count_data(); 3062 // Empty mark stack 3063 reset_marking_state(); 3064 for (int i = 0; i < (int)_max_task_num; ++i) { 3065 _tasks[i]->clear_region_fields(); 3066 } 3067 _has_aborted = true; 3068 3069 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3070 satb_mq_set.abandon_partial_marking(); 3071 // This can be called either during or outside marking, we'll read 3072 // the expected_active value from the SATB queue set. 3073 satb_mq_set.set_active_all_threads( 3074 false, /* new active value */ 3075 satb_mq_set.is_active() /* expected_active */); 3076 3077 _g1h->trace_heap_after_concurrent_cycle(); 3078 _g1h->register_concurrent_cycle_end(); 3079 } 3080 3081 static void print_ms_time_info(const char* prefix, const char* name, 3082 NumberSeq& ns) { 3083 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 3084 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 3085 if (ns.num() > 0) { 3086 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]", 3087 prefix, ns.sd(), ns.maximum()); 3088 } 3089 } 3090 3091 void ConcurrentMark::print_summary_info() { 3092 gclog_or_tty->print_cr(" Concurrent marking:"); 3093 print_ms_time_info(" ", "init marks", _init_times); 3094 print_ms_time_info(" ", "remarks", _remark_times); 3095 { 3096 print_ms_time_info(" ", "final marks", _remark_mark_times); 3097 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 3098 3099 } 3100 print_ms_time_info(" ", "cleanups", _cleanup_times); 3101 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).", 3102 _total_counting_time, 3103 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / 3104 (double)_cleanup_times.num() 3105 : 0.0)); 3106 if (G1ScrubRemSets) { 3107 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 3108 _total_rs_scrub_time, 3109 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / 3110 (double)_cleanup_times.num() 3111 : 0.0)); 3112 } 3113 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.", 3114 (_init_times.sum() + _remark_times.sum() + 3115 _cleanup_times.sum())/1000.0); 3116 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s " 3117 "(%8.2f s marking).", 3118 cmThread()->vtime_accum(), 3119 cmThread()->vtime_mark_accum()); 3120 } 3121 3122 void ConcurrentMark::print_worker_threads_on(outputStream* st) const { 3123 if (use_parallel_marking_threads()) { 3124 _parallel_workers->print_worker_threads_on(st); 3125 } 3126 } 3127 3128 // We take a break if someone is trying to stop the world. 3129 bool ConcurrentMark::do_yield_check(uint worker_id) { 3130 if (should_yield()) { 3131 if (worker_id == 0) { 3132 _g1h->g1_policy()->record_concurrent_pause(); 3133 } 3134 cmThread()->yield(); 3135 return true; 3136 } else { 3137 return false; 3138 } 3139 } 3140 3141 bool ConcurrentMark::should_yield() { 3142 return cmThread()->should_yield(); 3143 } 3144 3145 bool ConcurrentMark::containing_card_is_marked(void* p) { 3146 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1); 3147 return _card_bm.at(offset >> CardTableModRefBS::card_shift); 3148 } 3149 3150 bool ConcurrentMark::containing_cards_are_marked(void* start, 3151 void* last) { 3152 return containing_card_is_marked(start) && 3153 containing_card_is_marked(last); 3154 } 3155 3156 #ifndef PRODUCT 3157 // for debugging purposes 3158 void ConcurrentMark::print_finger() { 3159 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT, 3160 _heap_start, _heap_end, _finger); 3161 for (int i = 0; i < (int) _max_task_num; ++i) { 3162 gclog_or_tty->print(" %d: "PTR_FORMAT, i, _tasks[i]->finger()); 3163 } 3164 gclog_or_tty->print_cr(""); 3165 } 3166 #endif 3167 3168 void CMTask::scan_object(oop obj) { 3169 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant"); 3170 3171 if (_cm->verbose_high()) { 3172 gclog_or_tty->print_cr("[%d] we're scanning object "PTR_FORMAT, 3173 _task_id, (void*) obj); 3174 } 3175 3176 size_t obj_size = obj->size(); 3177 _words_scanned += obj_size; 3178 3179 obj->oop_iterate(_cm_oop_closure); 3180 statsOnly( ++_objs_scanned ); 3181 check_limits(); 3182 } 3183 3184 // Closure for iteration over bitmaps 3185 class CMBitMapClosure : public BitMapClosure { 3186 private: 3187 // the bitmap that is being iterated over 3188 CMBitMap* _nextMarkBitMap; 3189 ConcurrentMark* _cm; 3190 CMTask* _task; 3191 3192 public: 3193 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) : 3194 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } 3195 3196 bool do_bit(size_t offset) { 3197 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); 3198 assert(_nextMarkBitMap->isMarked(addr), "invariant"); 3199 assert( addr < _cm->finger(), "invariant"); 3200 3201 statsOnly( _task->increase_objs_found_on_bitmap() ); 3202 assert(addr >= _task->finger(), "invariant"); 3203 3204 // We move that task's local finger along. 3205 _task->move_finger_to(addr); 3206 3207 _task->scan_object(oop(addr)); 3208 // we only partially drain the local queue and global stack 3209 _task->drain_local_queue(true); 3210 _task->drain_global_stack(true); 3211 3212 // if the has_aborted flag has been raised, we need to bail out of 3213 // the iteration 3214 return !_task->has_aborted(); 3215 } 3216 }; 3217 3218 // Closure for iterating over objects, currently only used for 3219 // processing SATB buffers. 3220 class CMObjectClosure : public ObjectClosure { 3221 private: 3222 CMTask* _task; 3223 3224 public: 3225 void do_object(oop obj) { 3226 _task->deal_with_reference(obj); 3227 } 3228 3229 CMObjectClosure(CMTask* task) : _task(task) { } 3230 }; 3231 3232 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 3233 ConcurrentMark* cm, 3234 CMTask* task) 3235 : _g1h(g1h), _cm(cm), _task(task) { 3236 assert(_ref_processor == NULL, "should be initialized to NULL"); 3237 3238 if (G1UseConcMarkReferenceProcessing) { 3239 _ref_processor = g1h->ref_processor_cm(); 3240 assert(_ref_processor != NULL, "should not be NULL"); 3241 } 3242 } 3243 3244 void CMTask::setup_for_region(HeapRegion* hr) { 3245 // Separated the asserts so that we know which one fires. 