1 /* 2 * Copyright (c) 2001, 2014, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/symbolTable.hpp" 27 #include "code/codeCache.hpp" 28 #include "gc_implementation/g1/concurrentMark.inline.hpp" 29 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 30 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 31 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 32 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 33 #include "gc_implementation/g1/g1Log.hpp" 34 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 35 #include "gc_implementation/g1/g1RemSet.hpp" 36 #include "gc_implementation/g1/heapRegion.inline.hpp" 37 #include "gc_implementation/g1/heapRegionManager.inline.hpp" 38 #include "gc_implementation/g1/heapRegionRemSet.hpp" 39 #include "gc_implementation/g1/heapRegionSet.inline.hpp" 40 #include "gc_implementation/shared/vmGCOperations.hpp" 41 #include "gc_implementation/shared/gcTimer.hpp" 42 #include "gc_implementation/shared/gcTrace.hpp" 43 #include "gc_implementation/shared/gcTraceTime.hpp" 44 #include "memory/allocation.hpp" 45 #include "memory/genOopClosures.inline.hpp" 46 #include "memory/referencePolicy.hpp" 47 #include "memory/resourceArea.hpp" 48 #include "oops/oop.inline.hpp" 49 #include "runtime/handles.inline.hpp" 50 #include "runtime/java.hpp" 51 #include "runtime/atomic.inline.hpp" 52 #include "runtime/prefetch.inline.hpp" 53 #include "services/memTracker.hpp" 54 55 // Concurrent marking bit map wrapper 56 57 CMBitMapRO::CMBitMapRO(int shifter) : 58 _bm(), 59 _shifter(shifter) { 60 _bmStartWord = 0; 61 _bmWordSize = 0; 62 } 63 64 HeapWord* CMBitMapRO::getNextMarkedWordAddress(const HeapWord* addr, 65 const HeapWord* limit) const { 66 // First we must round addr *up* to a possible object boundary. 67 addr = (HeapWord*)align_size_up((intptr_t)addr, 68 HeapWordSize << _shifter); 69 size_t addrOffset = heapWordToOffset(addr); 70 if (limit == NULL) { 71 limit = _bmStartWord + _bmWordSize; 72 } 73 size_t limitOffset = heapWordToOffset(limit); 74 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); 75 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 76 assert(nextAddr >= addr, "get_next_one postcondition"); 77 assert(nextAddr == limit || isMarked(nextAddr), 78 "get_next_one postcondition"); 79 return nextAddr; 80 } 81 82 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(const HeapWord* addr, 83 const HeapWord* limit) const { 84 size_t addrOffset = heapWordToOffset(addr); 85 if (limit == NULL) { 86 limit = _bmStartWord + _bmWordSize; 87 } 88 size_t limitOffset = heapWordToOffset(limit); 89 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset); 90 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 91 assert(nextAddr >= addr, "get_next_one postcondition"); 92 assert(nextAddr == limit || !isMarked(nextAddr), 93 "get_next_one postcondition"); 94 return nextAddr; 95 } 96 97 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const { 98 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check"); 99 return (int) (diff >> _shifter); 100 } 101 102 #ifndef PRODUCT 103 bool CMBitMapRO::covers(MemRegion heap_rs) const { 104 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 105 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize, 106 "size inconsistency"); 107 return _bmStartWord == (HeapWord*)(heap_rs.start()) && 108 _bmWordSize == heap_rs.word_size(); 109 } 110 #endif 111 112 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const { 113 _bm.print_on_error(st, prefix); 114 } 115 116 size_t CMBitMap::compute_size(size_t heap_size) { 117 return heap_size / mark_distance(); 118 } 119 120 size_t CMBitMap::mark_distance() { 121 return MinObjAlignmentInBytes * BitsPerByte; 122 } 123 124 void CMBitMap::initialize(MemRegion heap, G1RegionToSpaceMapper* storage) { 125 _bmStartWord = heap.start(); 126 _bmWordSize = heap.word_size(); 127 128 _bm.set_map((BitMap::bm_word_t*) storage->reserved().start()); 129 _bm.set_size(_bmWordSize >> _shifter); 130 131 storage->set_mapping_changed_listener(&_listener); 132 } 133 134 void CMBitMapMappingChangedListener::on_commit(uint start_region, size_t num_regions) { 135 // We need to clear the bitmap on commit, removing any existing information. 136 MemRegion mr(G1CollectedHeap::heap()->bottom_addr_for_region(start_region), num_regions * HeapRegion::GrainWords); 137 _bm->clearRange(mr); 138 } 139 140 // Closure used for clearing the given mark bitmap. 141 class ClearBitmapHRClosure : public HeapRegionClosure { 142 private: 143 ConcurrentMark* _cm; 144 CMBitMap* _bitmap; 145 bool _may_yield; // The closure may yield during iteration. If yielded, abort the iteration. 146 public: 147 ClearBitmapHRClosure(ConcurrentMark* cm, CMBitMap* bitmap, bool may_yield) : HeapRegionClosure(), _cm(cm), _bitmap(bitmap), _may_yield(may_yield) { 148 assert(!may_yield || cm != NULL, "CM must be non-NULL if this closure is expected to yield."); 149 } 150 151 virtual bool doHeapRegion(HeapRegion* r) { 152 size_t const chunk_size_in_words = M / HeapWordSize; 153 154 HeapWord* cur = r->bottom(); 155 HeapWord* const end = r->end(); 156 157 while (cur < end) { 158 MemRegion mr(cur, MIN2(cur + chunk_size_in_words, end)); 159 _bitmap->clearRange(mr); 160 161 cur += chunk_size_in_words; 162 163 // Abort iteration if after yielding the marking has been aborted. 164 if (_may_yield && _cm->do_yield_check() && _cm->has_aborted()) { 165 return true; 166 } 167 // Repeat the asserts from before the start of the closure. We will do them 168 // as asserts here to minimize their overhead on the product. However, we 169 // will have them as guarantees at the beginning / end of the bitmap 170 // clearing to get some checking in the product. 171 assert(!_may_yield || _cm->cmThread()->during_cycle(), "invariant"); 172 assert(!_may_yield || !G1CollectedHeap::heap()->mark_in_progress(), "invariant"); 173 } 174 175 return false; 176 } 177 }; 178 179 void CMBitMap::clearAll() { 180 ClearBitmapHRClosure cl(NULL, this, false /* may_yield */); 181 G1CollectedHeap::heap()->heap_region_iterate(&cl); 182 guarantee(cl.complete(), "Must have completed iteration."); 183 return; 184 } 185 186 void CMBitMap::markRange(MemRegion mr) { 187 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 188 assert(!mr.is_empty(), "unexpected empty region"); 189 assert((offsetToHeapWord(heapWordToOffset(mr.end())) == 190 ((HeapWord *) mr.end())), 191 "markRange memory region end is not card aligned"); 192 // convert address range into offset range 193 _bm.at_put_range(heapWordToOffset(mr.start()), 194 heapWordToOffset(mr.end()), true); 195 } 196 197 void CMBitMap::clearRange(MemRegion mr) { 198 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 199 assert(!mr.is_empty(), "unexpected empty region"); 200 // convert address range into offset range 201 _bm.at_put_range(heapWordToOffset(mr.start()), 202 heapWordToOffset(mr.end()), false); 203 } 204 205 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr, 206 HeapWord* end_addr) { 207 HeapWord* start = getNextMarkedWordAddress(addr); 208 start = MIN2(start, end_addr); 209 HeapWord* end = getNextUnmarkedWordAddress(start); 210 end = MIN2(end, end_addr); 211 assert(start <= end, "Consistency check"); 212 MemRegion mr(start, end); 213 if (!mr.is_empty()) { 214 clearRange(mr); 215 } 216 return mr; 217 } 218 219 CMMarkStack::CMMarkStack(ConcurrentMark* cm) : 220 _base(NULL), _cm(cm) 221 #ifdef ASSERT 222 , _drain_in_progress(false) 223 , _drain_in_progress_yields(false) 224 #endif 225 {} 226 227 bool CMMarkStack::allocate(size_t capacity) { 228 // allocate a stack of the requisite depth 229 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop))); 230 if (!rs.is_reserved()) { 231 warning("ConcurrentMark MarkStack allocation failure"); 232 return false; 233 } 234 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 235 if (!_virtual_space.initialize(rs, rs.size())) { 236 warning("ConcurrentMark MarkStack backing store failure"); 237 // Release the virtual memory reserved for the marking stack 238 rs.release(); 239 return false; 240 } 241 assert(_virtual_space.committed_size() == rs.size(), 242 "Didn't reserve backing store for all of ConcurrentMark stack?"); 243 _base = (oop*) _virtual_space.low(); 244 setEmpty(); 245 _capacity = (jint) capacity; 246 _saved_index = -1; 247 _should_expand = false; 248 NOT_PRODUCT(_max_depth = 0); 249 return true; 250 } 251 252 void CMMarkStack::expand() { 253 // Called, during remark, if we've overflown the marking stack during marking. 254 assert(isEmpty(), "stack should been emptied while handling overflow"); 255 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted"); 256 // Clear expansion flag 257 _should_expand = false; 258 if (_capacity == (jint) MarkStackSizeMax) { 259 if (PrintGCDetails && Verbose) { 260 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit"); 261 } 262 return; 263 } 264 // Double capacity if possible 265 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax); 266 // Do not give up existing stack until we have managed to 267 // get the double capacity that we desired. 268 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity * 269 sizeof(oop))); 270 if (rs.is_reserved()) { 271 // Release the backing store associated with old stack 272 _virtual_space.release(); 273 // Reinitialize virtual space for new stack 274 if (!_virtual_space.initialize(rs, rs.size())) { 275 fatal("Not enough swap for expanded marking stack capacity"); 276 } 277 _base = (oop*)(_virtual_space.low()); 278 _index = 0; 279 _capacity = new_capacity; 280 } else { 281 if (PrintGCDetails && Verbose) { 282 // Failed to double capacity, continue; 283 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from " 284 SIZE_FORMAT"K to " SIZE_FORMAT"K", 285 _capacity / K, new_capacity / K); 286 } 287 } 288 } 289 290 void CMMarkStack::set_should_expand() { 291 // If we're resetting the marking state because of an 292 // marking stack overflow, record that we should, if 293 // possible, expand the stack. 294 _should_expand = _cm->has_overflown(); 295 } 296 297 CMMarkStack::~CMMarkStack() { 298 if (_base != NULL) { 299 _base = NULL; 300 _virtual_space.release(); 301 } 302 } 303 304 void CMMarkStack::par_push(oop ptr) { 305 while (true) { 306 if (isFull()) { 307 _overflow = true; 308 return; 309 } 310 // Otherwise... 311 jint index = _index; 312 jint next_index = index+1; 313 jint res = Atomic::cmpxchg(next_index, &_index, index); 314 if (res == index) { 315 _base[index] = ptr; 316 // Note that we don't maintain this atomically. We could, but it 317 // doesn't seem necessary. 318 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 319 return; 320 } 321 // Otherwise, we need to try again. 322 } 323 } 324 325 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) { 326 while (true) { 327 if (isFull()) { 328 _overflow = true; 329 return; 330 } 331 // Otherwise... 332 jint index = _index; 333 jint next_index = index + n; 334 if (next_index > _capacity) { 335 _overflow = true; 336 return; 337 } 338 jint res = Atomic::cmpxchg(next_index, &_index, index); 339 if (res == index) { 340 for (int i = 0; i < n; i++) { 341 int ind = index + i; 342 assert(ind < _capacity, "By overflow test above."); 343 _base[ind] = ptr_arr[i]; 344 } 345 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 346 return; 347 } 348 // Otherwise, we need to try again. 349 } 350 } 351 352 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { 353 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 354 jint start = _index; 355 jint next_index = start + n; 356 if (next_index > _capacity) { 357 _overflow = true; 358 return; 359 } 360 // Otherwise. 361 _index = next_index; 362 for (int i = 0; i < n; i++) { 363 int ind = start + i; 364 assert(ind < _capacity, "By overflow test above."); 365 _base[ind] = ptr_arr[i]; 366 } 367 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 368 } 369 370 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { 371 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 372 jint index = _index; 373 if (index == 0) { 374 *n = 0; 375 return false; 376 } else { 377 int k = MIN2(max, index); 378 jint new_ind = index - k; 379 for (int j = 0; j < k; j++) { 380 ptr_arr[j] = _base[new_ind + j]; 381 } 382 _index = new_ind; 383 *n = k; 384 return true; 385 } 386 } 387 388 template<class OopClosureClass> 389 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) { 390 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after 391 || SafepointSynchronize::is_at_safepoint(), 392 "Drain recursion must be yield-safe."); 393 bool res = true; 394 debug_only(_drain_in_progress = true); 395 debug_only(_drain_in_progress_yields = yield_after); 396 while (!isEmpty()) { 397 oop newOop = pop(); 398 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop"); 399 assert(newOop->is_oop(), "Expected an oop"); 400 assert(bm == NULL || bm->isMarked((HeapWord*)newOop), 401 "only grey objects on this stack"); 402 newOop->oop_iterate(cl); 403 if (yield_after && _cm->do_yield_check()) { 404 res = false; 405 break; 406 } 407 } 408 debug_only(_drain_in_progress = false); 409 return res; 410 } 411 412 void CMMarkStack::note_start_of_gc() { 413 assert(_saved_index == -1, 414 "note_start_of_gc()/end_of_gc() bracketed incorrectly"); 415 _saved_index = _index; 416 } 417 418 void CMMarkStack::note_end_of_gc() { 419 // This is intentionally a guarantee, instead of an assert. If we 420 // accidentally add something to the mark stack during GC, it 421 // will be a correctness issue so it's better if we crash. we'll 422 // only check this once per GC anyway, so it won't be a performance 423 // issue in any way. 424 guarantee(_saved_index == _index, 425 err_msg("saved index: %d index: %d", _saved_index, _index)); 426 _saved_index = -1; 427 } 428 429 void CMMarkStack::oops_do(OopClosure* f) { 430 assert(_saved_index == _index, 431 err_msg("saved index: %d index: %d", _saved_index, _index)); 432 for (int i = 0; i < _index; i += 1) { 433 f->do_oop(&_base[i]); 434 } 435 } 436 437 CMRootRegions::CMRootRegions() : 438 _young_list(NULL), _cm(NULL), _scan_in_progress(false), 439 _should_abort(false), _next_survivor(NULL) { } 440 441 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) { 442 _young_list = g1h->young_list(); 443 _cm = cm; 444 } 445 446 void CMRootRegions::prepare_for_scan() { 447 assert(!scan_in_progress(), "pre-condition"); 448 449 // Currently, only survivors can be root regions. 450 assert(_next_survivor == NULL, "pre-condition"); 451 _next_survivor = _young_list->first_survivor_region(); 452 _scan_in_progress = (_next_survivor != NULL); 453 _should_abort = false; 454 } 455 456 HeapRegion* CMRootRegions::claim_next() { 457 if (_should_abort) { 458 // If someone has set the should_abort flag, we return NULL to 459 // force the caller to bail out of their loop. 460 return NULL; 461 } 462 463 // Currently, only survivors can be root regions. 464 HeapRegion* res = _next_survivor; 465 if (res != NULL) { 466 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 467 // Read it again in case it changed while we were waiting for the lock. 468 res = _next_survivor; 469 if (res != NULL) { 470 if (res == _young_list->last_survivor_region()) { 471 // We just claimed the last survivor so store NULL to indicate 472 // that we're done. 473 _next_survivor = NULL; 474 } else { 475 _next_survivor = res->get_next_young_region(); 476 } 477 } else { 478 // Someone else claimed the last survivor while we were trying 479 // to take the lock so nothing else to do. 480 } 481 } 482 assert(res == NULL || res->is_survivor(), "post-condition"); 483 484 return res; 485 } 486 487 void CMRootRegions::scan_finished() { 488 assert(scan_in_progress(), "pre-condition"); 489 490 // Currently, only survivors can be root regions. 491 if (!_should_abort) { 492 assert(_next_survivor == NULL, "we should have claimed all survivors"); 493 } 494 _next_survivor = NULL; 495 496 { 497 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 498 _scan_in_progress = false; 499 RootRegionScan_lock->notify_all(); 500 } 501 } 502 503 bool CMRootRegions::wait_until_scan_finished() { 504 if (!scan_in_progress()) return false; 505 506 { 507 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 508 while (scan_in_progress()) { 509 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 510 } 511 } 512 return true; 513 } 514 515 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 516 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 517 #endif // _MSC_VER 518 519 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 520 return MAX2((n_par_threads + 2) / 4, 1U); 521 } 522 523 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, G1RegionToSpaceMapper* prev_bitmap_storage, G1RegionToSpaceMapper* next_bitmap_storage) : 524 _g1h(g1h), 525 _markBitMap1(), 526 _markBitMap2(), 527 _parallel_marking_threads(0), 528 _max_parallel_marking_threads(0), 529 _sleep_factor(0.0), 530 _marking_task_overhead(1.0), 531 _cleanup_sleep_factor(0.0), 532 _cleanup_task_overhead(1.0), 533 _cleanup_list("Cleanup List"), 534 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/), 535 _card_bm((g1h->reserved_region().byte_size() + CardTableModRefBS::card_size - 1) >> 536 CardTableModRefBS::card_shift, 537 false /* in_resource_area*/), 538 539 _prevMarkBitMap(&_markBitMap1), 540 _nextMarkBitMap(&_markBitMap2), 541 542 _markStack(this), 543 // _finger set in set_non_marking_state 544 545 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)), 546 // _active_tasks set in set_non_marking_state 547 // _tasks set inside the constructor 548 _task_queues(new CMTaskQueueSet((int) _max_worker_id)), 549 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)), 550 551 _has_overflown(false), 552 _concurrent(false), 553 _has_aborted(false), 554 _aborted_gc_id(GCId::undefined()), 555 _restart_for_overflow(false), 556 _concurrent_marking_in_progress(false), 557 558 // _verbose_level set below 559 560 _init_times(), 561 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 562 _cleanup_times(), 563 _total_counting_time(0.0), 564 _total_rs_scrub_time(0.0), 565 566 _parallel_workers(NULL), 567 568 _count_card_bitmaps(NULL), 569 _count_marked_bytes(NULL), 570 _completed_initialization(false) { 571 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel; 572 if (verbose_level < no_verbose) { 573 verbose_level = no_verbose; 574 } 575 if (verbose_level > high_verbose) { 576 verbose_level = high_verbose; 577 } 578 _verbose_level = verbose_level; 579 580 if (verbose_low()) { 581 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", " 582 "heap end = " PTR_FORMAT, p2i(_heap_start), p2i(_heap_end)); 583 } 584 585 _markBitMap1.initialize(g1h->reserved_region(), prev_bitmap_storage); 586 _markBitMap2.initialize(g1h->reserved_region(), next_bitmap_storage); 587 588 // Create & start a ConcurrentMark thread. 