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/g1CollectorPolicy.hpp" 32 #include "gc/g1/g1CollectorState.hpp" 33 #include "gc/g1/g1ConcurrentMark.inline.hpp" 34 #include "gc/g1/g1HeapVerifier.hpp" 35 #include "gc/g1/g1OopClosures.inline.hpp" 36 #include "gc/g1/g1CardLiveData.inline.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 61 // Concurrent marking bit map wrapper 62 63 G1CMBitMapRO::G1CMBitMapRO(int shifter) : 64 _bm(), 65 _shifter(shifter) { 66 _bmStartWord = 0; 67 _bmWordSize = 0; 68 } 69 70 HeapWord* G1CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr, 71 const HeapWord* limit) const { 72 // First we must round addr *up* to a possible object boundary. 73 addr = (HeapWord*)align_size_up((intptr_t)addr, 74 HeapWordSize << _shifter); 75 size_t addrOffset = heapWordToOffset(addr); 76 assert(limit != NULL, "limit must not be NULL"); 77 size_t limitOffset = heapWordToOffset(limit); 78 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); 79 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 80 assert(nextAddr >= addr, "get_next_one postcondition"); 81 assert(nextAddr == limit || isMarked(nextAddr), 82 "get_next_one postcondition"); 83 return nextAddr; 84 } 85 86 #ifndef PRODUCT 87 bool G1CMBitMapRO::covers(MemRegion heap_rs) const { 88 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 89 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize, 90 "size inconsistency"); 91 return _bmStartWord == (HeapWord*)(heap_rs.start()) && 92 _bmWordSize == heap_rs.word_size(); 93 } 94 #endif 95 96 void G1CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const { 97 _bm.print_on_error(st, prefix); 98 } 99 100 size_t G1CMBitMap::compute_size(size_t heap_size) { 101 return ReservedSpace::allocation_align_size_up(heap_size / mark_distance()); 102 } 103 104 size_t G1CMBitMap::mark_distance() { 105 return MinObjAlignmentInBytes * BitsPerByte; 106 } 107 108 void G1CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) { 109 _bmStartWord = heap.start(); 110 _bmWordSize = heap.word_size(); 111 112 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start()); 113 _bm.set_size(_bmWordSize >> _shifter); 114 115 storage->set_mapping_changed_listener(&_listener); 116 } 117 118 void 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 _young_list(NULL), _cm(NULL), _scan_in_progress(false), 264 _should_abort(false), _next_survivor(NULL) { } 265 266 void G1CMRootRegions::init(G1CollectedHeap* g1h, G1ConcurrentMark* cm) { 267 _young_list = g1h->young_list(); 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 assert(_next_survivor == NULL, "pre-condition"); 276 _next_survivor = _young_list->first_survivor_region(); 277 _scan_in_progress = (_next_survivor != NULL); 278 _should_abort = false; 279 } 280 281 HeapRegion* G1CMRootRegions::claim_next() { 282 if (_should_abort) { 283 // If someone has set the should_abort flag, we return NULL to 284 // force the caller to bail out of their loop. 285 return NULL; 286 } 287 288 // Currently, only survivors can be root regions. 289 HeapRegion* res = _next_survivor; 290 if (res != NULL) { 291 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 292 // Read it again in case it changed while we were waiting for the lock. 293 res = _next_survivor; 294 if (res != NULL) { 295 if (res == _young_list->last_survivor_region()) { 296 // We just claimed the last survivor so store NULL to indicate 297 // that we're done. 298 _next_survivor = NULL; 299 } else { 300 _next_survivor = res->get_next_young_region(); 301 } 302 } else { 303 // Someone else claimed the last survivor while we were trying 304 // to take the lock so nothing else to do. 305 } 306 } 307 assert(res == NULL || res->is_survivor(), "post-condition"); 308 309 return res; 310 } 311 312 void G1CMRootRegions::notify_scan_done() { 313 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 314 _scan_in_progress = false; 315 RootRegionScan_lock->notify_all(); 316 } 317 318 void G1CMRootRegions::cancel_scan() { 319 notify_scan_done(); 320 } 321 322 void G1CMRootRegions::scan_finished() { 323 assert(scan_in_progress(), "pre-condition"); 324 325 // Currently, only survivors can be root regions. 326 if (!_should_abort) { 327 assert(_next_survivor == NULL, "we should have claimed all survivors"); 328 } 329 _next_survivor = NULL; 330 331 notify_scan_done(); 332 } 333 334 bool G1CMRootRegions::wait_until_scan_finished() { 335 if (!scan_in_progress()) return false; 336 337 { 338 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 339 while (scan_in_progress()) { 340 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 341 } 342 } 343 return true; 344 } 345 346 uint G1ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 347 return MAX2((n_par_threads + 2) / 4, 1U); 348 } 349 350 G1ConcurrentMark::G1ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) : 351 _g1h(g1h), 352 _markBitMap1(), 353 _markBitMap2(), 354 _parallel_marking_threads(0), 355 _max_parallel_marking_threads(0), 356 _sleep_factor(0.0), 357 _marking_task_overhead(1.0), 358 _cleanup_list("Cleanup List"), 359 360 _prevMarkBitMap(&_markBitMap1), 361 _nextMarkBitMap(&_markBitMap2), 362 363 _markStack(this), 364 // _finger set in set_non_marking_state 365 366 _max_worker_id(ParallelGCThreads), 367 // _active_tasks set in set_non_marking_state 368 // _tasks set inside the constructor 369 _task_queues(new G1CMTaskQueueSet((int) _max_worker_id)), 370 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)), 371 372 _has_overflown(false), 373 _concurrent(false), 374 _has_aborted(false), 375 _restart_for_overflow(false), 376 _concurrent_marking_in_progress(false), 377 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 378 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) G1OldTracer()), 379 380 // _verbose_level set below 381 382 _init_times(), 383 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 384 _cleanup_times(), 385 _total_counting_time(0.0), 386 _total_rs_scrub_time(0.0), 387 388 _parallel_workers(NULL), 389 390 _completed_initialization(false) { 391 392 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage); 393 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage); 394 395 // Create & start a ConcurrentMark thread. 396 _cmThread = new ConcurrentMarkThread(this); 397 assert(cmThread() != NULL, "CM Thread should have been created"); 398 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 399 if (_cmThread->osthread() == NULL) { 400 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread"); 401 } 402 403 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 404 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency"); 405 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency"); 406 407 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 408 satb_qs.set_buffer_size(G1SATBBufferSize); 409 410 _root_regions.init(_g1h, this); 411 412 if (ConcGCThreads > ParallelGCThreads) { 413 log_warning(gc)("Can't have more ConcGCThreads (%u) than ParallelGCThreads (%u).", 414 ConcGCThreads, ParallelGCThreads); 415 return; 416 } 417 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 418 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 419 // if both are set 420 _sleep_factor = 0.0; 421 _marking_task_overhead = 1.0; 422 } else if (G1MarkingOverheadPercent > 0) { 423 // We will calculate the number of parallel marking threads based 424 // on a target overhead with respect to the soft real-time goal 425 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 426 double overall_cm_overhead = 427 (double) MaxGCPauseMillis * marking_overhead / 428 (double) GCPauseIntervalMillis; 429 double cpu_ratio = 1.0 / (double) os::processor_count(); 430 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 431 double marking_task_overhead = 432 overall_cm_overhead / marking_thread_num * 433 (double) os::processor_count(); 434 double sleep_factor = 435 (1.0 - marking_task_overhead) / marking_task_overhead; 436 437 FLAG_SET_ERGO(uint, ConcGCThreads, (uint) marking_thread_num); 438 _sleep_factor = sleep_factor; 439 _marking_task_overhead = marking_task_overhead; 440 } else { 441 // Calculate the number of parallel marking threads by scaling 442 // the number of parallel GC threads. 