1 /* 2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/metadataOnStackMark.hpp" 27 #include "classfile/symbolTable.hpp" 28 #include "code/codeCache.hpp" 29 #include "gc/g1/concurrentMarkThread.inline.hpp" 30 #include "gc/g1/g1CollectedHeap.inline.hpp" 31 #include "gc/g1/g1CollectorState.hpp" 32 #include "gc/g1/g1ConcurrentMark.inline.hpp" 33 #include "gc/g1/g1HeapVerifier.hpp" 34 #include "gc/g1/g1OopClosures.inline.hpp" 35 #include "gc/g1/g1CardLiveData.inline.hpp" 36 #include "gc/g1/g1Policy.hpp" 37 #include "gc/g1/g1StringDedup.hpp" 38 #include "gc/g1/heapRegion.inline.hpp" 39 #include "gc/g1/heapRegionRemSet.hpp" 40 #include "gc/g1/heapRegionSet.inline.hpp" 41 #include "gc/g1/suspendibleThreadSet.hpp" 42 #include "gc/shared/gcId.hpp" 43 #include "gc/shared/gcTimer.hpp" 44 #include "gc/shared/gcTrace.hpp" 45 #include "gc/shared/gcTraceTime.inline.hpp" 46 #include "gc/shared/genOopClosures.inline.hpp" 47 #include "gc/shared/referencePolicy.hpp" 48 #include "gc/shared/strongRootsScope.hpp" 49 #include "gc/shared/taskqueue.inline.hpp" 50 #include "gc/shared/vmGCOperations.hpp" 51 #include "logging/log.hpp" 52 #include "memory/allocation.hpp" 53 #include "memory/resourceArea.hpp" 54 #include "oops/oop.inline.hpp" 55 #include "runtime/atomic.inline.hpp" 56 #include "runtime/handles.inline.hpp" 57 #include "runtime/java.hpp" 58 #include "runtime/prefetch.inline.hpp" 59 #include "services/memTracker.hpp" 60 #include "utilities/growableArray.hpp" 61 62 // Concurrent marking bit map wrapper 63 64 G1CMBitMapRO::G1CMBitMapRO(int shifter) : 65 _bm(), 66 _shifter(shifter) { 67 _bmStartWord = 0; 68 _bmWordSize = 0; 69 } 70 71 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr, 72 const HeapWord* limit) const { 73 // First we must round addr *up* to a possible object boundary. 74 addr = (HeapWord*)align_size_up((intptr_t)addr, 75 HeapWordSize << _shifter); 76 size_t addrOffset = heapWordToOffset(addr); 77 assert(limit != NULL, "limit must not be NULL"); 78 size_t limitOffset = heapWordToOffset(limit); 79 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); 80 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 81 assert(nextAddr >= addr, "get_next_one postcondition"); 82 assert(nextAddr == limit || isMarked(nextAddr), 83 "get_next_one postcondition"); 84 return nextAddr; 85 } 86 87 #ifndef PRODUCT 88 bool G1CMBitMapRO::covers(MemRegion heap_rs) const { 89 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 90 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize, 91 "size inconsistency"); 92 return _bmStartWord == (HeapWord*)(heap_rs.start()) && 93 _bmWordSize == heap_rs.word_size(); 94 } 95 #endif 96 97 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const { 98 _bm.print_on_error(st, prefix); 99 } 100 101 size_t G1CMBitMap::compute_size(size_t heap_size) { 102 return ReservedSpace::allocation_align_size_up(heap_size / mark_distance()); 103 } 104 105 size_t G1CMBitMap::mark_distance() { 106 return MinObjAlignmentInBytes * BitsPerByte; 107 } 108 109 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) { 110 _bmStartWord = heap.start(); 111 _bmWordSize = heap.word_size(); 112 113 _bm = BitMapView((BitMap::bm_word_t*) storage->reserved().start(), _bmWordSize >> _shifter); 114 115 storage->set_mapping_changed_listener(&_listener); 116 } 117 118 void G1CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions, bool zero_filled) { 119 if (zero_filled) { 120 return; 121 } 122 // We need to clear the bitmap on commit, removing any existing information. 123 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords); 124 _bm->clear_range(mr); 125 } 126 127 void G1CMBitMap::clear_range(MemRegion mr) { 128 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 129 assert(!mr.is_empty(), "unexpected empty region"); 130 // convert address range into offset range 131 _bm.at_put_range(heapWordToOffset(mr.start()), 132 heapWordToOffset(mr.end()), false); 133 } 134 135 G1CMMarkStack::G1CMMarkStack(G1ConcurrentMark* cm) : 136 _base(NULL), _cm(cm) 137 {} 138 139 bool G1CMMarkStack::allocate(size_t capacity) { 140 // allocate a stack of the requisite depth 141 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop))); 142 if (!rs.is_reserved()) { 143 log_warning(gc)("ConcurrentMark MarkStack allocation failure"); 144 return false; 145 } 146 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 147 if (!_virtual_space.initialize(rs, rs.size())) { 148 log_warning(gc)("ConcurrentMark MarkStack backing store failure"); 149 // Release the virtual memory reserved for the marking stack 150 rs.release(); 151 return false; 152 } 153 assert(_virtual_space.committed_size() == rs.size(), 154 "Didn't reserve backing store for all of G1ConcurrentMark stack?"); 155 _base = (oop*) _virtual_space.low(); 156 setEmpty(); 157 _capacity = (jint) capacity; 158 _saved_index = -1; 159 _should_expand = false; 160 return true; 161 } 162 163 void G1CMMarkStack::expand() { 164 // Called, during remark, if we've overflown the marking stack during marking. 165 assert(isEmpty(), "stack should been emptied while handling overflow"); 166 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted"); 167 // Clear expansion flag 168 _should_expand = false; 169 if (_capacity == (jint) MarkStackSizeMax) { 170 log_trace(gc)("(benign) Can't expand marking stack capacity, at max size limit"); 171 return; 172 } 173 // Double capacity if possible 174 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax); 175 // Do not give up existing stack until we have managed to 176 // get the double capacity that we desired. 177 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity * 178 sizeof(oop))); 179 if (rs.is_reserved()) { 180 // Release the backing store associated with old stack 181 _virtual_space.release(); 182 // Reinitialize virtual space for new stack 183 if (!_virtual_space.initialize(rs, rs.size())) { 184 fatal("Not enough swap for expanded marking stack capacity"); 185 } 186 _base = (oop*)(_virtual_space.low()); 187 _index = 0; 188 _capacity = new_capacity; 189 } else { 190 // Failed to double capacity, continue; 191 log_trace(gc)("(benign) Failed to expand marking stack capacity from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 192 _capacity / K, new_capacity / K); 193 } 194 } 195 196 void G1CMMarkStack::set_should_expand() { 197 // If we're resetting the marking state because of an 198 // marking stack overflow, record that we should, if 199 // possible, expand the stack. 200 _should_expand = _cm->has_overflown(); 201 } 202 203 G1CMMarkStack::~G1CMMarkStack() { 204 if (_base != NULL) { 205 _base = NULL; 206 _virtual_space.release(); 207 } 208 } 209 210 void G1CMMarkStack::par_push_arr(oop* ptr_arr, int n) { 211 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 212 jint start = _index; 213 jint next_index = start + n; 214 if (next_index > _capacity) { 215 _overflow = true; 216 return; 217 } 218 // Otherwise. 219 _index = next_index; 220 for (int i = 0; i < n; i++) { 221 int ind = start + i; 222 assert(ind < _capacity, "By overflow test above."); 223 _base[ind] = ptr_arr[i]; 224 } 225 } 226 227 bool G1CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { 228 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 229 jint index = _index; 230 if (index == 0) { 231 *n = 0; 232 return false; 233 } else { 234 int k = MIN2(max, index); 235 jint new_ind = index - k; 236 for (int j = 0; j < k; j++) { 237 ptr_arr[j] = _base[new_ind + j]; 238 } 239 _index = new_ind; 240 *n = k; 241 return true; 242 } 243 } 244 245 void G1CMMarkStack::note_start_of_gc() { 246 assert(_saved_index == -1, 247 "note_start_of_gc()/end_of_gc() bracketed incorrectly"); 248 _saved_index = _index; 249 } 250 251 void G1CMMarkStack::note_end_of_gc() { 252 // This is intentionally a guarantee, instead of an assert. If we 253 // accidentally add something to the mark stack during GC, it 254 // will be a correctness issue so it's better if we crash. we'll 255 // only check this once per GC anyway, so it won't be a performance 256 // issue in any way. 257 guarantee(_saved_index == _index, 258 "saved index: %d index: %d", _saved_index, _index); 259 _saved_index = -1; 260 } 261 262 G1CMRootRegions::G1CMRootRegions() : 263 _cm(NULL), _scan_in_progress(false), 264 _should_abort(false), _claimed_survivor_index(0) { } 265 266 void G1CMRootRegions::init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm) { 267 _survivors = survivors; 268 _cm = cm; 269 } 270 271 void G1CMRootRegions::prepare_for_scan() { 272 assert(!scan_in_progress(), "pre-condition"); 273 274 // Currently, only survivors can be root regions. 275 _claimed_survivor_index = 0; 276 _scan_in_progress = true; 277 _should_abort = false; 278 } 279 280 HeapRegion* G1CMRootRegions::claim_next() { 281 if (_should_abort) { 282 // If someone has set the should_abort flag, we return NULL to 283 // force the caller to bail out of their loop. 284 return NULL; 285 } 286 287 // Currently, only survivors can be root regions. 288 const GrowableArray<HeapRegion*>* survivor_regions = _survivors->regions(); 289 290 int claimed_index = Atomic::add(1, &_claimed_survivor_index) - 1; 291 if (claimed_index < survivor_regions->length()) { 292 return survivor_regions->at(claimed_index); 293 } 294 return NULL; 295 } 296 297 void G1CMRootRegions::notify_scan_done() { 298 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 299 _scan_in_progress = false; 300 RootRegionScan_lock->notify_all(); 301 } 302 303 void G1CMRootRegions::cancel_scan() { 304 notify_scan_done(); 305 } 306 307 void G1CMRootRegions::scan_finished() { 308 assert(scan_in_progress(), "pre-condition"); 309 310 // Currently, only survivors can be root regions. 311 if (!_should_abort) { 312 assert(_claimed_survivor_index >= 0, "otherwise comparison is invalid: %d", _claimed_survivor_index); 313 assert((uint)_claimed_survivor_index >= _survivors->length(), 314 "we should have claimed all survivors, claimed index = %u, length = %u", 315 (uint)_claimed_survivor_index, _survivors->length()); 316 } 317 318 notify_scan_done(); 319 } 320 321 bool G1CMRootRegions::wait_until_scan_finished() { 322 if (!scan_in_progress()) return false; 323 324 { 325 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 326 while (scan_in_progress()) { 327 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 328 } 329 } 330 return true; 331 } 332 333 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 334 return MAX2((n_par_threads + 2) / 4, 1U); 335 } 336 337 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) : 338 _g1h(g1h), 339 _markBitMap1(), 340 _markBitMap2(), 341 _parallel_marking_threads(0), 342 _max_parallel_marking_threads(0), 343 _sleep_factor(0.0), 344 _marking_task_overhead(1.0), 345 _cleanup_list("Cleanup List"), 346 347 _prevMarkBitMap(&_markBitMap1), 348 _nextMarkBitMap(&_markBitMap2), 349 350 _markStack(this), 351 // _finger set in set_non_marking_state 352 353 _max_worker_id(ParallelGCThreads), 354 // _active_tasks set in set_non_marking_state 355 // _tasks set inside the constructor 356 _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)), 357 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)), 358 359 _has_overflown(false), 360 _concurrent(false), 361 _has_aborted(false), 362 _restart_for_overflow(false), 363 _concurrent_marking_in_progress(false), 364 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 365 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()), 366 367 // _verbose_level set below 368 369 _init_times(), 370 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 371 _cleanup_times(), 372 _total_counting_time(0.0), 373 _total_rs_scrub_time(0.0), 374 375 _parallel_workers(NULL), 376 377 _completed_initialization(false) { 378 379 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage); 380 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage); 381 382 // Create & start a ConcurrentMark thread. 383 _cmThread = new ConcurrentMarkThread(this); 384 assert(cmThread() != NULL, "CM Thread should have been created"); 385 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 386 if (_cmThread->osthread() == NULL) { 387 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread"); 388 } 389 390 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 391 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency"); 392 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency"); 393 394 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 395 satb_qs.set_buffer_size(G1SATBBufferSize); 396 397 _root_regions.init(_g1h->survivor(), this); 398 399 if (ConcGCThreads > ParallelGCThreads) { 400 log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).", 401 ConcGCThreads, ParallelGCThreads); 402 return; 403 } 404 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 405 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 406 // if both are set 407 _sleep_factor = 0.0; 408 _marking_task_overhead = 1.0; 409 } else if (G1MarkingOverheadPercent > 0) { 410 // We will calculate the number of parallel marking threads based 411 // on a target overhead with respect to the soft real-time goal 412 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 413 double overall_cm_overhead = 414 (double) MaxGCPauseMillis * marking_overhead / 415 (double) GCPauseIntervalMillis; 416 double cpu_ratio = 1.0 / (double) os::processor_count(); 417 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 418 double marking_task_overhead = 419 overall_cm_overhead / marking_thread_num * 420 (double) os::processor_count(); 421 double sleep_factor = 422 (1.0 - marking_task_overhead) / marking_task_overhead; 423 424 FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num); 425 _sleep_factor = sleep_factor; 426 _marking_task_overhead = marking_task_overhead; 427 } else { 428 // Calculate the number of parallel marking threads by scaling 429 // the number of parallel GC threads. 