589 _cmThread = new ConcurrentMarkThread(this); 590 assert(cmThread() != NULL, "CM Thread should have been created"); 591 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 592 if (_cmThread->osthread() == NULL) { 593 vm_shutdown_during_initialization("Could not create ConcurrentMarkThread"); 594 } 595 596 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 597 assert(_markBitMap1.covers(g1h->reserved_region()), "_markBitMap1 inconsistency"); 598 assert(_markBitMap2.covers(g1h->reserved_region()), "_markBitMap2 inconsistency"); 599 600 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 601 satb_qs.set_buffer_size(G1SATBBufferSize); 602 603 _root_regions.init(_g1h, this); 604 605 if (ConcGCThreads > ParallelGCThreads) { 606 warning("Can't have more ConcGCThreads (" UINTX_FORMAT ") " 607 "than ParallelGCThreads (" UINTX_FORMAT ").", 608 ConcGCThreads, ParallelGCThreads); 609 return; 610 } 611 if (ParallelGCThreads == 0) { 612 // if we are not running with any parallel GC threads we will not 613 // spawn any marking threads either 614 _parallel_marking_threads = 0; 615 _max_parallel_marking_threads = 0; 616 _sleep_factor = 0.0; 617 _marking_task_overhead = 1.0; 618 } else { 619 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 620 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 621 // if both are set 622 _sleep_factor = 0.0; 623 _marking_task_overhead = 1.0; 624 } else if (G1MarkingOverheadPercent > 0) { 625 // We will calculate the number of parallel marking threads based 626 // on a target overhead with respect to the soft real-time goal 627 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 628 double overall_cm_overhead = 629 (double) MaxGCPauseMillis * marking_overhead / 630 (double) GCPauseIntervalMillis; 631 double cpu_ratio = 1.0 / (double) os::processor_count(); 632 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 633 double marking_task_overhead = 634 overall_cm_overhead / marking_thread_num * 635 (double) os::processor_count(); 636 double sleep_factor = 637 (1.0 - marking_task_overhead) / marking_task_overhead; 638 639 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num); 640 _sleep_factor = sleep_factor; 641 _marking_task_overhead = marking_task_overhead; 642 } else { 643 // Calculate the number of parallel marking threads by scaling 644 // the number of parallel GC threads. 645 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads); 646 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num); 647 _sleep_factor = 0.0; 648 _marking_task_overhead = 1.0; 649 } 650 651 assert(ConcGCThreads > 0, "Should have been set"); 652 _parallel_marking_threads = (uint) ConcGCThreads; 653 _max_parallel_marking_threads = _parallel_marking_threads; 654 655 if (parallel_marking_threads() > 1) { 656 _cleanup_task_overhead = 1.0; 657 } else { 658 _cleanup_task_overhead = marking_task_overhead(); 659 } 660 _cleanup_sleep_factor = 661 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead(); 662 663 #if 0 664 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads()); 665 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead()); 666 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor()); 667 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead()); 668 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor()); 669 #endif 670 671 guarantee(parallel_marking_threads() > 0, "peace of mind"); 672 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads", 673 _max_parallel_marking_threads, false, true); 674 if (_parallel_workers == NULL) { 675 vm_exit_during_initialization("Failed necessary allocation."); 676 } else { 677 _parallel_workers->initialize_workers(); 678 } 679 } 680 681 if (FLAG_IS_DEFAULT(MarkStackSize)) { 682 uintx mark_stack_size = 683 MIN2(MarkStackSizeMax, 684 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE))); 685 // Verify that the calculated value for MarkStackSize is in range. 686 // It would be nice to use the private utility routine from Arguments. 687 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 688 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): " 689 "must be between " UINTX_FORMAT " and " UINTX_FORMAT, 690 mark_stack_size, (uintx) 1, MarkStackSizeMax); 691 return; 692 } 693 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size); 694 } else { 695 // Verify MarkStackSize is in range. 696 if (FLAG_IS_CMDLINE(MarkStackSize)) { 697 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 698 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 699 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): " 700 "must be between " UINTX_FORMAT " and " UINTX_FORMAT, 701 MarkStackSize, (uintx) 1, MarkStackSizeMax); 702 return; 703 } 704 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 705 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 706 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")" 707 " or for MarkStackSizeMax (" UINTX_FORMAT ")", 708 MarkStackSize, MarkStackSizeMax); 709 return; 710 } 711 } 712 } 713 } 714 715 if (!_markStack.allocate(MarkStackSize)) { 716 warning("Failed to allocate CM marking stack"); 717 return; 718 } 719 720 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC); 721 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC); 722 723 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC); 724 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC); 725 726 BitMap::idx_t card_bm_size = _card_bm.size(); 727 728 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 729 _active_tasks = _max_worker_id; 730 731 size_t max_regions = (size_t) _g1h->max_regions(); 732 for (uint i = 0; i < _max_worker_id; ++i) { 733 CMTaskQueue* task_queue = new CMTaskQueue(); 734 task_queue->initialize(); 735 _task_queues->register_queue(i, task_queue); 736 737 _count_card_bitmaps[i] = BitMap(card_bm_size, false); 738 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC); 739 740 _tasks[i] = new CMTask(i, this, 741 _count_marked_bytes[i], 742 &_count_card_bitmaps[i], 743 task_queue, _task_queues); 744 745 _accum_task_vtime[i] = 0.0; 746 } 747 748 // Calculate the card number for the bottom of the heap. Used 749 // in biasing indexes into the accounting card bitmaps. 750 _heap_bottom_card_num = 751 intptr_t(uintptr_t(_g1h->reserved_region().start()) >> 752 CardTableModRefBS::card_shift); 753 754 // Clear all the liveness counting data 755 clear_all_count_data(); 756 757 // so that the call below can read a sensible value 758 _heap_start = g1h->reserved_region().start(); 759 set_non_marking_state(); 760 _completed_initialization = true; 761 } 762 763 void ConcurrentMark::reset() { 764 // Starting values for these two. This should be called in a STW 765 // phase. 766 MemRegion reserved = _g1h->g1_reserved(); 767 _heap_start = reserved.start(); 768 _heap_end = reserved.end(); 769 770 // Separated the asserts so that we know which one fires. 771 assert(_heap_start != NULL, "heap bounds should look ok"); 772 assert(_heap_end != NULL, "heap bounds should look ok"); 773 assert(_heap_start < _heap_end, "heap bounds should look ok"); 774 775 // Reset all the marking data structures and any necessary flags 776 reset_marking_state(); 777 778 if (verbose_low()) { 779 gclog_or_tty->print_cr("[global] resetting"); 780 } 781 782 // We do reset all of them, since different phases will use 783 // different number of active threads. So, it's easiest to have all 784 // of them ready. 785 for (uint i = 0; i < _max_worker_id; ++i) { 786 _tasks[i]->reset(_nextMarkBitMap); 787 } 788 789 // we need this to make sure that the flag is on during the evac 790 // pause with initial mark piggy-backed 791 set_concurrent_marking_in_progress(); 792 } 793 794 795 void ConcurrentMark::reset_marking_state(bool clear_overflow) { 796 _markStack.set_should_expand(); 797 _markStack.setEmpty(); // Also clears the _markStack overflow flag 798 if (clear_overflow) { 799 clear_has_overflown(); 800 } else { 801 assert(has_overflown(), "pre-condition"); 802 } 803 _finger = _heap_start; 804 805 for (uint i = 0; i < _max_worker_id; ++i) { 806 CMTaskQueue* queue = _task_queues->queue(i); 807 queue->set_empty(); 808 } 809 } 810 811 void ConcurrentMark::set_concurrency(uint active_tasks) { 812 assert(active_tasks <= _max_worker_id, "we should not have more"); 813 814 _active_tasks = active_tasks; 815 // Need to update the three data structures below according to the 816 // number of active threads for this phase. 817 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 818 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 819 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 820 } 821 822 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 823 set_concurrency(active_tasks); 824 825 _concurrent = concurrent; 826 // We propagate this to all tasks, not just the active ones. 827 for (uint i = 0; i < _max_worker_id; ++i) 828 _tasks[i]->set_concurrent(concurrent); 829 830 if (concurrent) { 831 set_concurrent_marking_in_progress(); 832 } else { 833 // We currently assume that the concurrent flag has been set to 834 // false before we start remark. At this point we should also be 835 // in a STW phase. 836 assert(!concurrent_marking_in_progress(), "invariant"); 837 assert(out_of_regions(), 838 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT, 839 p2i(_finger), p2i(_heap_end))); 840 } 841 } 842 843 void ConcurrentMark::set_non_marking_state() { 844 // We set the global marking state to some default values when we're 845 // not doing marking. 846 reset_marking_state(); 847 _active_tasks = 0; 848 clear_concurrent_marking_in_progress(); 849 } 850 851 ConcurrentMark::~ConcurrentMark() { 852 // The ConcurrentMark instance is never freed. 853 ShouldNotReachHere(); 854 } 855 856 void ConcurrentMark::clearNextBitmap() { 857 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 858 859 // Make sure that the concurrent mark thread looks to still be in 860 // the current cycle. 861 guarantee(cmThread()->during_cycle(), "invariant"); 862 863 // We are finishing up the current cycle by clearing the next 864 // marking bitmap and getting it ready for the next cycle. During 865 // this time no other cycle can start. So, let's make sure that this 866 // is the case. 867 guarantee(!g1h->mark_in_progress(), "invariant"); 868 869 ClearBitmapHRClosure cl(this, _nextMarkBitMap, true /* may_yield */); 870 g1h->heap_region_iterate(&cl); 871 872 // Clear the liveness counting data. If the marking has been aborted, the abort() 873 // call already did that. 874 if (cl.complete()) { 875 clear_all_count_data(); 876 } 877 878 // Repeat the asserts from above. 879 guarantee(cmThread()->during_cycle(), "invariant"); 880 guarantee(!g1h->mark_in_progress(), "invariant"); 881 } 882 883 class CheckBitmapClearHRClosure : public HeapRegionClosure { 884 CMBitMap* _bitmap; 885 bool _error; 886 public: 887 CheckBitmapClearHRClosure(CMBitMap* bitmap) : _bitmap(bitmap) { 888 } 889 890 virtual bool doHeapRegion(HeapRegion* r) { 891 // This closure can be called concurrently to the mutator, so we must make sure 892 // that the result of the getNextMarkedWordAddress() call is compared to the 893 // value passed to it as limit to detect any found bits. 894 // We can use the region's orig_end() for the limit and the comparison value 895 // as it always contains the "real" end of the region that never changes and 896 // has no side effects. 897 // Due to the latter, there can also be no problem with the compiler generating 898 // reloads of the orig_end() call. 899 HeapWord* end = r->orig_end(); 900 return _bitmap->getNextMarkedWordAddress(r->bottom(), end) != end; 901 } 902 }; 903 904 bool ConcurrentMark::nextMarkBitmapIsClear() { 905 CheckBitmapClearHRClosure cl(_nextMarkBitMap); 906 _g1h->heap_region_iterate(&cl); 907 return cl.complete(); 908 } 909 910 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 911 public: 912 bool doHeapRegion(HeapRegion* r) { 913 if (!r->is_continues_humongous()) { 914 r->note_start_of_marking(); 915 } 916 return false; 917 } 918 }; 919 920 void ConcurrentMark::checkpointRootsInitialPre() { 921 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 922 G1CollectorPolicy* g1p = g1h->g1_policy(); 923 924 _has_aborted = false; 925 926 #ifndef PRODUCT 927 if (G1PrintReachableAtInitialMark) { 928 print_reachable("at-cycle-start", 929 VerifyOption_G1UsePrevMarking, true /* all */); 930 } 931 #endif 932 933 // Initialize marking structures. This has to be done in a STW phase. 934 reset(); 935 936 // For each region note start of marking. 937 NoteStartOfMarkHRClosure startcl; 938 g1h->heap_region_iterate(&startcl); 939 } 940 941 942 void ConcurrentMark::checkpointRootsInitialPost() { 943 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 944 945 // If we force an overflow during remark, the remark operation will 946 // actually abort and we'll restart concurrent marking. If we always 947 // force an overflow during remark we'll never actually complete the 948 // marking phase. So, we initialize this here, at the start of the 949 // cycle, so that at the remaining overflow number will decrease at 950 // every remark and we'll eventually not need to cause one. 951 force_overflow_stw()->init(); 952 953 // Start Concurrent Marking weak-reference discovery. 954 ReferenceProcessor* rp = g1h->ref_processor_cm(); 955 // enable ("weak") refs discovery 956 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 957 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 958 959 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 960 // This is the start of the marking cycle, we're expected all 961 // threads to have SATB queues with active set to false. 962 satb_mq_set.set_active_all_threads(true, /* new active value */ 963 false /* expected_active */); 964 965 _root_regions.prepare_for_scan(); 966 967 // update_g1_committed() will be called at the end of an evac pause 968 // when marking is on. So, it's also called at the end of the 969 // initial-mark pause to update the heap end, if the heap expands 970 // during it. No need to call it here. 971 } 972 973 /* 974 * Notice that in the next two methods, we actually leave the STS 975 * during the barrier sync and join it immediately afterwards. If we 976 * do not do this, the following deadlock can occur: one thread could 977 * be in the barrier sync code, waiting for the other thread to also 978 * sync up, whereas another one could be trying to yield, while also 979 * waiting for the other threads to sync up too. 980 * 981 * Note, however, that this code is also used during remark and in 982 * this case we should not attempt to leave / enter the STS, otherwise 983 * we'll either hit an assert (debug / fastdebug) or deadlock 984 * (product). So we should only leave / enter the STS if we are 985 * operating concurrently. 986 * 987 * Because the thread that does the sync barrier has left the STS, it 988 * is possible to be suspended for a Full GC or an evacuation pause 989 * could occur. This is actually safe, since the entering the sync 990 * barrier is one of the last things do_marking_step() does, and it 991 * doesn't manipulate any data structures afterwards. 992 */ 993 994 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 995 if (verbose_low()) { 996 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id); 997 } 998 999 if (concurrent()) { 1000 SuspendibleThreadSet::leave(); 1001 } 1002 1003 bool barrier_aborted = !_first_overflow_barrier_sync.enter(); 1004 1005 if (concurrent()) { 1006 SuspendibleThreadSet::join(); 1007 } 1008 // at this point everyone should have synced up and not be doing any 1009 // more work 1010 1011 if (verbose_low()) { 1012 if (barrier_aborted) { 1013 gclog_or_tty->print_cr("[%u] aborted first barrier", worker_id); 1014 } else { 1015 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id); 1016 } 1017 } 1018 1019 if (barrier_aborted) { 1020 // If the barrier aborted we ignore the overflow condition and 1021 // just abort the whole marking phase as quickly as possible. 1022 return; 1023 } 1024 1025 // If we're executing the concurrent phase of marking, reset the marking 1026 // state; otherwise the marking state is reset after reference processing, 1027 // during the remark pause. 1028 // If we reset here as a result of an overflow during the remark we will 1029 // see assertion failures from any subsequent set_concurrency_and_phase() 1030 // calls. 1031 if (concurrent()) { 1032 // let the task associated with with worker 0 do this 1033 if (worker_id == 0) { 1034 // task 0 is responsible for clearing the global data structures 1035 // We should be here because of an overflow. During STW we should 1036 // not clear the overflow flag since we rely on it being true when 1037 // we exit this method to abort the pause and restart concurrent 1038 // marking. 1039 reset_marking_state(true /* clear_overflow */); 1040 force_overflow()->update(); 1041 1042 if (G1Log::fine()) { 1043 gclog_or_tty->gclog_stamp(concurrent_gc_id()); 1044 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]"); 1045 } 1046 } 1047 } 1048 1049 // after this, each task should reset its own data structures then 1050 // then go into the second barrier 1051 } 1052 1053 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 1054 if (verbose_low()) { 1055 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id); 1056 } 1057 1058 if (concurrent()) { 1059 SuspendibleThreadSet::leave(); 1060 } 1061 1062 bool barrier_aborted = !_second_overflow_barrier_sync.enter(); 1063 1064 if (concurrent()) { 1065 SuspendibleThreadSet::join(); 1066 } 1067 // at this point everything should be re-initialized and ready to go 1068 1069 if (verbose_low()) { 1070 if (barrier_aborted) { 1071 gclog_or_tty->print_cr("[%u] aborted second barrier", worker_id); 1072 } else { 1073 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id); 1074 } 1075 } 1076 } 1077 1078 #ifndef PRODUCT 1079 void ForceOverflowSettings::init() { 1080 _num_remaining = G1ConcMarkForceOverflow; 1081 _force = false; 1082 update(); 1083 } 1084 1085 void ForceOverflowSettings::update() { 1086 if (_num_remaining > 0) { 1087 _num_remaining -= 1; 1088 _force = true; 1089 } else { 1090 _force = false; 1091 } 1092 } 1093 1094 bool ForceOverflowSettings::should_force() { 1095 if (_force) { 1096 _force = false; 1097 return true; 1098 } else { 1099 return false; 1100 } 1101 } 1102 #endif // !PRODUCT 1103 1104 class CMConcurrentMarkingTask: public AbstractGangTask { 1105 private: 1106 ConcurrentMark* _cm; 1107 ConcurrentMarkThread* _cmt; 1108 1109 public: 1110 void work(uint worker_id) { 1111 assert(Thread::current()->is_ConcurrentGC_thread(), 1112 "this should only be done by a conc GC thread"); 1113 ResourceMark rm; 1114 1115 double start_vtime = os::elapsedVTime(); 1116 1117 SuspendibleThreadSet::join(); 1118 1119 assert(worker_id < _cm->active_tasks(), "invariant"); 1120 CMTask* the_task = _cm->task(worker_id); 1121 the_task->record_start_time(); 1122 if (!_cm->has_aborted()) { 1123 do { 1124 double start_vtime_sec = os::elapsedVTime(); 1125 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1126 1127 the_task->do_marking_step(mark_step_duration_ms, 1128 true /* do_termination */, 1129 false /* is_serial*/); 1130 1131 double end_vtime_sec = os::elapsedVTime(); 1132 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 1133 _cm->clear_has_overflown(); 1134 1135 _cm->do_yield_check(worker_id); 1136 1137 jlong sleep_time_ms; 1138 if (!_cm->has_aborted() && the_task->has_aborted()) { 1139 sleep_time_ms = 1140 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 1141 SuspendibleThreadSet::leave(); 1142 os::sleep(Thread::current(), sleep_time_ms, false); 1143 SuspendibleThreadSet::join(); 1144 } 1145 } while (!_cm->has_aborted() && the_task->has_aborted()); 1146 } 1147 the_task->record_end_time(); 1148 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 1149 1150 SuspendibleThreadSet::leave(); 1151 1152 double end_vtime = os::elapsedVTime(); 1153 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 1154 } 1155 1156 CMConcurrentMarkingTask(ConcurrentMark* cm, 1157 ConcurrentMarkThread* cmt) : 1158 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 1159 1160 ~CMConcurrentMarkingTask() { } 1161 }; 1162 1163 // Calculates the number of active workers for a concurrent 1164 // phase. 1165 uint ConcurrentMark::calc_parallel_marking_threads() { 1166 if (G1CollectedHeap::use_parallel_gc_threads()) { 1167 uint n_conc_workers = 0; 1168 if (!UseDynamicNumberOfGCThreads || 1169 (!FLAG_IS_DEFAULT(ConcGCThreads) && 1170 !ForceDynamicNumberOfGCThreads)) { 1171 n_conc_workers = max_parallel_marking_threads(); 1172 } else { 1173 n_conc_workers = 1174 AdaptiveSizePolicy::calc_default_active_workers( 1175 max_parallel_marking_threads(), 1176 1, /* Minimum workers */ 1177 parallel_marking_threads(), 1178 Threads::number_of_non_daemon_threads()); 1179 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 1180 // that scaling has already gone into "_max_parallel_marking_threads". 1181 } 1182 assert(n_conc_workers > 0, "Always need at least 1"); 1183 return n_conc_workers; 1184 } 1185 // If we are not running with any parallel GC threads we will not 1186 // have spawned any marking threads either. Hence the number of 1187 // concurrent workers should be 0. 