443 uint marking_thread_num = scale_parallel_threads(ParallelGCThreads); 444 FLAG_SET_ERGO(uint, ConcGCThreads, marking_thread_num); 445 _sleep_factor = 0.0; 446 _marking_task_overhead = 1.0; 447 } 448 449 assert(ConcGCThreads > 0, "Should have been set"); 450 _parallel_marking_threads = ConcGCThreads; 451 _max_parallel_marking_threads = _parallel_marking_threads; 452 453 _parallel_workers = new WorkGang("G1 Marker", 454 _max_parallel_marking_threads, false, true); 455 if (_parallel_workers == NULL) { 456 vm_exit_during_initialization("Failed necessary allocation."); 457 } else { 458 _parallel_workers->initialize_workers(); 459 } 460 461 if (FLAG_IS_DEFAULT(MarkStackSize)) { 462 size_t mark_stack_size = 463 MIN2(MarkStackSizeMax, 464 MAX2(MarkStackSize, (size_t) (parallel_marking_threads() * TASKQUEUE_SIZE))); 465 // Verify that the calculated value for MarkStackSize is in range. 466 // It would be nice to use the private utility routine from Arguments. 467 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 468 log_warning(gc)("Invalid value calculated for MarkStackSize (" SIZE_FORMAT "): " 469 "must be between 1 and " SIZE_FORMAT, 470 mark_stack_size, MarkStackSizeMax); 471 return; 472 } 473 FLAG_SET_ERGO(size_t, MarkStackSize, mark_stack_size); 474 } else { 475 // Verify MarkStackSize is in range. 476 if (FLAG_IS_CMDLINE(MarkStackSize)) { 477 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 478 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 479 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT "): " 480 "must be between 1 and " SIZE_FORMAT, 481 MarkStackSize, MarkStackSizeMax); 482 return; 483 } 484 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 485 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 486 log_warning(gc)("Invalid value specified for MarkStackSize (" SIZE_FORMAT ")" 487 " or for MarkStackSizeMax (" SIZE_FORMAT ")", 488 MarkStackSize, MarkStackSizeMax); 489 return; 490 } 491 } 492 } 493 } 494 495 if (!_markStack.allocate(MarkStackSize)) { 496 log_warning(gc)("Failed to allocate CM marking stack"); 497 return; 498 } 499 500 _tasks = NEW_C_HEAP_ARRAY(G1CMTask*, _max_worker_id, mtGC); 501 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC); 502 503 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 504 _active_tasks = _max_worker_id; 505 506 for (uint i = 0; i < _max_worker_id; ++i) { 507 G1CMTaskQueue* task_queue = new G1CMTaskQueue(); 508 task_queue->initialize(); 509 _task_queues->register_queue(i, task_queue); 510 511 _tasks[i] = new G1CMTask(i, this, task_queue, _task_queues); 512 513 _accum_task_vtime[i] = 0.0; 514 } 515 516 // so that the call below can read a sensible value 517 _heap_start = g1h->reserved_region().start(); 518 set_non_marking_state(); 519 _completed_initialization = true; 520 } 521 522 void G1ConcurrentMark::reset() { 523 // Starting values for these two. This should be called in a STW 524 // phase. 525 MemRegion reserved = _g1h->g1_reserved(); 526 _heap_start = reserved.start(); 527 _heap_end = reserved.end(); 528 529 // Separated the asserts so that we know which one fires. 530 assert(_heap_start != NULL, "heap bounds should look ok"); 531 assert(_heap_end != NULL, "heap bounds should look ok"); 532 assert(_heap_start < _heap_end, "heap bounds should look ok"); 533 534 // Reset all the marking data structures and any necessary flags 535 reset_marking_state(); 536 537 // We do reset all of them, since different phases will use 538 // different number of active threads. So, it's easiest to have all 539 // of them ready. 540 for (uint i = 0; i < _max_worker_id; ++i) { 541 _tasks[i]->reset(_nextMarkBitMap); 542 } 543 544 // we need this to make sure that the flag is on during the evac 545 // pause with initial mark piggy-backed 546 set_concurrent_marking_in_progress(); 547 } 548 549 550 void G1ConcurrentMark::reset_marking_state(bool clear_overflow) { 551 _markStack.set_should_expand(); 552 _markStack.setEmpty(); // Also clears the _markStack overflow flag 553 if (clear_overflow) { 554 clear_has_overflown(); 555 } else { 556 assert(has_overflown(), "pre-condition"); 557 } 558 _finger = _heap_start; 559 560 for (uint i = 0; i < _max_worker_id; ++i) { 561 G1CMTaskQueue* queue = _task_queues->queue(i); 562 queue->set_empty(); 563 } 564 } 565 566 void G1ConcurrentMark::set_concurrency(uint active_tasks) { 567 assert(active_tasks <= _max_worker_id, "we should not have more"); 568 569 _active_tasks = active_tasks; 570 // Need to update the three data structures below according to the 571 // number of active threads for this phase. 572 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 573 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 574 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 575 } 576 577 void G1ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 578 set_concurrency(active_tasks); 579 580 _concurrent = concurrent; 581 // We propagate this to all tasks, not just the active ones. 582 for (uint i = 0; i < _max_worker_id; ++i) 583 _tasks[i]->set_concurrent(concurrent); 584 585 if (concurrent) { 586 set_concurrent_marking_in_progress(); 587 } else { 588 // We currently assume that the concurrent flag has been set to 589 // false before we start remark. At this point we should also be 590 // in a STW phase. 591 assert(!concurrent_marking_in_progress(), "invariant"); 592 assert(out_of_regions(), 593 "only way to get here: _finger: " PTR_FORMAT ", _heap_end: " PTR_FORMAT, 594 p2i(_finger), p2i(_heap_end)); 595 } 596 } 597 598 void G1ConcurrentMark::set_non_marking_state() { 599 // We set the global marking state to some default values when we're 600 // not doing marking. 601 reset_marking_state(); 602 _active_tasks = 0; 603 clear_concurrent_marking_in_progress(); 604 } 605 606 G1ConcurrentMark::~G1ConcurrentMark() { 607 // The G1ConcurrentMark instance is never freed. 608 ShouldNotReachHere(); 609 } 610 611 class G1ClearBitMapTask : public AbstractGangTask { 612 // Heap region closure used for clearing the given mark bitmap. 613 class G1ClearBitmapHRClosure : public HeapRegionClosure { 614 private: 615 G1CMBitMap* _bitmap; 616 G1ConcurrentMark* _cm; 617 public: 618 G1ClearBitmapHRClosure(G1CMBitMap* bitmap, G1ConcurrentMark* cm) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap) { 619 } 620 621 virtual bool doHeapRegion(HeapRegion* r) { 622 size_t const chunk_size_in_words = M / HeapWordSize; 623 624 HeapWord* cur = r->bottom(); 625 HeapWord* const end = r->end(); 626 627 while (cur < end) { 628 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end)); 629 _bitmap->clear_range(mr); 630 631 cur += chunk_size_in_words; 632 633 // Abort iteration if after yielding the marking has been aborted. 634 if (_cm != NULL && _cm->do_yield_check() && _cm->has_aborted()) { 635 return true; 636 } 637 // Repeat the asserts from before the start of the closure. We will do them 638 // as asserts here to minimize their overhead on the product. However, we 639 // will have them as guarantees at the beginning / end of the bitmap 640 // clearing to get some checking in the product. 641 assert(_cm == NULL || _cm->cmThread()->during_cycle(), "invariant"); 642 assert(_cm == NULL || !G1CollectedHeap::heap()->collector_state()->mark_in_progress(), "invariant"); 643 } 644 assert(cur == end, "Must have completed iteration over the bitmap for region %u.", r->hrm_index()); 645 646 return false; 647 } 648 }; 649 650 G1ClearBitmapHRClosure _cl; 651 HeapRegionClaimer _hr_claimer; 652 bool _suspendible; // If the task is suspendible, workers must join the STS. 653 654 public: 655 G1ClearBitMapTask(G1CMBitMap* bitmap, G1ConcurrentMark* cm, uint n_workers, bool suspendible) : 656 AbstractGangTask("Parallel Clear Bitmap Task"), 657 _cl(bitmap, suspendible ? cm : NULL), 658 _hr_claimer(n_workers), 659 _suspendible(suspendible) 660 { } 661 662 void work(uint worker_id) { 663 SuspendibleThreadSetJoiner sts_join(_suspendible); 664 G1CollectedHeap::heap()->heap_region_par_iterate(&_cl, worker_id, &_hr_claimer, true); 665 } 666 667 bool is_complete() { 668 return _cl.complete(); 669 } 670 }; 671 672 void G1ConcurrentMark::clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield) { 673 assert(may_yield || SafepointSynchronize::is_at_safepoint(), "Non-yielding bitmap clear only allowed at safepoint."); 674 675 G1ClearBitMapTask task(bitmap, this, workers->active_workers(), may_yield); 676 workers->run_task(&task); 677 guarantee(!may_yield || task.is_complete(), "Must have completed iteration when not yielding."); 678 } 679 680 void G1ConcurrentMark::cleanup_for_next_mark() { 681 // Make sure that the concurrent mark thread looks to still be in 682 // the current cycle. 683 guarantee(cmThread()->during_cycle(), "invariant"); 684 685 // We are finishing up the current cycle by clearing the next 686 // marking bitmap and getting it ready for the next cycle. During 687 // this time no other cycle can start. So, let's make sure that this 688 // is the case. 689 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant"); 690 691 clear_bitmap(_nextMarkBitMap, _parallel_workers, true); 692 693 // Clear the live count data. If the marking has been aborted, the abort() 694 // call already did that. 695 if (!has_aborted()) { 696 clear_live_data(_parallel_workers); 697 DEBUG_ONLY(verify_live_data_clear()); 698 } 699 700 // Repeat the asserts from above. 