430 uint marking_thread_num = scale_parallel_threads(ParallelGCThreads); 431 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num); 432 _sleep_factor = 0.0; 433 _marking_task_overhead = 1.0; 434 } 435 436 assert(ConcGCThreads > 0, "Should have been set"); 437 _parallel_marking_threads = ConcGCThreads; 438 _max_parallel_marking_threads = _parallel_marking_threads; 439 440 _parallel_workers = new WorkGang("G1 Marker", 441 _max_parallel_marking_threads, false, true); 442 if (_parallel_workers == NULL) { 443 vm_exit_during_initialization("Failed necessary allocation."); 444 } else { 445 _parallel_workers->initialize_workers(); 446 } 447 448 if (FLAG_IS_DEFAULT(MarkStackSize)) { 449 size_t mark_stack_size = 450 MIN2(MarkStackSizeMax, 451 MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE))); 452 // Verify that the calculated value for MarkStackSize is in range. 453 // It would be nice to use the private utility routine from Arguments. 454 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 455 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): " 456 "must be between 1 and " SIZE_FORMAT, 457 mark_stack_size, MarkStackSizeMax); 458 return; 459 } 460 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size); 461 } else { 462 // Verify MarkStackSize is in range. 463 if (FLAG_IS_CMDLINE(MarkStackSize)) { 464 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 465 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 466 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): " 467 "must be between 1 and " SIZE_FORMAT, 468 MarkStackSize, MarkStackSizeMax); 469 return; 470 } 471 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 472 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 473 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")" 474 " or for MarkStackSizeMax (" SIZE_FORMAT ")", 475 MarkStackSize, MarkStackSizeMax); 476 return; 477 } 478 } 479 } 480 } 481 482 if (!_markStack.allocate(MarkStackSize)) { 483 log_warning(gc)("Failed to allocate CM marking stack"); 484 return; 485 } 486 487 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC); 488 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC); 489 490 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 491 _active_tasks = _max_worker_id; 492 493 for (uint i = 0; i < _max_worker_id; ++i) { 494 G1CMTaskQueue* task_queue = new G1CMTaskQueue(); 495 task_queue->initialize(); 496 _task_queues->register_queue(i, task_queue); 497 498 _tasks[i] = new G1CMTask(i, this, task_queue, _task_queues); 499 500 _accum_task_vtime[i] = 0.0; 501 } 502 503 // so that the call below can read a sensible value 504 _heap_start = g1h->reserved_region().start(); 505 set_non_marking_state(); 506 _completed_initialization = true; 507 } 508 509 void G1ConcurrentMark::reset() { 510 // Starting values for these two. This should be called in a STW 511 // phase. 512 MemRegion reserved = _g1h->g1_reserved(); 513 _heap_start = reserved.start(); 514 _heap_end = reserved.end(); 515 516 // Separated the asserts so that we know which one fires. 517 assert(_heap_start != NULL, "heap bounds should look ok"); 518 assert(_heap_end != NULL, "heap bounds should look ok"); 519 assert(_heap_start < _heap_end, "heap bounds should look ok"); 520 521 // Reset all the marking data structures and any necessary flags 522 reset_marking_state(); 523 524 // We do reset all of them, since different phases will use 525 // different number of active threads. So, it's easiest to have all 526 // of them ready. 527 for (uint i = 0; i < _max_worker_id; ++i) { 528 _tasks[i]->reset(_nextMarkBitMap); 529 } 530 531 // we need this to make sure that the flag is on during the evac 532 // pause with initial mark piggy-backed 533 set_concurrent_marking_in_progress(); 534 } 535 536 537 void G1ConcurrentMark::reset_marking_state(bool clear_overflow) { 538 _markStack.set_should_expand(); 539 _markStack.setEmpty(); // Also clears the _markStack overflow flag 540 if (clear_overflow) { 541 clear_has_overflown(); 542 } else { 543 assert(has_overflown(), "pre-condition"); 544 } 545 _finger = _heap_start; 546 547 for (uint i = 0; i < _max_worker_id; ++i) { 548 G1CMTaskQueue* queue = _task_queues->queue(i); 549 queue->set_empty(); 550 } 551 } 552 553 void G1ConcurrentMark::set_concurrency(uint active_tasks) { 554 assert(active_tasks <= _max_worker_id, "we should not have more"); 555 556 _active_tasks = active_tasks; 557 // Need to update the three data structures below according to the 558 // number of active threads for this phase. 559 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 560 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 561 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 562 } 563 564 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 565 set_concurrency(active_tasks); 566 567 _concurrent = concurrent; 568 // We propagate this to all tasks, not just the active ones. 569 for (uint i = 0; i < _max_worker_id; ++i) 570 _tasks[i]->set_concurrent(concurrent); 571 572 if (concurrent) { 573 set_concurrent_marking_in_progress(); 574 } else { 575 // We currently assume that the concurrent flag has been set to 576 // false before we start remark. At this point we should also be 577 // in a STW phase. 578 assert(!concurrent_marking_in_progress(), "invariant"); 579 assert(out_of_regions(), 580 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT, 581 p2i(_finger), p2i(_heap_end)); 582 } 583 } 584 585 void G1ConcurrentMark::set_non_marking_state() { 586 // We set the global marking state to some default values when we're 587 // not doing marking. 588 reset_marking_state(); 589 _active_tasks = 0; 590 clear_concurrent_marking_in_progress(); 591 } 592 593 G1ConcurrentMark::~G1ConcurrentMark() { 594 // The G1ConcurrentMark instance is never freed. 595 ShouldNotReachHere(); 596 } 597 598 class G1ClearBitMapTask : public AbstractGangTask { 599 public: 600 static size_t chunk_size() { return M; } 601 602 private: 603 // Heap region closure used for clearing the given mark bitmap. 604 class G1ClearBitmapHRClosure : public HeapRegionClosure { 605 private: 606 G1CMBitMap* _bitmap; 607 G1ConcurrentMark* _cm; 608 public: 609 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) { 610 } 611 612 virtual bool doHeapRegion(HeapRegion* r) { 613 size_t const chunk_size_in_words = G1ClearBitMapTask::chunk_size() / HeapWordSize; 614 615 HeapWord* cur = r->bottom(); 616 HeapWord* const end = r->end(); 617 618 while (cur < end) { 619 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end)); 620 _bitmap->clear_range(mr); 621 622 cur += chunk_size_in_words; 623 624 // Abort iteration if after yielding the marking has been aborted. 625 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) { 626 return true; 627 } 628 // Repeat the asserts from before the start of the closure. We will do them 629 // as asserts here to minimize their overhead on the product. However, we 630 // will have them as guarantees at the beginning / end of the bitmap 631 // clearing to get some checking in the product. 632 assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant"); 633 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant"); 634 } 635 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index()); 636 637 return false; 638 } 639 }; 640 641 G1ClearBitmapHRClosure _cl; 642 HeapRegionClaimer _hr_claimer; 643 bool _suspendible; // If the task is suspendible, workers must join the STS. 644 645 public: 646 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) : 647 AbstractGangTask("G1 Clear Bitmap"), 648 _cl(bitmap, suspendible ? cm : NULL), 649 _hr_claimer(n_workers), 650 _suspendible(suspendible) 651 { } 652 653 void work(uint worker_id) { 654 SuspendibleThreadSetJoiner sts_join(_suspendible); 655 G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer, true); 656 } 657 658 bool is_complete() { 659 return _cl.complete(); 660 } 661 }; 662 663 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) { 664 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint."); 665 666 size_t const num_bytes_to_clear = (HeapRegion::GrainBytes * _g1h->num_regions()) / G1CMBitMap::heap_map_factor(); 667 size_t const num_chunks = align_size_up(num_bytes_to_clear, G1ClearBitMapTask::chunk_size()) / G1ClearBitMapTask::chunk_size(); 668 669 uint const num_workers = (uint)MIN2(num_chunks, (size_t)workers->active_workers()); 670 671 G1ClearBitMapTask cl(bitmap, this, num_workers, may_yield); 672 673 log_debug(gc, ergo)("Running %s with %u workers for " SIZE_FORMAT " work units.", cl.name(), num_workers, num_chunks); 674 workers->run_task(&cl, num_workers); 675 guarantee(!may_yield || cl.is_complete(), "Must have completed iteration when not yielding."); 676 } 677 678 void G1ConcurrentMark::cleanup_for_next_mark() { 679 // Make sure that the concurrent mark thread looks to still be in 680 // the current cycle. 681 guarantee(cmThread()->during_cycle(), "invariant"); 682 683 // We are finishing up the current cycle by clearing the next 684 // marking bitmap and getting it ready for the next cycle. During 685 // this time no other cycle can start. So, let's make sure that this 686 // is the case. 687 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant"); 688 689 clear_bitmap(_nextMarkBitMap, _parallel_workers, true); 690 691 // Clear the live count data. If the marking has been aborted, the abort() 692 // call already did that. 693 if (!has_aborted()) { 694 clear_live_data(_parallel_workers); 695 DEBUG_ONLY(verify_live_data_clear()); 696 } 697 698 // Repeat the asserts from above. 699 guarantee(cmThread()->during_cycle(), "invariant"); 700 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant"); 701 } 702 703 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) { 704 assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint."); 705 clear_bitmap((G1CMBitMap*)_prevMarkBitMap, workers, false); 706 } 707 708 class CheckBitmapClearHRClosure : public HeapRegionClosure { 709 G1CMBitMap* _bitmap; 710 bool _error; 711 public: 712 CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) { 713 } 714 715 virtual bool doHeapRegion(HeapRegion* r) { 716 // This closure can be called concurrently to the mutator, so we must make sure 717 // that the result of the getNextMarkedWordAddress() call is compared to the 718 // value passed to it as limit to detect any found bits. 719 // end never changes in G1. 720 HeapWord* end = r->end(); 721 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end; 722 } 723 }; 724 725 bool G1ConcurrentMark::nextMarkBitmapIsClear() { 726 CheckBitmapClearHRClosure cl(_nextMarkBitMap); 727 _g1h->heap_region_iterate(&cl); 728 return cl.complete(); 729 } 730 731 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 732 public: 733 bool doHeapRegion(HeapRegion* r) { 734 r->note_start_of_marking(); 735 return false; 736 } 737 }; 738 739 void G1ConcurrentMark::checkpointRootsInitialPre() { 740 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 741 G1Policy* g1p = g1h->g1_policy(); 742 743 _has_aborted = false; 744 745 // Initialize marking structures. This has to be done in a STW phase. 746 reset(); 747 748 // For each region note start of marking. 749 NoteStartOfMarkHRClosure startcl; 750 g1h->heap_region_iterate(&startcl); 751 } 752 753 754 void G1ConcurrentMark::checkpointRootsInitialPost() { 755 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 756 757 // Start Concurrent Marking weak-reference discovery. 758 ReferenceProcessor* rp = g1h->ref_processor_cm(); 759 // enable ("weak") refs discovery 760 rp->enable_discovery(); 761 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 762 763 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 764 // This is the start of the marking cycle, we're expected all 765 // threads to have SATB queues with active set to false. 766 satb_mq_set.set_active_all_threads(true, /* new active value */ 767 false /* expected_active */); 768 769 _root_regions.prepare_for_scan(); 770 771 // update_g1_committed() will be called at the end of an evac pause 772 // when marking is on. So, it's also called at the end of the 773 // initial-mark pause to update the heap end, if the heap expands 774 // during it. No need to call it here. 775 } 776 777 /* 778 * Notice that in the next two methods, we actually leave the STS 779 * during the barrier sync and join it immediately afterwards. If we 780 * do not do this, the following deadlock can occur: one thread could 781 * be in the barrier sync code, waiting for the other thread to also 782 * sync up, whereas another one could be trying to yield, while also 783 * waiting for the other threads to sync up too. 784 * 785 * Note, however, that this code is also used during remark and in 786 * this case we should not attempt to leave / enter the STS, otherwise 787 * we'll either hit an assert (debug / fastdebug) or deadlock 788 * (product). So we should only leave / enter the STS if we are 789 * operating concurrently. 790 * 791 * Because the thread that does the sync barrier has left the STS, it 792 * is possible to be suspended for a Full GC or an evacuation pause 793 * could occur. This is actually safe, since the entering the sync 794 * barrier is one of the last things do_marking_step() does, and it 795 * doesn't manipulate any data structures afterwards. 796 */ 797 798 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 799 bool barrier_aborted; 800 { 801 SuspendibleThreadSetLeaver sts_leave(concurrent()); 802 barrier_aborted = !_first_overflow_barrier_sync.enter(); 803 } 804 805 // at this point everyone should have synced up and not be doing any 806 // more work 807 808 if (barrier_aborted) { 809 // If the barrier aborted we ignore the overflow condition and 810 // just abort the whole marking phase as quickly as possible. 811 return; 812 } 813 814 // If we're executing the concurrent phase of marking, reset the marking 815 // state; otherwise the marking state is reset after reference processing, 816 // during the remark pause. 817 // If we reset here as a result of an overflow during the remark we will 818 // see assertion failures from any subsequent set_concurrency_and_phase() 819 // calls. 820 if (concurrent()) { 821 // let the task associated with with worker 0 do this 822 if (worker_id == 0) { 823 // task 0 is responsible for clearing the global data structures 824 // We should be here because of an overflow. During STW we should 825 // not clear the overflow flag since we rely on it being true when 826 // we exit this method to abort the pause and restart concurrent 827 // marking. 