1188 return 0; 1189 } 1190 1191 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) { 1192 // Currently, only survivors can be root regions. 1193 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 1194 G1RootRegionScanClosure cl(_g1h, this, worker_id); 1195 1196 const uintx interval = PrefetchScanIntervalInBytes; 1197 HeapWord* curr = hr->bottom(); 1198 const HeapWord* end = hr->top(); 1199 while (curr < end) { 1200 Prefetch::read(curr, interval); 1201 oop obj = oop(curr); 1202 int size = obj->oop_iterate(&cl); 1203 assert(size == obj->size(), "sanity"); 1204 curr += size; 1205 } 1206 } 1207 1208 class CMRootRegionScanTask : public AbstractGangTask { 1209 private: 1210 ConcurrentMark* _cm; 1211 1212 public: 1213 CMRootRegionScanTask(ConcurrentMark* cm) : 1214 AbstractGangTask("Root Region Scan"), _cm(cm) { } 1215 1216 void work(uint worker_id) { 1217 assert(Thread::current()->is_ConcurrentGC_thread(), 1218 "this should only be done by a conc GC thread"); 1219 1220 CMRootRegions* root_regions = _cm->root_regions(); 1221 HeapRegion* hr = root_regions->claim_next(); 1222 while (hr != NULL) { 1223 _cm->scanRootRegion(hr, worker_id); 1224 hr = root_regions->claim_next(); 1225 } 1226 } 1227 }; 1228 1229 void ConcurrentMark::scanRootRegions() { 1230 // Start of concurrent marking. 1231 ClassLoaderDataGraph::clear_claimed_marks(); 1232 1233 // scan_in_progress() will have been set to true only if there was 1234 // at least one root region to scan. So, if it's false, we 1235 // should not attempt to do any further work. 1236 if (root_regions()->scan_in_progress()) { 1237 _parallel_marking_threads = calc_parallel_marking_threads(); 1238 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1239 "Maximum number of marking threads exceeded"); 1240 uint active_workers = MAX2(1U, parallel_marking_threads()); 1241 1242 CMRootRegionScanTask task(this); 1243 if (use_parallel_marking_threads()) { 1244 _parallel_workers->set_active_workers((int) active_workers); 1245 _parallel_workers->run_task(&task); 1246 } else { 1247 task.work(0); 1248 } 1249 1250 // It's possible that has_aborted() is true here without actually 1251 // aborting the survivor scan earlier. This is OK as it's 1252 // mainly used for sanity checking. 1253 root_regions()->scan_finished(); 1254 } 1255 } 1256 1257 void ConcurrentMark::markFromRoots() { 1258 // we might be tempted to assert that: 1259 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1260 // "inconsistent argument?"); 1261 // However that wouldn't be right, because it's possible that 1262 // a safepoint is indeed in progress as a younger generation 1263 // stop-the-world GC happens even as we mark in this generation. 1264 1265 _restart_for_overflow = false; 1266 force_overflow_conc()->init(); 1267 1268 // _g1h has _n_par_threads 1269 _parallel_marking_threads = calc_parallel_marking_threads(); 1270 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1271 "Maximum number of marking threads exceeded"); 1272 1273 uint active_workers = MAX2(1U, parallel_marking_threads()); 1274 1275 // Parallel task terminator is set in "set_concurrency_and_phase()" 1276 set_concurrency_and_phase(active_workers, true /* concurrent */); 1277 1278 CMConcurrentMarkingTask markingTask(this, cmThread()); 1279 if (use_parallel_marking_threads()) { 1280 _parallel_workers->set_active_workers((int)active_workers); 1281 // Don't set _n_par_threads because it affects MT in process_roots() 1282 // and the decisions on that MT processing is made elsewhere. 1283 assert(_parallel_workers->active_workers() > 0, "Should have been set"); 1284 _parallel_workers->run_task(&markingTask); 1285 } else { 1286 markingTask.work(0); 1287 } 1288 print_stats(); 1289 } 1290 1291 // Helper class to get rid of some boilerplate code. 1292 class G1CMTraceTime : public GCTraceTime { 1293 static bool doit_and_prepend(bool doit) { 1294 if (doit) { 1295 gclog_or_tty->put(' '); 1296 } 1297 return doit; 1298 } 1299 1300 public: 1301 G1CMTraceTime(const char* title, bool doit) 1302 : GCTraceTime(title, doit_and_prepend(doit), false, G1CollectedHeap::heap()->gc_timer_cm(), 1303 G1CollectedHeap::heap()->concurrent_mark()->concurrent_gc_id()) { 1304 } 1305 }; 1306 1307 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1308 // world is stopped at this checkpoint 1309 assert(SafepointSynchronize::is_at_safepoint(), 1310 "world should be stopped"); 1311 1312 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1313 1314 // If a full collection has happened, we shouldn't do this. 1315 if (has_aborted()) { 1316 g1h->set_marking_complete(); // So bitmap clearing isn't confused 1317 return; 1318 } 1319 1320 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1321 1322 if (VerifyDuringGC) { 1323 HandleMark hm; // handle scope 1324 Universe::heap()->prepare_for_verify(); 1325 Universe::verify(VerifyOption_G1UsePrevMarking, 1326 " VerifyDuringGC:(before)"); 1327 } 1328 g1h->check_bitmaps("Remark Start"); 1329 1330 G1CollectorPolicy* g1p = g1h->g1_policy(); 1331 g1p->record_concurrent_mark_remark_start(); 1332 1333 double start = os::elapsedTime(); 1334 1335 checkpointRootsFinalWork(); 1336 1337 double mark_work_end = os::elapsedTime(); 1338 1339 weakRefsWork(clear_all_soft_refs); 1340 1341 if (has_overflown()) { 1342 // Oops. We overflowed. Restart concurrent marking. 1343 _restart_for_overflow = true; 1344 if (G1TraceMarkStackOverflow) { 1345 gclog_or_tty->print_cr("\nRemark led to restart for overflow."); 1346 } 1347 1348 // Verify the heap w.r.t. the previous marking bitmap. 1349 if (VerifyDuringGC) { 1350 HandleMark hm; // handle scope 1351 Universe::heap()->prepare_for_verify(); 1352 Universe::verify(VerifyOption_G1UsePrevMarking, 1353 " VerifyDuringGC:(overflow)"); 1354 } 1355 1356 // Clear the marking state because we will be restarting 1357 // marking due to overflowing the global mark stack. 1358 reset_marking_state(); 1359 } else { 1360 { 1361 G1CMTraceTime trace("GC aggregate-data", G1Log::finer()); 1362 1363 // Aggregate the per-task counting data that we have accumulated 1364 // while marking. 1365 aggregate_count_data(); 1366 } 1367 1368 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1369 // We're done with marking. 1370 // This is the end of the marking cycle, we're expected all 1371 // threads to have SATB queues with active set to true. 1372 satb_mq_set.set_active_all_threads(false, /* new active value */ 1373 true /* expected_active */); 1374 1375 if (VerifyDuringGC) { 1376 HandleMark hm; // handle scope 1377 Universe::heap()->prepare_for_verify(); 1378 Universe::verify(VerifyOption_G1UseNextMarking, 1379 " VerifyDuringGC:(after)"); 1380 } 1381 g1h->check_bitmaps("Remark End"); 1382 assert(!restart_for_overflow(), "sanity"); 1383 // Completely reset the marking state since marking completed 1384 set_non_marking_state(); 1385 } 1386 1387 // Expand the marking stack, if we have to and if we can. 1388 if (_markStack.should_expand()) { 1389 _markStack.expand(); 1390 } 1391 1392 // Statistics 1393 double now = os::elapsedTime(); 1394 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1395 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1396 _remark_times.add((now - start) * 1000.0); 1397 1398 g1p->record_concurrent_mark_remark_end(); 1399 1400 G1CMIsAliveClosure is_alive(g1h); 1401 g1h->gc_tracer_cm()->report_object_count_after_gc(&is_alive); 1402 } 1403 1404 // Base class of the closures that finalize and verify the 1405 // liveness counting data. 1406 class CMCountDataClosureBase: public HeapRegionClosure { 1407 protected: 1408 G1CollectedHeap* _g1h; 1409 ConcurrentMark* _cm; 1410 CardTableModRefBS* _ct_bs; 1411 1412 BitMap* _region_bm; 1413 BitMap* _card_bm; 1414 1415 // Takes a region that's not empty (i.e., it has at least one 1416 // live object in it and sets its corresponding bit on the region 1417 // bitmap to 1. If the region is "starts humongous" it will also set 1418 // to 1 the bits on the region bitmap that correspond to its 1419 // associated "continues humongous" regions. 1420 void set_bit_for_region(HeapRegion* hr) { 1421 assert(!hr->is_continues_humongous(), "should have filtered those out"); 1422 1423 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index(); 1424 if (!hr->is_starts_humongous()) { 1425 // Normal (non-humongous) case: just set the bit. 1426 _region_bm->par_at_put(index, true); 1427 } else { 1428 // Starts humongous case: calculate how many regions are part of 1429 // this humongous region and then set the bit range. 1430 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index(); 1431 _region_bm->par_at_put_range(index, end_index, true); 1432 } 1433 } 1434 1435 public: 1436 CMCountDataClosureBase(G1CollectedHeap* g1h, 1437 BitMap* region_bm, BitMap* card_bm): 1438 _g1h(g1h), _cm(g1h->concurrent_mark()), 1439 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 1440 _region_bm(region_bm), _card_bm(card_bm) { } 1441 }; 1442 1443 // Closure that calculates the # live objects per region. Used 1444 // for verification purposes during the cleanup pause. 1445 class CalcLiveObjectsClosure: public CMCountDataClosureBase { 1446 CMBitMapRO* _bm; 1447 size_t _region_marked_bytes; 1448 1449 public: 1450 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h, 1451 BitMap* region_bm, BitMap* card_bm) : 1452 CMCountDataClosureBase(g1h, region_bm, card_bm), 1453 _bm(bm), _region_marked_bytes(0) { } 1454 1455 bool doHeapRegion(HeapRegion* hr) { 1456 1457 if (hr->is_continues_humongous()) { 1458 // We will ignore these here and process them when their 1459 // associated "starts humongous" region is processed (see 1460 // set_bit_for_heap_region()). Note that we cannot rely on their 1461 // associated "starts humongous" region to have their bit set to 1462 // 1 since, due to the region chunking in the parallel region 1463 // iteration, a "continues humongous" region might be visited 1464 // before its associated "starts humongous". 1465 return false; 1466 } 1467 1468 HeapWord* ntams = hr->next_top_at_mark_start(); 1469 HeapWord* start = hr->bottom(); 1470 1471 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(), 1472 err_msg("Preconditions not met - " 1473 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT, 1474 p2i(start), p2i(ntams), p2i(hr->end()))); 1475 1476 // Find the first marked object at or after "start". 1477 start = _bm->getNextMarkedWordAddress(start, ntams); 1478 1479 size_t marked_bytes = 0; 1480 1481 while (start < ntams) { 1482 oop obj = oop(start); 1483 int obj_sz = obj->size(); 1484 HeapWord* obj_end = start + obj_sz; 1485 1486 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 1487 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end); 1488 1489 // Note: if we're looking at the last region in heap - obj_end 1490 // could be actually just beyond the end of the heap; end_idx 1491 // will then correspond to a (non-existent) card that is also 1492 // just beyond the heap. 1493 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) { 1494 // end of object is not card aligned - increment to cover 1495 // all the cards spanned by the object 1496 end_idx += 1; 1497 } 1498 1499 // Set the bits in the card BM for the cards spanned by this object. 1500 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1501 1502 // Add the size of this object to the number of marked bytes. 1503 marked_bytes += (size_t)obj_sz * HeapWordSize; 1504 1505 // Find the next marked object after this one. 1506 start = _bm->getNextMarkedWordAddress(obj_end, ntams); 1507 } 1508 1509 // Mark the allocated-since-marking portion... 1510 HeapWord* top = hr->top(); 1511 if (ntams < top) { 1512 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1513 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1514 1515 // Note: if we're looking at the last region in heap - top 1516 // could be actually just beyond the end of the heap; end_idx 1517 // will then correspond to a (non-existent) card that is also 1518 // just beyond the heap. 1519 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1520 // end of object is not card aligned - increment to cover 1521 // all the cards spanned by the object 1522 end_idx += 1; 1523 } 1524 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1525 1526 // This definitely means the region has live objects. 1527 set_bit_for_region(hr); 1528 } 1529 1530 // Update the live region bitmap. 1531 if (marked_bytes > 0) { 1532 set_bit_for_region(hr); 1533 } 1534 1535 // Set the marked bytes for the current region so that 1536 // it can be queried by a calling verification routine 1537 _region_marked_bytes = marked_bytes; 1538 1539 return false; 1540 } 1541 1542 size_t region_marked_bytes() const { return _region_marked_bytes; } 1543 }; 1544 1545 // Heap region closure used for verifying the counting data 1546 // that was accumulated concurrently and aggregated during 1547 // the remark pause. This closure is applied to the heap 1548 // regions during the STW cleanup pause. 1549 1550 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure { 1551 G1CollectedHeap* _g1h; 1552 ConcurrentMark* _cm; 1553 CalcLiveObjectsClosure _calc_cl; 1554 BitMap* _region_bm; // Region BM to be verified 1555 BitMap* _card_bm; // Card BM to be verified 1556 bool _verbose; // verbose output? 1557 1558 BitMap* _exp_region_bm; // Expected Region BM values 1559 BitMap* _exp_card_bm; // Expected card BM values 1560 1561 int _failures; 1562 1563 public: 1564 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h, 1565 BitMap* region_bm, 1566 BitMap* card_bm, 1567 BitMap* exp_region_bm, 1568 BitMap* exp_card_bm, 1569 bool verbose) : 1570 _g1h(g1h), _cm(g1h->concurrent_mark()), 1571 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm), 1572 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose), 1573 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm), 1574 _failures(0) { } 1575 1576 int failures() const { return _failures; } 1577 1578 bool doHeapRegion(HeapRegion* hr) { 1579 if (hr->is_continues_humongous()) { 1580 // We will ignore these here and process them when their 1581 // associated "starts humongous" region is processed (see 1582 // set_bit_for_heap_region()). Note that we cannot rely on their 1583 // associated "starts humongous" region to have their bit set to 1584 // 1 since, due to the region chunking in the parallel region 1585 // iteration, a "continues humongous" region might be visited 1586 // before its associated "starts humongous". 1587 return false; 1588 } 1589 1590 int failures = 0; 1591 1592 // Call the CalcLiveObjectsClosure to walk the marking bitmap for 1593 // this region and set the corresponding bits in the expected region 1594 // and card bitmaps. 1595 bool res = _calc_cl.doHeapRegion(hr); 1596 assert(res == false, "should be continuing"); 1597 1598 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL), 1599 Mutex::_no_safepoint_check_flag); 1600 1601 // Verify the marked bytes for this region. 1602 size_t exp_marked_bytes = _calc_cl.region_marked_bytes(); 1603 size_t act_marked_bytes = hr->next_marked_bytes(); 1604 1605 // We're not OK if expected marked bytes > actual marked bytes. It means 1606 // we have missed accounting some objects during the actual marking. 1607 if (exp_marked_bytes > act_marked_bytes) { 1608 if (_verbose) { 1609 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: " 1610 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT, 1611 hr->hrm_index(), exp_marked_bytes, act_marked_bytes); 1612 } 1613 failures += 1; 1614 } 1615 1616 // Verify the bit, for this region, in the actual and expected 1617 // (which was just calculated) region bit maps. 1618 // We're not OK if the bit in the calculated expected region 1619 // bitmap is set and the bit in the actual region bitmap is not. 1620 BitMap::idx_t index = (BitMap::idx_t) hr->hrm_index(); 1621 1622 bool expected = _exp_region_bm->at(index); 1623 bool actual = _region_bm->at(index); 1624 if (expected && !actual) { 1625 if (_verbose) { 1626 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: " 1627 "expected: %s, actual: %s", 1628 hr->hrm_index(), 1629 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1630 } 1631 failures += 1; 1632 } 1633 1634 // Verify that the card bit maps for the cards spanned by the current 1635 // region match. We have an error if we have a set bit in the expected 1636 // bit map and the corresponding bit in the actual bitmap is not set. 1637 1638 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom()); 1639 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top()); 1640 1641 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) { 1642 expected = _exp_card_bm->at(i); 1643 actual = _card_bm->at(i); 1644 1645 if (expected && !actual) { 1646 if (_verbose) { 1647 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": " 1648 "expected: %s, actual: %s", 1649 hr->hrm_index(), i, 1650 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1651 } 1652 failures += 1; 1653 } 1654 } 1655 1656 if (failures > 0 && _verbose) { 1657 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", " 1658 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT, 1659 HR_FORMAT_PARAMS(hr), p2i(hr->next_top_at_mark_start()), 1660 _calc_cl.region_marked_bytes(), hr->next_marked_bytes()); 1661 } 1662 1663 _failures += failures; 1664 1665 // We could stop iteration over the heap when we 1666 // find the first violating region by returning true. 1667 return false; 1668 } 1669 }; 1670 1671 class G1ParVerifyFinalCountTask: public AbstractGangTask { 1672 protected: 1673 G1CollectedHeap* _g1h; 1674 ConcurrentMark* _cm; 1675 BitMap* _actual_region_bm; 1676 BitMap* _actual_card_bm; 1677 1678 uint _n_workers; 1679 1680 BitMap* _expected_region_bm; 1681 BitMap* _expected_card_bm; 1682 1683 int _failures; 1684 bool _verbose; 1685 1686 HeapRegionClaimer _hrclaimer; 1687 1688 public: 1689 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h, 1690 BitMap* region_bm, BitMap* card_bm, 1691 BitMap* expected_region_bm, BitMap* expected_card_bm) 1692 : AbstractGangTask("G1 verify final counting"), 1693 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1694 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1695 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm), 1696 _failures(0), _verbose(false), 1697 _n_workers(0) { 1698 assert(VerifyDuringGC, "don't call this otherwise"); 1699 1700 // Use the value already set as the number of active threads 1701 // in the call to run_task(). 1702 if (G1CollectedHeap::use_parallel_gc_threads()) { 1703 assert( _g1h->workers()->active_workers() > 0, 1704 "Should have been previously set"); 1705 _n_workers = _g1h->workers()->active_workers(); 1706 _hrclaimer.initialize(_n_workers); 1707 } else { 1708 _n_workers = 1; 1709 } 1710 1711 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity"); 1712 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity"); 1713 1714 _verbose = _cm->verbose_medium(); 1715 } 1716 1717 void work(uint worker_id) { 1718 assert(worker_id < _n_workers, "invariant"); 1719 1720 VerifyLiveObjectDataHRClosure verify_cl(_g1h, 1721 _actual_region_bm, _actual_card_bm, 1722 _expected_region_bm, 1723 _expected_card_bm, 1724 _verbose); 1725 1726 if (G1CollectedHeap::use_parallel_gc_threads()) { 1727 _g1h->heap_region_par_iterate(&verify_cl, worker_id, &_hrclaimer); 1728 } else { 1729 _g1h->heap_region_iterate(&verify_cl); 1730 } 1731 1732 Atomic::add(verify_cl.failures(), &_failures); 1733 } 1734 1735 int failures() const { return _failures; } 1736 }; 1737 1738 // Closure that finalizes the liveness counting data. 1739 // Used during the cleanup pause. 1740 // Sets the bits corresponding to the interval [NTAMS, top] 1741 // (which contains the implicitly live objects) in the 1742 // card liveness bitmap. Also sets the bit for each region, 1743 // containing live data, in the region liveness bitmap. 