701 guarantee(cmThread()->during_cycle(), "invariant"); 702 guarantee(!_g1h->collector_state()->mark_in_progress(), "invariant"); 703 } 704 705 void G1ConcurrentMark::clear_prev_bitmap(WorkGang* workers) { 706 assert(SafepointSynchronize::is_at_safepoint(), "Should only clear the entire prev bitmap at a safepoint."); 707 clear_bitmap((G1CMBitMap*)_prevMarkBitMap, workers, false); 708 } 709 710 class CheckBitmapClearHRClosure : public HeapRegionClosure { 711 G1CMBitMap* _bitmap; 712 bool _error; 713 public: 714 CheckBitmapClearHRClosure(G1CMBitMap* bitmap) : _bitmap(bitmap) { 715 } 716 717 virtual bool doHeapRegion(HeapRegion* r) { 718 // This closure can be called concurrently to the mutator, so we must make sure 719 // that the result of the getNextMarkedWordAddress() call is compared to the 720 // value passed to it as limit to detect any found bits. 721 // end never changes in G1. 722 HeapWord* end = r->end(); 723 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end; 724 } 725 }; 726 727 bool G1ConcurrentMark::nextMarkBitmapIsClear() { 728 CheckBitmapClearHRClosure cl(_nextMarkBitMap); 729 _g1h->heap_region_iterate(&cl); 730 return cl.complete(); 731 } 732 733 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 734 public: 735 bool doHeapRegion(HeapRegion* r) { 736 r->note_start_of_marking(); 737 return false; 738 } 739 }; 740 741 void G1ConcurrentMark::checkpointRootsInitialPre() { 742 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 743 G1CollectorPolicy* g1p = g1h->g1_policy(); 744 745 _has_aborted = false; 746 747 // Initialize marking structures. This has to be done in a STW phase. 748 reset(); 749 750 // For each region note start of marking. 751 NoteStartOfMarkHRClosure startcl; 752 g1h->heap_region_iterate(&startcl); 753 } 754 755 756 void G1ConcurrentMark::checkpointRootsInitialPost() { 757 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 758 759 // Start Concurrent Marking weak-reference discovery. 760 ReferenceProcessor* rp = g1h->ref_processor_cm(); 761 // enable ("weak") refs discovery 762 rp->enable_discovery(); 763 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 764 765 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 766 // This is the start of the marking cycle, we're expected all 767 // threads to have SATB queues with active set to false. 768 satb_mq_set.set_active_all_threads(true, /* new active value */ 769 false /* expected_active */); 770 771 _root_regions.prepare_for_scan(); 772 773 // update_g1_committed() will be called at the end of an evac pause 774 // when marking is on. So, it's also called at the end of the 775 // initial-mark pause to update the heap end, if the heap expands 776 // during it. No need to call it here. 777 } 778 779 /* 780 * Notice that in the next two methods, we actually leave the STS 781 * during the barrier sync and join it immediately afterwards. If we 782 * do not do this, the following deadlock can occur: one thread could 783 * be in the barrier sync code, waiting for the other thread to also 784 * sync up, whereas another one could be trying to yield, while also 785 * waiting for the other threads to sync up too. 786 * 787 * Note, however, that this code is also used during remark and in 788 * this case we should not attempt to leave / enter the STS, otherwise 789 * we'll either hit an assert (debug / fastdebug) or deadlock 790 * (product). So we should only leave / enter the STS if we are 791 * operating concurrently. 792 * 793 * Because the thread that does the sync barrier has left the STS, it 794 * is possible to be suspended for a Full GC or an evacuation pause 795 * could occur. This is actually safe, since the entering the sync 796 * barrier is one of the last things do_marking_step() does, and it 797 * doesn't manipulate any data structures afterwards. 798 */ 799 800 void G1ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 801 bool barrier_aborted; 802 { 803 SuspendibleThreadSetLeaver sts_leave(concurrent()); 804 barrier_aborted = !_first_overflow_barrier_sync.enter(); 805 } 806 807 // at this point everyone should have synced up and not be doing any 808 // more work 809 810 if (barrier_aborted) { 811 // If the barrier aborted we ignore the overflow condition and 812 // just abort the whole marking phase as quickly as possible. 813 return; 814 } 815 816 // If we're executing the concurrent phase of marking, reset the marking 817 // state; otherwise the marking state is reset after reference processing, 818 // during the remark pause. 819 // If we reset here as a result of an overflow during the remark we will 820 // see assertion failures from any subsequent set_concurrency_and_phase() 821 // calls. 822 if (concurrent()) { 823 // let the task associated with with worker 0 do this 824 if (worker_id == 0) { 825 // task 0 is responsible for clearing the global data structures 826 // We should be here because of an overflow. During STW we should 827 // not clear the overflow flag since we rely on it being true when 828 // we exit this method to abort the pause and restart concurrent 829 // marking. 830 reset_marking_state(true /* clear_overflow */); 831 832 log_info(gc, marking)("Concurrent Mark reset for overflow"); 833 } 834 } 835 836 // after this, each task should reset its own data structures then 837 // then go into the second barrier 838 } 839 840 void G1ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 841 SuspendibleThreadSetLeaver sts_leave(concurrent()); 842 _second_overflow_barrier_sync.enter(); 843 844 // at this point everything should be re-initialized and ready to go 845 } 846 847 class G1CMConcurrentMarkingTask: public AbstractGangTask { 848 private: 849 G1ConcurrentMark* _cm; 850 ConcurrentMarkThread* _cmt; 851 852 public: 853 void work(uint worker_id) { 854 assert(Thread::current()->is_ConcurrentGC_thread(), 855 "this should only be done by a conc GC thread"); 856 ResourceMark rm; 857 858 double start_vtime = os::elapsedVTime(); 859 860 { 861 SuspendibleThreadSetJoiner sts_join; 862 863 assert(worker_id < _cm->active_tasks(), "invariant"); 864 G1CMTask* the_task = _cm->task(worker_id); 865 the_task->record_start_time(); 866 if (!_cm->has_aborted()) { 867 do { 868 double start_vtime_sec = os::elapsedVTime(); 869 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 870 871 the_task->do_marking_step(mark_step_duration_ms, 872 true /* do_termination */, 873 false /* is_serial*/); 874 875 double end_vtime_sec = os::elapsedVTime(); 876 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 877 _cm->clear_has_overflown(); 878 879 _cm->do_yield_check(); 880 881 jlong sleep_time_ms; 882 if (!_cm->has_aborted() && the_task->has_aborted()) { 883 sleep_time_ms = 884 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 885 { 886 SuspendibleThreadSetLeaver sts_leave; 887 os::sleep(Thread::current(), sleep_time_ms, false); 888 } 889 } 890 } while (!_cm->has_aborted() && the_task->has_aborted()); 891 } 892 the_task->record_end_time(); 893 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 894 } 895 896 double end_vtime = os::elapsedVTime(); 897 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 898 } 899 900 G1CMConcurrentMarkingTask(G1ConcurrentMark* cm, 901 ConcurrentMarkThread* cmt) : 902 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 903 904 ~G1CMConcurrentMarkingTask() { } 905 }; 906 907 // Calculates the number of active workers for a concurrent 908 // phase. 909 uint G1ConcurrentMark::calc_parallel_marking_threads() { 910 uint n_conc_workers = 0; 911 if (!UseDynamicNumberOfGCThreads || 912 (!FLAG_IS_DEFAULT(ConcGCThreads) && 913 !ForceDynamicNumberOfGCThreads)) { 914 n_conc_workers = max_parallel_marking_threads(); 915 } else { 916 n_conc_workers = 917 AdaptiveSizePolicy::calc_default_active_workers( 918 max_parallel_marking_threads(), 919 1, /* Minimum workers */ 920 parallel_marking_threads(), 921 Threads::number_of_non_daemon_threads()); 922 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 923 // that scaling has already gone into "_max_parallel_marking_threads". 924 } 925 assert(n_conc_workers > 0, "Always need at least 1"); 926 return n_conc_workers; 927 } 928 929 void G1ConcurrentMark::scanRootRegion(HeapRegion* hr) { 930 // Currently, only survivors can be root regions. 