3246 assert(hr != NULL, 3247 "claim_region() should have filtered out continues humongous regions"); 3248 assert(!hr->continuesHumongous(), 3249 "claim_region() should have filtered out continues humongous regions"); 3250 3251 if (_cm->verbose_low()) { 3252 gclog_or_tty->print_cr("[%d] setting up for region "PTR_FORMAT, 3253 _task_id, hr); 3254 } 3255 3256 _curr_region = hr; 3257 _finger = hr->bottom(); 3258 update_region_limit(); 3259 } 3260 3261 void CMTask::update_region_limit() { 3262 HeapRegion* hr = _curr_region; 3263 HeapWord* bottom = hr->bottom(); 3264 HeapWord* limit = hr->next_top_at_mark_start(); 3265 3266 if (limit == bottom) { 3267 if (_cm->verbose_low()) { 3268 gclog_or_tty->print_cr("[%d] found an empty region " 3269 "["PTR_FORMAT", "PTR_FORMAT")", 3270 _task_id, bottom, limit); 3271 } 3272 // The region was collected underneath our feet. 3273 // We set the finger to bottom to ensure that the bitmap 3274 // iteration that will follow this will not do anything. 3275 // (this is not a condition that holds when we set the region up, 3276 // as the region is not supposed to be empty in the first place) 3277 _finger = bottom; 3278 } else if (limit >= _region_limit) { 3279 assert(limit >= _finger, "peace of mind"); 3280 } else { 3281 assert(limit < _region_limit, "only way to get here"); 3282 // This can happen under some pretty unusual circumstances. An 3283 // evacuation pause empties the region underneath our feet (NTAMS 3284 // at bottom). We then do some allocation in the region (NTAMS 3285 // stays at bottom), followed by the region being used as a GC 3286 // alloc region (NTAMS will move to top() and the objects 3287 // originally below it will be grayed). All objects now marked in 3288 // the region are explicitly grayed, if below the global finger, 3289 // and we do not need in fact to scan anything else. So, we simply 3290 // set _finger to be limit to ensure that the bitmap iteration 3291 // doesn't do anything. 3292 _finger = limit; 3293 } 3294 3295 _region_limit = limit; 3296 } 3297 3298 void CMTask::giveup_current_region() { 3299 assert(_curr_region != NULL, "invariant"); 3300 if (_cm->verbose_low()) { 3301 gclog_or_tty->print_cr("[%d] giving up region "PTR_FORMAT, 3302 _task_id, _curr_region); 3303 } 3304 clear_region_fields(); 3305 } 3306 3307 void CMTask::clear_region_fields() { 3308 // Values for these three fields that indicate that we're not 3309 // holding on to a region. 3310 _curr_region = NULL; 3311 _finger = NULL; 3312 _region_limit = NULL; 3313 } 3314 3315 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 3316 if (cm_oop_closure == NULL) { 3317 assert(_cm_oop_closure != NULL, "invariant"); 3318 } else { 3319 assert(_cm_oop_closure == NULL, "invariant"); 3320 } 3321 _cm_oop_closure = cm_oop_closure; 3322 } 3323 3324 void CMTask::reset(CMBitMap* nextMarkBitMap) { 3325 guarantee(nextMarkBitMap != NULL, "invariant"); 3326 3327 if (_cm->verbose_low()) { 3328 gclog_or_tty->print_cr("[%d] resetting", _task_id); 3329 } 3330 3331 _nextMarkBitMap = nextMarkBitMap; 3332 clear_region_fields(); 3333 3334 _calls = 0; 3335 _elapsed_time_ms = 0.0; 3336 _termination_time_ms = 0.0; 3337 _termination_start_time_ms = 0.0; 3338 3339 #if _MARKING_STATS_ 3340 _local_pushes = 0; 3341 _local_pops = 0; 3342 _local_max_size = 0; 3343 _objs_scanned = 0; 3344 _global_pushes = 0; 3345 _global_pops = 0; 3346 _global_max_size = 0; 3347 _global_transfers_to = 0; 3348 _global_transfers_from = 0; 3349 _regions_claimed = 0; 3350 _objs_found_on_bitmap = 0; 3351 _satb_buffers_processed = 0; 3352 _steal_attempts = 0; 3353 _steals = 0; 3354 _aborted = 0; 3355 _aborted_overflow = 0; 3356 _aborted_cm_aborted = 0; 3357 _aborted_yield = 0; 3358 _aborted_timed_out = 0; 3359 _aborted_satb = 0; 3360 _aborted_termination = 0; 3361 #endif // _MARKING_STATS_ 3362 } 3363 3364 bool CMTask::should_exit_termination() { 3365 regular_clock_call(); 3366 // This is called when we are in the termination protocol. We should 3367 // quit if, for some reason, this task wants to abort or the global 3368 // stack is not empty (this means that we can get work from it). 3369 return !_cm->mark_stack_empty() || has_aborted(); 3370 } 3371 3372 void CMTask::reached_limit() { 3373 assert(_words_scanned >= _words_scanned_limit || 3374 _refs_reached >= _refs_reached_limit , 3375 "shouldn't have been called otherwise"); 3376 regular_clock_call(); 3377 } 3378 3379 void CMTask::regular_clock_call() { 3380 if (has_aborted()) return; 3381 3382 // First, we need to recalculate the words scanned and refs reached 3383 // limits for the next clock call. 3384 recalculate_limits(); 3385 3386 // During the regular clock call we do the following 3387 3388 // (1) If an overflow has been flagged, then we abort. 3389 if (_cm->has_overflown()) { 3390 set_has_aborted(); 3391 return; 3392 } 3393 3394 // If we are not concurrent (i.e. we're doing remark) we don't need 3395 // to check anything else. The other steps are only needed during 3396 // the concurrent marking phase. 3397 if (!concurrent()) return; 3398 3399 // (2) If marking has been aborted for Full GC, then we also abort. 3400 if (_cm->has_aborted()) { 3401 set_has_aborted(); 3402 statsOnly( ++_aborted_cm_aborted ); 3403 return; 3404 } 3405 3406 double curr_time_ms = os::elapsedVTime() * 1000.0; 3407 3408 // (3) If marking stats are enabled, then we update the step history. 3409 #if _MARKING_STATS_ 3410 if (_words_scanned >= _words_scanned_limit) { 3411 ++_clock_due_to_scanning; 3412 } 3413 if (_refs_reached >= _refs_reached_limit) { 3414 ++_clock_due_to_marking; 3415 } 3416 3417 double last_interval_ms = curr_time_ms - _interval_start_time_ms; 3418 _interval_start_time_ms = curr_time_ms; 3419 _all_clock_intervals_ms.add(last_interval_ms); 3420 3421 if (_cm->verbose_medium()) { 3422 gclog_or_tty->print_cr("[%d] regular clock, interval = %1.2lfms, " 3423 "scanned = %d%s, refs reached = %d%s", 3424 _task_id, last_interval_ms, 3425 _words_scanned, 3426 (_words_scanned >= _words_scanned_limit) ? " (*)" : "", 3427 _refs_reached, 3428 (_refs_reached >= _refs_reached_limit) ? " (*)" : ""); 3429 } 3430 #endif // _MARKING_STATS_ 3431 3432 // (4) We check whether we should yield. If we have to, then we abort. 3433 if (_cm->should_yield()) { 3434 // We should yield. To do this we abort the task. The caller is 3435 // responsible for yielding. 3436 set_has_aborted(); 3437 statsOnly( ++_aborted_yield ); 3438 return; 3439 } 3440 3441 // (5) We check whether we've reached our time quota. If we have, 3442 // then we abort. 3443 double elapsed_time_ms = curr_time_ms - _start_time_ms; 3444 if (elapsed_time_ms > _time_target_ms) { 3445 set_has_aborted(); 3446 _has_timed_out = true; 3447 statsOnly( ++_aborted_timed_out ); 3448 return; 3449 } 3450 3451 // (6) Finally, we check whether there are enough completed STAB 3452 // buffers available for processing. If there are, we abort. 