1744 1745 class FinalCountDataUpdateClosure: public CMCountDataClosureBase { 1746 public: 1747 FinalCountDataUpdateClosure(G1CollectedHeap* g1h, 1748 BitMap* region_bm, 1749 BitMap* card_bm) : 1750 CMCountDataClosureBase(g1h, region_bm, card_bm) { } 1751 1752 bool doHeapRegion(HeapRegion* hr) { 1753 1754 if (hr->is_continues_humongous()) { 1755 // We will ignore these here and process them when their 1756 // associated "starts humongous" region is processed (see 1757 // set_bit_for_heap_region()). Note that we cannot rely on their 1758 // associated "starts humongous" region to have their bit set to 1759 // 1 since, due to the region chunking in the parallel region 1760 // iteration, a "continues humongous" region might be visited 1761 // before its associated "starts humongous". 1762 return false; 1763 } 1764 1765 HeapWord* ntams = hr->next_top_at_mark_start(); 1766 HeapWord* top = hr->top(); 1767 1768 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions."); 1769 1770 // Mark the allocated-since-marking portion... 1771 if (ntams < top) { 1772 // This definitely means the region has live objects. 1773 set_bit_for_region(hr); 1774 1775 // Now set the bits in the card bitmap for [ntams, top) 1776 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1777 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1778 1779 // Note: if we're looking at the last region in heap - top 1780 // could be actually just beyond the end of the heap; end_idx 1781 // will then correspond to a (non-existent) card that is also 1782 // just beyond the heap. 1783 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1784 // end of object is not card aligned - increment to cover 1785 // all the cards spanned by the object 1786 end_idx += 1; 1787 } 1788 1789 assert(end_idx <= _card_bm->size(), 1790 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1791 end_idx, _card_bm->size())); 1792 assert(start_idx < _card_bm->size(), 1793 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1794 start_idx, _card_bm->size())); 1795 1796 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1797 } 1798 1799 // Set the bit for the region if it contains live data 1800 if (hr->next_marked_bytes() > 0) { 1801 set_bit_for_region(hr); 1802 } 1803 1804 return false; 1805 } 1806 }; 1807 1808 class G1ParFinalCountTask: public AbstractGangTask { 1809 protected: 1810 G1CollectedHeap* _g1h; 1811 ConcurrentMark* _cm; 1812 BitMap* _actual_region_bm; 1813 BitMap* _actual_card_bm; 1814 1815 uint _n_workers; 1816 HeapRegionClaimer _hrclaimer; 1817 1818 public: 1819 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm) 1820 : AbstractGangTask("G1 final counting"), 1821 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1822 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1823 _n_workers(0) { 1824 // Use the value already set as the number of active threads 1825 // in the call to run_task(). 1826 if (G1CollectedHeap::use_parallel_gc_threads()) { 1827 assert( _g1h->workers()->active_workers() > 0, 1828 "Should have been previously set"); 1829 _n_workers = _g1h->workers()->active_workers(); 1830 _hrclaimer.initialize(_n_workers); 1831 } else { 1832 _n_workers = 1; 1833 } 1834 } 1835 1836 void work(uint worker_id) { 1837 assert(worker_id < _n_workers, "invariant"); 1838 1839 FinalCountDataUpdateClosure final_update_cl(_g1h, 1840 _actual_region_bm, 1841 _actual_card_bm); 1842 1843 if (G1CollectedHeap::use_parallel_gc_threads()) { 1844 _g1h->heap_region_par_iterate(&final_update_cl, worker_id, &_hrclaimer); 1845 } else { 1846 _g1h->heap_region_iterate(&final_update_cl); 1847 } 1848 } 1849 }; 1850 1851 class G1ParNoteEndTask; 1852 1853 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1854 G1CollectedHeap* _g1; 1855 size_t _max_live_bytes; 1856 uint _regions_claimed; 1857 size_t _freed_bytes; 1858 FreeRegionList* _local_cleanup_list; 1859 HeapRegionSetCount _old_regions_removed; 1860 HeapRegionSetCount _humongous_regions_removed; 1861 HRRSCleanupTask* _hrrs_cleanup_task; 1862 double _claimed_region_time; 1863 double _max_region_time; 1864 1865 public: 1866 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1867 FreeRegionList* local_cleanup_list, 1868 HRRSCleanupTask* hrrs_cleanup_task) : 1869 _g1(g1), 1870 _max_live_bytes(0), _regions_claimed(0), 1871 _freed_bytes(0), 1872 _claimed_region_time(0.0), _max_region_time(0.0), 1873 _local_cleanup_list(local_cleanup_list), 1874 _old_regions_removed(), 1875 _humongous_regions_removed(), 1876 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1877 1878 size_t freed_bytes() { return _freed_bytes; } 1879 const HeapRegionSetCount& old_regions_removed() { return _old_regions_removed; } 1880 const HeapRegionSetCount& humongous_regions_removed() { return _humongous_regions_removed; } 1881 1882 bool doHeapRegion(HeapRegion *hr) { 1883 if (hr->is_continues_humongous()) { 1884 return false; 1885 } 1886 // We use a claim value of zero here because all regions 1887 // were claimed with value 1 in the FinalCount task. 1888 _g1->reset_gc_time_stamps(hr); 1889 double start = os::elapsedTime(); 1890 _regions_claimed++; 1891 hr->note_end_of_marking(); 1892 _max_live_bytes += hr->max_live_bytes(); 1893 1894 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 1895 _freed_bytes += hr->used(); 1896 hr->set_containing_set(NULL); 1897 if (hr->is_humongous()) { 1898 assert(hr->is_starts_humongous(), "we should only see starts humongous"); 1899 _humongous_regions_removed.increment(1u, hr->capacity()); 1900 _g1->free_humongous_region(hr, _local_cleanup_list, true); 1901 } else { 1902 _old_regions_removed.increment(1u, hr->capacity()); 1903 _g1->free_region(hr, _local_cleanup_list, true); 1904 } 1905 } else { 1906 hr->rem_set()->do_cleanup_work(_hrrs_cleanup_task); 1907 } 1908 1909 double region_time = (os::elapsedTime() - start); 1910 _claimed_region_time += region_time; 1911 if (region_time > _max_region_time) { 1912 _max_region_time = region_time; 1913 } 1914 return false; 1915 } 1916 1917 size_t max_live_bytes() { return _max_live_bytes; } 1918 uint regions_claimed() { return _regions_claimed; } 1919 double claimed_region_time_sec() { return _claimed_region_time; } 1920 double max_region_time_sec() { return _max_region_time; } 1921 }; 1922 1923 class G1ParNoteEndTask: public AbstractGangTask { 1924 friend class G1NoteEndOfConcMarkClosure; 1925 1926 protected: 1927 G1CollectedHeap* _g1h; 1928 size_t _max_live_bytes; 1929 size_t _freed_bytes; 1930 FreeRegionList* _cleanup_list; 1931 HeapRegionClaimer _hrclaimer; 1932 1933 public: 1934 G1ParNoteEndTask(G1CollectedHeap* g1h, FreeRegionList* cleanup_list, uint n_workers) : 1935 AbstractGangTask("G1 note end"), _g1h(g1h), _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { 1936 if (G1CollectedHeap::use_parallel_gc_threads()) { 1937 _hrclaimer.initialize(n_workers); 1938 } 1939 } 1940 1941 void work(uint worker_id) { 1942 double start = os::elapsedTime(); 1943 FreeRegionList local_cleanup_list("Local Cleanup List"); 1944 HRRSCleanupTask hrrs_cleanup_task; 1945 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, &local_cleanup_list, 1946 &hrrs_cleanup_task); 1947 if (G1CollectedHeap::use_parallel_gc_threads()) { 1948 _g1h->heap_region_par_iterate(&g1_note_end, worker_id, &_hrclaimer); 1949 } else { 1950 _g1h->heap_region_iterate(&g1_note_end); 1951 } 1952 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1953 1954 // Now update the lists 1955 _g1h->remove_from_old_sets(g1_note_end.old_regions_removed(), g1_note_end.humongous_regions_removed()); 1956 { 1957 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1958 _g1h->decrement_summary_bytes(g1_note_end.freed_bytes()); 1959 _max_live_bytes += g1_note_end.max_live_bytes(); 1960 _freed_bytes += g1_note_end.freed_bytes(); 1961 1962 // If we iterate over the global cleanup list at the end of 1963 // cleanup to do this printing we will not guarantee to only 1964 // generate output for the newly-reclaimed regions (the list 1965 // might not be empty at the beginning of cleanup; we might 1966 // still be working on its previous contents). So we do the 1967 // printing here, before we append the new regions to the global 1968 // cleanup list. 1969 1970 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1971 if (hr_printer->is_active()) { 1972 FreeRegionListIterator iter(&local_cleanup_list); 1973 while (iter.more_available()) { 1974 HeapRegion* hr = iter.get_next(); 1975 hr_printer->cleanup(hr); 1976 } 1977 } 1978 1979 _cleanup_list->add_ordered(&local_cleanup_list); 1980 assert(local_cleanup_list.is_empty(), "post-condition"); 1981 1982 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1983 } 1984 } 1985 size_t max_live_bytes() { return _max_live_bytes; } 1986 size_t freed_bytes() { return _freed_bytes; } 1987 }; 1988 1989 class G1ParScrubRemSetTask: public AbstractGangTask { 1990 protected: 1991 G1RemSet* _g1rs; 1992 BitMap* _region_bm; 1993 BitMap* _card_bm; 1994 HeapRegionClaimer _hrclaimer; 1995 1996 public: 1997 G1ParScrubRemSetTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm, uint n_workers) : 1998 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), _region_bm(region_bm), _card_bm(card_bm) { 1999 if (G1CollectedHeap::use_parallel_gc_threads()) { 2000 _hrclaimer.initialize(n_workers); 2001 } 2002 } 2003 2004 void work(uint worker_id) { 2005 if (G1CollectedHeap::use_parallel_gc_threads()) { 2006 _g1rs->scrub_par(_region_bm, _card_bm, worker_id, &_hrclaimer); 2007 } else { 2008 _g1rs->scrub(_region_bm, _card_bm); 2009 } 2010 } 2011 2012 }; 2013 2014 void ConcurrentMark::cleanup() { 2015 // world is stopped at this checkpoint 2016 assert(SafepointSynchronize::is_at_safepoint(), 2017 "world should be stopped"); 2018 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2019 2020 // If a full collection has happened, we shouldn't do this. 2021 if (has_aborted()) { 2022 g1h->set_marking_complete(); // So bitmap clearing isn't confused 2023 return; 2024 } 2025 2026 g1h->verify_region_sets_optional(); 2027 2028 if (VerifyDuringGC) { 2029 HandleMark hm; // handle scope 2030 Universe::heap()->prepare_for_verify(); 2031 Universe::verify(VerifyOption_G1UsePrevMarking, 2032 " VerifyDuringGC:(before)"); 2033 } 2034 g1h->check_bitmaps("Cleanup Start"); 2035 2036 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); 2037 g1p->record_concurrent_mark_cleanup_start(); 2038 2039 double start = os::elapsedTime(); 2040 2041 HeapRegionRemSet::reset_for_cleanup_tasks(); 2042 2043 uint n_workers; 2044 2045 // Do counting once more with the world stopped for good measure. 2046 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm); 2047 2048 if (G1CollectedHeap::use_parallel_gc_threads()) { 2049 g1h->set_par_threads(); 2050 n_workers = g1h->n_par_threads(); 2051 assert(g1h->n_par_threads() == n_workers, 2052 "Should not have been reset"); 2053 g1h->workers()->run_task(&g1_par_count_task); 2054 // Done with the parallel phase so reset to 0. 2055 g1h->set_par_threads(0); 2056 } else { 2057 n_workers = 1; 2058 g1_par_count_task.work(0); 2059 } 2060 2061 if (VerifyDuringGC) { 2062 // Verify that the counting data accumulated during marking matches 2063 // that calculated by walking the marking bitmap. 2064 2065 // Bitmaps to hold expected values 2066 BitMap expected_region_bm(_region_bm.size(), true); 2067 BitMap expected_card_bm(_card_bm.size(), true); 2068 2069 G1ParVerifyFinalCountTask g1_par_verify_task(g1h, 2070 &_region_bm, 2071 &_card_bm, 2072 &expected_region_bm, 2073 &expected_card_bm); 2074 2075 if (G1CollectedHeap::use_parallel_gc_threads()) { 2076 g1h->set_par_threads((int)n_workers); 2077 g1h->workers()->run_task(&g1_par_verify_task); 2078 // Done with the parallel phase so reset to 0. 2079 g1h->set_par_threads(0); 2080 } else { 2081 g1_par_verify_task.work(0); 2082 } 2083 2084 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures"); 2085 } 2086 2087 size_t start_used_bytes = g1h->used(); 2088 g1h->set_marking_complete(); 2089 2090 double count_end = os::elapsedTime(); 2091 double this_final_counting_time = (count_end - start); 2092 _total_counting_time += this_final_counting_time; 2093 2094 if (G1PrintRegionLivenessInfo) { 2095 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking"); 2096 _g1h->heap_region_iterate(&cl); 2097 } 2098 2099 // Install newly created mark bitMap as "prev". 2100 swapMarkBitMaps(); 2101 2102 g1h->reset_gc_time_stamp(); 2103 2104 // Note end of marking in all heap regions. 2105 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list, n_workers); 2106 if (G1CollectedHeap::use_parallel_gc_threads()) { 2107 g1h->set_par_threads((int)n_workers); 2108 g1h->workers()->run_task(&g1_par_note_end_task); 2109 g1h->set_par_threads(0); 2110 } else { 2111 g1_par_note_end_task.work(0); 2112 } 2113 g1h->check_gc_time_stamps(); 2114 2115 if (!cleanup_list_is_empty()) { 2116 // The cleanup list is not empty, so we'll have to process it 2117 // concurrently. Notify anyone else that might be wanting free 2118 // regions that there will be more free regions coming soon. 2119 g1h->set_free_regions_coming(); 2120 } 2121 2122 // call below, since it affects the metric by which we sort the heap 2123 // regions. 2124 if (G1ScrubRemSets) { 2125 double rs_scrub_start = os::elapsedTime(); 2126 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm, n_workers); 2127 if (G1CollectedHeap::use_parallel_gc_threads()) { 2128 g1h->set_par_threads((int)n_workers); 2129 g1h->workers()->run_task(&g1_par_scrub_rs_task); 2130 g1h->set_par_threads(0); 2131 } else { 2132 g1_par_scrub_rs_task.work(0); 2133 } 2134 2135 double rs_scrub_end = os::elapsedTime(); 2136 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); 2137 _total_rs_scrub_time += this_rs_scrub_time; 2138 } 2139 2140 // this will also free any regions totally full of garbage objects, 2141 // and sort the regions. 2142 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers); 2143 2144 // Statistics. 2145 double end = os::elapsedTime(); 2146 _cleanup_times.add((end - start) * 1000.0); 2147 2148 if (G1Log::fine()) { 2149 g1h->print_size_transition(gclog_or_tty, 2150 start_used_bytes, 2151 g1h->used(), 2152 g1h->capacity()); 2153 } 2154 2155 // Clean up will have freed any regions completely full of garbage. 2156 // Update the soft reference policy with the new heap occupancy. 2157 Universe::update_heap_info_at_gc(); 2158 2159 if (VerifyDuringGC) { 2160 HandleMark hm; // handle scope 2161 Universe::heap()->prepare_for_verify(); 2162 Universe::verify(VerifyOption_G1UsePrevMarking, 2163 " VerifyDuringGC:(after)"); 2164 } 2165 2166 g1h->check_bitmaps("Cleanup End"); 2167 2168 g1h->verify_region_sets_optional(); 2169 2170 // We need to make this be a "collection" so any collection pause that 2171 // races with it goes around and waits for completeCleanup to finish. 2172 g1h->increment_total_collections(); 2173 2174 // Clean out dead classes and update Metaspace sizes. 2175 if (ClassUnloadingWithConcurrentMark) { 2176 ClassLoaderDataGraph::purge(); 2177 } 2178 MetaspaceGC::compute_new_size(); 2179 2180 // We reclaimed old regions so we should calculate the sizes to make 2181 // sure we update the old gen/space data. 2182 g1h->g1mm()->update_sizes(); 2183 2184 g1h->trace_heap_after_concurrent_cycle(); 2185 } 2186 2187 void ConcurrentMark::completeCleanup() { 2188 if (has_aborted()) return; 2189 2190 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2191 2192 _cleanup_list.verify_optional(); 2193 FreeRegionList tmp_free_list("Tmp Free List"); 2194 2195 if (G1ConcRegionFreeingVerbose) { 2196 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2197 "cleanup list has %u entries", 2198 _cleanup_list.length()); 2199 } 2200 2201 // No one else should be accessing the _cleanup_list at this point, 2202 // so it is not necessary to take any locks 2203 while (!_cleanup_list.is_empty()) { 2204 HeapRegion* hr = _cleanup_list.remove_region(true /* from_head */); 2205 assert(hr != NULL, "Got NULL from a non-empty list"); 2206 hr->par_clear(); 2207 tmp_free_list.add_ordered(hr); 2208 2209 // Instead of adding one region at a time to the secondary_free_list, 2210 // we accumulate them in the local list and move them a few at a 2211 // time. This also cuts down on the number of notify_all() calls 2212 // we do during this process. We'll also append the local list when 2213 // _cleanup_list is empty (which means we just removed the last 2214 // region from the _cleanup_list). 2215 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 2216 _cleanup_list.is_empty()) { 2217 if (G1ConcRegionFreeingVerbose) { 2218 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2219 "appending %u entries to the secondary_free_list, " 2220 "cleanup list still has %u entries", 2221 tmp_free_list.length(), 2222 _cleanup_list.length()); 2223 } 2224 2225 { 2226 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 2227 g1h->secondary_free_list_add(&tmp_free_list); 2228 SecondaryFreeList_lock->notify_all(); 2229 } 2230 2231 if (G1StressConcRegionFreeing) { 2232 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 2233 os::sleep(Thread::current(), (jlong) 1, false); 2234 } 2235 } 2236 } 2237 } 2238 assert(tmp_free_list.is_empty(), "post-condition"); 2239 } 2240 2241 // Supporting Object and Oop closures for reference discovery 2242 // and processing in during marking 2243 2244 bool G1CMIsAliveClosure::do_object_b(oop obj) { 2245 HeapWord* addr = (HeapWord*)obj; 2246 return addr != NULL && 2247 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 2248 } 2249 2250 // 'Keep Alive' oop closure used by both serial parallel reference processing. 2251 // Uses the CMTask associated with a worker thread (for serial reference 2252 // processing the CMTask for worker 0 is used) to preserve (mark) and 2253 // trace referent objects. 2254 // 2255 // Using the CMTask and embedded local queues avoids having the worker 2256 // threads operating on the global mark stack. This reduces the risk 2257 // of overflowing the stack - which we would rather avoid at this late 2258 // state. Also using the tasks' local queues removes the potential 2259 // of the workers interfering with each other that could occur if 2260 // operating on the global stack. 2261 2262 class G1CMKeepAliveAndDrainClosure: public OopClosure { 2263 ConcurrentMark* _cm; 2264 CMTask* _task; 2265 int _ref_counter_limit; 2266 int _ref_counter; 2267 bool _is_serial; 2268 public: 2269 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) : 2270 _cm(cm), _task(task), _is_serial(is_serial), 2271 _ref_counter_limit(G1RefProcDrainInterval) { 2272 assert(_ref_counter_limit > 0, "sanity"); 2273 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 2274 _ref_counter = _ref_counter_limit; 2275 } 2276 2277 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 2278 virtual void do_oop( oop* p) { do_oop_work(p); } 2279 2280 template <class T> void do_oop_work(T* p) { 2281 if (!_cm->has_overflown()) { 2282 oop obj = oopDesc::load_decode_heap_oop(p); 2283 if (_cm->verbose_high()) { 2284 gclog_or_tty->print_cr("\t[%u] we're looking at location " 2285 "*"PTR_FORMAT" = "PTR_FORMAT, 2286 _task->worker_id(), p2i(p), p2i((void*) obj)); 2287 } 2288 2289 _task->deal_with_reference(obj); 2290 _ref_counter--; 2291 2292 if (_ref_counter == 0) { 2293 // We have dealt with _ref_counter_limit references, pushing them 2294 // and objects reachable from them on to the local stack (and 2295 // possibly the global stack). Call CMTask::do_marking_step() to 2296 // process these entries. 2297 // 2298 // We call CMTask::do_marking_step() in a loop, which we'll exit if 2299 // there's nothing more to do (i.e. we're done with the entries that 2300 // were pushed as a result of the CMTask::deal_with_reference() calls 2301 // above) or we overflow. 2302 // 2303 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2304 // flag while there may still be some work to do. (See the comment at 2305 // the beginning of CMTask::do_marking_step() for those conditions - 2306 // one of which is reaching the specified time target.) It is only 2307 // when CMTask::do_marking_step() returns without setting the 2308 // has_aborted() flag that the marking step has completed. 