931 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 932 G1RootRegionScanClosure cl(_g1h, this); 933 934 const uintx interval = PrefetchScanIntervalInBytes; 935 HeapWord* curr = hr->bottom(); 936 const HeapWord* end = hr->top(); 937 while (curr < end) { 938 Prefetch::read(curr, interval); 939 oop obj = oop(curr); 940 int size = obj->oop_iterate_size(&cl); 941 assert(size == obj->size(), "sanity"); 942 curr += size; 943 } 944 } 945 946 class G1CMRootRegionScanTask : public AbstractGangTask { 947 private: 948 G1ConcurrentMark* _cm; 949 950 public: 951 G1CMRootRegionScanTask(G1ConcurrentMark* cm) : 952 AbstractGangTask("Root Region Scan"), _cm(cm) { } 953 954 void work(uint worker_id) { 955 assert(Thread::current()->is_ConcurrentGC_thread(), 956 "this should only be done by a conc GC thread"); 957 958 G1CMRootRegions* root_regions = _cm->root_regions(); 959 HeapRegion* hr = root_regions->claim_next(); 960 while (hr != NULL) { 961 _cm->scanRootRegion(hr); 962 hr = root_regions->claim_next(); 963 } 964 } 965 }; 966 967 void G1ConcurrentMark::scan_root_regions() { 968 // scan_in_progress() will have been set to true only if there was 969 // at least one root region to scan. So, if it's false, we 970 // should not attempt to do any further work. 971 if (root_regions()->scan_in_progress()) { 972 assert(!has_aborted(), "Aborting before root region scanning is finished not supported."); 973 974 _parallel_marking_threads = calc_parallel_marking_threads(); 975 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 976 "Maximum number of marking threads exceeded"); 977 uint active_workers = MAX2(1U, parallel_marking_threads()); 978 979 G1CMRootRegionScanTask task(this); 980 _parallel_workers->set_active_workers(active_workers); 981 _parallel_workers->run_task(&task); 982 983 // It's possible that has_aborted() is true here without actually 984 // aborting the survivor scan earlier. This is OK as it's 985 // mainly used for sanity checking. 986 root_regions()->scan_finished(); 987 } 988 } 989 990 void G1ConcurrentMark::concurrent_cycle_start() { 991 _gc_timer_cm->register_gc_start(); 992 993 _gc_tracer_cm->report_gc_start(GCCause::_no_gc /* first parameter is not used */, _gc_timer_cm->gc_start()); 994 995 _g1h->trace_heap_before_gc(_gc_tracer_cm); 996 } 997 998 void G1ConcurrentMark::concurrent_cycle_end() { 999 _g1h->trace_heap_after_gc(_gc_tracer_cm); 1000 1001 if (has_aborted()) { 1002 _gc_tracer_cm->report_concurrent_mode_failure(); 1003 } 1004 1005 _gc_timer_cm->register_gc_end(); 1006 1007 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1008 } 1009 1010 void G1ConcurrentMark::mark_from_roots() { 1011 // we might be tempted to assert that: 1012 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1013 // "inconsistent argument?"); 1014 // However that wouldn't be right, because it's possible that 1015 // a safepoint is indeed in progress as a younger generation 1016 // stop-the-world GC happens even as we mark in this generation. 1017 1018 _restart_for_overflow = false; 1019 1020 // _g1h has _n_par_threads 1021 _parallel_marking_threads = calc_parallel_marking_threads(); 1022 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1023 "Maximum number of marking threads exceeded"); 1024 1025 uint active_workers = MAX2(1U, parallel_marking_threads()); 1026 assert(active_workers > 0, "Should have been set"); 1027 1028 // Parallel task terminator is set in "set_concurrency_and_phase()" 1029 set_concurrency_and_phase(active_workers, true /* concurrent */); 1030 1031 G1CMConcurrentMarkingTask markingTask(this, cmThread()); 1032 _parallel_workers->set_active_workers(active_workers); 1033 _parallel_workers->run_task(&markingTask); 1034 print_stats(); 1035 } 1036 1037 void G1ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1038 // world is stopped at this checkpoint 1039 assert(SafepointSynchronize::is_at_safepoint(), 1040 "world should be stopped"); 1041 1042 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1043 1044 // If a full collection has happened, we shouldn't do this. 1045 if (has_aborted()) { 1046 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1047 return; 1048 } 1049 1050 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1051 1052 if (VerifyDuringGC) { 1053 HandleMark hm; // handle scope 1054 g1h->prepare_for_verify(); 1055 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1056 } 1057 g1h->verifier()->check_bitmaps("Remark Start"); 1058 1059 G1CollectorPolicy* g1p = g1h->g1_policy(); 1060 g1p->record_concurrent_mark_remark_start(); 1061 1062 double start = os::elapsedTime(); 1063 1064 checkpointRootsFinalWork(); 1065 1066 double mark_work_end = os::elapsedTime(); 1067 1068 weakRefsWork(clear_all_soft_refs); 1069 1070 if (has_overflown()) { 1071 // Oops. We overflowed. Restart concurrent marking. 1072 _restart_for_overflow = true; 1073 1074 // Verify the heap w.r.t. the previous marking bitmap. 1075 if (VerifyDuringGC) { 1076 HandleMark hm; // handle scope 1077 g1h->prepare_for_verify(); 1078 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (overflow)"); 1079 } 1080 1081 // Clear the marking state because we will be restarting 1082 // marking due to overflowing the global mark stack. 1083 reset_marking_state(); 1084 } else { 1085 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1086 // We're done with marking. 1087 // This is the end of the marking cycle, we're expected all 1088 // threads to have SATB queues with active set to true. 1089 satb_mq_set.set_active_all_threads(false, /* new active value */ 1090 true /* expected_active */); 1091 1092 if (VerifyDuringGC) { 1093 HandleMark hm; // handle scope 1094 g1h->prepare_for_verify(); 1095 Universe::verify(VerifyOption_G1UseNextMarking, "During GC (after)"); 1096 } 1097 g1h->verifier()->check_bitmaps("Remark End"); 1098 assert(!restart_for_overflow(), "sanity"); 1099 // Completely reset the marking state since marking completed 1100 set_non_marking_state(); 1101 } 1102 1103 // Expand the marking stack, if we have to and if we can. 1104 if (_markStack.should_expand()) { 1105 _markStack.expand(); 1106 } 1107 1108 // Statistics 1109 double now = os::elapsedTime(); 1110 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1111 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1112 _remark_times.add((now - start) * 1000.0); 1113 1114 g1p->record_concurrent_mark_remark_end(); 1115 1116 G1CMIsAliveClosure is_alive(g1h); 1117 _gc_tracer_cm->report_object_count_after_gc(&is_alive); 1118 } 1119 1120 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1121 G1CollectedHeap* _g1; 1122 size_t _freed_bytes; 1123 FreeRegionList* _local_cleanup_list; 1124 uint _old_regions_removed; 1125 uint _humongous_regions_removed; 1126 HRRSCleanupTask* _hrrs_cleanup_task; 1127 1128 public: 1129 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1130 FreeRegionList* local_cleanup_list, 1131 HRRSCleanupTask* hrrs_cleanup_task) : 1132 _g1(g1), 1133 _freed_bytes(0), 1134 _local_cleanup_list(local_cleanup_list), 1135 _old_regions_removed(0), 1136 _humongous_regions_removed(0), 1137 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1138 1139 size_t freed_bytes() { return _freed_bytes; } 1140 const uint old_regions_removed() { return _old_regions_removed; } 1141 const uint humongous_regions_removed() { return _humongous_regions_removed; } 1142 1143 bool doHeapRegion(HeapRegion *hr) { 1144 if (hr->is_archive()) { 1145 return false; 1146 } 1147 // We use a claim value of zero here because all regions 1148 // were claimed with value 1 in the FinalCount task. 1149 _g1->reset_gc_time_stamps(hr); 1150 hr->note_end_of_marking(); 1151 1152 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 1153 _freed_bytes += hr->used(); 1154 hr->set_containing_set(NULL); 1155 if (hr->is_humongous()) { 1156 _humongous_regions_removed++; 1157 _g1->free_humongous_region(hr, _local_cleanup_list, true); 1158 } else { 1159 _old_regions_removed++; 1160 _g1->free_region(hr, _local_cleanup_list, true); 1161 } 1162 } else { 1163 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task); 1164 } 1165 1166 return false; 1167 } 1168 }; 1169 1170 class G1ParNoteEndTask: public AbstractGangTask { 1171 friend class G1NoteEndOfConcMarkClosure; 1172 1173 protected: 1174 G1CollectedHeap* _g1h; 1175 FreeRegionList* _cleanup_list; 1176 HeapRegionClaimer _hrclaimer; 1177 1178 public: 1179 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) : 1180 AbstractGangTask("G1 note end"), _g1h(g1h), _cleanup_list(cleanup_list), _hrclaimer(n_workers) { 1181 } 1182 1183 void work(uint worker_id) { 1184 FreeRegionList local_cleanup_list("Local Cleanup List"); 1185 HRRSCleanupTask hrrs_cleanup_task; 1186 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list, 