828 reset_marking_state(true /* clear_overflow */); 829 830 log_info(gc, marking)("Concurrent Mark reset for overflow"); 831 } 832 } 833 834 // after this, each task should reset its own data structures then 835 // then go into the second barrier 836 } 837 838 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 839 SuspendibleThreadSetLeaver sts_leave(concurrent()); 840 _second_overflow_barrier_sync.enter(); 841 842 // at this point everything should be re-initialized and ready to go 843 } 844 845 class G1CMConcurrentMarkingTask: public AbstractGangTask { 846 private: 847 G1ConcurrentMark* _cm; 848 ConcurrentMarkThread* _cmt; 849 850 public: 851 void work(uint worker_id) { 852 assert(Thread::current()->is_ConcurrentGC_thread(), 853 "this should only be done by a conc GC thread"); 854 ResourceMark rm; 855 856 double start_vtime = os::elapsedVTime(); 857 858 { 859 SuspendibleThreadSetJoiner sts_join; 860 861 assert(worker_id < _cm->active_tasks(), "invariant"); 862 G1CMTask* the_task = _cm->task(worker_id); 863 the_task->record_start_time(); 864 if (!_cm->has_aborted()) { 865 do { 866 double start_vtime_sec = os::elapsedVTime(); 867 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 868 869 the_task->do_marking_step(mark_step_duration_ms, 870 true /* do_termination */, 871 false /* is_serial*/); 872 873 double end_vtime_sec = os::elapsedVTime(); 874 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 875 _cm->clear_has_overflown(); 876 877 _cm->do_yield_check(); 878 879 jlong sleep_time_ms; 880 if (!_cm->has_aborted() && the_task->has_aborted()) { 881 sleep_time_ms = 882 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 883 { 884 SuspendibleThreadSetLeaver sts_leave; 885 os::sleep(Thread::current(), sleep_time_ms, false); 886 } 887 } 888 } while (!_cm->has_aborted() && the_task->has_aborted()); 889 } 890 the_task->record_end_time(); 891 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 892 } 893 894 double end_vtime = os::elapsedVTime(); 895 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 896 } 897 898 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm, 899 ConcurrentMarkThread* cmt) : 900 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 901 902 ~G1CMConcurrentMarkingTask() { } 903 }; 904 905 // Calculates the number of active workers for a concurrent 906 // phase. 907 uint G1ConcurrentMark::calc_parallel_marking_threads() { 908 uint n_conc_workers = 0; 909 if (!UseDynamicNumberOfGCThreads || 910 (!FLAG_IS_DEFAULT(ConcGCThreads) && 911 !ForceDynamicNumberOfGCThreads)) { 912 n_conc_workers = max_parallel_marking_threads(); 913 } else { 914 n_conc_workers = 915 AdaptiveSizePolicy::calc_default_active_workers( 916 max_parallel_marking_threads(), 917 1, /* Minimum workers */ 918 parallel_marking_threads(), 919 Threads::number_of_non_daemon_threads()); 920 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 921 // that scaling has already gone into "_max_parallel_marking_threads". 922 } 923 assert(n_conc_workers > 0, "Always need at least 1"); 924 return n_conc_workers; 925 } 926 927 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr) { 928 // Currently, only survivors can be root regions. 929 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 930 G1RootRegionScanClosure cl(_g1h, this); 931 932 const uintx interval = PrefetchScanIntervalInBytes; 933 HeapWord* curr = hr->bottom(); 934 const HeapWord* end = hr->top(); 935 while (curr < end) { 936 Prefetch::read(curr, interval); 937 oop obj = oop(curr); 938 int size = obj->oop_iterate_size(&cl); 939 assert(size == obj->size(), "sanity"); 940 curr += size; 941 } 942 } 943 944 class G1CMRootRegionScanTask : public AbstractGangTask { 945 private: 946 G1ConcurrentMark* _cm; 947 948 public: 949 G1CMRootRegionScanTask(G1ConcurrentMark* cm) : 950 AbstractGangTask("Root Region Scan"), _cm(cm) { } 951 952 void work(uint worker_id) { 953 assert(Thread::current()->is_ConcurrentGC_thread(), 954 "this should only be done by a conc GC thread"); 955 956 G1CMRootRegions* root_regions = _cm->root_regions(); 957 HeapRegion* hr = root_regions->claim_next(); 958 while (hr != NULL) { 959 _cm->scanRootRegion(hr); 960 hr = root_regions->claim_next(); 961 } 962 } 963 }; 964 965 void G1ConcurrentMark::scan_root_regions() { 966 // scan_in_progress() will have been set to true only if there was 967 // at least one root region to scan. So, if it's false, we 968 // should not attempt to do any further work. 969 if (root_regions()->scan_in_progress()) { 970 assert(!has_aborted(), "Aborting before root region scanning is finished not supported."); 971 972 _parallel_marking_threads = calc_parallel_marking_threads(); 973 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 974 "Maximum number of marking threads exceeded"); 975 uint active_workers = MAX2(1U, parallel_marking_threads()); 976 977 G1CMRootRegionScanTask task(this); 978 _parallel_workers->set_active_workers(active_workers); 979 _parallel_workers->run_task(&task); 980 981 // It's possible that has_aborted() is true here without actually 982 // aborting the survivor scan earlier. This is OK as it's 983 // mainly used for sanity checking. 984 root_regions()->scan_finished(); 985 } 986 } 987 988 void G1ConcurrentMark::concurrent_cycle_start() { 989 _gc_timer_cm->register_gc_start(); 990 991 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start()); 992 993 _g1h->trace_heap_before_gc(_gc_tracer_cm); 994 } 995 996 void G1ConcurrentMark::concurrent_cycle_end() { 997 _g1h->trace_heap_after_gc(_gc_tracer_cm); 998 999 if (has_aborted()) { 1000 _gc_tracer_cm->report_concurrent_mode_failure(); 1001 } 1002 1003 _gc_timer_cm->register_gc_end(); 1004 1005 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1006 } 1007 1008 void G1ConcurrentMark::mark_from_roots() { 1009 // we might be tempted to assert that: 1010 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1011 // "inconsistent argument?"); 1012 // However that wouldn't be right, because it's possible that 1013 // a safepoint is indeed in progress as a younger generation 1014 // stop-the-world GC happens even as we mark in this generation. 1015 1016 _restart_for_overflow = false; 1017 1018 // _g1h has _n_par_threads 1019 _parallel_marking_threads = calc_parallel_marking_threads(); 1020 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1021 "Maximum number of marking threads exceeded"); 1022 1023 uint active_workers = MAX2(1U, parallel_marking_threads()); 1024 assert(active_workers > 0, "Should have been set"); 1025 1026 // Parallel task terminator is set in "set_concurrency_and_phase()" 1027 set_concurrency_and_phase(active_workers, true /* concurrent */); 1028 1029 G1CMConcurrentMarkingTask markingTask(this, cmThread()); 1030 _parallel_workers->set_active_workers(active_workers); 1031 _parallel_workers->run_task(&markingTask); 1032 print_stats(); 1033 } 1034 1035 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1036 // world is stopped at this checkpoint 1037 assert(SafepointSynchronize::is_at_safepoint(), 1038 "world should be stopped"); 1039 1040 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1041 1042 // If a full collection has happened, we shouldn't do this. 1043 if (has_aborted()) { 1044 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1045 return; 1046 } 1047 1048 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1049 1050 if (VerifyDuringGC) { 1051 HandleMark hm; // handle scope 1052 g1h->prepare_for_verify(); 1053 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1054 } 1055 g1h->verifier()->check_bitmaps("Remark Start"); 1056 1057 G1Policy* g1p = g1h->g1_policy(); 1058 g1p->record_concurrent_mark_remark_start(); 1059 1060 double start = os::elapsedTime(); 1061 1062 checkpointRootsFinalWork(); 1063 1064 double mark_work_end = os::elapsedTime(); 1065 1066 weakRefsWork(clear_all_soft_refs); 1067 1068 if (has_overflown()) { 1069 // Oops. We overflowed. Restart concurrent marking. 1070 _restart_for_overflow = true; 1071 1072 // Verify the heap w.r.t. the previous marking bitmap. 1073 if (VerifyDuringGC) { 1074 HandleMark hm; // handle scope 1075 g1h->prepare_for_verify(); 1076 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)"); 1077 } 1078 1079 // Clear the marking state because we will be restarting 1080 // marking due to overflowing the global mark stack. 1081 reset_marking_state(); 1082 } else { 1083 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1084 // We're done with marking. 1085 // This is the end of the marking cycle, we're expected all 1086 // threads to have SATB queues with active set to true. 1087 satb_mq_set.set_active_all_threads(false, /* new active value */ 1088 true /* expected_active */); 1089 1090 if (VerifyDuringGC) { 1091 HandleMark hm; // handle scope 1092 g1h->prepare_for_verify(); 1093 Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)"); 1094 } 1095 g1h->verifier()->check_bitmaps("Remark End"); 1096 assert(!restart_for_overflow(), "sanity"); 1097 // Completely reset the marking state since marking completed 1098 set_non_marking_state(); 1099 } 1100 1101 // Expand the marking stack, if we have to and if we can. 1102 if (_markStack.should_expand()) { 1103 _markStack.expand(); 1104 } 1105 1106 // Statistics 1107 double now = os::elapsedTime(); 1108 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1109 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1110 _remark_times.add((now - start) * 1000.0); 1111 1112 g1p->record_concurrent_mark_remark_end(); 1113 1114 G1CMIsAliveClosure is_alive(g1h); 1115 _gc_tracer_cm->report_object_count_after_gc(&is_alive); 1116 } 1117 1118 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1119 G1CollectedHeap* _g1; 1120 size_t _freed_bytes; 1121 FreeRegionList* _local_cleanup_list; 1122 uint _old_regions_removed; 1123 uint _humongous_regions_removed; 1124 HRRSCleanupTask* _hrrs_cleanup_task; 1125 1126 public: 1127 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1128 FreeRegionList* local_cleanup_list, 1129 HRRSCleanupTask* hrrs_cleanup_task) : 1130 _g1(g1), 1131 _freed_bytes(0), 1132 _local_cleanup_list(local_cleanup_list), 1133 _old_regions_removed(0), 1134 _humongous_regions_removed(0), 1135 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1136 1137 size_t freed_bytes() { return _freed_bytes; } 1138 const uint old_regions_removed() { return _old_regions_removed; } 1139 const uint humongous_regions_removed() { return _humongous_regions_removed; } 1140 1141 bool doHeapRegion(HeapRegion *hr) { 1142 if (hr->is_archive()) { 1143 return false; 1144 } 1145 _g1->reset_gc_time_stamps(hr); 1146 hr->note_end_of_marking(); 1147 1148 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 1149 _freed_bytes += hr->used(); 1150 hr->set_containing_set(NULL); 1151 if (hr->is_humongous()) { 1152 _humongous_regions_removed++; 1153 _g1->free_humongous_region(hr, _local_cleanup_list, true); 1154 } else { 1155 _old_regions_removed++; 1156 _g1->free_region(hr, _local_cleanup_list, true); 1157 } 1158 } else { 1159 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task); 1160 } 1161 1162 return false; 1163 } 1164 }; 1165 1166 class G1ParNoteEndTask: public AbstractGangTask { 1167 friend class G1NoteEndOfConcMarkClosure; 1168 1169 protected: 1170 G1CollectedHeap* _g1h; 1171 FreeRegionList* _cleanup_list; 1172 HeapRegionClaimer _hrclaimer; 1173 1174 public: 1175 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) : 1176 AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) { 1177 } 1178 1179 void work(uint worker_id) { 1180 FreeRegionList local_cleanup_list("Local Cleanup List"); 1181 HRRSCleanupTask hrrs_cleanup_task; 1182 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list, 1183 &hrrs_cleanup_task); 1184 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer); 1185 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1186 1187 // Now update the lists 1188 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed()); 1189 { 1190 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1191 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes()); 1192 1193 // If we iterate over the global cleanup list at the end of 1194 // cleanup to do this printing we will not guarantee to only 1195 // generate output for the newly-reclaimed regions (the list 1196 // might not be empty at the beginning of cleanup; we might 1197 // still be working on its previous contents). So we do the 1198 // printing here, before we append the new regions to the global 1199 // cleanup list. 1200 1201 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1202 if (hr_printer->is_active()) { 1203 FreeRegionListIterator iter(&local_cleanup_list); 1204 while (iter.more_available()) { 1205 HeapRegion* hr = iter.get_next(); 1206 hr_printer->cleanup(hr); 1207 } 1208 } 1209 1210 _cleanup_list->add_ordered(&local_cleanup_list); 1211 assert(local_cleanup_list.is_empty(), "post-condition"); 1212 1213 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1214 } 1215 } 1216 }; 1217 1218 void G1ConcurrentMark::cleanup() { 1219 // world is stopped at this checkpoint 1220 assert(SafepointSynchronize::is_at_safepoint(), 1221 "world should be stopped"); 1222 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1223 1224 // If a full collection has happened, we shouldn't do this. 1225 if (has_aborted()) { 1226 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1227 return; 1228 } 1229 1230 g1h->verifier()->verify_region_sets_optional(); 1231 1232 if (VerifyDuringGC) { 1233 HandleMark hm; // handle scope 1234 g1h->prepare_for_verify(); 1235 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1236 } 1237 g1h->verifier()->check_bitmaps("Cleanup Start"); 1238 1239 G1Policy* g1p = g1h->g1_policy(); 1240 g1p->record_concurrent_mark_cleanup_start(); 1241 1242 double start = os::elapsedTime(); 1243 1244 HeapRegionRemSet::reset_for_cleanup_tasks(); 1245 1246 { 1247 GCTraceTime(Debug, gc)("Finalize Live Data"); 1248 finalize_live_data(); 1249 } 1250 1251 if (VerifyDuringGC) { 1252 GCTraceTime(Debug, gc)("Verify Live Data"); 1253 verify_live_data(); 1254 } 1255 1256 g1h->collector_state()->set_mark_in_progress(false); 1257 1258 double count_end = os::elapsedTime(); 1259 double this_final_counting_time = (count_end - start); 1260 _total_counting_time += this_final_counting_time; 1261 1262 if (log_is_enabled(Trace, gc, liveness)) { 1263 G1PrintRegionLivenessInfoClosure cl("Post-Marking"); 1264 _g1h->heap_region_iterate(&cl); 1265 } 1266 1267 // Install newly created mark bitMap as "prev". 