2309 do { 2310 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 2311 _task->do_marking_step(mark_step_duration_ms, 2312 false /* do_termination */, 2313 _is_serial); 2314 } while (_task->has_aborted() && !_cm->has_overflown()); 2315 _ref_counter = _ref_counter_limit; 2316 } 2317 } else { 2318 if (_cm->verbose_high()) { 2319 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id()); 2320 } 2321 } 2322 } 2323 }; 2324 2325 // 'Drain' oop closure used by both serial and parallel reference processing. 2326 // Uses the CMTask associated with a given worker thread (for serial 2327 // reference processing the CMtask for worker 0 is used). Calls the 2328 // do_marking_step routine, with an unbelievably large timeout value, 2329 // to drain the marking data structures of the remaining entries 2330 // added by the 'keep alive' oop closure above. 2331 2332 class G1CMDrainMarkingStackClosure: public VoidClosure { 2333 ConcurrentMark* _cm; 2334 CMTask* _task; 2335 bool _is_serial; 2336 public: 2337 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) : 2338 _cm(cm), _task(task), _is_serial(is_serial) { 2339 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 2340 } 2341 2342 void do_void() { 2343 do { 2344 if (_cm->verbose_high()) { 2345 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s", 2346 _task->worker_id(), BOOL_TO_STR(_is_serial)); 2347 } 2348 2349 // We call CMTask::do_marking_step() to completely drain the local 2350 // and global marking stacks of entries pushed by the 'keep alive' 2351 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 2352 // 2353 // CMTask::do_marking_step() is called in a loop, which we'll exit 2354 // if there's nothing more to do (i.e. we've completely drained the 2355 // entries that were pushed as a a result of applying the 'keep alive' 2356 // closure to the entries on the discovered ref lists) or we overflow 2357 // the global marking stack. 2358 // 2359 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2360 // flag while there may still be some work to do. (See the comment at 2361 // the beginning of CMTask::do_marking_step() for those conditions - 2362 // one of which is reaching the specified time target.) It is only 2363 // when CMTask::do_marking_step() returns without setting the 2364 // has_aborted() flag that the marking step has completed. 2365 2366 _task->do_marking_step(1000000000.0 /* something very large */, 2367 true /* do_termination */, 2368 _is_serial); 2369 } while (_task->has_aborted() && !_cm->has_overflown()); 2370 } 2371 }; 2372 2373 // Implementation of AbstractRefProcTaskExecutor for parallel 2374 // reference processing at the end of G1 concurrent marking 2375 2376 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2377 private: 2378 G1CollectedHeap* _g1h; 2379 ConcurrentMark* _cm; 2380 WorkGang* _workers; 2381 int _active_workers; 2382 2383 public: 2384 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 2385 ConcurrentMark* cm, 2386 WorkGang* workers, 2387 int n_workers) : 2388 _g1h(g1h), _cm(cm), 2389 _workers(workers), _active_workers(n_workers) { } 2390 2391 // Executes the given task using concurrent marking worker threads. 2392 virtual void execute(ProcessTask& task); 2393 virtual void execute(EnqueueTask& task); 2394 }; 2395 2396 class G1CMRefProcTaskProxy: public AbstractGangTask { 2397 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2398 ProcessTask& _proc_task; 2399 G1CollectedHeap* _g1h; 2400 ConcurrentMark* _cm; 2401 2402 public: 2403 G1CMRefProcTaskProxy(ProcessTask& proc_task, 2404 G1CollectedHeap* g1h, 2405 ConcurrentMark* cm) : 2406 AbstractGangTask("Process reference objects in parallel"), 2407 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 2408 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 2409 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 2410 } 2411 2412 virtual void work(uint worker_id) { 2413 ResourceMark rm; 2414 HandleMark hm; 2415 CMTask* task = _cm->task(worker_id); 2416 G1CMIsAliveClosure g1_is_alive(_g1h); 2417 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 2418 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 2419 2420 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 2421 } 2422 }; 2423 2424 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 2425 assert(_workers != NULL, "Need parallel worker threads."); 2426 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2427 2428 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 2429 2430 // We need to reset the concurrency level before each 2431 // proxy task execution, so that the termination protocol 2432 // and overflow handling in CMTask::do_marking_step() knows 2433 // how many workers to wait for. 2434 _cm->set_concurrency(_active_workers); 2435 _g1h->set_par_threads(_active_workers); 2436 _workers->run_task(&proc_task_proxy); 2437 _g1h->set_par_threads(0); 2438 } 2439 2440 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 2441 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 2442 EnqueueTask& _enq_task; 2443 2444 public: 2445 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 2446 AbstractGangTask("Enqueue reference objects in parallel"), 2447 _enq_task(enq_task) { } 2448 2449 virtual void work(uint worker_id) { 2450 _enq_task.work(worker_id); 2451 } 2452 }; 2453 2454 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 2455 assert(_workers != NULL, "Need parallel worker threads."); 2456 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2457 2458 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 2459 2460 // Not strictly necessary but... 2461 // 2462 // We need to reset the concurrency level before each 2463 // proxy task execution, so that the termination protocol 2464 // and overflow handling in CMTask::do_marking_step() knows 2465 // how many workers to wait for. 2466 _cm->set_concurrency(_active_workers); 2467 _g1h->set_par_threads(_active_workers); 2468 _workers->run_task(&enq_task_proxy); 2469 _g1h->set_par_threads(0); 2470 } 2471 2472 void ConcurrentMark::weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes) { 2473 G1CollectedHeap::heap()->parallel_cleaning(is_alive, true, true, purged_classes); 2474 } 2475 2476 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 2477 if (has_overflown()) { 2478 // Skip processing the discovered references if we have 2479 // overflown the global marking stack. Reference objects 2480 // only get discovered once so it is OK to not 2481 // de-populate the discovered reference lists. We could have, 2482 // but the only benefit would be that, when marking restarts, 2483 // less reference objects are discovered. 2484 return; 2485 } 2486 2487 ResourceMark rm; 2488 HandleMark hm; 2489 2490 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2491 2492 // Is alive closure. 2493 G1CMIsAliveClosure g1_is_alive(g1h); 2494 2495 // Inner scope to exclude the cleaning of the string and symbol 2496 // tables from the displayed time. 2497 { 2498 G1CMTraceTime t("GC ref-proc", G1Log::finer()); 2499 2500 ReferenceProcessor* rp = g1h->ref_processor_cm(); 2501 2502 // See the comment in G1CollectedHeap::ref_processing_init() 2503 // about how reference processing currently works in G1. 2504 2505 // Set the soft reference policy 2506 rp->setup_policy(clear_all_soft_refs); 2507 assert(_markStack.isEmpty(), "mark stack should be empty"); 2508 2509 // Instances of the 'Keep Alive' and 'Complete GC' closures used 2510 // in serial reference processing. Note these closures are also 2511 // used for serially processing (by the the current thread) the 2512 // JNI references during parallel reference processing. 2513 // 2514 // These closures do not need to synchronize with the worker 2515 // threads involved in parallel reference processing as these 2516 // instances are executed serially by the current thread (e.g. 2517 // reference processing is not multi-threaded and is thus 2518 // performed by the current thread instead of a gang worker). 2519 // 2520 // The gang tasks involved in parallel reference processing create 2521 // their own instances of these closures, which do their own 2522 // synchronization among themselves. 2523 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 2524 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 2525 2526 // We need at least one active thread. If reference processing 2527 // is not multi-threaded we use the current (VMThread) thread, 2528 // otherwise we use the work gang from the G1CollectedHeap and 2529 // we utilize all the worker threads we can. 2530 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL; 2531 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U); 2532 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U); 2533 2534 // Parallel processing task executor. 2535 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 2536 g1h->workers(), active_workers); 2537 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 2538 2539 // Set the concurrency level. The phase was already set prior to 2540 // executing the remark task. 2541 set_concurrency(active_workers); 2542 2543 // Set the degree of MT processing here. If the discovery was done MT, 2544 // the number of threads involved during discovery could differ from 2545 // the number of active workers. This is OK as long as the discovered 2546 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 2547 rp->set_active_mt_degree(active_workers); 2548 2549 // Process the weak references. 2550 const ReferenceProcessorStats& stats = 2551 rp->process_discovered_references(&g1_is_alive, 2552 &g1_keep_alive, 2553 &g1_drain_mark_stack, 2554 executor, 2555 g1h->gc_timer_cm(), 2556 concurrent_gc_id()); 2557 g1h->gc_tracer_cm()->report_gc_reference_stats(stats); 2558 2559 // The do_oop work routines of the keep_alive and drain_marking_stack 2560 // oop closures will set the has_overflown flag if we overflow the 2561 // global marking stack. 2562 2563 assert(_markStack.overflow() || _markStack.isEmpty(), 2564 "mark stack should be empty (unless it overflowed)"); 2565 2566 if (_markStack.overflow()) { 2567 // This should have been done already when we tried to push an 2568 // entry on to the global mark stack. But let's do it again. 2569 set_has_overflown(); 2570 } 2571 2572 assert(rp->num_q() == active_workers, "why not"); 2573 2574 rp->enqueue_discovered_references(executor); 2575 2576 rp->verify_no_references_recorded(); 2577 assert(!rp->discovery_enabled(), "Post condition"); 2578 } 2579 2580 if (has_overflown()) { 2581 // We can not trust g1_is_alive if the marking stack overflowed 2582 return; 2583 } 2584 2585 assert(_markStack.isEmpty(), "Marking should have completed"); 2586 2587 // Unload Klasses, String, Symbols, Code Cache, etc. 2588 { 2589 G1CMTraceTime trace("Unloading", G1Log::finer()); 2590 2591 if (ClassUnloadingWithConcurrentMark) { 2592 bool purged_classes; 2593 2594 { 2595 G1CMTraceTime trace("System Dictionary Unloading", G1Log::finest()); 2596 purged_classes = SystemDictionary::do_unloading(&g1_is_alive); 2597 } 2598 2599 { 2600 G1CMTraceTime trace("Parallel Unloading", G1Log::finest()); 2601 weakRefsWorkParallelPart(&g1_is_alive, purged_classes); 2602 } 2603 } 2604 2605 if (G1StringDedup::is_enabled()) { 2606 G1CMTraceTime trace("String Deduplication Unlink", G1Log::finest()); 2607 G1StringDedup::unlink(&g1_is_alive); 2608 } 2609 } 2610 } 2611 2612 void ConcurrentMark::swapMarkBitMaps() { 2613 CMBitMapRO* temp = _prevMarkBitMap; 2614 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; 2615 _nextMarkBitMap = (CMBitMap*) temp; 2616 } 2617 2618 class CMObjectClosure; 2619 2620 // Closure for iterating over objects, currently only used for 2621 // processing SATB buffers. 2622 class CMObjectClosure : public ObjectClosure { 2623 private: 2624 CMTask* _task; 2625 2626 public: 2627 void do_object(oop obj) { 2628 _task->deal_with_reference(obj); 2629 } 2630 2631 CMObjectClosure(CMTask* task) : _task(task) { } 2632 }; 2633 2634 class G1RemarkThreadsClosure : public ThreadClosure { 2635 CMObjectClosure _cm_obj; 2636 G1CMOopClosure _cm_cl; 2637 MarkingCodeBlobClosure _code_cl; 2638 int _thread_parity; 2639 bool _is_par; 2640 2641 public: 2642 G1RemarkThreadsClosure(G1CollectedHeap* g1h, CMTask* task, bool is_par) : 2643 _cm_obj(task), _cm_cl(g1h, g1h->concurrent_mark(), task), _code_cl(&_cm_cl, !CodeBlobToOopClosure::FixRelocations), 2644 _thread_parity(SharedHeap::heap()->strong_roots_parity()), _is_par(is_par) {} 2645 2646 void do_thread(Thread* thread) { 2647 if (thread->is_Java_thread()) { 2648 if (thread->claim_oops_do(_is_par, _thread_parity)) { 2649 JavaThread* jt = (JavaThread*)thread; 2650 2651 // In theory it should not be neccessary to explicitly walk the nmethods to find roots for concurrent marking 2652 // however the liveness of oops reachable from nmethods have very complex lifecycles: 2653 // * Alive if on the stack of an executing method 2654 // * Weakly reachable otherwise 2655 // Some objects reachable from nmethods, such as the class loader (or klass_holder) of the receiver should be 2656 // live by the SATB invariant but other oops recorded in nmethods may behave differently. 2657 jt->nmethods_do(&_code_cl); 2658 2659 jt->satb_mark_queue().apply_closure_and_empty(&_cm_obj); 2660 } 2661 } else if (thread->is_VM_thread()) { 2662 if (thread->claim_oops_do(_is_par, _thread_parity)) { 2663 JavaThread::satb_mark_queue_set().shared_satb_queue()->apply_closure_and_empty(&_cm_obj); 2664 } 2665 } 2666 } 2667 }; 2668 2669 class CMRemarkTask: public AbstractGangTask { 2670 private: 2671 ConcurrentMark* _cm; 2672 bool _is_serial; 2673 public: 2674 void work(uint worker_id) { 2675 // Since all available tasks are actually started, we should 2676 // only proceed if we're supposed to be active. 2677 if (worker_id < _cm->active_tasks()) { 2678 CMTask* task = _cm->task(worker_id); 2679 task->record_start_time(); 2680 { 2681 ResourceMark rm; 2682 HandleMark hm; 2683 2684 G1RemarkThreadsClosure threads_f(G1CollectedHeap::heap(), task, !_is_serial); 2685 Threads::threads_do(&threads_f); 2686 } 2687 2688 do { 2689 task->do_marking_step(1000000000.0 /* something very large */, 2690 true /* do_termination */, 2691 _is_serial); 2692 } while (task->has_aborted() && !_cm->has_overflown()); 2693 // If we overflow, then we do not want to restart. We instead 2694 // want to abort remark and do concurrent marking again. 2695 task->record_end_time(); 2696 } 2697 } 2698 2699 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) : 2700 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) { 2701 _cm->terminator()->reset_for_reuse(active_workers); 2702 } 2703 }; 2704 2705 void ConcurrentMark::checkpointRootsFinalWork() { 2706 ResourceMark rm; 2707 HandleMark hm; 2708 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2709 2710 G1CMTraceTime trace("Finalize Marking", G1Log::finer()); 2711 2712 g1h->ensure_parsability(false); 2713 2714 if (G1CollectedHeap::use_parallel_gc_threads()) { 2715 G1CollectedHeap::StrongRootsScope srs(g1h); 2716 // this is remark, so we'll use up all active threads 2717 uint active_workers = g1h->workers()->active_workers(); 2718 if (active_workers == 0) { 2719 assert(active_workers > 0, "Should have been set earlier"); 2720 active_workers = (uint) ParallelGCThreads; 2721 g1h->workers()->set_active_workers(active_workers); 2722 } 2723 set_concurrency_and_phase(active_workers, false /* concurrent */); 2724 // Leave _parallel_marking_threads at it's 2725 // value originally calculated in the ConcurrentMark 2726 // constructor and pass values of the active workers 2727 // through the gang in the task. 2728 2729 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */); 2730 // We will start all available threads, even if we decide that the 2731 // active_workers will be fewer. The extra ones will just bail out 2732 // immediately. 2733 g1h->set_par_threads(active_workers); 2734 g1h->workers()->run_task(&remarkTask); 2735 g1h->set_par_threads(0); 2736 } else { 2737 G1CollectedHeap::StrongRootsScope srs(g1h); 2738 uint active_workers = 1; 2739 set_concurrency_and_phase(active_workers, false /* concurrent */); 2740 2741 // Note - if there's no work gang then the VMThread will be 2742 // the thread to execute the remark - serially. We have 2743 // to pass true for the is_serial parameter so that 2744 // CMTask::do_marking_step() doesn't enter the sync 2745 // barriers in the event of an overflow. Doing so will 2746 // cause an assert that the current thread is not a 2747 // concurrent GC thread. 2748 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/); 2749 remarkTask.work(0); 2750 } 2751 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2752 guarantee(has_overflown() || 2753 satb_mq_set.completed_buffers_num() == 0, 2754 err_msg("Invariant: has_overflown = %s, num buffers = %d", 2755 BOOL_TO_STR(has_overflown()), 2756 satb_mq_set.completed_buffers_num())); 2757 2758 print_stats(); 2759 } 2760 2761 #ifndef PRODUCT 2762 2763 class PrintReachableOopClosure: public OopClosure { 2764 private: 2765 G1CollectedHeap* _g1h; 2766 outputStream* _out; 2767 VerifyOption _vo; 2768 bool _all; 2769 2770 public: 2771 PrintReachableOopClosure(outputStream* out, 2772 VerifyOption vo, 2773 bool all) : 2774 _g1h(G1CollectedHeap::heap()), 2775 _out(out), _vo(vo), _all(all) { } 2776 2777 void do_oop(narrowOop* p) { do_oop_work(p); } 2778 void do_oop( oop* p) { do_oop_work(p); } 2779 2780 template <class T> void do_oop_work(T* p) { 2781 oop obj = oopDesc::load_decode_heap_oop(p); 2782 const char* str = NULL; 2783 const char* str2 = ""; 2784 2785 if (obj == NULL) { 2786 str = ""; 2787 } else if (!_g1h->is_in_g1_reserved(obj)) { 2788 str = " O"; 2789 } else { 2790 HeapRegion* hr = _g1h->heap_region_containing(obj); 2791 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo); 2792 bool marked = _g1h->is_marked(obj, _vo); 2793 2794 if (over_tams) { 2795 str = " >"; 2796 if (marked) { 2797 str2 = " AND MARKED"; 2798 } 2799 } else if (marked) { 2800 str = " M"; 2801 } else { 2802 str = " NOT"; 2803 } 2804 } 2805 2806 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s", 2807 p2i(p), p2i((void*) obj), str, str2); 2808 } 2809 }; 2810 2811 class PrintReachableObjectClosure : public ObjectClosure { 2812 private: 2813 G1CollectedHeap* _g1h; 2814 outputStream* _out; 2815 VerifyOption _vo; 2816 bool _all; 2817 HeapRegion* _hr; 2818 2819 public: 2820 PrintReachableObjectClosure(outputStream* out, 2821 VerifyOption vo, 2822 bool all, 2823 HeapRegion* hr) : 2824 _g1h(G1CollectedHeap::heap()), 2825 _out(out), _vo(vo), _all(all), _hr(hr) { } 2826 2827 void do_object(oop o) { 2828 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo); 2829 bool marked = _g1h->is_marked(o, _vo); 2830 bool print_it = _all || over_tams || marked; 2831 2832 if (print_it) { 2833 _out->print_cr(" "PTR_FORMAT"%s", 2834 p2i((void *)o), (over_tams) ? " >" : (marked) ? " M" : ""); 2835 PrintReachableOopClosure oopCl(_out, _vo, _all); 2836 o->oop_iterate_no_header(&oopCl); 2837 } 2838 } 2839 }; 2840 2841 class PrintReachableRegionClosure : public HeapRegionClosure { 2842 private: 2843 G1CollectedHeap* _g1h; 2844 outputStream* _out; 2845 VerifyOption _vo; 2846 bool _all; 2847 2848 public: 2849 bool doHeapRegion(HeapRegion* hr) { 2850 HeapWord* b = hr->bottom(); 2851 HeapWord* e = hr->end(); 2852 HeapWord* t = hr->top(); 2853 HeapWord* p = _g1h->top_at_mark_start(hr, _vo); 2854 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" " 2855 "TAMS: " PTR_FORMAT, p2i(b), p2i(e), p2i(t), p2i(p)); 2856 _out->cr(); 2857 2858 HeapWord* from = b; 2859 HeapWord* to = t; 2860 2861 if (to > from) { 2862 _out->print_cr("Objects in [" PTR_FORMAT ", " PTR_FORMAT "]", p2i(from), p2i(to)); 2863 _out->cr(); 2864 PrintReachableObjectClosure ocl(_out, _vo, _all, hr); 2865 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl); 2866 _out->cr(); 2867 } 2868 2869 return false; 2870 } 2871 2872 PrintReachableRegionClosure(outputStream* out, 2873 VerifyOption vo, 2874 bool all) : 2875 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { } 2876 }; 2877 2878 void ConcurrentMark::print_reachable(const char* str, 2879 VerifyOption vo, 2880 bool all) { 2881 gclog_or_tty->cr(); 2882 gclog_or_tty->print_cr("== Doing heap dump... "); 2883 2884 if (G1PrintReachableBaseFile == NULL) { 2885 gclog_or_tty->print_cr(" #### error: no base file defined"); 2886 return; 2887 } 2888 2889 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) > 2890 (JVM_MAXPATHLEN - 1)) { 2891 gclog_or_tty->print_cr(" #### error: file name too long"); 2892 return; 2893 } 2894 2895 char file_name[JVM_MAXPATHLEN]; 2896 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str); 2897 gclog_or_tty->print_cr(" dumping to file %s", file_name); 2898 2899 fileStream fout(file_name); 2900 if (!fout.is_open()) { 2901 gclog_or_tty->print_cr(" #### error: could not open file"); 2902 return; 2903 } 2904 2905 outputStream* out = &fout; 2906 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo)); 2907 out->cr(); 2908 2909 out->print_cr("--- ITERATING OVER REGIONS"); 2910 out->cr(); 2911 PrintReachableRegionClosure rcl(out, vo, all); 2912 _g1h->heap_region_iterate(&rcl); 2913 out->cr(); 2914 2915 gclog_or_tty->print_cr(" done"); 2916 gclog_or_tty->flush(); 2917 } 2918 2919 #endif // PRODUCT 2920 2921 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 2922 // Note we are overriding the read-only view of the prev map here, via 2923 // the cast. 2924 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); 2925 } 2926 2927 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) { 2928 _nextMarkBitMap->clearRange(mr); 2929 } 2930 2931 HeapRegion* 2932 ConcurrentMark::claim_region(uint worker_id) { 2933 // "checkpoint" the finger 2934 HeapWord* finger = _finger; 2935 2936 // _heap_end will not change underneath our feet; it only changes at 2937 // yield points. 