1187 &hrrs_cleanup_task); 1188 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer); 1189 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1190 1191 // Now update the lists 1192 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed()); 1193 { 1194 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1195 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes()); 1196 1197 // If we iterate over the global cleanup list at the end of 1198 // cleanup to do this printing we will not guarantee to only 1199 // generate output for the newly-reclaimed regions (the list 1200 // might not be empty at the beginning of cleanup; we might 1201 // still be working on its previous contents). So we do the 1202 // printing here, before we append the new regions to the global 1203 // cleanup list. 1204 1205 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1206 if (hr_printer->is_active()) { 1207 FreeRegionListIterator iter(&local_cleanup_list); 1208 while (iter.more_available()) { 1209 HeapRegion* hr = iter.get_next(); 1210 hr_printer->cleanup(hr); 1211 } 1212 } 1213 1214 _cleanup_list->add_ordered(&local_cleanup_list); 1215 assert(local_cleanup_list.is_empty(), "post-condition"); 1216 1217 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1218 } 1219 } 1220 }; 1221 1222 void G1ConcurrentMark::cleanup() { 1223 // world is stopped at this checkpoint 1224 assert(SafepointSynchronize::is_at_safepoint(), 1225 "world should be stopped"); 1226 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1227 1228 // If a full collection has happened, we shouldn't do this. 1229 if (has_aborted()) { 1230 g1h->collector_state()->set_mark_in_progress(false); // So bitmap clearing isn't confused 1231 return; 1232 } 1233 1234 g1h->verifier()->verify_region_sets_optional(); 1235 1236 if (VerifyDuringGC) { 1237 HandleMark hm; // handle scope 1238 g1h->prepare_for_verify(); 1239 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (before)"); 1240 } 1241 g1h->verifier()->check_bitmaps("Cleanup Start"); 1242 1243 G1CollectorPolicy* g1p = g1h->g1_policy(); 1244 g1p->record_concurrent_mark_cleanup_start(); 1245 1246 double start = os::elapsedTime(); 1247 1248 HeapRegionRemSet::reset_for_cleanup_tasks(); 1249 1250 { 1251 GCTraceTime(Debug, gc)("Finalize Live Data"); 1252 finalize_live_data(); 1253 } 1254 1255 if (VerifyDuringGC) { 1256 GCTraceTime(Debug, gc)("Verify Live Data"); 1257 verify_live_data(); 1258 } 1259 1260 g1h->collector_state()->set_mark_in_progress(false); 1261 1262 double count_end = os::elapsedTime(); 1263 double this_final_counting_time = (count_end - start); 1264 _total_counting_time += this_final_counting_time; 1265 1266 if (log_is_enabled(Trace, gc, liveness)) { 1267 G1PrintRegionLivenessInfoClosure cl("Post-Marking"); 1268 _g1h->heap_region_iterate(&cl); 1269 } 1270 1271 // Install newly created mark bitMap as "prev". 1272 swapMarkBitMaps(); 1273 1274 g1h->reset_gc_time_stamp(); 1275 1276 uint n_workers = _g1h->workers()->active_workers(); 1277 1278 // Note end of marking in all heap regions. 1279 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers); 1280 g1h->workers()->run_task(&g1_par_note_end_task); 1281 g1h->check_gc_time_stamps(); 1282 1283 if (!cleanup_list_is_empty()) { 1284 // The cleanup list is not empty, so we'll have to process it 1285 // concurrently. Notify anyone else that might be wanting free 1286 // regions that there will be more free regions coming soon. 1287 g1h->set_free_regions_coming(); 1288 } 1289 1290 // call below, since it affects the metric by which we sort the heap 1291 // regions. 1292 if (G1ScrubRemSets) { 1293 double rs_scrub_start = os::elapsedTime(); 1294 g1h->scrub_rem_set(); 1295 _total_rs_scrub_time += (os::elapsedTime() - rs_scrub_start); 1296 } 1297 1298 // this will also free any regions totally full of garbage objects, 1299 // and sort the regions. 1300 g1h->g1_policy()->record_concurrent_mark_cleanup_end(); 1301 1302 // Statistics. 1303 double end = os::elapsedTime(); 1304 _cleanup_times.add((end - start) * 1000.0); 1305 1306 // Clean up will have freed any regions completely full of garbage. 1307 // Update the soft reference policy with the new heap occupancy. 1308 Universe::update_heap_info_at_gc(); 1309 1310 if (VerifyDuringGC) { 1311 HandleMark hm; // handle scope 1312 g1h->prepare_for_verify(); 1313 Universe::verify(VerifyOption_G1UsePrevMarking, "During GC (after)"); 1314 } 1315 1316 g1h->verifier()->check_bitmaps("Cleanup End"); 1317 1318 g1h->verifier()->verify_region_sets_optional(); 1319 1320 // We need to make this be a "collection" so any collection pause that 1321 // races with it goes around and waits for completeCleanup to finish. 1322 g1h->increment_total_collections(); 1323 1324 // Clean out dead classes and update Metaspace sizes. 1325 if (ClassUnloadingWithConcurrentMark) { 1326 ClassLoaderDataGraph::purge(); 1327 } 1328 MetaspaceGC::compute_new_size(); 1329 1330 // We reclaimed old regions so we should calculate the sizes to make 1331 // sure we update the old gen/space data. 1332 g1h->g1mm()->update_sizes(); 1333 g1h->allocation_context_stats().update_after_mark(); 1334 } 1335 1336 void G1ConcurrentMark::complete_cleanup() { 1337 if (has_aborted()) return; 1338 1339 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1340 1341 _cleanup_list.verify_optional(); 1342 FreeRegionList tmp_free_list("Tmp Free List"); 1343 1344 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1345 "cleanup list has %u entries", 1346 _cleanup_list.length()); 1347 1348 // No one else should be accessing the _cleanup_list at this point, 1349 // so it is not necessary to take any locks 1350 while (!_cleanup_list.is_empty()) { 1351 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */); 1352 assert(hr != NULL, "Got NULL from a non-empty list"); 1353 hr->par_clear(); 1354 tmp_free_list.add_ordered(hr); 1355 1356 // Instead of adding one region at a time to the secondary_free_list, 1357 // we accumulate them in the local list and move them a few at a 1358 // time. This also cuts down on the number of notify_all() calls 1359 // we do during this process. We'll also append the local list when 1360 // _cleanup_list is empty (which means we just removed the last 1361 // region from the _cleanup_list). 1362 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 1363 _cleanup_list.is_empty()) { 1364 log_develop_trace(gc, freelist)("G1ConcRegionFreeing [complete cleanup] : " 1365 "appending %u entries to the secondary_free_list, " 1366 "cleanup list still has %u entries", 1367 tmp_free_list.length(), 1368 _cleanup_list.length()); 1369 1370 { 1371 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 1372 g1h->secondary_free_list_add(&tmp_free_list); 1373 SecondaryFreeList_lock->notify_all(); 1374 } 1375 #ifndef PRODUCT 1376 if (G1StressConcRegionFreeing) { 1377 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 1378 os::sleep(Thread::current(), (jlong) 1, false); 1379 } 1380 } 1381 #endif 1382 } 1383 } 1384 assert(tmp_free_list.is_empty(), "post-condition"); 1385 } 1386 1387 // Supporting Object and Oop closures for reference discovery 1388 // and processing in during marking 1389 1390 bool G1CMIsAliveClosure::do_object_b(oop obj) { 1391 HeapWord* addr = (HeapWord*)obj; 1392 return addr != NULL && 1393 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 1394 } 1395 1396 // 'Keep Alive' oop closure used by both serial parallel reference processing. 1397 // Uses the G1CMTask associated with a worker thread (for serial reference 1398 // processing the G1CMTask for worker 0 is used) to preserve (mark) and 1399 // trace referent objects. 1400 // 1401 // Using the G1CMTask and embedded local queues avoids having the worker 1402 // threads operating on the global mark stack. This reduces the risk 1403 // of overflowing the stack - which we would rather avoid at this late 1404 // state. Also using the tasks' local queues removes the potential 1405 // of the workers interfering with each other that could occur if 1406 // operating on the global stack. 1407 1408 class G1CMKeepAliveAndDrainClosure: public OopClosure { 1409 G1ConcurrentMark* _cm; 1410 G1CMTask* _task; 1411 int _ref_counter_limit; 1412 int _ref_counter; 1413 bool _is_serial; 1414 public: 1415 G1CMKeepAliveAndDrainClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1416 _cm(cm), _task(task), _is_serial(is_serial), 1417 _ref_counter_limit(G1RefProcDrainInterval) { 1418 assert(_ref_counter_limit > 0, "sanity"); 1419 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1420 _ref_counter = _ref_counter_limit; 1421 } 1422 1423 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 1424 virtual void do_oop( oop* p) { do_oop_work(p); } 1425 1426 template <class T> void do_oop_work(T* p) { 1427 if (!