1268 swapMarkBitMaps(); 1269 1270 g1h->reset_gc_time_stamp(); 1271 1272 uint n_workers = _g1h->workers()->active_workers(); 1273 1274 // Note end of marking in all heap regions. 1275 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers); 1276 g1h->workers()->run_task(&g1_par_note_end_task); 1277 g1h->check_gc_time_stamps(); 1278 1279 if (!cleanup_list_is_empty()) { 1280 // The cleanup list is not empty, so we'll have to process it 1281 // concurrently. Notify anyone else that might be wanting free 1282 // regions that there will be more free regions coming soon. 1283 g1h->set_free_regions_coming(); 1284 } 1285 1286 // call below, since it affects the metric by which we sort the heap 1287 // regions. 1288 if (G1ScrubRemSets) { 1289 double rs_scrub_start = os::elapsedTime(); 1290 g1h->scrub_rem_set(); 1291 _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start); 1292 } 1293 1294 // this will also free any regions totally full of garbage objects, 1295 // and sort the regions. 1296 g1h->g1_policy()->record_concurrent_mark_cleanup_end(); 1297 1298 // Statistics. 1299 double end = os::elapsedTime(); 1300 _cleanup_times.add((end - start) * 1000.0); 1301 1302 // Clean up will have freed any regions completely full of garbage. 1303 // Update the soft reference policy with the new heap occupancy. 1304 Universe::update_heap_info_at_gc(); 1305 1306 if (VerifyDuringGC) { 1307 HandleMark hm; // handle scope 1308 g1h->prepare_for_verify(); 1309 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)"); 1310 } 1311 1312 g1h->verifier()->check_bitmaps("Cleanup End"); 1313 1314 g1h->verifier()->verify_region_sets_optional(); 1315 1316 // We need to make this be a "collection" so any collection pause that 1317 // races with it goes around and waits for completeCleanup to finish. 1318 g1h->increment_total_collections(); 1319 1320 // Clean out dead classes and update Metaspace sizes. 1321 if (ClassUnloadingWithConcurrentMark) { 1322 ClassLoaderDataGraph::purge(); 1323 } 1324 MetaspaceGC::compute_new_size(); 1325 1326 // We reclaimed old regions so we should calculate the sizes to make 1327 // sure we update the old gen/space data. 1328 g1h->g1mm()->update_sizes(); 1329 g1h->allocation_context_stats().update_after_mark(); 1330 } 1331 1332 void G1ConcurrentMark::complete_cleanup() { 1333 if (has_aborted()) return; 1334 1335 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1336 1337 _cleanup_list.verify_optional(); 1338 FreeRegionList tmp_free_list("Tmp Free List"); 1339 1340 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1341 "cleanup list has %u entries", 1342 _cleanup_list.length()); 1343 1344 // No one else should be accessing the _cleanup_list at this point, 1345 // so it is not necessary to take any locks 1346 while (!_cleanup_list.is_empty()) { 1347 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */); 1348 assert(hr != NULL, "Got NULL from a non-empty list"); 1349 hr->par_clear(); 1350 tmp_free_list.add_ordered(hr); 1351 1352 // Instead of adding one region at a time to the secondary_free_list, 1353 // we accumulate them in the local list and move them a few at a 1354 // time. This also cuts down on the number of notify_all() calls 1355 // we do during this process. We'll also append the local list when 1356 // _cleanup_list is empty (which means we just removed the last 1357 // region from the _cleanup_list). 1358 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 1359 _cleanup_list.is_empty()) { 1360 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1361 "appending %u entries to the secondary_free_list, " 1362 "cleanup list still has %u entries", 1363 tmp_free_list.length(), 1364 _cleanup_list.length()); 1365 1366 { 1367 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 1368 g1h->secondary_free_list_add(&tmp_free_list); 1369 SecondaryFreeList_lock->notify_all(); 1370 } 1371 #ifndef PRODUCT 1372 if (G1StressConcRegionFreeing) { 1373 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 1374 os::sleep(Thread::current(), (jlong) 1, false); 1375 } 1376 } 1377 #endif 1378 } 1379 } 1380 assert(tmp_free_list.is_empty(), "post-condition"); 1381 } 1382 1383 // Supporting Object and Oop closures for reference discovery 1384 // and processing in during marking 1385 1386 bool G1CMIsAliveClosure::do_object_b(oop obj) { 1387 HeapWord* addr = (HeapWord*)obj; 1388 return addr != NULL && 1389 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 1390 } 1391 1392 // 'Keep Alive' oop closure used by both serial parallel reference processing. 1393 // Uses the G1CMTask associated with a worker thread (for serial reference 1394 // processing the G1CMTask for worker 0 is used) to preserve (mark) and 1395 // trace referent objects. 1396 // 1397 // Using the G1CMTask and embedded local queues avoids having the worker 1398 // threads operating on the global mark stack. This reduces the risk 1399 // of overflowing the stack - which we would rather avoid at this late 1400 // state. Also using the tasks' local queues removes the potential 1401 // of the workers interfering with each other that could occur if 1402 // operating on the global stack. 1403 1404 class G1CMKeepAliveAndDrainClosure: public OopClosure { 1405 G1ConcurrentMark* _cm; 1406 G1CMTask* _task; 1407 int _ref_counter_limit; 1408 int _ref_counter; 1409 bool _is_serial; 1410 public: 1411 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1412 _cm(cm), _task(task), _is_serial(is_serial), 1413 _ref_counter_limit(G1RefProcDrainInterval) { 1414 assert(_ref_counter_limit > 0, "sanity"); 1415 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1416 _ref_counter = _ref_counter_limit; 1417 } 1418 1419 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 1420 virtual void do_oop( oop* p) { do_oop_work(p); } 1421 1422 template <class T> void do_oop_work(T* p) { 1423 if (!_cm->has_overflown()) { 1424 oop obj = oopDesc::load_decode_heap_oop(p); 1425 _task->deal_with_reference(obj); 1426 _ref_counter--; 1427 1428 if (_ref_counter == 0) { 1429 // We have dealt with _ref_counter_limit references, pushing them 1430 // and objects reachable from them on to the local stack (and 1431 // possibly the global stack). Call G1CMTask::do_marking_step() to 1432 // process these entries. 1433 // 1434 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if 1435 // there's nothing more to do (i.e. we're done with the entries that 1436 // were pushed as a result of the G1CMTask::deal_with_reference() calls 1437 // above) or we overflow. 1438 // 1439 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1440 // flag while there may still be some work to do. (See the comment at 1441 // the beginning of G1CMTask::do_marking_step() for those conditions - 1442 // one of which is reaching the specified time target.) It is only 1443 // when G1CMTask::do_marking_step() returns without setting the 1444 // has_aborted() flag that the marking step has completed. 1445 do { 1446 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1447 _task->do_marking_step(mark_step_duration_ms, 1448 false /* do_termination */, 1449 _is_serial); 1450 } while (_task->has_aborted() && !_cm->has_overflown()); 1451 _ref_counter = _ref_counter_limit; 1452 } 1453 } 1454 } 1455 }; 1456 1457 // 'Drain' oop closure used by both serial and parallel reference processing. 1458 // Uses the G1CMTask associated with a given worker thread (for serial 1459 // reference processing the G1CMtask for worker 0 is used). Calls the 1460 // do_marking_step routine, with an unbelievably large timeout value, 1461 // to drain the marking data structures of the remaining entries 1462 // added by the 'keep alive' oop closure above. 1463 1464 class G1CMDrainMarkingStackClosure: public VoidClosure { 1465 G1ConcurrentMark* _cm; 1466 G1CMTask* _task; 1467 bool _is_serial; 1468 public: 1469 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1470 _cm(cm), _task(task), _is_serial(is_serial) { 1471 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1472 } 1473 1474 void do_void() { 1475 do { 1476 // We call G1CMTask::do_marking_step() to completely drain the local 1477 // and global marking stacks of entries pushed by the 'keep alive' 1478 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 1479 // 1480 // G1CMTask::do_marking_step() is called in a loop, which we'll exit 1481 // if there's nothing more to do (i.e. we've completely drained the 1482 // entries that were pushed as a a result of applying the 'keep alive' 1483 // closure to the entries on the discovered ref lists) or we overflow 1484 // the global marking stack. 1485 // 1486 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1487 // flag while there may still be some work to do. (See the comment at 1488 // the beginning of G1CMTask::do_marking_step() for those conditions - 1489 // one of which is reaching the specified time target.) It is only 1490 // when G1CMTask::do_marking_step() returns without setting the 1491 // has_aborted() flag that the marking step has completed. 1492 1493 _task->do_marking_step(1000000000.0 /* something very large */, 1494 true /* do_termination */, 1495 _is_serial); 1496 } while (_task->has_aborted() && !_cm->has_overflown()); 1497 } 1498 }; 1499 1500 // Implementation of AbstractRefProcTaskExecutor for parallel 1501 // reference processing at the end of G1 concurrent marking 1502 1503 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 1504 private: 1505 G1CollectedHeap* _g1h; 1506 G1ConcurrentMark* _cm; 1507 WorkGang* _workers; 1508 uint _active_workers; 1509 1510 public: 1511 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 1512 G1ConcurrentMark* cm, 1513 WorkGang* workers, 1514 uint n_workers) : 1515 _g1h(g1h), _cm(cm), 1516 _workers(workers), _active_workers(n_workers) { } 1517 1518 // Executes the given task using concurrent marking worker threads. 1519 virtual void execute(ProcessTask& task); 1520 virtual void execute(EnqueueTask& task); 1521 }; 1522 1523 class G1CMRefProcTaskProxy: public AbstractGangTask { 1524 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 1525 ProcessTask& _proc_task; 1526 G1CollectedHeap* _g1h; 1527 G1ConcurrentMark* _cm; 1528 1529 public: 1530 G1CMRefProcTaskProxy(ProcessTask& proc_task, 1531 G1CollectedHeap* g1h, 1532 G1ConcurrentMark* cm) : 1533 AbstractGangTask("Process reference objects in parallel"), 1534 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 1535 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1536 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 1537 } 1538 1539 virtual void work(uint worker_id) { 1540 ResourceMark rm; 1541 HandleMark hm; 1542 G1CMTask* task = _cm->task(worker_id); 1543 G1CMIsAliveClosure g1_is_alive(_g1h); 1544 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 1545 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 1546 1547 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 1548 } 1549 }; 1550 1551 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 1552 assert(_workers != NULL, "Need parallel worker threads."); 1553 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1554 1555 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 1556 1557 // We need to reset the concurrency level before each 1558 // proxy task execution, so that the termination protocol 1559 // and overflow handling in G1CMTask::do_marking_step() knows 1560 // how many workers to wait for. 1561 _cm->set_concurrency(_active_workers); 1562 _workers->run_task(&proc_task_proxy); 1563 } 1564 1565 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 1566 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 1567 EnqueueTask& _enq_task; 1568 1569 public: 1570 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 1571 AbstractGangTask("Enqueue reference objects in parallel"), 1572 _enq_task(enq_task) { } 1573 1574 virtual void work(uint worker_id) { 1575 _enq_task.work(worker_id); 1576 } 1577 }; 1578 1579 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 1580 assert(_workers != NULL, "Need parallel worker threads."); 1581 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1582 1583 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 1584 1585 // Not strictly necessary but... 1586 // 1587 // We need to reset the concurrency level before each 1588 // proxy task execution, so that the termination protocol 1589 // and overflow handling in G1CMTask::do_marking_step() knows 1590 // how many workers to wait for. 1591 _cm->set_concurrency(_active_workers); 1592 _workers->run_task(&enq_task_proxy); 1593 } 1594 1595 void G1ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) { 1596 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes); 1597 } 1598 1599 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 1600 if (has_overflown()) { 1601 // Skip processing the discovered references if we have 1602 // overflown the global marking stack. Reference objects 1603 // only get discovered once so it is OK to not 1604 // de-populate the discovered reference lists. We could have, 1605 // but the only benefit would be that, when marking restarts, 1606 // less reference objects are discovered. 1607 return; 1608 } 1609 1610 ResourceMark rm; 1611 HandleMark hm; 1612 1613 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1614 1615 // Is alive closure. 1616 G1CMIsAliveClosure g1_is_alive(g1h); 1617 1618 // Inner scope to exclude the cleaning of the string and symbol 1619 // tables from the displayed time. 1620 { 1621 GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm); 1622 1623 ReferenceProcessor* rp = g1h->ref_processor_cm(); 1624 1625 // See the comment in G1CollectedHeap::ref_processing_init() 1626 // about how reference processing currently works in G1. 