3453 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3454 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 3455 if (_cm->verbose_low()) { 3456 gclog_or_tty->print_cr("[%d] aborting to deal with pending SATB buffers", 3457 _task_id); 3458 } 3459 // we do need to process SATB buffers, we'll abort and restart 3460 // the marking task to do so 3461 set_has_aborted(); 3462 statsOnly( ++_aborted_satb ); 3463 return; 3464 } 3465 } 3466 3467 void CMTask::recalculate_limits() { 3468 _real_words_scanned_limit = _words_scanned + words_scanned_period; 3469 _words_scanned_limit = _real_words_scanned_limit; 3470 3471 _real_refs_reached_limit = _refs_reached + refs_reached_period; 3472 _refs_reached_limit = _real_refs_reached_limit; 3473 } 3474 3475 void CMTask::decrease_limits() { 3476 // This is called when we believe that we're going to do an infrequent 3477 // operation which will increase the per byte scanned cost (i.e. move 3478 // entries to/from the global stack). It basically tries to decrease the 3479 // scanning limit so that the clock is called earlier. 3480 3481 if (_cm->verbose_medium()) { 3482 gclog_or_tty->print_cr("[%d] decreasing limits", _task_id); 3483 } 3484 3485 _words_scanned_limit = _real_words_scanned_limit - 3486 3 * words_scanned_period / 4; 3487 _refs_reached_limit = _real_refs_reached_limit - 3488 3 * refs_reached_period / 4; 3489 } 3490 3491 void CMTask::move_entries_to_global_stack() { 3492 // local array where we'll store the entries that will be popped 3493 // from the local queue 3494 oop buffer[global_stack_transfer_size]; 3495 3496 int n = 0; 3497 oop obj; 3498 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { 3499 buffer[n] = obj; 3500 ++n; 3501 } 3502 3503 if (n > 0) { 3504 // we popped at least one entry from the local queue 3505 3506 statsOnly( ++_global_transfers_to; _local_pops += n ); 3507 3508 if (!_cm->mark_stack_push(buffer, n)) { 3509 if (_cm->verbose_low()) { 3510 gclog_or_tty->print_cr("[%d] aborting due to global stack overflow", 3511 _task_id); 3512 } 3513 set_has_aborted(); 3514 } else { 3515 // the transfer was successful 3516 3517 if (_cm->verbose_medium()) { 3518 gclog_or_tty->print_cr("[%d] pushed %d entries to the global stack", 3519 _task_id, n); 3520 } 3521 statsOnly( int tmp_size = _cm->mark_stack_size(); 3522 if (tmp_size > _global_max_size) { 3523 _global_max_size = tmp_size; 3524 } 3525 _global_pushes += n ); 3526 } 3527 } 3528 3529 // this operation was quite expensive, so decrease the limits 3530 decrease_limits(); 3531 } 3532 3533 void CMTask::get_entries_from_global_stack() { 3534 // local array where we'll store the entries that will be popped 3535 // from the global stack. 3536 oop buffer[global_stack_transfer_size]; 3537 int n; 3538 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); 3539 assert(n <= global_stack_transfer_size, 3540 "we should not pop more than the given limit"); 3541 if (n > 0) { 3542 // yes, we did actually pop at least one entry 3543 3544 statsOnly( ++_global_transfers_from; _global_pops += n ); 3545 if (_cm->verbose_medium()) { 3546 gclog_or_tty->print_cr("[%d] popped %d entries from the global stack", 3547 _task_id, n); 3548 } 3549 for (int i = 0; i < n; ++i) { 3550 bool success = _task_queue->push(buffer[i]); 3551 // We only call this when the local queue is empty or under a 3552 // given target limit. So, we do not expect this push to fail. 3553 assert(success, "invariant"); 3554 } 3555 3556 statsOnly( int tmp_size = _task_queue->size(); 3557 if (tmp_size > _local_max_size) { 3558 _local_max_size = tmp_size; 3559 } 3560 _local_pushes += n ); 3561 } 3562 3563 // this operation was quite expensive, so decrease the limits 3564 decrease_limits(); 3565 } 3566 3567 void CMTask::drain_local_queue(bool partially) { 3568 if (has_aborted()) return; 3569 3570 // Decide what the target size is, depending whether we're going to 3571 // drain it partially (so that other tasks can steal if they run out 3572 // of things to do) or totally (at the very end). 3573 size_t target_size; 3574 if (partially) { 3575 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 3576 } else { 3577 target_size = 0; 3578 } 3579 3580 if (_task_queue->size() > target_size) { 3581 if (_cm->verbose_high()) { 3582 gclog_or_tty->print_cr("[%d] draining local queue, target size = %d", 3583 _task_id, target_size); 3584 } 3585 3586 oop obj; 3587 bool ret = _task_queue->pop_local(obj); 3588 while (ret) { 3589 statsOnly( ++_local_pops ); 3590 3591 if (_cm->verbose_high()) { 3592 gclog_or_tty->print_cr("[%d] popped "PTR_FORMAT, _task_id, 3593 (void*) obj); 3594 } 3595 3596 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); 3597 assert(!_g1h->is_on_master_free_list( 3598 _g1h->heap_region_containing((HeapWord*) obj)), "invariant"); 3599 3600 scan_object(obj); 3601 3602 if (_task_queue->size() <= target_size || has_aborted()) { 3603 ret = false; 3604 } else { 3605 ret = _task_queue->pop_local(obj); 3606 } 3607 } 3608 3609 if (_cm->verbose_high()) { 3610 gclog_or_tty->print_cr("[%d] drained local queue, size = %d", 3611 _task_id, _task_queue->size()); 3612 } 3613 } 3614 } 3615 3616 void CMTask::drain_global_stack(bool partially) { 3617 if (has_aborted()) return; 3618 3619 // We have a policy to drain the local queue before we attempt to 3620 // drain the global stack. 3621 assert(partially || _task_queue->size() == 0, "invariant"); 3622 3623 // Decide what the target size is, depending whether we're going to 3624 // drain it partially (so that other tasks can steal if they run out 3625 // of things to do) or totally (at the very end). Notice that, 3626 // because we move entries from the global stack in chunks or 3627 // because another task might be doing the same, we might in fact 3628 // drop below the target. But, this is not a problem. 3629 size_t target_size; 3630 if (partially) { 3631 target_size = _cm->partial_mark_stack_size_target(); 3632 } else { 3633 target_size = 0; 3634 } 3635 3636 if (_cm->mark_stack_size() > target_size) { 3637 if (_cm->verbose_low()) { 3638 gclog_or_tty->print_cr("[%d] draining global_stack, target size %d", 3639 _task_id, target_size); 3640 } 3641 3642 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 3643 get_entries_from_global_stack(); 3644 drain_local_queue(partially); 3645 } 3646 3647 if (_cm->verbose_low()) { 3648 gclog_or_tty->print_cr("[%d] drained global stack, size = %d", 3649 _task_id, _cm->mark_stack_size()); 3650 } 3651 } 3652 } 3653 3654 // SATB Queue has several assumptions on whether to call the par or 3655 // non-par versions of the methods. this is why some of the code is 3656 // replicated. We should really get rid of the single-threaded version 3657 // of the code to simplify things. 