_cm->has_overflown()) { 1428 oop obj = oopDesc::load_decode_heap_oop(p); 1429 _task->deal_with_reference(obj); 1430 _ref_counter--; 1431 1432 if (_ref_counter == 0) { 1433 // We have dealt with _ref_counter_limit references, pushing them 1434 // and objects reachable from them on to the local stack (and 1435 // possibly the global stack). Call G1CMTask::do_marking_step() to 1436 // process these entries. 1437 // 1438 // We call G1CMTask::do_marking_step() in a loop, which we'll exit if 1439 // there's nothing more to do (i.e. we're done with the entries that 1440 // were pushed as a result of the G1CMTask::deal_with_reference() calls 1441 // above) or we overflow. 1442 // 1443 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1444 // flag while there may still be some work to do. (See the comment at 1445 // the beginning of G1CMTask::do_marking_step() for those conditions - 1446 // one of which is reaching the specified time target.) It is only 1447 // when G1CMTask::do_marking_step() returns without setting the 1448 // has_aborted() flag that the marking step has completed. 1449 do { 1450 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1451 _task->do_marking_step(mark_step_duration_ms, 1452 false /* do_termination */, 1453 _is_serial); 1454 } while (_task->has_aborted() && !_cm->has_overflown()); 1455 _ref_counter = _ref_counter_limit; 1456 } 1457 } 1458 } 1459 }; 1460 1461 // 'Drain' oop closure used by both serial and parallel reference processing. 1462 // Uses the G1CMTask associated with a given worker thread (for serial 1463 // reference processing the G1CMtask for worker 0 is used). Calls the 1464 // do_marking_step routine, with an unbelievably large timeout value, 1465 // to drain the marking data structures of the remaining entries 1466 // added by the 'keep alive' oop closure above. 1467 1468 class G1CMDrainMarkingStackClosure: public VoidClosure { 1469 G1ConcurrentMark* _cm; 1470 G1CMTask* _task; 1471 bool _is_serial; 1472 public: 1473 G1CMDrainMarkingStackClosure(G1ConcurrentMark* cm, G1CMTask* task, bool is_serial) : 1474 _cm(cm), _task(task), _is_serial(is_serial) { 1475 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 1476 } 1477 1478 void do_void() { 1479 do { 1480 // We call G1CMTask::do_marking_step() to completely drain the local 1481 // and global marking stacks of entries pushed by the 'keep alive' 1482 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 1483 // 1484 // G1CMTask::do_marking_step() is called in a loop, which we'll exit 1485 // if there's nothing more to do (i.e. we've completely drained the 1486 // entries that were pushed as a a result of applying the 'keep alive' 1487 // closure to the entries on the discovered ref lists) or we overflow 1488 // the global marking stack. 1489 // 1490 // Note: G1CMTask::do_marking_step() can set the G1CMTask::has_aborted() 1491 // flag while there may still be some work to do. (See the comment at 1492 // the beginning of G1CMTask::do_marking_step() for those conditions - 1493 // one of which is reaching the specified time target.) It is only 1494 // when G1CMTask::do_marking_step() returns without setting the 1495 // has_aborted() flag that the marking step has completed. 1496 1497 _task->do_marking_step(1000000000.0 /* something very large */, 1498 true /* do_termination */, 1499 _is_serial); 1500 } while (_task->has_aborted() && !_cm->has_overflown()); 1501 } 1502 }; 1503 1504 // Implementation of AbstractRefProcTaskExecutor for parallel 1505 // reference processing at the end of G1 concurrent marking 1506 1507 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 1508 private: 1509 G1CollectedHeap* _g1h; 1510 G1ConcurrentMark* _cm; 1511 WorkGang* _workers; 1512 uint _active_workers; 1513 1514 public: 1515 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 1516 G1ConcurrentMark* cm, 1517 WorkGang* workers, 1518 uint n_workers) : 1519 _g1h(g1h), _cm(cm), 1520 _workers(workers), _active_workers(n_workers) { } 1521 1522 // Executes the given task using concurrent marking worker threads. 1523 virtual void execute(ProcessTask& task); 1524 virtual void execute(EnqueueTask& task); 1525 }; 1526 1527 class G1CMRefProcTaskProxy: public AbstractGangTask { 1528 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 1529 ProcessTask& _proc_task; 1530 G1CollectedHeap* _g1h; 1531 G1ConcurrentMark* _cm; 1532 1533 public: 1534 G1CMRefProcTaskProxy(ProcessTask& proc_task, 1535 G1CollectedHeap* g1h, 1536 G1ConcurrentMark* cm) : 1537 AbstractGangTask("Process reference objects in parallel"), 1538 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 1539 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 1540 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 1541 } 1542 1543 virtual void work(uint worker_id) { 1544 ResourceMark rm; 1545 HandleMark hm; 1546 G1CMTask* task = _cm->task(worker_id); 1547 G1CMIsAliveClosure g1_is_alive(_g1h); 1548 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 1549 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 1550 1551 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 1552 } 1553 }; 1554 1555 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 1556 assert(_workers != NULL, "Need parallel worker threads."); 1557 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1558 1559 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 1560 1561 // We need to reset the concurrency level before each 1562 // proxy task execution, so that the termination protocol 1563 // and overflow handling in G1CMTask::do_marking_step() knows 1564 // how many workers to wait for. 1565 _cm->set_concurrency(_active_workers); 1566 _workers->run_task(&proc_task_proxy); 1567 } 1568 1569 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 1570 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 1571 EnqueueTask& _enq_task; 1572 1573 public: 1574 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 1575 AbstractGangTask("Enqueue reference objects in parallel"), 1576 _enq_task(enq_task) { } 1577 1578 virtual void work(uint worker_id) { 1579 _enq_task.work(worker_id); 1580 } 1581 }; 1582 1583 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 1584 assert(_workers != NULL, "Need parallel worker threads."); 1585 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 1586 1587 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 1588 1589 // Not strictly necessary but... 1590 // 1591 // We need to reset the concurrency level before each 1592 // proxy task execution, so that the termination protocol 1593 // and overflow handling in G1CMTask::do_marking_step() knows 1594 // how many workers to wait for. 1595 _cm->set_concurrency(_active_workers); 1596 _workers->run_task(&enq_task_proxy); 1597 } 1598 1599 void G1ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) { 1600 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes); 1601 } 1602 1603 void G1ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 1604 if (has_overflown()) { 1605 // Skip processing the discovered references if we have 1606 // overflown the global marking stack. Reference objects 1607 // only get discovered once so it is OK to not 1608 // de-populate the discovered reference lists. We could have, 1609 // but the only benefit would be that, when marking restarts, 1610 // less reference objects are discovered. 1611 return; 1612 } 1613 1614 ResourceMark rm; 1615 HandleMark hm; 1616 1617 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1618 1619 // Is alive closure. 1620 G1CMIsAliveClosure g1_is_alive(g1h); 1621 1622 // Inner scope to exclude the cleaning of the string and symbol 1623 // tables from the displayed time. 1624 { 1625 GCTraceTime(Debug, gc, phases) trace("Reference Processing", _gc_timer_cm); 1626 1627 ReferenceProcessor* rp = g1h->ref_processor_cm(); 1628 1629 // See the comment in G1CollectedHeap::ref_processing_init() 1630 // about how reference processing currently works in G1. 1631 1632 // Set the soft reference policy 1633 rp->setup_policy(clear_all_soft_refs); 1634 assert(_markStack.isEmpty(), "mark stack should be empty"); 1635 1636 // Instances of the 'Keep Alive' and 'Complete GC' closures used 1637 // in serial reference processing. Note these closures are also 1638 // used for serially processing (by the the current thread) the 1639 // JNI references during parallel reference processing. 1640 // 1641 // These closures do not need to synchronize with the worker 1642 // threads involved in parallel reference processing as these 1643 // instances are executed serially by the current thread (e.g. 1644 // reference processing is not multi-threaded and is thus 1645 // performed by the current thread instead of a gang worker). 