1627 1628 // Set the soft reference policy 1629 rp->setup_policy(clear_all_soft_refs); 1630 assert(_markStack.isEmpty(), "mark stack should be empty"); 1631 1632 // Instances of the 'Keep Alive' and 'Complete GC' closures used 1633 // in serial reference processing. Note these closures are also 1634 // used for serially processing (by the the current thread) the 1635 // JNI references during parallel reference processing. 1636 // 1637 // These closures do not need to synchronize with the worker 1638 // threads involved in parallel reference processing as these 1639 // instances are executed serially by the current thread (e.g. 1640 // reference processing is not multi-threaded and is thus 1641 // performed by the current thread instead of a gang worker). 1642 // 1643 // The gang tasks involved in parallel reference processing create 1644 // their own instances of these closures, which do their own 1645 // synchronization among themselves. 1646 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 1647 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 1648 1649 // We need at least one active thread. If reference processing 1650 // is not multi-threaded we use the current (VMThread) thread, 1651 // otherwise we use the work gang from the G1CollectedHeap and 1652 // we utilize all the worker threads we can. 1653 bool processing_is_mt = rp->processing_is_mt(); 1654 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U); 1655 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U); 1656 1657 // Parallel processing task executor. 1658 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 1659 g1h->workers(), active_workers); 1660 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 1661 1662 // Set the concurrency level. The phase was already set prior to 1663 // executing the remark task. 1664 set_concurrency(active_workers); 1665 1666 // Set the degree of MT processing here. If the discovery was done MT, 1667 // the number of threads involved during discovery could differ from 1668 // the number of active workers. This is OK as long as the discovered 1669 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 1670 rp->set_active_mt_degree(active_workers); 1671 1672 // Process the weak references. 1673 const ReferenceProcessorStats& stats = 1674 rp->process_discovered_references(&g1_is_alive, 1675 &g1_keep_alive, 1676 &g1_drain_mark_stack, 1677 executor, 1678 _gc_timer_cm); 1679 _gc_tracer_cm->report_gc_reference_stats(stats); 1680 1681 // The do_oop work routines of the keep_alive and drain_marking_stack 1682 // oop closures will set the has_overflown flag if we overflow the 1683 // global marking stack. 1684 1685 assert(_markStack.overflow() || _markStack.isEmpty(), 1686 "mark stack should be empty (unless it overflowed)"); 1687 1688 if (_markStack.overflow()) { 1689 // This should have been done already when we tried to push an 1690 // entry on to the global mark stack. But let's do it again. 1691 set_has_overflown(); 1692 } 1693 1694 assert(rp->num_q() == active_workers, "why not"); 1695 1696 rp->enqueue_discovered_references(executor); 1697 1698 rp->verify_no_references_recorded(); 1699 assert(!rp->discovery_enabled(), "Post condition"); 1700 } 1701 1702 if (has_overflown()) { 1703 // We can not trust g1_is_alive if the marking stack overflowed 1704 return; 1705 } 1706 1707 assert(_markStack.isEmpty(), "Marking should have completed"); 1708 1709 // Unload Klasses, String, Symbols, Code Cache, etc. 1710 if (ClassUnloadingWithConcurrentMark) { 1711 bool purged_classes; 1712 1713 { 1714 GCTraceTime(Debug, gc, phases) trace("System Dictionary Unloading", _gc_timer_cm); 1715 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */); 1716 } 1717 1718 { 1719 GCTraceTime(Debug, gc, phases) trace("Parallel Unloading", _gc_timer_cm); 1720 weakRefsWorkParallelPart(&g1_is_alive, purged_classes); 1721 } 1722 } 1723 1724 if (G1StringDedup::is_enabled()) { 1725 GCTraceTime(Debug, gc, phases) trace("String Deduplication Unlink", _gc_timer_cm); 1726 G1StringDedup::unlink(&g1_is_alive); 1727 } 1728 } 1729 1730 void G1ConcurrentMark::swapMarkBitMaps() { 1731 G1CMBitMapRO* temp = _prevMarkBitMap; 1732 _prevMarkBitMap = (G1CMBitMapRO*)_nextMarkBitMap; 1733 _nextMarkBitMap = (G1CMBitMap*) temp; 1734 } 1735 1736 // Closure for marking entries in SATB buffers. 1737 class G1CMSATBBufferClosure : public SATBBufferClosure { 1738 private: 1739 G1CMTask* _task; 1740 G1CollectedHeap* _g1h; 1741 1742 // This is very similar to G1CMTask::deal_with_reference, but with 1743 // more relaxed requirements for the argument, so this must be more 1744 // circumspect about treating the argument as an object. 1745 void do_entry(void* entry) const { 1746 _task->increment_refs_reached(); 1747 HeapRegion* hr = _g1h->heap_region_containing(entry); 1748 if (entry < hr->next_top_at_mark_start()) { 1749 // Until we get here, we don't know whether entry refers to a valid 1750 // object; it could instead have been a stale reference. 1751 oop obj = static_cast<oop>(entry); 1752 assert(obj->is_oop(true /* ignore mark word */), 1753 "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj)); 1754 _task->make_reference_grey(obj); 1755 } 1756 } 1757 1758 public: 1759 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h) 1760 : _task(task), _g1h(g1h) { } 1761 1762 virtual void do_buffer(void** buffer, size_t size) { 1763 for (size_t i = 0; i < size; ++i) { 1764 do_entry(buffer[i]); 1765 } 1766 } 1767 }; 1768 1769 class G1RemarkThreadsClosure : public ThreadClosure { 1770 G1CMSATBBufferClosure _cm_satb_cl; 1771 G1CMOopClosure _cm_cl; 1772 MarkingCodeBlobClosure _code_cl; 1773 int _thread_parity; 1774 1775 public: 1776 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) : 1777 _cm_satb_cl(task, g1h), 1778 _cm_cl(g1h, g1h->concurrent_mark(), task), 1779 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations), 1780 _thread_parity(Threads::thread_claim_parity()) {} 1781 1782 void do_thread(Thread* thread) { 1783 if (thread->is_Java_thread()) { 1784 if (thread->claim_oops_do(true, _thread_parity)) { 1785 JavaThread* jt = (JavaThread*)thread; 1786 1787 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking 1788 // however the liveness of oops reachable from nmethods have very complex lifecycles: 1789 // * Alive if on the stack of an executing method 1790 // * Weakly reachable otherwise 1791 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be 1792 // live by the SATB invariant but other oops recorded in nmethods may behave differently. 1793 jt->nmethods_do(&_code_cl); 1794 1795 jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl); 1796 } 1797 } else if (thread->is_VM_thread()) { 1798 if (thread->claim_oops_do(true, _thread_parity)) { 1799 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl); 1800 } 1801 } 1802 } 1803 }; 1804 1805 class G1CMRemarkTask: public AbstractGangTask { 1806 private: 1807 G1ConcurrentMark* _cm; 1808 public: 1809 void work(uint worker_id) { 1810 // Since all available tasks are actually started, we should 1811 // only proceed if we're supposed to be active. 1812 if (worker_id < _cm->active_tasks()) { 1813 G1CMTask* task = _cm->task(worker_id); 1814 task->record_start_time(); 1815 { 1816 ResourceMark rm; 1817 HandleMark hm; 1818 1819 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task); 1820 Threads::threads_do(&threads_f); 1821 } 1822 1823 do { 1824 task->do_marking_step(1000000000.0 /* something very large */, 1825 true /* do_termination */, 1826 false /* is_serial */); 1827 } while (task->has_aborted() && !_cm->has_overflown()); 1828 // If we overflow, then we do not want to restart. We instead 1829 // want to abort remark and do concurrent marking again. 1830 task->record_end_time(); 1831 } 1832 } 1833 1834 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) : 1835 AbstractGangTask("Par Remark"), _cm(cm) { 1836 _cm->terminator()->reset_for_reuse(active_workers); 1837 } 1838 }; 1839 1840 void G1ConcurrentMark::checkpointRootsFinalWork() { 1841 ResourceMark rm; 1842 HandleMark hm; 1843 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1844 1845 GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm); 1846 1847 g1h->ensure_parsability(false); 1848 1849 // this is remark, so we'll use up all active threads 1850 uint active_workers = g1h->workers()->active_workers(); 1851 set_concurrency_and_phase(active_workers, false /* concurrent */); 1852 // Leave _parallel_marking_threads at it's 1853 // value originally calculated in the G1ConcurrentMark 1854 // constructor and pass values of the active workers 1855 // through the gang in the task. 1856 1857 { 1858 StrongRootsScope srs(active_workers); 1859 1860 G1CMRemarkTask remarkTask(this, active_workers); 1861 // We will start all available threads, even if we decide that the 1862 // active_workers will be fewer. The extra ones will just bail out 1863 // immediately. 1864 g1h->workers()->run_task(&remarkTask); 1865 } 1866 1867 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1868 guarantee(has_overflown() || 1869 satb_mq_set.completed_buffers_num() == 0, 1870 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT, 1871 BOOL_TO_STR(has_overflown()), 1872 satb_mq_set.completed_buffers_num()); 1873 1874 print_stats(); 1875 } 1876 1877 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 1878 // Note we are overriding the read-only view of the prev map here, via 1879 // the cast. 1880 ((G1CMBitMap*)_prevMarkBitMap)->clear_range(mr); 1881 } 1882 1883 HeapRegion* 1884 G1ConcurrentMark::claim_region(uint worker_id) { 1885 // "checkpoint" the finger 1886 HeapWord* finger = _finger; 1887 1888 // _heap_end will not change underneath our feet; it only changes at 1889 // yield points. 1890 while (finger < _heap_end) { 1891 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 1892 1893 HeapRegion* curr_region = _g1h->heap_region_containing(finger); 1894 1895 // Above heap_region_containing may return NULL as we always scan claim 1896 // until the end of the heap. In this case, just jump to the next region. 1897 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords; 1898 1899 // Is the gap between reading the finger and doing the CAS too long? 1900 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 1901 if (res == finger && curr_region != NULL) { 1902 // we succeeded 1903 HeapWord* bottom = curr_region->bottom(); 1904 HeapWord* limit = curr_region->next_top_at_mark_start(); 1905 1906 // notice that _finger == end cannot be guaranteed here since, 1907 // someone else might have moved the finger even further 1908 assert(_finger >= end, "the finger should have moved forward"); 1909 1910 if (limit > bottom) { 1911 return curr_region; 1912 } else { 1913 assert(limit == bottom, 1914 "the region limit should be at bottom"); 1915 // we return NULL and the caller should try calling 1916 // claim_region() again. 1917 return NULL; 1918 } 1919 } else { 1920 assert(_finger > finger, "the finger should have moved forward"); 1921 // read it again 1922 finger = _finger; 1923 } 1924 } 1925 1926 return NULL; 1927 } 1928 1929 #ifndef PRODUCT 1930 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC { 1931 private: 1932 G1CollectedHeap* _g1h; 1933 const char* _phase; 1934 int _info; 1935 1936 public: 1937 VerifyNoCSetOops(const char* phase, int info = -1) : 1938 _g1h(G1CollectedHeap::heap()), 1939 _phase(phase), 1940 _info(info) 1941 { } 1942 1943 void operator()(oop obj) const { 1944 guarantee(obj->is_oop(), 1945 "Non-oop " PTR_FORMAT ", phase: %s, info: %d", 1946 p2i(obj), _phase, _info); 1947 guarantee(!_g1h->obj_in_cs(obj), 1948 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d", 1949 p2i(obj), _phase, _info); 1950 } 1951 }; 1952 1953 void G1ConcurrentMark::verify_no_cset_oops() { 1954 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1955 if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) { 1956 return; 1957 } 1958 1959 // Verify entries on the global mark stack 1960 _markStack.iterate(VerifyNoCSetOops("Stack")); 1961 1962 // Verify entries on the task queues 1963 for (uint i = 0; i < _max_worker_id; ++i) { 1964 G1CMTaskQueue* queue = _task_queues->queue(i); 1965 queue->iterate(VerifyNoCSetOops("Queue", i)); 1966 } 1967 1968 // Verify the global finger 1969 HeapWord* global_finger = finger(); 1970 if (global_finger != NULL && global_finger < _heap_end) { 1971 // Since we always iterate over all regions, we might get a NULL HeapRegion 1972 // here. 1973 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger); 1974 guarantee(global_hr == NULL || global_finger == global_hr->bottom(), 1975 "global finger: " PTR_FORMAT " region: " HR_FORMAT, 1976 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)); 1977 } 1978 1979 // Verify the task fingers 1980 assert(parallel_marking_threads() <= _max_worker_id, "sanity"); 1981 for (uint i = 0; i < parallel_marking_threads(); ++i) { 1982 G1CMTask* task = _tasks[i]; 1983 HeapWord* task_finger = task->finger(); 1984 if (task_finger != NULL && task_finger < _heap_end) { 1985 // See above note on the global finger verification. 1986 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger); 1987 guarantee(task_hr == NULL || task_finger == task_hr->bottom() || 1988 !task_hr->in_collection_set(), 1989 "task finger: " PTR_FORMAT " region: " HR_FORMAT, 1990 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)); 1991 } 1992 } 1993 } 1994 #endif // PRODUCT 1995 void G1ConcurrentMark::create_live_data() { 1996 _g1h->g1_rem_set()->create_card_live_data(_parallel_workers, _nextMarkBitMap); 1997 } 1998 1999 void G1ConcurrentMark::finalize_live_data() { 2000 _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _nextMarkBitMap); 2001 } 2002 2003 void G1ConcurrentMark::verify_live_data() { 2004 _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _nextMarkBitMap); 2005 } 2006 2007 void G1ConcurrentMark::clear_live_data(WorkGang* workers) { 2008 _g1h->g1_rem_set()->clear_card_live_data(workers); 2009 } 2010 2011 #ifdef ASSERT 2012 void G1ConcurrentMark::verify_live_data_clear() { 2013 _g1h->g1_rem_set()->verify_card_live_data_is_clear(); 2014 } 2015 #endif 2016 2017 void G1ConcurrentMark::print_stats() { 2018 if (!log_is_enabled(Debug, gc, stats)) { 2019 return; 2020 } 2021 log_debug(gc, stats)("---------------------------------------------------------------------"); 2022 for (size_t i = 0; i < _active_tasks; ++i) { 2023 _tasks[i]->print_stats(); 2024 log_debug(gc, stats)("---------------------------------------------------------------------"); 2025 } 2026 } 2027 2028 void G1ConcurrentMark::abort() { 2029 if (!cmThread()->during_cycle() || _has_aborted) { 2030 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything. 