2938 while (finger < _heap_end) { 2939 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 2940 2941 // Note on how this code handles humongous regions. In the 2942 // normal case the finger will reach the start of a "starts 2943 // humongous" (SH) region. Its end will either be the end of the 2944 // last "continues humongous" (CH) region in the sequence, or the 2945 // standard end of the SH region (if the SH is the only region in 2946 // the sequence). That way claim_region() will skip over the CH 2947 // regions. However, there is a subtle race between a CM thread 2948 // executing this method and a mutator thread doing a humongous 2949 // object allocation. The two are not mutually exclusive as the CM 2950 // thread does not need to hold the Heap_lock when it gets 2951 // here. So there is a chance that claim_region() will come across 2952 // a free region that's in the progress of becoming a SH or a CH 2953 // region. In the former case, it will either 2954 // a) Miss the update to the region's end, in which case it will 2955 // visit every subsequent CH region, will find their bitmaps 2956 // empty, and do nothing, or 2957 // b) Will observe the update of the region's end (in which case 2958 // it will skip the subsequent CH regions). 2959 // If it comes across a region that suddenly becomes CH, the 2960 // scenario will be similar to b). So, the race between 2961 // claim_region() and a humongous object allocation might force us 2962 // to do a bit of unnecessary work (due to some unnecessary bitmap 2963 // iterations) but it should not introduce and correctness issues. 2964 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger); 2965 2966 // Above heap_region_containing_raw may return NULL as we always scan claim 2967 // until the end of the heap. In this case, just jump to the next region. 2968 HeapWord* end = curr_region != NULL ? curr_region->end() : finger + HeapRegion::GrainWords; 2969 2970 // Is the gap between reading the finger and doing the CAS too long? 2971 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 2972 if (res == finger && curr_region != NULL) { 2973 // we succeeded 2974 HeapWord* bottom = curr_region->bottom(); 2975 HeapWord* limit = curr_region->next_top_at_mark_start(); 2976 2977 if (verbose_low()) { 2978 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" " 2979 "["PTR_FORMAT", "PTR_FORMAT"), " 2980 "limit = "PTR_FORMAT, 2981 worker_id, p2i(curr_region), p2i(bottom), p2i(end), p2i(limit)); 2982 } 2983 2984 // notice that _finger == end cannot be guaranteed here since, 2985 // someone else might have moved the finger even further 2986 assert(_finger >= end, "the finger should have moved forward"); 2987 2988 if (verbose_low()) { 2989 gclog_or_tty->print_cr("[%u] we were successful with region = " 2990 PTR_FORMAT, worker_id, p2i(curr_region)); 2991 } 2992 2993 if (limit > bottom) { 2994 if (verbose_low()) { 2995 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, " 2996 "returning it ", worker_id, p2i(curr_region)); 2997 } 2998 return curr_region; 2999 } else { 3000 assert(limit == bottom, 3001 "the region limit should be at bottom"); 3002 if (verbose_low()) { 3003 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, " 3004 "returning NULL", worker_id, p2i(curr_region)); 3005 } 3006 // we return NULL and the caller should try calling 3007 // claim_region() again. 3008 return NULL; 3009 } 3010 } else { 3011 assert(_finger > finger, "the finger should have moved forward"); 3012 if (verbose_low()) { 3013 if (curr_region == NULL) { 3014 gclog_or_tty->print_cr("[%u] found uncommitted region, moving finger, " 3015 "global finger = "PTR_FORMAT", " 3016 "our finger = "PTR_FORMAT, 3017 worker_id, p2i(_finger), p2i(finger)); 3018 } else { 3019 gclog_or_tty->print_cr("[%u] somebody else moved the finger, " 3020 "global finger = "PTR_FORMAT", " 3021 "our finger = "PTR_FORMAT, 3022 worker_id, p2i(_finger), p2i(finger)); 3023 } 3024 } 3025 3026 // read it again 3027 finger = _finger; 3028 } 3029 } 3030 3031 return NULL; 3032 } 3033 3034 #ifndef PRODUCT 3035 enum VerifyNoCSetOopsPhase { 3036 VerifyNoCSetOopsStack, 3037 VerifyNoCSetOopsQueues, 3038 VerifyNoCSetOopsSATBCompleted, 3039 VerifyNoCSetOopsSATBThread 3040 }; 3041 3042 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure { 3043 private: 3044 G1CollectedHeap* _g1h; 3045 VerifyNoCSetOopsPhase _phase; 3046 int _info; 3047 3048 const char* phase_str() { 3049 switch (_phase) { 3050 case VerifyNoCSetOopsStack: return "Stack"; 3051 case VerifyNoCSetOopsQueues: return "Queue"; 3052 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers"; 3053 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers"; 3054 default: ShouldNotReachHere(); 3055 } 3056 return NULL; 3057 } 3058 3059 void do_object_work(oop obj) { 3060 guarantee(!_g1h->obj_in_cs(obj), 3061 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d", 3062 p2i((void*) obj), phase_str(), _info)); 3063 } 3064 3065 public: 3066 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { } 3067 3068 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) { 3069 _phase = phase; 3070 _info = info; 3071 } 3072 3073 virtual void do_oop(oop* p) { 3074 oop obj = oopDesc::load_decode_heap_oop(p); 3075 do_object_work(obj); 3076 } 3077 3078 virtual void do_oop(narrowOop* p) { 3079 // We should not come across narrow oops while scanning marking 3080 // stacks and SATB buffers. 3081 ShouldNotReachHere(); 3082 } 3083 3084 virtual void do_object(oop obj) { 3085 do_object_work(obj); 3086 } 3087 }; 3088 3089 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks, 3090 bool verify_enqueued_buffers, 3091 bool verify_thread_buffers, 3092 bool verify_fingers) { 3093 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 3094 if (!G1CollectedHeap::heap()->mark_in_progress()) { 3095 return; 3096 } 3097 3098 VerifyNoCSetOopsClosure cl; 3099 3100 if (verify_stacks) { 3101 // Verify entries on the global mark stack 3102 cl.set_phase(VerifyNoCSetOopsStack); 3103 _markStack.oops_do(&cl); 3104 3105 // Verify entries on the task queues 3106 for (uint i = 0; i < _max_worker_id; i += 1) { 3107 cl.set_phase(VerifyNoCSetOopsQueues, i); 3108 CMTaskQueue* queue = _task_queues->queue(i); 3109 queue->oops_do(&cl); 3110 } 3111 } 3112 3113 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 3114 3115 // Verify entries on the enqueued SATB buffers 3116 if (verify_enqueued_buffers) { 3117 cl.set_phase(VerifyNoCSetOopsSATBCompleted); 3118 satb_qs.iterate_completed_buffers_read_only(&cl); 3119 } 3120 3121 // Verify entries on the per-thread SATB buffers 3122 if (verify_thread_buffers) { 3123 cl.set_phase(VerifyNoCSetOopsSATBThread); 3124 satb_qs.iterate_thread_buffers_read_only(&cl); 3125 } 3126 3127 if (verify_fingers) { 3128 // Verify the global finger 3129 HeapWord* global_finger = finger(); 3130 if (global_finger != NULL && global_finger < _heap_end) { 3131 // The global finger always points to a heap region boundary. We 3132 // use heap_region_containing_raw() to get the containing region 3133 // given that the global finger could be pointing to a free region 3134 // which subsequently becomes continues humongous. If that 3135 // happens, heap_region_containing() will return the bottom of the 3136 // corresponding starts humongous region and the check below will 3137 // not hold any more. 3138 // Since we always iterate over all regions, we might get a NULL HeapRegion 3139 // here. 3140 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger); 3141 guarantee(global_hr == NULL || global_finger == global_hr->bottom(), 3142 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT, 3143 p2i(global_finger), HR_FORMAT_PARAMS(global_hr))); 3144 } 3145 3146 // Verify the task fingers 3147 assert(parallel_marking_threads() <= _max_worker_id, "sanity"); 3148 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) { 3149 CMTask* task = _tasks[i]; 3150 HeapWord* task_finger = task->finger(); 3151 if (task_finger != NULL && task_finger < _heap_end) { 3152 // See above note on the global finger verification. 3153 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger); 3154 guarantee(task_hr == NULL || task_finger == task_hr->bottom() || 3155 !task_hr->in_collection_set(), 3156 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT, 3157 p2i(task_finger), HR_FORMAT_PARAMS(task_hr))); 3158 } 3159 } 3160 } 3161 } 3162 #endif // PRODUCT 3163 3164 // Aggregate the counting data that was constructed concurrently 3165 // with marking. 3166 class AggregateCountDataHRClosure: public HeapRegionClosure { 3167 G1CollectedHeap* _g1h; 3168 ConcurrentMark* _cm; 3169 CardTableModRefBS* _ct_bs; 3170 BitMap* _cm_card_bm; 3171 uint _max_worker_id; 3172 3173 public: 3174 AggregateCountDataHRClosure(G1CollectedHeap* g1h, 3175 BitMap* cm_card_bm, 3176 uint max_worker_id) : 3177 _g1h(g1h), _cm(g1h->concurrent_mark()), 3178 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 3179 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { } 3180 3181 bool doHeapRegion(HeapRegion* hr) { 3182 if (hr->is_continues_humongous()) { 3183 // We will ignore these here and process them when their 3184 // associated "starts humongous" region is processed. 3185 // Note that we cannot rely on their associated 3186 // "starts humongous" region to have their bit set to 1 3187 // since, due to the region chunking in the parallel region 3188 // iteration, a "continues humongous" region might be visited 3189 // before its associated "starts humongous". 3190 return false; 3191 } 3192 3193 HeapWord* start = hr->bottom(); 3194 HeapWord* limit = hr->next_top_at_mark_start(); 3195 HeapWord* end = hr->end(); 3196 3197 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(), 3198 err_msg("Preconditions not met - " 3199 "start: "PTR_FORMAT", limit: "PTR_FORMAT", " 3200 "top: "PTR_FORMAT", end: "PTR_FORMAT, 3201 p2i(start), p2i(limit), p2i(hr->top()), p2i(hr->end()))); 3202 3203 assert(hr->next_marked_bytes() == 0, "Precondition"); 3204 3205 if (start == limit) { 3206 // NTAMS of this region has not been set so nothing to do. 3207 return false; 3208 } 3209 3210 // 'start' should be in the heap. 3211 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity"); 3212 // 'end' *may* be just beyond the end of the heap (if hr is the last region) 3213 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity"); 3214 3215 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 3216 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit); 3217 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end); 3218 3219 // If ntams is not card aligned then we bump card bitmap index 3220 // for limit so that we get the all the cards spanned by 3221 // the object ending at ntams. 3222 // Note: if this is the last region in the heap then ntams 3223 // could be actually just beyond the end of the the heap; 3224 // limit_idx will then correspond to a (non-existent) card 3225 // that is also outside the heap. 3226 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) { 3227 limit_idx += 1; 3228 } 3229 3230 assert(limit_idx <= end_idx, "or else use atomics"); 3231 3232 // Aggregate the "stripe" in the count data associated with hr. 3233 uint hrm_index = hr->hrm_index(); 3234 size_t marked_bytes = 0; 3235 3236 for (uint i = 0; i < _max_worker_id; i += 1) { 3237 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i); 3238 BitMap* task_card_bm = _cm->count_card_bitmap_for(i); 3239 3240 // Fetch the marked_bytes in this region for task i and 3241 // add it to the running total for this region. 3242 marked_bytes += marked_bytes_array[hrm_index]; 3243 3244 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx) 3245 // into the global card bitmap. 3246 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx); 3247 3248 while (scan_idx < limit_idx) { 3249 assert(task_card_bm->at(scan_idx) == true, "should be"); 3250 _cm_card_bm->set_bit(scan_idx); 3251 assert(_cm_card_bm->at(scan_idx) == true, "should be"); 3252 3253 // BitMap::get_next_one_offset() can handle the case when 3254 // its left_offset parameter is greater than its right_offset 3255 // parameter. It does, however, have an early exit if 3256 // left_offset == right_offset. So let's limit the value 3257 // passed in for left offset here. 3258 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx); 3259 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx); 3260 } 3261 } 3262 3263 // Update the marked bytes for this region. 3264 hr->add_to_marked_bytes(marked_bytes); 3265 3266 // Next heap region 3267 return false; 3268 } 3269 }; 3270 3271 class G1AggregateCountDataTask: public AbstractGangTask { 3272 protected: 3273 G1CollectedHeap* _g1h; 3274 ConcurrentMark* _cm; 3275 BitMap* _cm_card_bm; 3276 uint _max_worker_id; 3277 int _active_workers; 3278 HeapRegionClaimer _hrclaimer; 3279 3280 public: 3281 G1AggregateCountDataTask(G1CollectedHeap* g1h, 3282 ConcurrentMark* cm, 3283 BitMap* cm_card_bm, 3284 uint max_worker_id, 3285 int n_workers) : 3286 AbstractGangTask("Count Aggregation"), 3287 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm), 3288 _max_worker_id(max_worker_id), 3289 _active_workers(n_workers) { 3290 if (G1CollectedHeap::use_parallel_gc_threads()) { 3291 _hrclaimer.initialize(_active_workers); 3292 } 3293 } 3294 3295 void work(uint worker_id) { 3296 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id); 3297 3298 if (G1CollectedHeap::use_parallel_gc_threads()) { 3299 _g1h->heap_region_par_iterate(&cl, worker_id, &_hrclaimer); 3300 } else { 3301 _g1h->heap_region_iterate(&cl); 3302 } 3303 } 3304 }; 3305 3306 3307 void ConcurrentMark::aggregate_count_data() { 3308 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3309 _g1h->workers()->active_workers() : 3310 1); 3311 3312 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm, 3313 _max_worker_id, n_workers); 3314 3315 if (G1CollectedHeap::use_parallel_gc_threads()) { 3316 _g1h->set_par_threads(n_workers); 3317 _g1h->workers()->run_task(&g1_par_agg_task); 3318 _g1h->set_par_threads(0); 3319 } else { 3320 g1_par_agg_task.work(0); 3321 } 3322 _g1h->allocation_context_stats().update_at_remark(); 3323 } 3324 3325 // Clear the per-worker arrays used to store the per-region counting data 3326 void ConcurrentMark::clear_all_count_data() { 3327 // Clear the global card bitmap - it will be filled during 3328 // liveness count aggregation (during remark) and the 3329 // final counting task. 3330 _card_bm.clear(); 3331 3332 // Clear the global region bitmap - it will be filled as part 3333 // of the final counting task. 3334 _region_bm.clear(); 3335 3336 uint max_regions = _g1h->max_regions(); 3337 assert(_max_worker_id > 0, "uninitialized"); 3338 3339 for (uint i = 0; i < _max_worker_id; i += 1) { 3340 BitMap* task_card_bm = count_card_bitmap_for(i); 3341 size_t* marked_bytes_array = count_marked_bytes_array_for(i); 3342 3343 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 3344 assert(marked_bytes_array != NULL, "uninitialized"); 3345 3346 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t)); 3347 task_card_bm->clear(); 3348 } 3349 } 3350 3351 void ConcurrentMark::print_stats() { 3352 if (verbose_stats()) { 3353 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3354 for (size_t i = 0; i < _active_tasks; ++i) { 3355 _tasks[i]->print_stats(); 3356 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3357 } 3358 } 3359 } 3360 3361 // abandon current marking iteration due to a Full GC 3362 void ConcurrentMark::abort() { 3363 // Clear all marks in the next bitmap for the next marking cycle. This will allow us to skip the next 3364 // concurrent bitmap clearing. 3365 _nextMarkBitMap->clearAll(); 3366 3367 // Note we cannot clear the previous marking bitmap here 3368 // since VerifyDuringGC verifies the objects marked during 3369 // a full GC against the previous bitmap. 3370 3371 // Clear the liveness counting data 3372 clear_all_count_data(); 3373 // Empty mark stack 3374 reset_marking_state(); 3375 for (uint i = 0; i < _max_worker_id; ++i) { 3376 _tasks[i]->clear_region_fields(); 3377 } 3378 _first_overflow_barrier_sync.abort(); 3379 _second_overflow_barrier_sync.abort(); 3380 const GCId& gc_id = _g1h->gc_tracer_cm()->gc_id(); 3381 if (!gc_id.is_undefined()) { 3382 // We can do multiple full GCs before ConcurrentMarkThread::run() gets a chance 3383 // to detect that it was aborted. Only keep track of the first GC id that we aborted. 3384 _aborted_gc_id = gc_id; 3385 } 3386 _has_aborted = true; 3387 3388 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3389 satb_mq_set.abandon_partial_marking(); 3390 // This can be called either during or outside marking, we'll read 3391 // the expected_active value from the SATB queue set. 3392 satb_mq_set.set_active_all_threads( 3393 false, /* new active value */ 3394 satb_mq_set.is_active() /* expected_active */); 3395 3396 _g1h->trace_heap_after_concurrent_cycle(); 3397 _g1h->register_concurrent_cycle_end(); 3398 } 3399 3400 const GCId& ConcurrentMark::concurrent_gc_id() { 3401 if (has_aborted()) { 3402 return _aborted_gc_id; 3403 } 3404 return _g1h->gc_tracer_cm()->gc_id(); 3405 } 3406 3407 static void print_ms_time_info(const char* prefix, const char* name, 3408 NumberSeq& ns) { 3409 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 3410 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 3411 if (ns.num() > 0) { 3412 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]", 3413 prefix, ns.sd(), ns.maximum()); 3414 } 3415 } 3416 3417 void ConcurrentMark::print_summary_info() { 3418 gclog_or_tty->print_cr(" Concurrent marking:"); 3419 print_ms_time_info(" ", "init marks", _init_times); 3420 print_ms_time_info(" ", "remarks", _remark_times); 3421 { 3422 print_ms_time_info(" ", "final marks", _remark_mark_times); 3423 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 3424 3425 } 3426 print_ms_time_info(" ", "cleanups", _cleanup_times); 3427 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).", 3428 _total_counting_time, 3429 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / 3430 (double)_cleanup_times.num() 3431 : 0.0)); 3432 if (G1ScrubRemSets) { 3433 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 3434 _total_rs_scrub_time, 3435 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / 3436 (double)_cleanup_times.num() 3437 : 0.0)); 3438 } 3439 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.", 3440 (_init_times.sum() + _remark_times.sum() + 3441 _cleanup_times.sum())/1000.0); 3442 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s " 3443 "(%8.2f s marking).", 3444 cmThread()->vtime_accum(), 3445 cmThread()->vtime_mark_accum()); 3446 } 3447 3448 void ConcurrentMark::print_worker_threads_on(outputStream* st) const { 3449 if (use_parallel_marking_threads()) { 3450 _parallel_workers->print_worker_threads_on(st); 3451 } 3452 } 3453 3454 void ConcurrentMark::print_on_error(outputStream* st) const { 3455 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT, 3456 p2i(_prevMarkBitMap), p2i(_nextMarkBitMap)); 3457 _prevMarkBitMap->print_on_error(st, " Prev Bits: "); 3458 _nextMarkBitMap->print_on_error(st, " Next Bits: "); 3459 } 3460 3461 // We take a break if someone is trying to stop the world. 