3658 void CMTask::drain_satb_buffers() { 3659 if (has_aborted()) return; 3660 3661 // We set this so that the regular clock knows that we're in the 3662 // middle of draining buffers and doesn't set the abort flag when it 3663 // notices that SATB buffers are available for draining. It'd be 3664 // very counter productive if it did that. :-) 3665 _draining_satb_buffers = true; 3666 3667 CMObjectClosure oc(this); 3668 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3669 if (G1CollectedHeap::use_parallel_gc_threads()) { 3670 satb_mq_set.set_par_closure(_task_id, &oc); 3671 } else { 3672 satb_mq_set.set_closure(&oc); 3673 } 3674 3675 // This keeps claiming and applying the closure to completed buffers 3676 // until we run out of buffers or we need to abort. 3677 if (G1CollectedHeap::use_parallel_gc_threads()) { 3678 while (!has_aborted() && 3679 satb_mq_set.par_apply_closure_to_completed_buffer(_task_id)) { 3680 if (_cm->verbose_medium()) { 3681 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); 3682 } 3683 statsOnly( ++_satb_buffers_processed ); 3684 regular_clock_call(); 3685 } 3686 } else { 3687 while (!has_aborted() && 3688 satb_mq_set.apply_closure_to_completed_buffer()) { 3689 if (_cm->verbose_medium()) { 3690 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); 3691 } 3692 statsOnly( ++_satb_buffers_processed ); 3693 regular_clock_call(); 3694 } 3695 } 3696 3697 if (!concurrent() && !has_aborted()) { 3698 // We should only do this during remark. 3699 if (G1CollectedHeap::use_parallel_gc_threads()) { 3700 satb_mq_set.par_iterate_closure_all_threads(_task_id); 3701 } else { 3702 satb_mq_set.iterate_closure_all_threads(); 3703 } 3704 } 3705 3706 _draining_satb_buffers = false; 3707 3708 assert(has_aborted() || 3709 concurrent() || 3710 satb_mq_set.completed_buffers_num() == 0, "invariant"); 3711 3712 if (G1CollectedHeap::use_parallel_gc_threads()) { 3713 satb_mq_set.set_par_closure(_task_id, NULL); 3714 } else { 3715 satb_mq_set.set_closure(NULL); 3716 } 3717 3718 // again, this was a potentially expensive operation, decrease the 3719 // limits to get the regular clock call early 3720 decrease_limits(); 3721 } 3722 3723 void CMTask::print_stats() { 3724 gclog_or_tty->print_cr("Marking Stats, task = %d, calls = %d", 3725 _task_id, _calls); 3726 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 3727 _elapsed_time_ms, _termination_time_ms); 3728 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 3729 _step_times_ms.num(), _step_times_ms.avg(), 3730 _step_times_ms.sd()); 3731 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 3732 _step_times_ms.maximum(), _step_times_ms.sum()); 3733 3734 #if _MARKING_STATS_ 3735 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 3736 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(), 3737 _all_clock_intervals_ms.sd()); 3738 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 3739 _all_clock_intervals_ms.maximum(), 3740 _all_clock_intervals_ms.sum()); 3741 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d", 3742 _clock_due_to_scanning, _clock_due_to_marking); 3743 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d", 3744 _objs_scanned, _objs_found_on_bitmap); 3745 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d", 3746 _local_pushes, _local_pops, _local_max_size); 3747 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d", 3748 _global_pushes, _global_pops, _global_max_size); 3749 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d", 3750 _global_transfers_to,_global_transfers_from); 3751 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed); 3752 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed); 3753 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d", 3754 _steal_attempts, _steals); 3755 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted); 3756 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d", 3757 _aborted_overflow, _aborted_cm_aborted, _aborted_yield); 3758 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d", 3759 _aborted_timed_out, _aborted_satb, _aborted_termination); 3760 #endif // _MARKING_STATS_ 3761 } 3762 3763 /***************************************************************************** 3764 3765 The do_marking_step(time_target_ms) method is the building block 3766 of the parallel marking framework. It can be called in parallel 3767 with other invocations of do_marking_step() on different tasks 3768 (but only one per task, obviously) and concurrently with the 3769 mutator threads, or during remark, hence it eliminates the need 3770 for two versions of the code. When called during remark, it will 3771 pick up from where the task left off during the concurrent marking 3772 phase. Interestingly, tasks are also claimable during evacuation 3773 pauses too, since do_marking_step() ensures that it aborts before 3774 it needs to yield. 3775 3776 The data structures that is uses to do marking work are the 3777 following: 3778 3779 (1) Marking Bitmap. If there are gray objects that appear only 3780 on the bitmap (this happens either when dealing with an overflow 3781 or when the initial marking phase has simply marked the roots 3782 and didn't push them on the stack), then tasks claim heap 3783 regions whose bitmap they then scan to find gray objects. A 3784 global finger indicates where the end of the last claimed region 3785 is. A local finger indicates how far into the region a task has 3786 scanned. The two fingers are used to determine how to gray an 3787 object (i.e. whether simply marking it is OK, as it will be 3788 visited by a task in the future, or whether it needs to be also 3789 pushed on a stack). 3790 3791 (2) Local Queue. The local queue of the task which is accessed 3792 reasonably efficiently by the task. Other tasks can steal from 3793 it when they run out of work. Throughout the marking phase, a 3794 task attempts to keep its local queue short but not totally 3795 empty, so that entries are available for stealing by other 3796 tasks. Only when there is no more work, a task will totally 3797 drain its local queue. 