2031 return; 2032 } 2033 2034 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next 2035 // concurrent bitmap clearing. 2036 { 2037 GCTraceTime(Debug, gc)("Clear Next Bitmap"); 2038 clear_bitmap(_nextMarkBitMap, _g1h->workers(), false); 2039 } 2040 // Note we cannot clear the previous marking bitmap here 2041 // since VerifyDuringGC verifies the objects marked during 2042 // a full GC against the previous bitmap. 2043 2044 { 2045 GCTraceTime(Debug, gc)("Clear Live Data"); 2046 clear_live_data(_g1h->workers()); 2047 } 2048 DEBUG_ONLY({ 2049 GCTraceTime(Debug, gc)("Verify Live Data Clear"); 2050 verify_live_data_clear(); 2051 }) 2052 // Empty mark stack 2053 reset_marking_state(); 2054 for (uint i = 0; i < _max_worker_id; ++i) { 2055 _tasks[i]->clear_region_fields(); 2056 } 2057 _first_overflow_barrier_sync.abort(); 2058 _second_overflow_barrier_sync.abort(); 2059 _has_aborted = true; 2060 2061 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2062 satb_mq_set.abandon_partial_marking(); 2063 // This can be called either during or outside marking, we'll read 2064 // the expected_active value from the SATB queue set. 2065 satb_mq_set.set_active_all_threads( 2066 false, /* new active value */ 2067 satb_mq_set.is_active() /* expected_active */); 2068 } 2069 2070 static void print_ms_time_info(const char* prefix, const char* name, 2071 NumberSeq& ns) { 2072 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 2073 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 2074 if (ns.num() > 0) { 2075 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]", 2076 prefix, ns.sd(), ns.maximum()); 2077 } 2078 } 2079 2080 void G1ConcurrentMark::print_summary_info() { 2081 Log(gc, marking) log; 2082 if (!log.is_trace()) { 2083 return; 2084 } 2085 2086 log.trace(" Concurrent marking:"); 2087 print_ms_time_info(" ", "init marks", _init_times); 2088 print_ms_time_info(" ", "remarks", _remark_times); 2089 { 2090 print_ms_time_info(" ", "final marks", _remark_mark_times); 2091 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 2092 2093 } 2094 print_ms_time_info(" ", "cleanups", _cleanup_times); 2095 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).", 2096 _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2097 if (G1ScrubRemSets) { 2098 log.trace(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 2099 _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2100 } 2101 log.trace(" Total stop_world time = %8.2f s.", 2102 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0); 2103 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).", 2104 cmThread()->vtime_accum(), cmThread()->vtime_mark_accum()); 2105 } 2106 2107 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const { 2108 _parallel_workers->print_worker_threads_on(st); 2109 } 2110 2111 void G1ConcurrentMark::threads_do(ThreadClosure* tc) const { 2112 _parallel_workers->threads_do(tc); 2113 } 2114 2115 void G1ConcurrentMark::print_on_error(outputStream* st) const { 2116 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT, 2117 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap)); 2118 _prevMarkBitMap->print_on_error(st, " Prev Bits: "); 2119 _nextMarkBitMap->print_on_error(st, " Next Bits: "); 2120 } 2121 2122 // Closure for iteration over bitmaps 2123 class G1CMBitMapClosure : public BitMapClosure { 2124 private: 2125 // the bitmap that is being iterated over 2126 G1CMBitMap* _nextMarkBitMap; 2127 G1ConcurrentMark* _cm; 2128 G1CMTask* _task; 2129 2130 public: 2131 G1CMBitMapClosure(G1CMTask *task, G1ConcurrentMark* cm, G1CMBitMap* nextMarkBitMap) : 2132 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } 2133 2134 bool do_bit(size_t offset) { 2135 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); 2136 assert(_nextMarkBitMap->isMarked(addr), "invariant"); 2137 assert( addr < _cm->finger(), "invariant"); 2138 assert(addr >= _task->finger(), "invariant"); 2139 2140 // We move that task's local finger along. 2141 _task->move_finger_to(addr); 2142 2143 _task->scan_object(oop(addr)); 2144 // we only partially drain the local queue and global stack 2145 _task->drain_local_queue(true); 2146 _task->drain_global_stack(true); 2147 2148 // if the has_aborted flag has been raised, we need to bail out of 2149 // the iteration 2150 return !_task->has_aborted(); 2151 } 2152 }; 2153 2154 static ReferenceProcessor* get_cm_oop_closure_ref_processor(G1CollectedHeap* g1h) { 2155 ReferenceProcessor* result = g1h->ref_processor_cm(); 2156 assert(result != NULL, "CM reference processor should not be NULL"); 2157 return result; 2158 } 2159 2160 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 2161 G1ConcurrentMark* cm, 2162 G1CMTask* task) 2163 : MetadataAwareOopClosure(get_cm_oop_closure_ref_processor(g1h)), 2164 _g1h(g1h), _cm(cm), _task(task) 2165 { } 2166 2167 void G1CMTask::setup_for_region(HeapRegion* hr) { 2168 assert(hr != NULL, 2169 "claim_region() should have filtered out NULL regions"); 2170 _curr_region = hr; 2171 _finger = hr->bottom(); 2172 update_region_limit(); 2173 } 2174 2175 void G1CMTask::update_region_limit() { 2176 HeapRegion* hr = _curr_region; 2177 HeapWord* bottom = hr->bottom(); 2178 HeapWord* limit = hr->next_top_at_mark_start(); 2179 2180 if (limit == bottom) { 2181 // The region was collected underneath our feet. 2182 // We set the finger to bottom to ensure that the bitmap 2183 // iteration that will follow this will not do anything. 2184 // (this is not a condition that holds when we set the region up, 2185 // as the region is not supposed to be empty in the first place) 2186 _finger = bottom; 2187 } else if (limit >= _region_limit) { 2188 assert(limit >= _finger, "peace of mind"); 2189 } else { 2190 assert(limit < _region_limit, "only way to get here"); 2191 // This can happen under some pretty unusual circumstances. An 2192 // evacuation pause empties the region underneath our feet (NTAMS 2193 // at bottom). We then do some allocation in the region (NTAMS 2194 // stays at bottom), followed by the region being used as a GC 2195 // alloc region (NTAMS will move to top() and the objects 2196 // originally below it will be grayed). All objects now marked in 2197 // the region are explicitly grayed, if below the global finger, 2198 // and we do not need in fact to scan anything else. So, we simply 2199 // set _finger to be limit to ensure that the bitmap iteration 2200 // doesn't do anything. 2201 _finger = limit; 2202 } 2203 2204 _region_limit = limit; 2205 } 2206 2207 void G1CMTask::giveup_current_region() { 2208 assert(_curr_region != NULL, "invariant"); 2209 clear_region_fields(); 2210 } 2211 2212 void G1CMTask::clear_region_fields() { 2213 // Values for these three fields that indicate that we're not 2214 // holding on to a region. 2215 _curr_region = NULL; 2216 _finger = NULL; 2217 _region_limit = NULL; 2218 } 2219 2220 void G1CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 2221 if (cm_oop_closure == NULL) { 2222 assert(_cm_oop_closure != NULL, "invariant"); 2223 } else { 2224 assert(_cm_oop_closure == NULL, "invariant"); 2225 } 2226 _cm_oop_closure = cm_oop_closure; 2227 } 2228 2229 void G1CMTask::reset(G1CMBitMap* nextMarkBitMap) { 2230 guarantee(nextMarkBitMap != NULL, "invariant"); 2231 _nextMarkBitMap = nextMarkBitMap; 2232 clear_region_fields(); 2233 2234 _calls = 0; 2235 _elapsed_time_ms = 0.0; 2236 _termination_time_ms = 0.0; 2237 _termination_start_time_ms = 0.0; 2238 } 2239 2240 bool G1CMTask::should_exit_termination() { 2241 regular_clock_call(); 2242 // This is called when we are in the termination protocol. We should 2243 // quit if, for some reason, this task wants to abort or the global 2244 // stack is not empty (this means that we can get work from it). 2245 return !_cm->mark_stack_empty() || has_aborted(); 2246 } 2247 2248 void G1CMTask::reached_limit() { 2249 assert(_words_scanned >= _words_scanned_limit || 2250 _refs_reached >= _refs_reached_limit , 2251 "shouldn't have been called otherwise"); 2252 regular_clock_call(); 2253 } 2254 2255 void G1CMTask::regular_clock_call() { 2256 if (has_aborted()) return; 2257 2258 // First, we need to recalculate the words scanned and refs reached 2259 // limits for the next clock call. 2260 recalculate_limits(); 2261 2262 // During the regular clock call we do the following 2263 2264 // (1) If an overflow has been flagged, then we abort. 2265 if (_cm->has_overflown()) { 2266 set_has_aborted(); 2267 return; 2268 } 2269 2270 // If we are not concurrent (i.e. we're doing remark) we don't need 2271 // to check anything else. The other steps are only needed during 2272 // the concurrent marking phase. 2273 if (!concurrent()) return; 2274 2275 // (2) If marking has been aborted for Full GC, then we also abort. 2276 if (_cm->has_aborted()) { 2277 set_has_aborted(); 2278 return; 2279 } 2280 2281 double curr_time_ms = os::elapsedVTime() * 1000.0; 2282 2283 // (4) We check whether we should yield. If we have to, then we abort. 2284 if (SuspendibleThreadSet::should_yield()) { 2285 // We should yield. To do this we abort the task. The caller is 2286 // responsible for yielding. 2287 set_has_aborted(); 2288 return; 2289 } 2290 2291 // (5) We check whether we've reached our time quota. If we have, 2292 // then we abort. 2293 double elapsed_time_ms = curr_time_ms - _start_time_ms; 2294 if (elapsed_time_ms > _time_target_ms) { 2295 set_has_aborted(); 2296 _has_timed_out = true; 2297 return; 2298 } 2299 2300 // (6) Finally, we check whether there are enough completed STAB 2301 // buffers available for processing. If there are, we abort. 2302 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2303 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 2304 // we do need to process SATB buffers, we'll abort and restart 2305 // the marking task to do so 2306 set_has_aborted(); 2307 return; 2308 } 2309 } 2310 2311 void G1CMTask::recalculate_limits() { 2312 _real_words_scanned_limit = _words_scanned + words_scanned_period; 2313 _words_scanned_limit = _real_words_scanned_limit; 2314 2315 _real_refs_reached_limit = _refs_reached + refs_reached_period; 2316 _refs_reached_limit = _real_refs_reached_limit; 2317 } 2318 2319 void G1CMTask::decrease_limits() { 2320 // This is called when we believe that we're going to do an infrequent 2321 // operation which will increase the per byte scanned cost (i.e. move 2322 // entries to/from the global stack). It basically tries to decrease the 2323 // scanning limit so that the clock is called earlier. 2324 2325 _words_scanned_limit = _real_words_scanned_limit - 2326 3 * words_scanned_period / 4; 2327 _refs_reached_limit = _real_refs_reached_limit - 2328 3 * refs_reached_period / 4; 2329 } 2330 2331 void G1CMTask::move_entries_to_global_stack() { 2332 // local array where we'll store the entries that will be popped 2333 // from the local queue 2334 oop buffer[global_stack_transfer_size]; 2335 2336 int n = 0; 2337 oop obj; 2338 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { 2339 buffer[n] = obj; 2340 ++n; 2341 } 2342 2343 if (n > 0) { 2344 // we popped at least one entry from the local queue 2345 2346 if (!_cm->mark_stack_push(buffer, n)) { 2347 set_has_aborted(); 2348 } 2349 } 2350 2351 // this operation was quite expensive, so decrease the limits 2352 decrease_limits(); 2353 } 2354 2355 void G1CMTask::get_entries_from_global_stack() { 2356 // local array where we'll store the entries that will be popped 2357 // from the global stack. 2358 oop buffer[global_stack_transfer_size]; 2359 int n; 2360 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); 2361 assert(n <= global_stack_transfer_size, 2362 "we should not pop more than the given limit"); 2363 if (n > 0) { 2364 // yes, we did actually pop at least one entry 2365 for (int i = 0; i < n; ++i) { 2366 bool success = _task_queue->push(buffer[i]); 2367 // We only call this when the local queue is empty or under a 2368 // given target limit. So, we do not expect this push to fail. 2369 assert(success, "invariant"); 2370 } 2371 } 2372 2373 // this operation was quite expensive, so decrease the limits 2374 decrease_limits(); 2375 } 2376 2377 void G1CMTask::drain_local_queue(bool partially) { 2378 if (has_aborted()) return; 2379 2380 // Decide what the target size is, depending whether we're going to 2381 // drain it partially (so that other tasks can steal if they run out 2382 // of things to do) or totally (at the very end). 2383 size_t target_size; 2384 if (partially) { 2385 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 2386 } else { 2387 target_size = 0; 2388 } 2389 2390 if (_task_queue->size() > target_size) { 2391 oop obj; 2392 bool ret = _task_queue->pop_local(obj); 2393 while (ret) { 2394 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); 2395 assert(!_g1h->is_on_master_free_list( 2396 _g1h->heap_region_containing((HeapWord*) obj)), "invariant"); 2397 2398 scan_object(obj); 2399 2400 if (_task_queue->size() <= target_size || has_aborted()) { 2401 ret = false; 2402 } else { 2403 ret = _task_queue->pop_local(obj); 2404 } 2405 } 2406 } 2407 } 2408 2409 void G1CMTask::drain_global_stack(bool partially) { 2410 if (has_aborted()) return; 2411 2412 // We have a policy to drain the local queue before we attempt to 2413 // drain the global stack. 2414 assert(partially || _task_queue->size() == 0, "invariant"); 2415 2416 // Decide what the target size is, depending whether we're going to 2417 // drain it partially (so that other tasks can steal if they run out 2418 // of things to do) or totally (at the very end). Notice that, 2419 // because we move entries from the global stack in chunks or 2420 // because another task might be doing the same, we might in fact 2421 // drop below the target. But, this is not a problem. 