3462 bool ConcurrentMark::do_yield_check(uint worker_id) { 3463 if (SuspendibleThreadSet::should_yield()) { 3464 if (worker_id == 0) { 3465 _g1h->g1_policy()->record_concurrent_pause(); 3466 } 3467 SuspendibleThreadSet::yield(); 3468 return true; 3469 } else { 3470 return false; 3471 } 3472 } 3473 3474 #ifndef PRODUCT 3475 // for debugging purposes 3476 void ConcurrentMark::print_finger() { 3477 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT, 3478 p2i(_heap_start), p2i(_heap_end), p2i(_finger)); 3479 for (uint i = 0; i < _max_worker_id; ++i) { 3480 gclog_or_tty->print(" %u: " PTR_FORMAT, i, p2i(_tasks[i]->finger())); 3481 } 3482 gclog_or_tty->cr(); 3483 } 3484 #endif 3485 3486 void CMTask::scan_object(oop obj) { 3487 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant"); 3488 3489 if (_cm->verbose_high()) { 3490 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT, 3491 _worker_id, p2i((void*) obj)); 3492 } 3493 3494 size_t obj_size = obj->size(); 3495 _words_scanned += obj_size; 3496 3497 obj->oop_iterate(_cm_oop_closure); 3498 statsOnly( ++_objs_scanned ); 3499 check_limits(); 3500 } 3501 3502 // Closure for iteration over bitmaps 3503 class CMBitMapClosure : public BitMapClosure { 3504 private: 3505 // the bitmap that is being iterated over 3506 CMBitMap* _nextMarkBitMap; 3507 ConcurrentMark* _cm; 3508 CMTask* _task; 3509 3510 public: 3511 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) : 3512 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } 3513 3514 bool do_bit(size_t offset) { 3515 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); 3516 assert(_nextMarkBitMap->isMarked(addr), "invariant"); 3517 assert( addr < _cm->finger(), "invariant"); 3518 3519 statsOnly( _task->increase_objs_found_on_bitmap() ); 3520 assert(addr >= _task->finger(), "invariant"); 3521 3522 // We move that task's local finger along. 3523 _task->move_finger_to(addr); 3524 3525 _task->scan_object(oop(addr)); 3526 // we only partially drain the local queue and global stack 3527 _task->drain_local_queue(true); 3528 _task->drain_global_stack(true); 3529 3530 // if the has_aborted flag has been raised, we need to bail out of 3531 // the iteration 3532 return !_task->has_aborted(); 3533 } 3534 }; 3535 3536 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 3537 ConcurrentMark* cm, 3538 CMTask* task) 3539 : _g1h(g1h), _cm(cm), _task(task) { 3540 assert(_ref_processor == NULL, "should be initialized to NULL"); 3541 3542 if (G1UseConcMarkReferenceProcessing) { 3543 _ref_processor = g1h->ref_processor_cm(); 3544 assert(_ref_processor != NULL, "should not be NULL"); 3545 } 3546 } 3547 3548 void CMTask::setup_for_region(HeapRegion* hr) { 3549 assert(hr != NULL, 3550 "claim_region() should have filtered out NULL regions"); 3551 assert(!hr->is_continues_humongous(), 3552 "claim_region() should have filtered out continues humongous regions"); 3553 3554 if (_cm->verbose_low()) { 3555 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT, 3556 _worker_id, p2i(hr)); 3557 } 3558 3559 _curr_region = hr; 3560 _finger = hr->bottom(); 3561 update_region_limit(); 3562 } 3563 3564 void CMTask::update_region_limit() { 3565 HeapRegion* hr = _curr_region; 3566 HeapWord* bottom = hr->bottom(); 3567 HeapWord* limit = hr->next_top_at_mark_start(); 3568 3569 if (limit == bottom) { 3570 if (_cm->verbose_low()) { 3571 gclog_or_tty->print_cr("[%u] found an empty region " 3572 "["PTR_FORMAT", "PTR_FORMAT")", 3573 _worker_id, p2i(bottom), p2i(limit)); 3574 } 3575 // The region was collected underneath our feet. 3576 // We set the finger to bottom to ensure that the bitmap 3577 // iteration that will follow this will not do anything. 3578 // (this is not a condition that holds when we set the region up, 3579 // as the region is not supposed to be empty in the first place) 3580 _finger = bottom; 3581 } else if (limit >= _region_limit) { 3582 assert(limit >= _finger, "peace of mind"); 3583 } else { 3584 assert(limit < _region_limit, "only way to get here"); 3585 // This can happen under some pretty unusual circumstances. An 3586 // evacuation pause empties the region underneath our feet (NTAMS 3587 // at bottom). We then do some allocation in the region (NTAMS 3588 // stays at bottom), followed by the region being used as a GC 3589 // alloc region (NTAMS will move to top() and the objects 3590 // originally below it will be grayed). All objects now marked in 3591 // the region are explicitly grayed, if below the global finger, 3592 // and we do not need in fact to scan anything else. So, we simply 3593 // set _finger to be limit to ensure that the bitmap iteration 3594 // doesn't do anything. 3595 _finger = limit; 3596 } 3597 3598 _region_limit = limit; 3599 } 3600 3601 void CMTask::giveup_current_region() { 3602 assert(_curr_region != NULL, "invariant"); 3603 if (_cm->verbose_low()) { 3604 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT, 3605 _worker_id, p2i(_curr_region)); 3606 } 3607 clear_region_fields(); 3608 } 3609 3610 void CMTask::clear_region_fields() { 3611 // Values for these three fields that indicate that we're not 3612 // holding on to a region. 3613 _curr_region = NULL; 3614 _finger = NULL; 3615 _region_limit = NULL; 3616 } 3617 3618 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 3619 if (cm_oop_closure == NULL) { 3620 assert(_cm_oop_closure != NULL, "invariant"); 3621 } else { 3622 assert(_cm_oop_closure == NULL, "invariant"); 3623 } 3624 _cm_oop_closure = cm_oop_closure; 3625 } 3626 3627 void CMTask::reset(CMBitMap* nextMarkBitMap) { 3628 guarantee(nextMarkBitMap != NULL, "invariant"); 3629 3630 if (_cm->verbose_low()) { 3631 gclog_or_tty->print_cr("[%u] resetting", _worker_id); 3632 } 3633 3634 _nextMarkBitMap = nextMarkBitMap; 3635 clear_region_fields(); 3636 3637 _calls = 0; 3638 _elapsed_time_ms = 0.0; 3639 _termination_time_ms = 0.0; 3640 _termination_start_time_ms = 0.0; 3641 3642 #if _MARKING_STATS_ 3643 _local_pushes = 0; 3644 _local_pops = 0; 3645 _local_max_size = 0; 3646 _objs_scanned = 0; 3647 _global_pushes = 0; 3648 _global_pops = 0; 3649 _global_max_size = 0; 3650 _global_transfers_to = 0; 3651 _global_transfers_from = 0; 3652 _regions_claimed = 0; 3653 _objs_found_on_bitmap = 0; 3654 _satb_buffers_processed = 0; 3655 _steal_attempts = 0; 3656 _steals = 0; 3657 _aborted = 0; 3658 _aborted_overflow = 0; 3659 _aborted_cm_aborted = 0; 3660 _aborted_yield = 0; 3661 _aborted_timed_out = 0; 3662 _aborted_satb = 0; 3663 _aborted_termination = 0; 3664 #endif // _MARKING_STATS_ 3665 } 3666 3667 bool CMTask::should_exit_termination() { 3668 regular_clock_call(); 3669 // This is called when we are in the termination protocol. We should 3670 // quit if, for some reason, this task wants to abort or the global 3671 // stack is not empty (this means that we can get work from it). 3672 return !_cm->mark_stack_empty() || has_aborted(); 3673 } 3674 3675 void CMTask::reached_limit() { 3676 assert(_words_scanned >= _words_scanned_limit || 3677 _refs_reached >= _refs_reached_limit , 3678 "shouldn't have been called otherwise"); 3679 regular_clock_call(); 3680 } 3681 3682 void CMTask::regular_clock_call() { 3683 if (has_aborted()) return; 3684 3685 // First, we need to recalculate the words scanned and refs reached 3686 // limits for the next clock call. 3687 recalculate_limits(); 3688 3689 // During the regular clock call we do the following 3690 3691 // (1) If an overflow has been flagged, then we abort. 3692 if (_cm->has_overflown()) { 3693 set_has_aborted(); 3694 return; 3695 } 3696 3697 // If we are not concurrent (i.e. we're doing remark) we don't need 3698 // to check anything else. The other steps are only needed during 3699 // the concurrent marking phase. 3700 if (!concurrent()) return; 3701 3702 // (2) If marking has been aborted for Full GC, then we also abort. 3703 if (_cm->has_aborted()) { 3704 set_has_aborted(); 3705 statsOnly( ++_aborted_cm_aborted ); 3706 return; 3707 } 3708 3709 double curr_time_ms = os::elapsedVTime() * 1000.0; 3710 3711 // (3) If marking stats are enabled, then we update the step history. 3712 #if _MARKING_STATS_ 3713 if (_words_scanned >= _words_scanned_limit) { 3714 ++_clock_due_to_scanning; 3715 } 3716 if (_refs_reached >= _refs_reached_limit) { 3717 ++_clock_due_to_marking; 3718 } 3719 3720 double last_interval_ms = curr_time_ms - _interval_start_time_ms; 3721 _interval_start_time_ms = curr_time_ms; 3722 _all_clock_intervals_ms.add(last_interval_ms); 3723 3724 if (_cm->verbose_medium()) { 3725 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, " 3726 "scanned = "SIZE_FORMAT"%s, refs reached = "SIZE_FORMAT"%s", 3727 _worker_id, last_interval_ms, 3728 _words_scanned, 3729 (_words_scanned >= _words_scanned_limit) ? " (*)" : "", 3730 _refs_reached, 3731 (_refs_reached >= _refs_reached_limit) ? " (*)" : ""); 3732 } 3733 #endif // _MARKING_STATS_ 3734 3735 // (4) We check whether we should yield. If we have to, then we abort. 3736 if (SuspendibleThreadSet::should_yield()) { 3737 // We should yield. To do this we abort the task. The caller is 3738 // responsible for yielding. 3739 set_has_aborted(); 3740 statsOnly( ++_aborted_yield ); 3741 return; 3742 } 3743 3744 // (5) We check whether we've reached our time quota. If we have, 3745 // then we abort. 3746 double elapsed_time_ms = curr_time_ms - _start_time_ms; 3747 if (elapsed_time_ms > _time_target_ms) { 3748 set_has_aborted(); 3749 _has_timed_out = true; 3750 statsOnly( ++_aborted_timed_out ); 3751 return; 3752 } 3753 3754 // (6) Finally, we check whether there are enough completed STAB 3755 // buffers available for processing. If there are, we abort. 3756 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3757 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 3758 if (_cm->verbose_low()) { 3759 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers", 3760 _worker_id); 3761 } 3762 // we do need to process SATB buffers, we'll abort and restart 3763 // the marking task to do so 3764 set_has_aborted(); 3765 statsOnly( ++_aborted_satb ); 3766 return; 3767 } 3768 } 3769 3770 void CMTask::recalculate_limits() { 3771 _real_words_scanned_limit = _words_scanned + words_scanned_period; 3772 _words_scanned_limit = _real_words_scanned_limit; 3773 3774 _real_refs_reached_limit = _refs_reached + refs_reached_period; 3775 _refs_reached_limit = _real_refs_reached_limit; 3776 } 3777 3778 void CMTask::decrease_limits() { 3779 // This is called when we believe that we're going to do an infrequent 3780 // operation which will increase the per byte scanned cost (i.e. move 3781 // entries to/from the global stack). It basically tries to decrease the 3782 // scanning limit so that the clock is called earlier. 3783 3784 if (_cm->verbose_medium()) { 3785 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id); 3786 } 3787 3788 _words_scanned_limit = _real_words_scanned_limit - 3789 3 * words_scanned_period / 4; 3790 _refs_reached_limit = _real_refs_reached_limit - 3791 3 * refs_reached_period / 4; 3792 } 3793 3794 void CMTask::move_entries_to_global_stack() { 3795 // local array where we'll store the entries that will be popped 3796 // from the local queue 3797 oop buffer[global_stack_transfer_size]; 3798 3799 int n = 0; 3800 oop obj; 3801 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { 3802 buffer[n] = obj; 3803 ++n; 3804 } 3805 3806 if (n > 0) { 3807 // we popped at least one entry from the local queue 3808 3809 statsOnly( ++_global_transfers_to; _local_pops += n ); 3810 3811 if (!_cm->mark_stack_push(buffer, n)) { 3812 if (_cm->verbose_low()) { 3813 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow", 3814 _worker_id); 3815 } 3816 set_has_aborted(); 3817 } else { 3818 // the transfer was successful 3819 3820 if (_cm->verbose_medium()) { 3821 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack", 3822 _worker_id, n); 3823 } 3824 statsOnly( int tmp_size = _cm->mark_stack_size(); 3825 if (tmp_size > _global_max_size) { 3826 _global_max_size = tmp_size; 3827 } 3828 _global_pushes += n ); 3829 } 3830 } 3831 3832 // this operation was quite expensive, so decrease the limits 3833 decrease_limits(); 3834 } 3835 3836 void CMTask::get_entries_from_global_stack() { 3837 // local array where we'll store the entries that will be popped 3838 // from the global stack. 3839 oop buffer[global_stack_transfer_size]; 3840 int n; 3841 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); 3842 assert(n <= global_stack_transfer_size, 3843 "we should not pop more than the given limit"); 3844 if (n > 0) { 3845 // yes, we did actually pop at least one entry 3846 3847 statsOnly( ++_global_transfers_from; _global_pops += n ); 3848 if (_cm->verbose_medium()) { 3849 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack", 3850 _worker_id, n); 3851 } 3852 for (int i = 0; i < n; ++i) { 3853 bool success = _task_queue->push(buffer[i]); 3854 // We only call this when the local queue is empty or under a 3855 // given target limit. So, we do not expect this push to fail. 3856 assert(success, "invariant"); 3857 } 3858 3859 statsOnly( int tmp_size = _task_queue->size(); 3860 if (tmp_size > _local_max_size) { 3861 _local_max_size = tmp_size; 3862 } 3863 _local_pushes += n ); 3864 } 3865 3866 // this operation was quite expensive, so decrease the limits 3867 decrease_limits(); 3868 } 3869 3870 void CMTask::drain_local_queue(bool partially) { 3871 if (has_aborted()) return; 3872 3873 // Decide what the target size is, depending whether we're going to 3874 // drain it partially (so that other tasks can steal if they run out 3875 // of things to do) or totally (at the very end). 3876 size_t target_size; 3877 if (partially) { 3878 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 3879 } else { 3880 target_size = 0; 3881 } 3882 3883 if (_task_queue->size() > target_size) { 3884 if (_cm->verbose_high()) { 3885 gclog_or_tty->print_cr("[%u] draining local queue, target size = " SIZE_FORMAT, 3886 _worker_id, target_size); 3887 } 3888 3889 oop obj; 3890 bool ret = _task_queue->pop_local(obj); 3891 while (ret) { 3892 statsOnly( ++_local_pops ); 3893 3894 if (_cm->verbose_high()) { 3895 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id, 3896 p2i((void*) obj)); 3897 } 3898 3899 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); 3900 assert(!_g1h->is_on_master_free_list( 3901 _g1h->heap_region_containing((HeapWord*) obj)), "invariant"); 3902 3903 scan_object(obj); 3904 3905 if (_task_queue->size() <= target_size || has_aborted()) { 3906 ret = false; 3907 } else { 3908 ret = _task_queue->pop_local(obj); 3909 } 3910 } 3911 3912 if (_cm->verbose_high()) { 3913 gclog_or_tty->print_cr("[%u] drained local queue, size = %u", 3914 _worker_id, _task_queue->size()); 3915 } 3916 } 3917 } 3918 3919 void CMTask::drain_global_stack(bool partially) { 3920 if (has_aborted()) return; 3921 3922 // We have a policy to drain the local queue before we attempt to 3923 // drain the global stack. 3924 assert(partially || _task_queue->size() == 0, "invariant"); 3925 3926 // Decide what the target size is, depending whether we're going to 3927 // drain it partially (so that other tasks can steal if they run out 3928 // of things to do) or totally (at the very end). Notice that, 3929 // because we move entries from the global stack in chunks or 3930 // because another task might be doing the same, we might in fact 3931 // drop below the target. But, this is not a problem. 3932 size_t target_size; 3933 if (partially) { 3934 target_size = _cm->partial_mark_stack_size_target(); 3935 } else { 3936 target_size = 0; 3937 } 3938 3939 if (_cm->mark_stack_size() > target_size) { 3940 if (_cm->verbose_low()) { 3941 gclog_or_tty->print_cr("[%u] draining global_stack, target size " SIZE_FORMAT, 3942 _worker_id, target_size); 3943 } 3944 3945 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 3946 get_entries_from_global_stack(); 3947 drain_local_queue(partially); 3948 } 3949 3950 if (_cm->verbose_low()) { 3951 gclog_or_tty->print_cr("[%u] drained global stack, size = " SIZE_FORMAT, 3952 _worker_id, _cm->mark_stack_size()); 3953 } 3954 } 3955 } 3956 3957 // SATB Queue has several assumptions on whether to call the par or 3958 // non-par versions of the methods. this is why some of the code is 3959 // replicated. We should really get rid of the single-threaded version 3960 // of the code to simplify things. 3961 void CMTask::drain_satb_buffers() { 3962 if (has_aborted()) return; 3963 3964 // We set this so that the regular clock knows that we're in the 3965 // middle of draining buffers and doesn't set the abort flag when it 3966 // notices that SATB buffers are available for draining. It'd be 3967 // very counter productive if it did that. :-) 3968 _draining_satb_buffers = true; 3969 3970 CMObjectClosure oc(this); 3971 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3972 if (G1CollectedHeap::use_parallel_gc_threads()) { 3973 satb_mq_set.set_par_closure(_worker_id, &oc); 3974 } else { 3975 satb_mq_set.set_closure(&oc); 3976 } 3977 3978 // This keeps claiming and applying the closure to completed buffers 3979 // until we run out of buffers or we need to abort. 