3798 3799 (3) Global Mark Stack. This handles local queue overflow. During 3800 marking only sets of entries are moved between it and the local 3801 queues, as access to it requires a mutex and more fine-grain 3802 interaction with it which might cause contention. If it 3803 overflows, then the marking phase should restart and iterate 3804 over the bitmap to identify gray objects. Throughout the marking 3805 phase, tasks attempt to keep the global mark stack at a small 3806 length but not totally empty, so that entries are available for 3807 popping by other tasks. Only when there is no more work, tasks 3808 will totally drain the global mark stack. 3809 3810 (4) SATB Buffer Queue. This is where completed SATB buffers are 3811 made available. Buffers are regularly removed from this queue 3812 and scanned for roots, so that the queue doesn't get too 3813 long. During remark, all completed buffers are processed, as 3814 well as the filled in parts of any uncompleted buffers. 3815 3816 The do_marking_step() method tries to abort when the time target 3817 has been reached. There are a few other cases when the 3818 do_marking_step() method also aborts: 3819 3820 (1) When the marking phase has been aborted (after a Full GC). 3821 3822 (2) When a global overflow (on the global stack) has been 3823 triggered. Before the task aborts, it will actually sync up with 3824 the other tasks to ensure that all the marking data structures 3825 (local queues, stacks, fingers etc.) are re-initialised so that 3826 when do_marking_step() completes, the marking phase can 3827 immediately restart. 3828 3829 (3) When enough completed SATB buffers are available. The 3830 do_marking_step() method only tries to drain SATB buffers right 3831 at the beginning. So, if enough buffers are available, the 3832 marking step aborts and the SATB buffers are processed at 3833 the beginning of the next invocation. 3834 3835 (4) To yield. when we have to yield then we abort and yield 3836 right at the end of do_marking_step(). This saves us from a lot 3837 of hassle as, by yielding we might allow a Full GC. If this 3838 happens then objects will be compacted underneath our feet, the 3839 heap might shrink, etc. We save checking for this by just 3840 aborting and doing the yield right at the end. 3841 3842 From the above it follows that the do_marking_step() method should 3843 be called in a loop (or, otherwise, regularly) until it completes. 3844 3845 If a marking step completes without its has_aborted() flag being 3846 true, it means it has completed the current marking phase (and 3847 also all other marking tasks have done so and have all synced up). 3848 3849 A method called regular_clock_call() is invoked "regularly" (in 3850 sub ms intervals) throughout marking. It is this clock method that 3851 checks all the abort conditions which were mentioned above and 3852 decides when the task should abort. A work-based scheme is used to 3853 trigger this clock method: when the number of object words the 3854 marking phase has scanned or the number of references the marking 3855 phase has visited reach a given limit. Additional invocations to 3856 the method clock have been planted in a few other strategic places 3857 too. The initial reason for the clock method was to avoid calling 3858 vtime too regularly, as it is quite expensive. So, once it was in 3859 place, it was natural to piggy-back all the other conditions on it 3860 too and not constantly check them throughout the code. 3861 3862 *****************************************************************************/ 3863 3864 void CMTask::do_marking_step(double time_target_ms, 3865 bool do_stealing, 3866 bool do_termination) { 3867 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 3868 assert(concurrent() == _cm->concurrent(), "they should be the same"); 3869 3870 G1CollectorPolicy* g1_policy = _g1h->g1_policy(); 3871 assert(_task_queues != NULL, "invariant"); 3872 assert(_task_queue != NULL, "invariant"); 3873 assert(_task_queues->queue(_task_id) == _task_queue, "invariant"); 3874 3875 assert(!_claimed, 3876 "only one thread should claim this task at any one time"); 3877 3878 // OK, this doesn't safeguard again all possible scenarios, as it is 3879 // possible for two threads to set the _claimed flag at the same 3880 // time. But it is only for debugging purposes anyway and it will 3881 // catch most problems. 3882 _claimed = true; 3883 3884 _start_time_ms = os::elapsedVTime() * 1000.0; 3885 statsOnly( _interval_start_time_ms = _start_time_ms ); 3886 3887 double diff_prediction_ms = 3888 g1_policy->get_new_prediction(&_marking_step_diffs_ms); 3889 _time_target_ms = time_target_ms - diff_prediction_ms; 3890 3891 // set up the variables that are used in the work-based scheme to 3892 // call the regular clock method 3893 _words_scanned = 0; 3894 _refs_reached = 0; 3895 recalculate_limits(); 3896 3897 // clear all flags 3898 clear_has_aborted(); 3899 _has_timed_out = false; 3900 _draining_satb_buffers = false; 3901 3902 ++_calls; 3903 3904 if (_cm->verbose_low()) { 3905 gclog_or_tty->print_cr("[%d] >>>>>>>>>> START, call = %d, " 3906 "target = %1.2lfms >>>>>>>>>>", 3907 _task_id, _calls, _time_target_ms); 3908 } 3909 3910 // Set up the bitmap and oop closures. Anything that uses them is 3911 // eventually called from this method, so it is OK to allocate these 3912 // statically. 3913 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); 3914 G1CMOopClosure cm_oop_closure(_g1h, _cm, this); 3915 set_cm_oop_closure(&cm_oop_closure); 3916 3917 if (_cm->has_overflown()) { 3918 // This can happen if the mark stack overflows during a GC pause 3919 // and this task, after a yield point, restarts. We have to abort 3920 // as we need to get into the overflow protocol which happens 3921 // right at the end of this task. 3922 set_has_aborted(); 3923 } 3924 3925 // First drain any available SATB buffers. After this, we will not 3926 // look at SATB buffers before the next invocation of this method. 3927 // If enough completed SATB buffers are queued up, the regular clock 3928 // will abort this task so that it restarts. 3929 drain_satb_buffers(); 3930 // ...then partially drain the local queue and the global stack 3931 drain_local_queue(true); 3932 drain_global_stack(true); 3933 3934 do { 3935 if (!has_aborted() && _curr_region != NULL) { 3936 // This means that we're already holding on to a region. 3937 assert(_finger != NULL, "if region is not NULL, then the finger " 3938 "should not be NULL either"); 3939 3940 // We might have restarted this task after an evacuation pause 3941 // which might have evacuated the region we're holding on to 3942 // underneath our feet. Let's read its limit again to make sure 3943 // that we do not iterate over a region of the heap that 3944 // contains garbage (update_region_limit() will also move 3945 // _finger to the start of the region if it is found empty). 