2422 size_t target_size; 2423 if (partially) { 2424 target_size = _cm->partial_mark_stack_size_target(); 2425 } else { 2426 target_size = 0; 2427 } 2428 2429 if (_cm->mark_stack_size() > target_size) { 2430 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 2431 get_entries_from_global_stack(); 2432 drain_local_queue(partially); 2433 } 2434 } 2435 } 2436 2437 // SATB Queue has several assumptions on whether to call the par or 2438 // non-par versions of the methods. this is why some of the code is 2439 // replicated. We should really get rid of the single-threaded version 2440 // of the code to simplify things. 2441 void G1CMTask::drain_satb_buffers() { 2442 if (has_aborted()) return; 2443 2444 // We set this so that the regular clock knows that we're in the 2445 // middle of draining buffers and doesn't set the abort flag when it 2446 // notices that SATB buffers are available for draining. It'd be 2447 // very counter productive if it did that. :-) 2448 _draining_satb_buffers = true; 2449 2450 G1CMSATBBufferClosure satb_cl(this, _g1h); 2451 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2452 2453 // This keeps claiming and applying the closure to completed buffers 2454 // until we run out of buffers or we need to abort. 2455 while (!has_aborted() && 2456 satb_mq_set.apply_closure_to_completed_buffer(&satb_cl)) { 2457 regular_clock_call(); 2458 } 2459 2460 _draining_satb_buffers = false; 2461 2462 assert(has_aborted() || 2463 concurrent() || 2464 satb_mq_set.completed_buffers_num() == 0, "invariant"); 2465 2466 // again, this was a potentially expensive operation, decrease the 2467 // limits to get the regular clock call early 2468 decrease_limits(); 2469 } 2470 2471 void G1CMTask::print_stats() { 2472 log_debug(gc, stats)("Marking Stats, task = %u, calls = %d", 2473 _worker_id, _calls); 2474 log_debug(gc, stats)(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 2475 _elapsed_time_ms, _termination_time_ms); 2476 log_debug(gc, stats)(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 2477 _step_times_ms.num(), _step_times_ms.avg(), 2478 _step_times_ms.sd()); 2479 log_debug(gc, stats)(" max = %1.2lfms, total = %1.2lfms", 2480 _step_times_ms.maximum(), _step_times_ms.sum()); 2481 } 2482 2483 bool G1ConcurrentMark::try_stealing(uint worker_id, int* hash_seed, oop& obj) { 2484 return _task_queues->steal(worker_id, hash_seed, obj); 2485 } 2486 2487 /***************************************************************************** 2488 2489 The do_marking_step(time_target_ms, ...) method is the building 2490 block of the parallel marking framework. It can be called in parallel 2491 with other invocations of do_marking_step() on different tasks 2492 (but only one per task, obviously) and concurrently with the 2493 mutator threads, or during remark, hence it eliminates the need 2494 for two versions of the code. When called during remark, it will 2495 pick up from where the task left off during the concurrent marking 2496 phase. Interestingly, tasks are also claimable during evacuation 2497 pauses too, since do_marking_step() ensures that it aborts before 2498 it needs to yield. 2499 2500 The data structures that it uses to do marking work are the 2501 following: 2502 2503 (1) Marking Bitmap. If there are gray objects that appear only 2504 on the bitmap (this happens either when dealing with an overflow 2505 or when the initial marking phase has simply marked the roots 2506 and didn't push them on the stack), then tasks claim heap 2507 regions whose bitmap they then scan to find gray objects. A 2508 global finger indicates where the end of the last claimed region 2509 is. A local finger indicates how far into the region a task has 2510 scanned. The two fingers are used to determine how to gray an 2511 object (i.e. whether simply marking it is OK, as it will be 2512 visited by a task in the future, or whether it needs to be also 2513 pushed on a stack). 2514 2515 (2) Local Queue. The local queue of the task which is accessed 2516 reasonably efficiently by the task. Other tasks can steal from 2517 it when they run out of work. Throughout the marking phase, a 2518 task attempts to keep its local queue short but not totally 2519 empty, so that entries are available for stealing by other 2520 tasks. Only when there is no more work, a task will totally 2521 drain its local queue. 2522 2523 (3) Global Mark Stack. This handles local queue overflow. During 2524 marking only sets of entries are moved between it and the local 2525 queues, as access to it requires a mutex and more fine-grain 2526 interaction with it which might cause contention. If it 2527 overflows, then the marking phase should restart and iterate 2528 over the bitmap to identify gray objects. Throughout the marking 2529 phase, tasks attempt to keep the global mark stack at a small 2530 length but not totally empty, so that entries are available for 2531 popping by other tasks. Only when there is no more work, tasks 2532 will totally drain the global mark stack. 2533 2534 (4) SATB Buffer Queue. This is where completed SATB buffers are 2535 made available. Buffers are regularly removed from this queue 2536 and scanned for roots, so that the queue doesn't get too 2537 long. During remark, all completed buffers are processed, as 2538 well as the filled in parts of any uncompleted buffers. 2539 2540 The do_marking_step() method tries to abort when the time target 2541 has been reached. There are a few other cases when the 2542 do_marking_step() method also aborts: 2543 2544 (1) When the marking phase has been aborted (after a Full GC). 2545 2546 (2) When a global overflow (on the global stack) has been 2547 triggered. Before the task aborts, it will actually sync up with 2548 the other tasks to ensure that all the marking data structures 2549 (local queues, stacks, fingers etc.) are re-initialized so that 2550 when do_marking_step() completes, the marking phase can 2551 immediately restart. 2552 2553 (3) When enough completed SATB buffers are available. The 2554 do_marking_step() method only tries to drain SATB buffers right 2555 at the beginning. So, if enough buffers are available, the 2556 marking step aborts and the SATB buffers are processed at 2557 the beginning of the next invocation. 2558 2559 (4) To yield. when we have to yield then we abort and yield 2560 right at the end of do_marking_step(). This saves us from a lot 2561 of hassle as, by yielding we might allow a Full GC. If this 2562 happens then objects will be compacted underneath our feet, the 2563 heap might shrink, etc. We save checking for this by just 2564 aborting and doing the yield right at the end. 2565 2566 From the above it follows that the do_marking_step() method should 2567 be called in a loop (or, otherwise, regularly) until it completes. 2568 2569 If a marking step completes without its has_aborted() flag being 2570 true, it means it has completed the current marking phase (and 2571 also all other marking tasks have done so and have all synced up). 2572 2573 A method called regular_clock_call() is invoked "regularly" (in 2574 sub ms intervals) throughout marking. It is this clock method that 2575 checks all the abort conditions which were mentioned above and 2576 decides when the task should abort. A work-based scheme is used to 2577 trigger this clock method: when the number of object words the 2578 marking phase has scanned or the number of references the marking 2579 phase has visited reach a given limit. Additional invocations to 2580 the method clock have been planted in a few other strategic places 2581 too. The initial reason for the clock method was to avoid calling 2582 vtime too regularly, as it is quite expensive. So, once it was in 2583 place, it was natural to piggy-back all the other conditions on it 2584 too and not constantly check them throughout the code. 2585 2586 If do_termination is true then do_marking_step will enter its 2587 termination protocol. 2588 2589 The value of is_serial must be true when do_marking_step is being 2590 called serially (i.e. by the VMThread) and do_marking_step should 2591 skip any synchronization in the termination and overflow code. 2592 Examples include the serial remark code and the serial reference 2593 processing closures. 2594 2595 The value of is_serial must be false when do_marking_step is 2596 being called by any of the worker threads in a work gang. 2597 Examples include the concurrent marking code (CMMarkingTask), 2598 the MT remark code, and the MT reference processing closures. 2599 2600 *****************************************************************************/ 2601 2602 void G1CMTask::do_marking_step(double time_target_ms, 2603 bool do_termination, 2604 bool is_serial) { 2605 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 2606 assert(concurrent() == _cm->concurrent(), "they should be the same"); 2607 2608 G1Policy* g1_policy = _g1h->g1_policy(); 2609 assert(_task_queues != NULL, "invariant"); 2610 assert(_task_queue != NULL, "invariant"); 2611 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant"); 2612 2613 assert(!_claimed, 2614 "only one thread should claim this task at any one time"); 2615 2616 // OK, this doesn't safeguard again all possible scenarios, as it is 2617 // possible for two threads to set the _claimed flag at the same 2618 // time. But it is only for debugging purposes anyway and it will 2619 // catch most problems. 2620 _claimed = true; 2621 2622 _start_time_ms = os::elapsedVTime() * 1000.0; 2623 2624 // If do_stealing is true then do_marking_step will attempt to 2625 // steal work from the other G1CMTasks. It only makes sense to 2626 // enable stealing when the termination protocol is enabled 2627 // and do_marking_step() is not being called serially. 2628 bool do_stealing = do_termination && !is_serial; 2629 2630 double diff_prediction_ms = _g1h->g1_policy()->predictor().get_new_prediction(&_marking_step_diffs_ms); 2631 _time_target_ms = time_target_ms - diff_prediction_ms; 2632 2633 // set up the variables that are used in the work-based scheme to 2634 // call the regular clock method 2635 _words_scanned = 0; 2636 _refs_reached = 0; 2637 recalculate_limits(); 2638 2639 // clear all flags 2640 clear_has_aborted(); 2641 _has_timed_out = false; 2642 _draining_satb_buffers = false; 2643 2644 ++_calls; 2645 2646 // Set up the bitmap and oop closures. Anything that uses them is 2647 // eventually called from this method, so it is OK to allocate these 2648 // statically. 2649 G1CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); 2650 G1CMOopClosure cm_oop_closure(_g1h, _cm, this); 2651 set_cm_oop_closure(&cm_oop_closure); 2652 2653 if (_cm->has_overflown()) { 2654 // This can happen if the mark stack overflows during a GC pause 2655 // and this task, after a yield point, restarts. We have to abort 2656 // as we need to get into the overflow protocol which happens 2657 // right at the end of this task. 2658 set_has_aborted(); 2659 } 2660 2661 // First drain any available SATB buffers. After this, we will not 2662 // look at SATB buffers before the next invocation of this method. 2663 // If enough completed SATB buffers are queued up, the regular clock 2664 // will abort this task so that it restarts. 2665 drain_satb_buffers(); 2666 // ...then partially drain the local queue and the global stack 2667 drain_local_queue(true); 2668 drain_global_stack(true); 2669 2670 do { 2671 if (!has_aborted() && _curr_region != NULL) { 2672 // This means that we're already holding on to a region. 2673 assert(_finger != NULL, "if region is not NULL, then the finger " 2674 "should not be NULL either"); 2675 2676 // We might have restarted this task after an evacuation pause 2677 // which might have evacuated the region we're holding on to 2678 // underneath our feet. Let's read its limit again to make sure 2679 // that we do not iterate over a region of the heap that 2680 // contains garbage (update_region_limit() will also move 2681 // _finger to the start of the region if it is found empty). 2682 update_region_limit(); 2683 // We will start from _finger not from the start of the region, 2684 // as we might be restarting this task after aborting half-way 2685 // through scanning this region. In this case, _finger points to 2686 // the address where we last found a marked object. If this is a 2687 // fresh region, _finger points to start(). 2688 MemRegion mr = MemRegion(_finger, _region_limit); 2689 2690 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(), 2691 "humongous regions should go around loop once only"); 2692 2693 // Some special cases: 2694 // If the memory region is empty, we can just give up the region. 2695 // If the current region is humongous then we only need to check 2696 // the bitmap for the bit associated with the start of the object, 2697 // scan the object if it's live, and give up the region. 2698 // Otherwise, let's iterate over the bitmap of the part of the region 2699 // that is left. 2700 // If the iteration is successful, give up the region. 2701 if (mr.is_empty()) { 2702 giveup_current_region(); 2703 regular_clock_call(); 2704 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) { 2705 if (_nextMarkBitMap->isMarked(mr.start())) { 2706 // The object is marked - apply the closure 2707 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start()); 2708 bitmap_closure.do_bit(offset); 2709 } 2710 // Even if this task aborted while scanning the humongous object 2711 // we can (and should) give up the current region. 