1646 // 1647 // The gang tasks involved in parallel reference processing create 1648 // their own instances of these closures, which do their own 1649 // synchronization among themselves. 1650 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 1651 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 1652 1653 // We need at least one active thread. If reference processing 1654 // is not multi-threaded we use the current (VMThread) thread, 1655 // otherwise we use the work gang from the G1CollectedHeap and 1656 // we utilize all the worker threads we can. 1657 bool processing_is_mt = rp->processing_is_mt(); 1658 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U); 1659 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U); 1660 1661 // Parallel processing task executor. 1662 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 1663 g1h->workers(), active_workers); 1664 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 1665 1666 // Set the concurrency level. The phase was already set prior to 1667 // executing the remark task. 1668 set_concurrency(active_workers); 1669 1670 // Set the degree of MT processing here. If the discovery was done MT, 1671 // the number of threads involved during discovery could differ from 1672 // the number of active workers. This is OK as long as the discovered 1673 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 1674 rp->set_active_mt_degree(active_workers); 1675 1676 // Process the weak references. 1677 const ReferenceProcessorStats& stats = 1678 rp->process_discovered_references(&g1_is_alive, 1679 &g1_keep_alive, 1680 &g1_drain_mark_stack, 1681 executor, 1682 _gc_timer_cm); 1683 _gc_tracer_cm->report_gc_reference_stats(stats); 1684 1685 // The do_oop work routines of the keep_alive and drain_marking_stack 1686 // oop closures will set the has_overflown flag if we overflow the 1687 // global marking stack. 1688 1689 assert(_markStack.overflow() || _markStack.isEmpty(), 1690 "mark stack should be empty (unless it overflowed)"); 1691 1692 if (_markStack.overflow()) { 1693 // This should have been done already when we tried to push an 1694 // entry on to the global mark stack. But let's do it again. 1695 set_has_overflown(); 1696 } 1697 1698 assert(rp->num_q() == active_workers, "why not"); 1699 1700 rp->enqueue_discovered_references(executor); 1701 1702 rp->verify_no_references_recorded(); 1703 assert(!rp->discovery_enabled(), "Post condition"); 1704 } 1705 1706 if (has_overflown()) { 1707 // We can not trust g1_is_alive if the marking stack overflowed 1708 return; 1709 } 1710 1711 assert(_markStack.isEmpty(), "Marking should have completed"); 1712 1713 // Unload Klasses, String, Symbols, Code Cache, etc. 1714 if (ClassUnloadingWithConcurrentMark) { 1715 bool purged_classes; 1716 1717 { 1718 GCTraceTime(Debug, gc, phases) trace("System Dictionary Unloading", _gc_timer_cm); 1719 purged_classes = SystemDictionary::do_unloading(&g1_is_alive, false /* Defer klass cleaning */); 1720 } 1721 1722 { 1723 GCTraceTime(Debug, gc, phases) trace("Parallel Unloading", _gc_timer_cm); 1724 weakRefsWorkParallelPart(&g1_is_alive, purged_classes); 1725 } 1726 } 1727 1728 if (G1StringDedup::is_enabled()) { 1729 GCTraceTime(Debug, gc, phases) trace("String Deduplication Unlink", _gc_timer_cm); 1730 G1StringDedup::unlink(&g1_is_alive); 1731 } 1732 } 1733 1734 void G1ConcurrentMark::swapMarkBitMaps() { 1735 G1CMBitMapRO* temp = _prevMarkBitMap; 1736 _prevMarkBitMap = (G1CMBitMapRO*)_nextMarkBitMap; 1737 _nextMarkBitMap = (G1CMBitMap*) temp; 1738 } 1739 1740 // Closure for marking entries in SATB buffers. 1741 class G1CMSATBBufferClosure : public SATBBufferClosure { 1742 private: 1743 G1CMTask* _task; 1744 G1CollectedHeap* _g1h; 1745 1746 // This is very similar to G1CMTask::deal_with_reference, but with 1747 // more relaxed requirements for the argument, so this must be more 1748 // circumspect about treating the argument as an object. 1749 void do_entry(void* entry) const { 1750 _task->increment_refs_reached(); 1751 HeapRegion* hr = _g1h->heap_region_containing(entry); 1752 if (entry < hr->next_top_at_mark_start()) { 1753 // Until we get here, we don't know whether entry refers to a valid 1754 // object; it could instead have been a stale reference. 1755 oop obj = static_cast<oop>(entry); 1756 assert(obj->is_oop(true /* ignore mark word */), 1757 "Invalid oop in SATB buffer: " PTR_FORMAT, p2i(obj)); 1758 _task->make_reference_grey(obj); 1759 } 1760 } 1761 1762 public: 1763 G1CMSATBBufferClosure(G1CMTask* task, G1CollectedHeap* g1h) 1764 : _task(task), _g1h(g1h) { } 1765 1766 virtual void do_buffer(void** buffer, size_t size) { 1767 for (size_t i = 0; i < size; ++i) { 1768 do_entry(buffer[i]); 1769 } 1770 } 1771 }; 1772 1773 class G1RemarkThreadsClosure : public ThreadClosure { 1774 G1CMSATBBufferClosure _cm_satb_cl; 1775 G1CMOopClosure _cm_cl; 1776 MarkingCodeBlobClosure _code_cl; 1777 int _thread_parity; 1778 1779 public: 1780 G1RemarkThreadsClosure(G1CollectedHeap* g1h, G1CMTask* task) : 1781 _cm_satb_cl(task, g1h), 1782 _cm_cl(g1h, g1h->concurrent_mark(), task), 1783 _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations), 1784 _thread_parity(Threads::thread_claim_parity()) {} 1785 1786 void do_thread(Thread* thread) { 1787 if (thread->is_Java_thread()) { 1788 if (thread->claim_oops_do(true, _thread_parity)) { 1789 JavaThread* jt = (JavaThread*)thread; 1790 1791 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking 1792 // however the liveness of oops reachable from nmethods have very complex lifecycles: 1793 // * Alive if on the stack of an executing method 1794 // * Weakly reachable otherwise 1795 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be 1796 // live by the SATB invariant but other oops recorded in nmethods may behave differently. 1797 jt->nmethods_do(&_code_cl); 1798 1799 jt->satb_mark_queue().apply_closure_and_empty(&_cm_satb_cl); 1800 } 1801 } else if (thread->is_VM_thread()) { 1802 if (thread->claim_oops_do(true, _thread_parity)) { 1803 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_satb_cl); 1804 } 1805 } 1806 } 1807 }; 1808 1809 class G1CMRemarkTask: public AbstractGangTask { 1810 private: 1811 G1ConcurrentMark* _cm; 1812 public: 1813 void work(uint worker_id) { 1814 // Since all available tasks are actually started, we should 1815 // only proceed if we're supposed to be active. 1816 if (worker_id < _cm->active_tasks()) { 1817 G1CMTask* task = _cm->task(worker_id); 1818 task->record_start_time(); 1819 { 1820 ResourceMark rm; 1821 HandleMark hm; 1822 1823 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task); 1824 Threads::threads_do(&threads_f); 1825 } 1826 1827 do { 1828 task->do_marking_step(1000000000.0 /* something very large */, 1829 true /* do_termination */, 1830 false /* is_serial */); 1831 } while (task->has_aborted() && !_cm->has_overflown()); 1832 // If we overflow, then we do not want to restart. We instead 1833 // want to abort remark and do concurrent marking again. 1834 task->record_end_time(); 1835 } 1836 } 1837 1838 G1CMRemarkTask(G1ConcurrentMark* cm, uint active_workers) : 1839 AbstractGangTask("Par Remark"), _cm(cm) { 1840 _cm->terminator()->reset_for_reuse(active_workers); 1841 } 1842 }; 1843 1844 void G1ConcurrentMark::checkpointRootsFinalWork() { 1845 ResourceMark rm; 1846 HandleMark hm; 1847 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1848 1849 GCTraceTime(Debug, gc, phases) trace("Finalize Marking", _gc_timer_cm); 1850 1851 g1h->ensure_parsability(false); 1852 1853 // this is remark, so we'll use up all active threads 1854 uint active_workers = g1h->workers()->active_workers(); 1855 set_concurrency_and_phase(active_workers, false /* concurrent */); 1856 // Leave _parallel_marking_threads at it's 1857 // value originally calculated in the G1ConcurrentMark 1858 // constructor and pass values of the active workers 1859 // through the gang in the task. 1860 1861 { 1862 StrongRootsScope srs(active_workers); 1863 1864 G1CMRemarkTask remarkTask(this, active_workers); 1865 // We will start all available threads, even if we decide that the 1866 // active_workers will be fewer. The extra ones will just bail out 1867 // immediately. 