3946 update_region_limit(); 3947 // We will start from _finger not from the start of the region, 3948 // as we might be restarting this task after aborting half-way 3949 // through scanning this region. In this case, _finger points to 3950 // the address where we last found a marked object. If this is a 3951 // fresh region, _finger points to start(). 3952 MemRegion mr = MemRegion(_finger, _region_limit); 3953 3954 if (_cm->verbose_low()) { 3955 gclog_or_tty->print_cr("[%d] we're scanning part " 3956 "["PTR_FORMAT", "PTR_FORMAT") " 3957 "of region "PTR_FORMAT, 3958 _task_id, _finger, _region_limit, _curr_region); 3959 } 3960 3961 // Let's iterate over the bitmap of the part of the 3962 // region that is left. 3963 if (mr.is_empty() || _nextMarkBitMap->iterate(&bitmap_closure, mr)) { 3964 // We successfully completed iterating over the region. Now, 3965 // let's give up the region. 3966 giveup_current_region(); 3967 regular_clock_call(); 3968 } else { 3969 assert(has_aborted(), "currently the only way to do so"); 3970 // The only way to abort the bitmap iteration is to return 3971 // false from the do_bit() method. However, inside the 3972 // do_bit() method we move the _finger to point to the 3973 // object currently being looked at. So, if we bail out, we 3974 // have definitely set _finger to something non-null. 3975 assert(_finger != NULL, "invariant"); 3976 3977 // Region iteration was actually aborted. So now _finger 3978 // points to the address of the object we last scanned. If we 3979 // leave it there, when we restart this task, we will rescan 3980 // the object. It is easy to avoid this. We move the finger by 3981 // enough to point to the next possible object header (the 3982 // bitmap knows by how much we need to move it as it knows its 3983 // granularity). 3984 assert(_finger < _region_limit, "invariant"); 3985 HeapWord* new_finger = _nextMarkBitMap->nextWord(_finger); 3986 // Check if bitmap iteration was aborted while scanning the last object 3987 if (new_finger >= _region_limit) { 3988 giveup_current_region(); 3989 } else { 3990 move_finger_to(new_finger); 3991 } 3992 } 3993 } 3994 // At this point we have either completed iterating over the 3995 // region we were holding on to, or we have aborted. 3996 3997 // We then partially drain the local queue and the global stack. 3998 // (Do we really need this?) 3999 drain_local_queue(true); 4000 drain_global_stack(true); 4001 4002 // Read the note on the claim_region() method on why it might 4003 // return NULL with potentially more regions available for 4004 // claiming and why we have to check out_of_regions() to determine 4005 // whether we're done or not. 4006 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 4007 // We are going to try to claim a new region. We should have 4008 // given up on the previous one. 4009 // Separated the asserts so that we know which one fires. 4010 assert(_curr_region == NULL, "invariant"); 4011 assert(_finger == NULL, "invariant"); 4012 assert(_region_limit == NULL, "invariant"); 4013 if (_cm->verbose_low()) { 4014 gclog_or_tty->print_cr("[%d] trying to claim a new region", _task_id); 4015 } 4016 HeapRegion* claimed_region = _cm->claim_region(_task_id); 4017 if (claimed_region != NULL) { 4018 // Yes, we managed to claim one 4019 statsOnly( ++_regions_claimed ); 4020 4021 if (_cm->verbose_low()) { 4022 gclog_or_tty->print_cr("[%d] we successfully claimed " 4023 "region "PTR_FORMAT, 4024 _task_id, claimed_region); 4025 } 4026 4027 setup_for_region(claimed_region); 4028 assert(_curr_region == claimed_region, "invariant"); 4029 } 4030 // It is important to call the regular clock here. It might take 4031 // a while to claim a region if, for example, we hit a large 4032 // block of empty regions. So we need to call the regular clock 4033 // method once round the loop to make sure it's called 4034 // frequently enough. 4035 regular_clock_call(); 4036 } 4037 4038 if (!has_aborted() && _curr_region == NULL) { 4039 assert(_cm->out_of_regions(), 4040 "at this point we should be out of regions"); 4041 } 4042 } while ( _curr_region != NULL && !has_aborted()); 4043 4044 if (!has_aborted()) { 4045 // We cannot check whether the global stack is empty, since other 4046 // tasks might be pushing objects to it concurrently. 4047 assert(_cm->out_of_regions(), 4048 "at this point we should be out of regions"); 4049 4050 if (_cm->verbose_low()) { 4051 gclog_or_tty->print_cr("[%d] all regions claimed", _task_id); 4052 } 4053 4054 // Try to reduce the number of available SATB buffers so that 4055 // remark has less work to do. 4056 drain_satb_buffers(); 4057 } 4058 4059 // Since we've done everything else, we can now totally drain the 4060 // local queue and global stack. 4061 drain_local_queue(false); 4062 drain_global_stack(false); 4063 4064 // Attempt at work stealing from other task's queues. 4065 if (do_stealing && !has_aborted()) { 4066 // We have not aborted. This means that we have finished all that 4067 // we could. Let's try to do some stealing... 4068 4069 // We cannot check whether the global stack is empty, since other 4070 // tasks might be pushing objects to it concurrently. 4071 assert(_cm->out_of_regions() && _task_queue->size() == 0, 4072 "only way to reach here"); 4073 4074 if (_cm->verbose_low()) { 4075 gclog_or_tty->print_cr("[%d] starting to steal", _task_id); 4076 } 4077 4078 while (!has_aborted()) { 4079 oop obj; 4080 statsOnly( ++_steal_attempts ); 4081 4082 if (_cm->try_stealing(_task_id, &_hash_seed, obj)) { 4083 if (_cm->verbose_medium()) { 4084 gclog_or_tty->print_cr("[%d] stolen "PTR_FORMAT" successfully", 4085 _task_id, (void*) obj); 4086 } 4087 4088 statsOnly( ++_steals ); 4089 4090 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), 4091 "any stolen object should be marked"); 4092 scan_object(obj); 4093 4094 // And since we're towards the end, let's totally drain the 4095 // local queue and global stack. 4096 drain_local_queue(false); 4097 drain_global_stack(false); 4098 } else { 4099 break; 4100 } 4101 } 4102 } 4103 4104 // If we are about to wrap up and go into termination, check if we 4105 // should raise the overflow flag. 4106 if (do_termination && !has_aborted()) { 4107 if (_cm->force_overflow()->should_force()) { 4108 _cm->set_has_overflown(); 4109 regular_clock_call(); 4110 } 4111 } 4112 4113 // We still haven't aborted. Now, let's try to get into the 4114 // termination protocol. 