2712 giveup_current_region(); 2713 regular_clock_call(); 2714 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) { 2715 giveup_current_region(); 2716 regular_clock_call(); 2717 } else { 2718 assert(has_aborted(), "currently the only way to do so"); 2719 // The only way to abort the bitmap iteration is to return 2720 // false from the do_bit() method. However, inside the 2721 // do_bit() method we move the _finger to point to the 2722 // object currently being looked at. So, if we bail out, we 2723 // have definitely set _finger to something non-null. 2724 assert(_finger != NULL, "invariant"); 2725 2726 // Region iteration was actually aborted. So now _finger 2727 // points to the address of the object we last scanned. If we 2728 // leave it there, when we restart this task, we will rescan 2729 // the object. It is easy to avoid this. We move the finger by 2730 // enough to point to the next possible object header (the 2731 // bitmap knows by how much we need to move it as it knows its 2732 // granularity). 2733 assert(_finger < _region_limit, "invariant"); 2734 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger); 2735 // Check if bitmap iteration was aborted while scanning the last object 2736 if (new_finger >= _region_limit) { 2737 giveup_current_region(); 2738 } else { 2739 move_finger_to(new_finger); 2740 } 2741 } 2742 } 2743 // At this point we have either completed iterating over the 2744 // region we were holding on to, or we have aborted. 2745 2746 // We then partially drain the local queue and the global stack. 2747 // (Do we really need this?) 2748 drain_local_queue(true); 2749 drain_global_stack(true); 2750 2751 // Read the note on the claim_region() method on why it might 2752 // return NULL with potentially more regions available for 2753 // claiming and why we have to check out_of_regions() to determine 2754 // whether we're done or not. 2755 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 2756 // We are going to try to claim a new region. We should have 2757 // given up on the previous one. 2758 // Separated the asserts so that we know which one fires. 2759 assert(_curr_region == NULL, "invariant"); 2760 assert(_finger == NULL, "invariant"); 2761 assert(_region_limit == NULL, "invariant"); 2762 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 2763 if (claimed_region != NULL) { 2764 // Yes, we managed to claim one 2765 setup_for_region(claimed_region); 2766 assert(_curr_region == claimed_region, "invariant"); 2767 } 2768 // It is important to call the regular clock here. It might take 2769 // a while to claim a region if, for example, we hit a large 2770 // block of empty regions. So we need to call the regular clock 2771 // method once round the loop to make sure it's called 2772 // frequently enough. 2773 regular_clock_call(); 2774 } 2775 2776 if (!has_aborted() && _curr_region == NULL) { 2777 assert(_cm->out_of_regions(), 2778 "at this point we should be out of regions"); 2779 } 2780 } while ( _curr_region != NULL && !has_aborted()); 2781 2782 if (!has_aborted()) { 2783 // We cannot check whether the global stack is empty, since other 2784 // tasks might be pushing objects to it concurrently. 2785 assert(_cm->out_of_regions(), 2786 "at this point we should be out of regions"); 2787 // Try to reduce the number of available SATB buffers so that 2788 // remark has less work to do. 2789 drain_satb_buffers(); 2790 } 2791 2792 // Since we've done everything else, we can now totally drain the 2793 // local queue and global stack. 2794 drain_local_queue(false); 2795 drain_global_stack(false); 2796 2797 // Attempt at work stealing from other task's queues. 2798 if (do_stealing && !has_aborted()) { 2799 // We have not aborted. This means that we have finished all that 2800 // we could. Let's try to do some stealing... 2801 2802 // We cannot check whether the global stack is empty, since other 2803 // tasks might be pushing objects to it concurrently. 2804 assert(_cm->out_of_regions() && _task_queue->size() == 0, 2805 "only way to reach here"); 2806 while (!has_aborted()) { 2807 oop obj; 2808 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) { 2809 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), 2810 "any stolen object should be marked"); 2811 scan_object(obj); 2812 2813 // And since we're towards the end, let's totally drain the 2814 // local queue and global stack. 2815 drain_local_queue(false); 2816 drain_global_stack(false); 2817 } else { 2818 break; 2819 } 2820 } 2821 } 2822 2823 // We still haven't aborted. Now, let's try to get into the 2824 // termination protocol. 2825 if (do_termination && !has_aborted()) { 2826 // We cannot check whether the global stack is empty, since other 2827 // tasks might be concurrently pushing objects on it. 2828 // Separated the asserts so that we know which one fires. 2829 assert(_cm->out_of_regions(), "only way to reach here"); 2830 assert(_task_queue->size() == 0, "only way to reach here"); 2831 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 2832 2833 // The G1CMTask class also extends the TerminatorTerminator class, 2834 // hence its should_exit_termination() method will also decide 2835 // whether to exit the termination protocol or not. 2836 bool finished = (is_serial || 2837 _cm->terminator()->offer_termination(this)); 2838 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 2839 _termination_time_ms += 2840 termination_end_time_ms - _termination_start_time_ms; 2841 2842 if (finished) { 2843 // We're all done. 2844 2845 if (_worker_id == 0) { 2846 // let's allow task 0 to do this 2847 if (concurrent()) { 2848 assert(_cm->concurrent_marking_in_progress(), "invariant"); 2849 // we need to set this to false before the next 2850 // safepoint. This way we ensure that the marking phase 2851 // doesn't observe any more heap expansions. 2852 _cm->clear_concurrent_marking_in_progress(); 2853 } 2854 } 2855 2856 // We can now guarantee that the global stack is empty, since 2857 // all other tasks have finished. We separated the guarantees so 2858 // that, if a condition is false, we can immediately find out 2859 // which one. 2860 guarantee(_cm->out_of_regions(), "only way to reach here"); 2861 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 2862 guarantee(_task_queue->size() == 0, "only way to reach here"); 2863 guarantee(!_cm->has_overflown(), "only way to reach here"); 2864 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); 2865 } else { 2866 // Apparently there's more work to do. Let's abort this task. It 2867 // will restart it and we can hopefully find more things to do. 2868 set_has_aborted(); 2869 } 2870 } 2871 2872 // Mainly for debugging purposes to make sure that a pointer to the 2873 // closure which was statically allocated in this frame doesn't 2874 // escape it by accident. 2875 set_cm_oop_closure(NULL); 2876 double end_time_ms = os::elapsedVTime() * 1000.0; 2877 double elapsed_time_ms = end_time_ms - _start_time_ms; 2878 // Update the step history. 2879 _step_times_ms.add(elapsed_time_ms); 2880 2881 if (has_aborted()) { 2882 // The task was aborted for some reason. 2883 if (_has_timed_out) { 2884 double diff_ms = elapsed_time_ms - _time_target_ms; 2885 // Keep statistics of how well we did with respect to hitting 2886 // our target only if we actually timed out (if we aborted for 2887 // other reasons, then the results might get skewed). 2888 _marking_step_diffs_ms.add(diff_ms); 2889 } 2890 2891 if (_cm->has_overflown()) { 2892 // This is the interesting one. We aborted because a global 2893 // overflow was raised. This means we have to restart the 2894 // marking phase and start iterating over regions. However, in 2895 // order to do this we have to make sure that all tasks stop 2896 // what they are doing and re-initialize in a safe manner. We 2897 // will achieve this with the use of two barrier sync points. 2898 2899 if (!is_serial) { 2900 // We only need to enter the sync barrier if being called 2901 // from a parallel context 2902 _cm->enter_first_sync_barrier(_worker_id); 2903 2904 // When we exit this sync barrier we know that all tasks have 2905 // stopped doing marking work. So, it's now safe to 2906 // re-initialize our data structures. At the end of this method, 2907 // task 0 will clear the global data structures. 2908 } 2909 2910 // We clear the local state of this task... 2911 clear_region_fields(); 2912 2913 if (!is_serial) { 2914 // ...and enter the second barrier. 2915 _cm->enter_second_sync_barrier(_worker_id); 2916 } 2917 // At this point, if we're during the concurrent phase of 2918 // marking, everything has been re-initialized and we're 2919 // ready to restart. 2920 } 2921 } 2922 2923 _claimed = false; 2924 } 2925 2926 G1CMTask::G1CMTask(uint worker_id, 2927 G1ConcurrentMark* cm, 2928 G1CMTaskQueue* task_queue, 2929 G1CMTaskQueueSet* task_queues) 2930 : _g1h(G1CollectedHeap::heap()), 2931 _worker_id(worker_id), _cm(cm), 2932 _claimed(false), 2933 _nextMarkBitMap(NULL), _hash_seed(17), 2934 _task_queue(task_queue), 2935 _task_queues(task_queues), 2936 _cm_oop_closure(NULL) { 2937 guarantee(task_queue != NULL, "invariant"); 2938 guarantee(task_queues != NULL, "invariant"); 2939 2940 _marking_step_diffs_ms.add(0.5); 2941 } 2942 2943 // These are formatting macros that are used below to ensure 2944 // consistent formatting. The *_H_* versions are used to format the 2945 // header for a particular value and they should be kept consistent 2946 // with the corresponding macro. Also note that most of the macros add 2947 // the necessary white space (as a prefix) which makes them a bit 2948 // easier to compose. 2949 2950 // All the output lines are prefixed with this string to be able to 2951 // identify them easily in a large log file. 2952 #define G1PPRL_LINE_PREFIX "###" 2953 2954 #define G1PPRL_ADDR_BASE_FORMAT " " PTR_FORMAT "-" PTR_FORMAT 2955 #ifdef _LP64 2956 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 2957 #else // _LP64 2958 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 2959 #endif // _LP64 2960 2961 // For per-region info 2962 #define G1PPRL_TYPE_FORMAT " %-4s" 2963 #define G1PPRL_TYPE_H_FORMAT " %4s" 2964 #define G1PPRL_BYTE_FORMAT " " SIZE_FORMAT_W(9) 2965 #define G1PPRL_BYTE_H_FORMAT " %9s" 2966 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 2967 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 2968 2969 // For summary info 2970 #define G1PPRL_SUM_ADDR_FORMAT(tag) " " tag ":" G1PPRL_ADDR_BASE_FORMAT 2971 #define G1PPRL_SUM_BYTE_FORMAT(tag) " " tag ": " SIZE_FORMAT 2972 #define G1PPRL_SUM_MB_FORMAT(tag) " " tag ": %1.2f MB" 2973 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag) " / %1.2f %%" 2974 2975 G1PrintRegionLivenessInfoClosure:: 2976 G1PrintRegionLivenessInfoClosure(const char* phase_name) 2977 : _total_used_bytes(0), _total_capacity_bytes(0), 2978 _total_prev_live_bytes(0), _total_next_live_bytes(0), 2979 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) { 2980 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2981 MemRegion g1_reserved = g1h->g1_reserved(); 2982 double now = os::elapsedTime(); 2983 2984 // Print the header of the output. 2985 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 2986 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX" HEAP" 2987 G1PPRL_SUM_ADDR_FORMAT("reserved") 2988 G1PPRL_SUM_BYTE_FORMAT("region-size"), 2989 p2i(g1_reserved.start()), p2i(g1_reserved.end()), 2990 HeapRegion::GrainBytes); 2991 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 2992 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 2993 G1PPRL_TYPE_H_FORMAT 2994 G1PPRL_ADDR_BASE_H_FORMAT 2995 G1PPRL_BYTE_H_FORMAT 2996 G1PPRL_BYTE_H_FORMAT 2997 G1PPRL_BYTE_H_FORMAT 2998 G1PPRL_DOUBLE_H_FORMAT 2999 G1PPRL_BYTE_H_FORMAT 3000 G1PPRL_BYTE_H_FORMAT, 3001 "type", "address-range", 3002 "used", "prev-live", "next-live", "gc-eff", 3003 "remset", "code-roots"); 3004 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3005 G1PPRL_TYPE_H_FORMAT 3006 G1PPRL_ADDR_BASE_H_FORMAT 3007 G1PPRL_BYTE_H_FORMAT 3008 G1PPRL_BYTE_H_FORMAT 3009 G1PPRL_BYTE_H_FORMAT 3010 G1PPRL_DOUBLE_H_FORMAT 3011 G1PPRL_BYTE_H_FORMAT 3012 G1PPRL_BYTE_H_FORMAT, 3013 "", "", 3014 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", 3015 "(bytes)", "(bytes)"); 3016 } 3017 3018 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) { 3019 const char* type = r->get_type_str(); 3020 HeapWord* bottom = r->bottom(); 3021 HeapWord* end = r->end(); 3022 size_t capacity_bytes = r->capacity(); 3023 size_t used_bytes = r->used(); 3024 size_t prev_live_bytes = r->live_bytes(); 3025 size_t next_live_bytes = r->next_live_bytes(); 3026 double gc_eff = r->gc_efficiency(); 3027 size_t remset_bytes = r->rem_set()->mem_size(); 3028 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size(); 3029 3030 _total_used_bytes += used_bytes; 3031 _total_capacity_bytes += capacity_bytes; 3032 _total_prev_live_bytes += prev_live_bytes; 3033 _total_next_live_bytes += next_live_bytes; 3034 _total_remset_bytes += remset_bytes; 3035 _total_strong_code_roots_bytes += strong_code_roots_bytes; 3036 3037 // Print a line for this particular region. 3038 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3039 G1PPRL_TYPE_FORMAT 3040 G1PPRL_ADDR_BASE_FORMAT 3041 G1PPRL_BYTE_FORMAT 3042 G1PPRL_BYTE_FORMAT 3043 G1PPRL_BYTE_FORMAT 3044 G1PPRL_DOUBLE_FORMAT 3045 G1PPRL_BYTE_FORMAT 3046 G1PPRL_BYTE_FORMAT, 3047 type, p2i(bottom), p2i(end), 3048 used_bytes, prev_live_bytes, next_live_bytes, gc_eff, 3049 remset_bytes, strong_code_roots_bytes); 3050 3051 return false; 3052 } 3053 3054 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 3055 // add static memory usages to remembered set sizes 3056 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size(); 3057 // Print the footer of the output. 3058 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX); 3059 log_trace(gc, liveness)(G1PPRL_LINE_PREFIX 3060 " SUMMARY" 3061 G1PPRL_SUM_MB_FORMAT("capacity") 3062 G1PPRL_SUM_MB_PERC_FORMAT("used") 3063 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 3064 G1PPRL_SUM_MB_PERC_FORMAT("next-live") 3065 G1PPRL_SUM_MB_FORMAT("remset") 3066 G1PPRL_SUM_MB_FORMAT("code-roots"), 3067 bytes_to_mb(_total_capacity_bytes), 3068 bytes_to_mb(_total_used_bytes), 3069 perc(_total_used_bytes, _total_capacity_bytes), 3070 bytes_to_mb(_total_prev_live_bytes), 3071 perc(_total_prev_live_bytes, _total_capacity_bytes), 3072 bytes_to_mb(_total_next_live_bytes), 3073 perc(_total_next_live_bytes, _total_capacity_bytes), 3074 bytes_to_mb(_total_remset_bytes), 3075 bytes_to_mb(_total_strong_code_roots_bytes)); 3076 }