1868 g1h->workers()->run_task(&remarkTask); 1869 } 1870 1871 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1872 guarantee(has_overflown() || 1873 satb_mq_set.completed_buffers_num() == 0, 1874 "Invariant: has_overflown = %s, num buffers = " SIZE_FORMAT, 1875 BOOL_TO_STR(has_overflown()), 1876 satb_mq_set.completed_buffers_num()); 1877 1878 print_stats(); 1879 } 1880 1881 void G1ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 1882 // Note we are overriding the read-only view of the prev map here, via 1883 // the cast. 1884 ((G1CMBitMap*)_prevMarkBitMap)->clear_range(mr); 1885 } 1886 1887 HeapRegion* 1888 G1ConcurrentMark::claim_region(uint worker_id) { 1889 // "checkpoint" the finger 1890 HeapWord* finger = _finger; 1891 1892 // _heap_end will not change underneath our feet; it only changes at 1893 // yield points. 1894 while (finger < _heap_end) { 1895 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 1896 1897 HeapRegion* curr_region = _g1h->heap_region_containing(finger); 1898 1899 // Above heap_region_containing may return NULL as we always scan claim 1900 // until the end of the heap. In this case, just jump to the next region. 1901 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords; 1902 1903 // Is the gap between reading the finger and doing the CAS too long? 1904 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 1905 if (res == finger && curr_region != NULL) { 1906 // we succeeded 1907 HeapWord* bottom = curr_region->bottom(); 1908 HeapWord* limit = curr_region->next_top_at_mark_start(); 1909 1910 // notice that _finger == end cannot be guaranteed here since, 1911 // someone else might have moved the finger even further 1912 assert(_finger >= end, "the finger should have moved forward"); 1913 1914 if (limit > bottom) { 1915 return curr_region; 1916 } else { 1917 assert(limit == bottom, 1918 "the region limit should be at bottom"); 1919 // we return NULL and the caller should try calling 1920 // claim_region() again. 1921 return NULL; 1922 } 1923 } else { 1924 assert(_finger > finger, "the finger should have moved forward"); 1925 // read it again 1926 finger = _finger; 1927 } 1928 } 1929 1930 return NULL; 1931 } 1932 1933 #ifndef PRODUCT 1934 class VerifyNoCSetOops VALUE_OBJ_CLASS_SPEC { 1935 private: 1936 G1CollectedHeap* _g1h; 1937 const char* _phase; 1938 int _info; 1939 1940 public: 1941 VerifyNoCSetOops(const char* phase, int info = -1) : 1942 _g1h(G1CollectedHeap::heap()), 1943 _phase(phase), 1944 _info(info) 1945 { } 1946 1947 void operator()(oop obj) const { 1948 guarantee(obj->is_oop(), 1949 "Non-oop " PTR_FORMAT ", phase: %s, info: %d", 1950 p2i(obj), _phase, _info); 1951 guarantee(!_g1h->obj_in_cs(obj), 1952 "obj: " PTR_FORMAT " in CSet, phase: %s, info: %d", 1953 p2i(obj), _phase, _info); 1954 } 1955 }; 1956 1957 void G1ConcurrentMark::verify_no_cset_oops() { 1958 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 1959 if (!G1CollectedHeap::heap()->collector_state()->mark_in_progress()) { 1960 return; 1961 } 1962 1963 // Verify entries on the global mark stack 1964 _markStack.iterate(VerifyNoCSetOops("Stack")); 1965 1966 // Verify entries on the task queues 1967 for (uint i = 0; i < _max_worker_id; ++i) { 1968 G1CMTaskQueue* queue = _task_queues->queue(i); 1969 queue->iterate(VerifyNoCSetOops("Queue", i)); 1970 } 1971 1972 // Verify the global finger 1973 HeapWord* global_finger = finger(); 1974 if (global_finger != NULL && global_finger < _heap_end) { 1975 // Since we always iterate over all regions, we might get a NULL HeapRegion 1976 // here. 1977 HeapRegion* global_hr = _g1h->heap_region_containing(global_finger); 1978 guarantee(global_hr == NULL || global_finger == global_hr->bottom(), 1979 "global finger: " PTR_FORMAT " region: " HR_FORMAT, 1980 p2i(global_finger), HR_FORMAT_PARAMS(global_hr)); 1981 } 1982 1983 // Verify the task fingers 1984 assert(parallel_marking_threads() <= _max_worker_id, "sanity"); 1985 for (uint i = 0; i < parallel_marking_threads(); ++i) { 1986 G1CMTask* task = _tasks[i]; 1987 HeapWord* task_finger = task->finger(); 1988 if (task_finger != NULL && task_finger < _heap_end) { 1989 // See above note on the global finger verification. 1990 HeapRegion* task_hr = _g1h->heap_region_containing(task_finger); 1991 guarantee(task_hr == NULL || task_finger == task_hr->bottom() || 1992 !task_hr->in_collection_set(), 1993 "task finger: " PTR_FORMAT " region: " HR_FORMAT, 1994 p2i(task_finger), HR_FORMAT_PARAMS(task_hr)); 1995 } 1996 } 1997 } 1998 #endif // PRODUCT 1999 void G1ConcurrentMark::create_live_data() { 2000 _g1h->g1_rem_set()->create_card_live_data(_parallel_workers, _nextMarkBitMap); 2001 } 2002 2003 void G1ConcurrentMark::finalize_live_data() { 2004 _g1h->g1_rem_set()->finalize_card_live_data(_g1h->workers(), _nextMarkBitMap); 2005 } 2006 2007 void G1ConcurrentMark::verify_live_data() { 2008 _g1h->g1_rem_set()->verify_card_live_data(_g1h->workers(), _nextMarkBitMap); 2009 } 2010 2011 void G1ConcurrentMark::clear_live_data(WorkGang* workers) { 2012 _g1h->g1_rem_set()->clear_card_live_data(workers); 2013 } 2014 2015 #ifdef ASSERT 2016 void G1ConcurrentMark::verify_live_data_clear() { 2017 _g1h->g1_rem_set()->verify_card_live_data_is_clear(); 2018 } 2019 #endif 2020 2021 void G1ConcurrentMark::print_stats() { 2022 if (!log_is_enabled(Debug, gc, stats)) { 2023 return; 2024 } 2025 log_debug(gc, stats)("---------------------------------------------------------------------"); 2026 for (size_t i = 0; i < _active_tasks; ++i) { 2027 _tasks[i]->print_stats(); 2028 log_debug(gc, stats)("---------------------------------------------------------------------"); 2029 } 2030 } 2031 2032 void G1ConcurrentMark::abort() { 2033 if (!cmThread()->during_cycle() || _has_aborted) { 2034 // We haven't started a concurrent cycle or we have already aborted it. No need to do anything. 2035 return; 2036 } 2037 2038 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next 2039 // concurrent bitmap clearing. 2040 { 2041 GCTraceTime(Debug, gc)("Clear Next Bitmap"); 2042 clear_bitmap(_nextMarkBitMap, _g1h->workers(), false); 2043 } 2044 // Note we cannot clear the previous marking bitmap here 2045 // since VerifyDuringGC verifies the objects marked during 2046 // a full GC against the previous bitmap. 2047 2048 { 2049 GCTraceTime(Debug, gc)("Clear Live Data"); 2050 clear_live_data(_g1h->workers()); 2051 } 2052 DEBUG_ONLY({ 2053 GCTraceTime(Debug, gc)("Verify Live Data Clear"); 2054 verify_live_data_clear(); 2055 }) 2056 // Empty mark stack 2057 reset_marking_state(); 2058 for (uint i = 0; i < _max_worker_id; ++i) { 2059 _tasks[i]->clear_region_fields(); 2060 } 2061 _first_overflow_barrier_sync.abort(); 2062 _second_overflow_barrier_sync.abort(); 2063 _has_aborted = true; 2064 2065 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2066 satb_mq_set.abandon_partial_marking(); 2067 // This can be called either during or outside marking, we'll read 2068 // the expected_active value from the SATB queue set. 2069 satb_mq_set.set_active_all_threads( 2070 false, /* new active value */ 2071 satb_mq_set.is_active() /* expected_active */); 2072 } 2073 2074 static void print_ms_time_info(const char* prefix, const char* name, 2075 NumberSeq& ns) { 2076 log_trace(gc, marking)("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 2077 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 2078 if (ns.num() > 0) { 2079 log_trace(gc, marking)("%s [std. dev = %8.2f ms, max = %8.2f ms]", 2080 prefix, ns.sd(), ns.maximum()); 2081 } 2082 } 2083 2084 void G1ConcurrentMark::print_summary_info() { 2085 Log(gc, marking) log; 2086 if (!log.is_trace()) { 2087 return; 2088 } 2089 2090 log.trace(" Concurrent marking:"); 2091 print_ms_time_info(" ", "init marks", _init_times); 2092 print_ms_time_info(" ", "remarks", _remark_times); 2093 { 2094 print_ms_time_info(" ", "final marks", _remark_mark_times); 2095 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 2096 2097 } 2098 print_ms_time_info(" ", "cleanups", _cleanup_times); 2099 log.trace(" Finalize live data total time = %8.2f s (avg = %8.2f ms).", 2100 _total_counting_time, (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2101 if (G1ScrubRemSets) { 2102 log.trace(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 2103 _total_rs_scrub_time, (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / (double)_cleanup_times.num() : 0.0)); 2104 } 2105 log.trace(" Total stop_world time = %8.2f s.", 2106 (_init_times.sum() + _remark_times.sum() + _cleanup_times.sum())/1000.0); 2107 log.trace(" Total concurrent time = %8.2f s (%8.2f s marking).", 2108 cmThread()->vtime_accum(), cmThread()->vtime_mark_accum()); 2109 } 2110 2111 void G1ConcurrentMark::print_worker_threads_on(outputStream* st) const { 2112 _parallel_workers->print_worker_threads_on(st); 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 G1CollectorPolicy* 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 }