4115 if (do_termination && !has_aborted()) { 4116 // We cannot check whether the global stack is empty, since other 4117 // tasks might be concurrently pushing objects on it. 4118 // Separated the asserts so that we know which one fires. 4119 assert(_cm->out_of_regions(), "only way to reach here"); 4120 assert(_task_queue->size() == 0, "only way to reach here"); 4121 4122 if (_cm->verbose_low()) { 4123 gclog_or_tty->print_cr("[%d] starting termination protocol", _task_id); 4124 } 4125 4126 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 4127 // The CMTask class also extends the TerminatorTerminator class, 4128 // hence its should_exit_termination() method will also decide 4129 // whether to exit the termination protocol or not. 4130 bool finished = _cm->terminator()->offer_termination(this); 4131 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 4132 _termination_time_ms += 4133 termination_end_time_ms - _termination_start_time_ms; 4134 4135 if (finished) { 4136 // We're all done. 4137 4138 if (_task_id == 0) { 4139 // let's allow task 0 to do this 4140 if (concurrent()) { 4141 assert(_cm->concurrent_marking_in_progress(), "invariant"); 4142 // we need to set this to false before the next 4143 // safepoint. This way we ensure that the marking phase 4144 // doesn't observe any more heap expansions. 4145 _cm->clear_concurrent_marking_in_progress(); 4146 } 4147 } 4148 4149 // We can now guarantee that the global stack is empty, since 4150 // all other tasks have finished. We separated the guarantees so 4151 // that, if a condition is false, we can immediately find out 4152 // which one. 4153 guarantee(_cm->out_of_regions(), "only way to reach here"); 4154 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 4155 guarantee(_task_queue->size() == 0, "only way to reach here"); 4156 guarantee(!_cm->has_overflown(), "only way to reach here"); 4157 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); 4158 4159 if (_cm->verbose_low()) { 4160 gclog_or_tty->print_cr("[%d] all tasks terminated", _task_id); 4161 } 4162 } else { 4163 // Apparently there's more work to do. Let's abort this task. It 4164 // will restart it and we can hopefully find more things to do. 4165 4166 if (_cm->verbose_low()) { 4167 gclog_or_tty->print_cr("[%d] apparently there is more work to do", 4168 _task_id); 4169 } 4170 4171 set_has_aborted(); 4172 statsOnly( ++_aborted_termination ); 4173 } 4174 } 4175 4176 // Mainly for debugging purposes to make sure that a pointer to the 4177 // closure which was statically allocated in this frame doesn't 4178 // escape it by accident. 4179 set_cm_oop_closure(NULL); 4180 double end_time_ms = os::elapsedVTime() * 1000.0; 4181 double elapsed_time_ms = end_time_ms - _start_time_ms; 4182 // Update the step history. 4183 _step_times_ms.add(elapsed_time_ms); 4184 4185 if (has_aborted()) { 4186 // The task was aborted for some reason. 4187 4188 statsOnly( ++_aborted ); 4189 4190 if (_has_timed_out) { 4191 double diff_ms = elapsed_time_ms - _time_target_ms; 4192 // Keep statistics of how well we did with respect to hitting 4193 // our target only if we actually timed out (if we aborted for 4194 // other reasons, then the results might get skewed). 4195 _marking_step_diffs_ms.add(diff_ms); 4196 } 4197 4198 if (_cm->has_overflown()) { 4199 // This is the interesting one. We aborted because a global 4200 // overflow was raised. This means we have to restart the 4201 // marking phase and start iterating over regions. However, in 4202 // order to do this we have to make sure that all tasks stop 4203 // what they are doing and re-initialise in a safe manner. We 4204 // will achieve this with the use of two barrier sync points. 4205 4206 if (_cm->verbose_low()) { 4207 gclog_or_tty->print_cr("[%d] detected overflow", _task_id); 4208 } 4209 4210 _cm->enter_first_sync_barrier(_task_id); 4211 // When we exit this sync barrier we know that all tasks have 4212 // stopped doing marking work. So, it's now safe to 4213 // re-initialise our data structures. At the end of this method, 4214 // task 0 will clear the global data structures. 4215 4216 statsOnly( ++_aborted_overflow ); 4217 4218 // We clear the local state of this task... 4219 clear_region_fields(); 4220 4221 // ...and enter the second barrier. 4222 _cm->enter_second_sync_barrier(_task_id); 4223 // At this point everything has bee re-initialised and we're 4224 // ready to restart. 4225 } 4226 4227 if (_cm->verbose_low()) { 4228 gclog_or_tty->print_cr("[%d] <<<<<<<<<< ABORTING, target = %1.2lfms, " 4229 "elapsed = %1.2lfms <<<<<<<<<<", 4230 _task_id, _time_target_ms, elapsed_time_ms); 4231 if (_cm->has_aborted()) { 4232 gclog_or_tty->print_cr("[%d] ========== MARKING ABORTED ==========", 4233 _task_id); 4234 } 4235 } 4236 } else { 4237 if (_cm->verbose_low()) { 4238 gclog_or_tty->print_cr("[%d] <<<<<<<<<< FINISHED, target = %1.2lfms, " 4239 "elapsed = %1.2lfms <<<<<<<<<<", 4240 _task_id, _time_target_ms, elapsed_time_ms); 4241 } 4242 } 4243 4244 _claimed = false; 4245 } 4246 4247 CMTask::CMTask(int task_id, 4248 ConcurrentMark* cm, 4249 size_t* marked_bytes, 4250 BitMap* card_bm, 4251 CMTaskQueue* task_queue, 4252 CMTaskQueueSet* task_queues) 4253 : _g1h(G1CollectedHeap::heap()), 4254 _task_id(task_id), _cm(cm), 4255 _claimed(false), 4256 _nextMarkBitMap(NULL), _hash_seed(17), 4257 _task_queue(task_queue), 4258 _task_queues(task_queues), 4259 _cm_oop_closure(NULL), 4260 _marked_bytes_array(marked_bytes), 4261 _card_bm(card_bm) { 4262 guarantee(task_queue != NULL, "invariant"); 4263 guarantee(task_queues != NULL, "invariant"); 4264 4265 statsOnly( _clock_due_to_scanning = 0; 4266 _clock_due_to_marking = 0 ); 4267 4268 _marking_step_diffs_ms.add(0.5); 4269 } 4270 4271 // These are formatting macros that are used below to ensure 4272 // consistent formatting. The *_H_* versions are used to format the 4273 // header for a particular value and they should be kept consistent 4274 // with the corresponding macro. Also note that most of the macros add 4275 // the necessary white space (as a prefix) which makes them a bit 4276 // easier to compose. 4277 4278 // All the output lines are prefixed with this string to be able to 4279 // identify them easily in a large log file. 4280 #define G1PPRL_LINE_PREFIX "###" 4281 4282 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT 4283 #ifdef _LP64 4284 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 4285 #else // _LP64 4286 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 4287 #endif // _LP64 4288 4289 // For per-region info 4290 #define G1PPRL_TYPE_FORMAT " %-4s" 4291 #define G1PPRL_TYPE_H_FORMAT " %4s" 4292 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9) 4293 #define G1PPRL_BYTE_H_FORMAT " %9s" 4294 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 4295 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 4296 4297 // For summary info 4298 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT 4299 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT 4300 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB" 4301 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%" 4302 4303 G1PrintRegionLivenessInfoClosure:: 4304 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name) 4305 : _out(out), 4306 _total_used_bytes(0), _total_capacity_bytes(0), 4307 _total_prev_live_bytes(0), _total_next_live_bytes(0), 4308 _hum_used_bytes(0), _hum_capacity_bytes(0), 4309 _hum_prev_live_bytes(0), _hum_next_live_bytes(0) { 4310 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 4311 MemRegion g1_committed = g1h->g1_committed(); 4312 MemRegion g1_reserved = g1h->g1_reserved(); 4313 double now = os::elapsedTime(); 4314 4315 // Print the header of the output. 4316 _out->cr(); 4317 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 4318 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP" 4319 G1PPRL_SUM_ADDR_FORMAT("committed") 4320 G1PPRL_SUM_ADDR_FORMAT("reserved") 4321 G1PPRL_SUM_BYTE_FORMAT("region-size"), 4322 g1_committed.start(), g1_committed.end(), 4323 g1_reserved.start(), g1_reserved.end(), 4324 HeapRegion::GrainBytes); 4325 _out->print_cr(G1PPRL_LINE_PREFIX); 4326 _out->print_cr(G1PPRL_LINE_PREFIX 4327 G1PPRL_TYPE_H_FORMAT 4328 G1PPRL_ADDR_BASE_H_FORMAT 4329 G1PPRL_BYTE_H_FORMAT 4330 G1PPRL_BYTE_H_FORMAT 4331 G1PPRL_BYTE_H_FORMAT 4332 G1PPRL_DOUBLE_H_FORMAT, 4333 "type", "address-range", 4334 "used", "prev-live", "next-live", "gc-eff"); 4335 _out->print_cr(G1PPRL_LINE_PREFIX 4336 G1PPRL_TYPE_H_FORMAT 4337 G1PPRL_ADDR_BASE_H_FORMAT 4338 G1PPRL_BYTE_H_FORMAT 4339 G1PPRL_BYTE_H_FORMAT 4340 G1PPRL_BYTE_H_FORMAT 4341 G1PPRL_DOUBLE_H_FORMAT, 4342 "", "", 4343 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)"); 4344 } 4345 4346 // It takes as a parameter a reference to one of the _hum_* fields, it 4347 // deduces the corresponding value for a region in a humongous region 4348 // series (either the region size, or what's left if the _hum_* field 4349 // is < the region size), and updates the _hum_* field accordingly. 4350 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) { 4351 size_t bytes = 0; 4352 // The > 0 check is to deal with the prev and next live bytes which 4353 // could be 0. 4354 if (*hum_bytes > 0) { 4355 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes); 4356 *hum_bytes -= bytes; 4357 } 4358 return bytes; 4359 } 4360 4361 // It deduces the values for a region in a humongous region series 4362 // from the _hum_* fields and updates those accordingly. It assumes 4363 // that that _hum_* fields have already been set up from the "starts 4364 // humongous" region and we visit the regions in address order. 4365 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes, 4366 size_t* capacity_bytes, 4367 size_t* prev_live_bytes, 4368 size_t* next_live_bytes) { 4369 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition"); 4370 *used_bytes = get_hum_bytes(&_hum_used_bytes); 4371 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes); 4372 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes); 4373 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes); 4374 } 4375 4376 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) { 4377 const char* type = ""; 4378 HeapWord* bottom = r->bottom(); 4379 HeapWord* end = r->end(); 4380 size_t capacity_bytes = r->capacity(); 4381 size_t used_bytes = r->used(); 4382 size_t prev_live_bytes = r->live_bytes(); 4383 size_t next_live_bytes = r->next_live_bytes(); 4384 double gc_eff = r->gc_efficiency(); 4385 if (r->used() == 0) { 4386 type = "FREE"; 4387 } else if (r->is_survivor()) { 4388 type = "SURV"; 4389 } else if (r->is_young()) { 4390 type = "EDEN"; 4391 } else if (r->startsHumongous()) { 4392 type = "HUMS"; 4393 4394 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 && 4395 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0, 4396 "they should have been zeroed after the last time we used them"); 4397 // Set up the _hum_* fields. 4398 _hum_capacity_bytes = capacity_bytes; 4399 _hum_used_bytes = used_bytes; 4400 _hum_prev_live_bytes = prev_live_bytes; 4401 _hum_next_live_bytes = next_live_bytes; 4402 get_hum_bytes(&used_bytes, &capacity_bytes, 4403 &prev_live_bytes, &next_live_bytes); 4404 end = bottom + HeapRegion::GrainWords; 4405 } else if (r->continuesHumongous()) { 4406 type = "HUMC"; 4407 get_hum_bytes(&used_bytes, &capacity_bytes, 4408 &prev_live_bytes, &next_live_bytes); 4409 assert(end == bottom + HeapRegion::GrainWords, "invariant"); 4410 } else { 4411 type = "OLD"; 4412 } 4413 4414 _total_used_bytes += used_bytes; 4415 _total_capacity_bytes += capacity_bytes; 4416 _total_prev_live_bytes += prev_live_bytes; 4417 _total_next_live_bytes += next_live_bytes; 4418 4419 // Print a line for this particular region. 4420 _out->print_cr(G1PPRL_LINE_PREFIX 4421 G1PPRL_TYPE_FORMAT 4422 G1PPRL_ADDR_BASE_FORMAT 4423 G1PPRL_BYTE_FORMAT 4424 G1PPRL_BYTE_FORMAT 4425 G1PPRL_BYTE_FORMAT 4426 G1PPRL_DOUBLE_FORMAT, 4427 type, bottom, end, 4428 used_bytes, prev_live_bytes, next_live_bytes, gc_eff); 4429 4430 return false; 4431 } 4432 4433 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 4434 // Print the footer of the output. 4435 _out->print_cr(G1PPRL_LINE_PREFIX); 4436 _out->print_cr(G1PPRL_LINE_PREFIX 4437 " SUMMARY" 4438 G1PPRL_SUM_MB_FORMAT("capacity") 4439 G1PPRL_SUM_MB_PERC_FORMAT("used") 4440 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 4441 G1PPRL_SUM_MB_PERC_FORMAT("next-live"), 4442 bytes_to_mb(_total_capacity_bytes), 4443 bytes_to_mb(_total_used_bytes), 4444 perc(_total_used_bytes, _total_capacity_bytes), 4445 bytes_to_mb(_total_prev_live_bytes), 4446 perc(_total_prev_live_bytes, _total_capacity_bytes), 4447 bytes_to_mb(_total_next_live_bytes), 4448 perc(_total_next_live_bytes, _total_capacity_bytes)); 4449 _out->cr(); 4450 }