3980 if (G1CollectedHeap::use_parallel_gc_threads()) { 3981 while (!has_aborted() && 3982 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) { 3983 if (_cm->verbose_medium()) { 3984 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id); 3985 } 3986 statsOnly( ++_satb_buffers_processed ); 3987 regular_clock_call(); 3988 } 3989 } else { 3990 while (!has_aborted() && 3991 satb_mq_set.apply_closure_to_completed_buffer()) { 3992 if (_cm->verbose_medium()) { 3993 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id); 3994 } 3995 statsOnly( ++_satb_buffers_processed ); 3996 regular_clock_call(); 3997 } 3998 } 3999 4000 _draining_satb_buffers = false; 4001 4002 assert(has_aborted() || 4003 concurrent() || 4004 satb_mq_set.completed_buffers_num() == 0, "invariant"); 4005 4006 if (G1CollectedHeap::use_parallel_gc_threads()) { 4007 satb_mq_set.set_par_closure(_worker_id, NULL); 4008 } else { 4009 satb_mq_set.set_closure(NULL); 4010 } 4011 4012 // again, this was a potentially expensive operation, decrease the 4013 // limits to get the regular clock call early 4014 decrease_limits(); 4015 } 4016 4017 void CMTask::print_stats() { 4018 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d", 4019 _worker_id, _calls); 4020 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 4021 _elapsed_time_ms, _termination_time_ms); 4022 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 4023 _step_times_ms.num(), _step_times_ms.avg(), 4024 _step_times_ms.sd()); 4025 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 4026 _step_times_ms.maximum(), _step_times_ms.sum()); 4027 4028 #if _MARKING_STATS_ 4029 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 4030 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(), 4031 _all_clock_intervals_ms.sd()); 4032 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 4033 _all_clock_intervals_ms.maximum(), 4034 _all_clock_intervals_ms.sum()); 4035 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d", 4036 _clock_due_to_scanning, _clock_due_to_marking); 4037 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d", 4038 _objs_scanned, _objs_found_on_bitmap); 4039 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d", 4040 _local_pushes, _local_pops, _local_max_size); 4041 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d", 4042 _global_pushes, _global_pops, _global_max_size); 4043 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d", 4044 _global_transfers_to,_global_transfers_from); 4045 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed); 4046 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed); 4047 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d", 4048 _steal_attempts, _steals); 4049 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted); 4050 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d", 4051 _aborted_overflow, _aborted_cm_aborted, _aborted_yield); 4052 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d", 4053 _aborted_timed_out, _aborted_satb, _aborted_termination); 4054 #endif // _MARKING_STATS_ 4055 } 4056 4057 /***************************************************************************** 4058 4059 The do_marking_step(time_target_ms, ...) method is the building 4060 block of the parallel marking framework. It can be called in parallel 4061 with other invocations of do_marking_step() on different tasks 4062 (but only one per task, obviously) and concurrently with the 4063 mutator threads, or during remark, hence it eliminates the need 4064 for two versions of the code. When called during remark, it will 4065 pick up from where the task left off during the concurrent marking 4066 phase. Interestingly, tasks are also claimable during evacuation 4067 pauses too, since do_marking_step() ensures that it aborts before 4068 it needs to yield. 4069 4070 The data structures that it uses to do marking work are the 4071 following: 4072 4073 (1) Marking Bitmap. If there are gray objects that appear only 4074 on the bitmap (this happens either when dealing with an overflow 4075 or when the initial marking phase has simply marked the roots 4076 and didn't push them on the stack), then tasks claim heap 4077 regions whose bitmap they then scan to find gray objects. A 4078 global finger indicates where the end of the last claimed region 4079 is. A local finger indicates how far into the region a task has 4080 scanned. The two fingers are used to determine how to gray an 4081 object (i.e. whether simply marking it is OK, as it will be 4082 visited by a task in the future, or whether it needs to be also 4083 pushed on a stack). 4084 4085 (2) Local Queue. The local queue of the task which is accessed 4086 reasonably efficiently by the task. Other tasks can steal from 4087 it when they run out of work. Throughout the marking phase, a 4088 task attempts to keep its local queue short but not totally 4089 empty, so that entries are available for stealing by other 4090 tasks. Only when there is no more work, a task will totally 4091 drain its local queue. 4092 4093 (3) Global Mark Stack. This handles local queue overflow. During 4094 marking only sets of entries are moved between it and the local 4095 queues, as access to it requires a mutex and more fine-grain 4096 interaction with it which might cause contention. If it 4097 overflows, then the marking phase should restart and iterate 4098 over the bitmap to identify gray objects. Throughout the marking 4099 phase, tasks attempt to keep the global mark stack at a small 4100 length but not totally empty, so that entries are available for 4101 popping by other tasks. Only when there is no more work, tasks 4102 will totally drain the global mark stack. 4103 4104 (4) SATB Buffer Queue. This is where completed SATB buffers are 4105 made available. Buffers are regularly removed from this queue 4106 and scanned for roots, so that the queue doesn't get too 4107 long. During remark, all completed buffers are processed, as 4108 well as the filled in parts of any uncompleted buffers. 4109 4110 The do_marking_step() method tries to abort when the time target 4111 has been reached. There are a few other cases when the 4112 do_marking_step() method also aborts: 4113 4114 (1) When the marking phase has been aborted (after a Full GC). 4115 4116 (2) When a global overflow (on the global stack) has been 4117 triggered. Before the task aborts, it will actually sync up with 4118 the other tasks to ensure that all the marking data structures 4119 (local queues, stacks, fingers etc.) are re-initialized so that 4120 when do_marking_step() completes, the marking phase can 4121 immediately restart. 4122 4123 (3) When enough completed SATB buffers are available. The 4124 do_marking_step() method only tries to drain SATB buffers right 4125 at the beginning. So, if enough buffers are available, the 4126 marking step aborts and the SATB buffers are processed at 4127 the beginning of the next invocation. 4128 4129 (4) To yield. when we have to yield then we abort and yield 4130 right at the end of do_marking_step(). This saves us from a lot 4131 of hassle as, by yielding we might allow a Full GC. If this 4132 happens then objects will be compacted underneath our feet, the 4133 heap might shrink, etc. We save checking for this by just 4134 aborting and doing the yield right at the end. 4135 4136 From the above it follows that the do_marking_step() method should 4137 be called in a loop (or, otherwise, regularly) until it completes. 4138 4139 If a marking step completes without its has_aborted() flag being 4140 true, it means it has completed the current marking phase (and 4141 also all other marking tasks have done so and have all synced up). 4142 4143 A method called regular_clock_call() is invoked "regularly" (in 4144 sub ms intervals) throughout marking. It is this clock method that 4145 checks all the abort conditions which were mentioned above and 4146 decides when the task should abort. A work-based scheme is used to 4147 trigger this clock method: when the number of object words the 4148 marking phase has scanned or the number of references the marking 4149 phase has visited reach a given limit. Additional invocations to 4150 the method clock have been planted in a few other strategic places 4151 too. The initial reason for the clock method was to avoid calling 4152 vtime too regularly, as it is quite expensive. So, once it was in 4153 place, it was natural to piggy-back all the other conditions on it 4154 too and not constantly check them throughout the code. 4155 4156 If do_termination is true then do_marking_step will enter its 4157 termination protocol. 4158 4159 The value of is_serial must be true when do_marking_step is being 4160 called serially (i.e. by the VMThread) and do_marking_step should 4161 skip any synchronization in the termination and overflow code. 4162 Examples include the serial remark code and the serial reference 4163 processing closures. 4164 4165 The value of is_serial must be false when do_marking_step is 4166 being called by any of the worker threads in a work gang. 4167 Examples include the concurrent marking code (CMMarkingTask), 4168 the MT remark code, and the MT reference processing closures. 4169 4170 *****************************************************************************/ 4171 4172 void CMTask::do_marking_step(double time_target_ms, 4173 bool do_termination, 4174 bool is_serial) { 4175 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 4176 assert(concurrent() == _cm->concurrent(), "they should be the same"); 4177 4178 G1CollectorPolicy* g1_policy = _g1h->g1_policy(); 4179 assert(_task_queues != NULL, "invariant"); 4180 assert(_task_queue != NULL, "invariant"); 4181 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant"); 4182 4183 assert(!_claimed, 4184 "only one thread should claim this task at any one time"); 4185 4186 // OK, this doesn't safeguard again all possible scenarios, as it is 4187 // possible for two threads to set the _claimed flag at the same 4188 // time. But it is only for debugging purposes anyway and it will 4189 // catch most problems. 4190 _claimed = true; 4191 4192 _start_time_ms = os::elapsedVTime() * 1000.0; 4193 statsOnly( _interval_start_time_ms = _start_time_ms ); 4194 4195 // If do_stealing is true then do_marking_step will attempt to 4196 // steal work from the other CMTasks. It only makes sense to 4197 // enable stealing when the termination protocol is enabled 4198 // and do_marking_step() is not being called serially. 4199 bool do_stealing = do_termination && !is_serial; 4200 4201 double diff_prediction_ms = 4202 g1_policy->get_new_prediction(&_marking_step_diffs_ms); 4203 _time_target_ms = time_target_ms - diff_prediction_ms; 4204 4205 // set up the variables that are used in the work-based scheme to 4206 // call the regular clock method 4207 _words_scanned = 0; 4208 _refs_reached = 0; 4209 recalculate_limits(); 4210 4211 // clear all flags 4212 clear_has_aborted(); 4213 _has_timed_out = false; 4214 _draining_satb_buffers = false; 4215 4216 ++_calls; 4217 4218 if (_cm->verbose_low()) { 4219 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, " 4220 "target = %1.2lfms >>>>>>>>>>", 4221 _worker_id, _calls, _time_target_ms); 4222 } 4223 4224 // Set up the bitmap and oop closures. Anything that uses them is 4225 // eventually called from this method, so it is OK to allocate these 4226 // statically. 4227 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); 4228 G1CMOopClosure cm_oop_closure(_g1h, _cm, this); 4229 set_cm_oop_closure(&cm_oop_closure); 4230 4231 if (_cm->has_overflown()) { 4232 // This can happen if the mark stack overflows during a GC pause 4233 // and this task, after a yield point, restarts. We have to abort 4234 // as we need to get into the overflow protocol which happens 4235 // right at the end of this task. 4236 set_has_aborted(); 4237 } 4238 4239 // First drain any available SATB buffers. After this, we will not 4240 // look at SATB buffers before the next invocation of this method. 4241 // If enough completed SATB buffers are queued up, the regular clock 4242 // will abort this task so that it restarts. 4243 drain_satb_buffers(); 4244 // ...then partially drain the local queue and the global stack 4245 drain_local_queue(true); 4246 drain_global_stack(true); 4247 4248 do { 4249 if (!has_aborted() && _curr_region != NULL) { 4250 // This means that we're already holding on to a region. 4251 assert(_finger != NULL, "if region is not NULL, then the finger " 4252 "should not be NULL either"); 4253 4254 // We might have restarted this task after an evacuation pause 4255 // which might have evacuated the region we're holding on to 4256 // underneath our feet. Let's read its limit again to make sure 4257 // that we do not iterate over a region of the heap that 4258 // contains garbage (update_region_limit() will also move 4259 // _finger to the start of the region if it is found empty). 4260 update_region_limit(); 4261 // We will start from _finger not from the start of the region, 4262 // as we might be restarting this task after aborting half-way 4263 // through scanning this region. In this case, _finger points to 4264 // the address where we last found a marked object. If this is a 4265 // fresh region, _finger points to start(). 4266 MemRegion mr = MemRegion(_finger, _region_limit); 4267 4268 if (_cm->verbose_low()) { 4269 gclog_or_tty->print_cr("[%u] we're scanning part " 4270 "["PTR_FORMAT", "PTR_FORMAT") " 4271 "of region "HR_FORMAT, 4272 _worker_id, p2i(_finger), p2i(_region_limit), 4273 HR_FORMAT_PARAMS(_curr_region)); 4274 } 4275 4276 assert(!_curr_region->is_humongous() || mr.start() == _curr_region->bottom(), 4277 "humongous regions should go around loop once only"); 4278 4279 // Some special cases: 4280 // If the memory region is empty, we can just give up the region. 4281 // If the current region is humongous then we only need to check 4282 // the bitmap for the bit associated with the start of the object, 4283 // scan the object if it's live, and give up the region. 4284 // Otherwise, let's iterate over the bitmap of the part of the region 4285 // that is left. 4286 // If the iteration is successful, give up the region. 4287 if (mr.is_empty()) { 4288 giveup_current_region(); 4289 regular_clock_call(); 4290 } else if (_curr_region->is_humongous() && mr.start() == _curr_region->bottom()) { 4291 if (_nextMarkBitMap->isMarked(mr.start())) { 4292 // The object is marked - apply the closure 4293 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start()); 4294 bitmap_closure.do_bit(offset); 4295 } 4296 // Even if this task aborted while scanning the humongous object 4297 // we can (and should) give up the current region. 4298 giveup_current_region(); 4299 regular_clock_call(); 4300 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) { 4301 giveup_current_region(); 4302 regular_clock_call(); 4303 } else { 4304 assert(has_aborted(), "currently the only way to do so"); 4305 // The only way to abort the bitmap iteration is to return 4306 // false from the do_bit() method. However, inside the 4307 // do_bit() method we move the _finger to point to the 4308 // object currently being looked at. So, if we bail out, we 4309 // have definitely set _finger to something non-null. 4310 assert(_finger != NULL, "invariant"); 4311 4312 // Region iteration was actually aborted. So now _finger 4313 // points to the address of the object we last scanned. If we 4314 // leave it there, when we restart this task, we will rescan 4315 // the object. It is easy to avoid this. We move the finger by 4316 // enough to point to the next possible object header (the 4317 // bitmap knows by how much we need to move it as it knows its 4318 // granularity). 4319 assert(_finger < _region_limit, "invariant"); 4320 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger); 4321 // Check if bitmap iteration was aborted while scanning the last object 4322 if (new_finger >= _region_limit) { 4323 giveup_current_region(); 4324 } else { 4325 move_finger_to(new_finger); 4326 } 4327 } 4328 } 4329 // At this point we have either completed iterating over the 4330 // region we were holding on to, or we have aborted. 4331 4332 // We then partially drain the local queue and the global stack. 4333 // (Do we really need this?) 4334 drain_local_queue(true); 4335 drain_global_stack(true); 4336 4337 // Read the note on the claim_region() method on why it might 4338 // return NULL with potentially more regions available for 4339 // claiming and why we have to check out_of_regions() to determine 4340 // whether we're done or not. 4341 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 4342 // We are going to try to claim a new region. We should have 4343 // given up on the previous one. 4344 // Separated the asserts so that we know which one fires. 4345 assert(_curr_region == NULL, "invariant"); 4346 assert(_finger == NULL, "invariant"); 4347 assert(_region_limit == NULL, "invariant"); 4348 if (_cm->verbose_low()) { 4349 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id); 4350 } 4351 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 4352 if (claimed_region != NULL) { 4353 // Yes, we managed to claim one 4354 statsOnly( ++_regions_claimed ); 4355 4356 if (_cm->verbose_low()) { 4357 gclog_or_tty->print_cr("[%u] we successfully claimed " 4358 "region "PTR_FORMAT, 4359 _worker_id, p2i(claimed_region)); 4360 } 4361 4362 setup_for_region(claimed_region); 4363 assert(_curr_region == claimed_region, "invariant"); 4364 } 4365 // It is important to call the regular clock here. It might take 4366 // a while to claim a region if, for example, we hit a large 4367 // block of empty regions. So we need to call the regular clock 4368 // method once round the loop to make sure it's called 4369 // frequently enough. 4370 regular_clock_call(); 4371 } 4372 4373 if (!has_aborted() && _curr_region == NULL) { 4374 assert(_cm->out_of_regions(), 4375 "at this point we should be out of regions"); 4376 } 4377 } while ( _curr_region != NULL && !has_aborted()); 4378 4379 if (!has_aborted()) { 4380 // We cannot check whether the global stack is empty, since other 4381 // tasks might be pushing objects to it concurrently. 4382 assert(_cm->out_of_regions(), 4383 "at this point we should be out of regions"); 4384 4385 if (_cm->verbose_low()) { 4386 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id); 4387 } 4388 4389 // Try to reduce the number of available SATB buffers so that 4390 // remark has less work to do. 4391 drain_satb_buffers(); 4392 } 4393 4394 // Since we've done everything else, we can now totally drain the 4395 // local queue and global stack. 4396 drain_local_queue(false); 4397 drain_global_stack(false); 4398 4399 // Attempt at work stealing from other task's queues. 4400 if (do_stealing && !has_aborted()) { 4401 // We have not aborted. This means that we have finished all that 4402 // we could. Let's try to do some stealing... 4403 4404 // We cannot check whether the global stack is empty, since other 4405 // tasks might be pushing objects to it concurrently. 4406 assert(_cm->out_of_regions() && _task_queue->size() == 0, 4407 "only way to reach here"); 4408 4409 if (_cm->verbose_low()) { 4410 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id); 4411 } 4412 4413 while (!has_aborted()) { 4414 oop obj; 4415 statsOnly( ++_steal_attempts ); 4416 4417 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) { 4418 if (_cm->verbose_medium()) { 4419 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully", 4420 _worker_id, p2i((void*) obj)); 4421 } 4422 4423 statsOnly( ++_steals ); 4424 4425 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), 4426 "any stolen object should be marked"); 4427 scan_object(obj); 4428 4429 // And since we're towards the end, let's totally drain the 4430 // local queue and global stack. 4431 drain_local_queue(false); 4432 drain_global_stack(false); 4433 } else { 4434 break; 4435 } 4436 } 4437 } 4438 4439 // If we are about to wrap up and go into termination, check if we 4440 // should raise the overflow flag. 4441 if (do_termination && !has_aborted()) { 4442 if (_cm->force_overflow()->should_force()) { 4443 _cm->set_has_overflown(); 4444 regular_clock_call(); 4445 } 4446 } 4447 4448 // We still haven't aborted. Now, let's try to get into the 4449 // termination protocol. 4450 if (do_termination && !has_aborted()) { 4451 // We cannot check whether the global stack is empty, since other 4452 // tasks might be concurrently pushing objects on it. 4453 // Separated the asserts so that we know which one fires. 4454 assert(_cm->out_of_regions(), "only way to reach here"); 4455 assert(_task_queue->size() == 0, "only way to reach here"); 4456 4457 if (_cm->verbose_low()) { 4458 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id); 4459 } 4460 4461 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 4462 4463 // The CMTask class also extends the TerminatorTerminator class, 4464 // hence its should_exit_termination() method will also decide 4465 // whether to exit the termination protocol or not. 4466 bool finished = (is_serial || 4467 _cm->terminator()->offer_termination(this)); 4468 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 4469 _termination_time_ms += 4470 termination_end_time_ms - _termination_start_time_ms; 4471 4472 if (finished) { 4473 // We're all done. 4474 4475 if (_worker_id == 0) { 4476 // let's allow task 0 to do this 4477 if (concurrent()) { 4478 assert(_cm->concurrent_marking_in_progress(), "invariant"); 4479 // we need to set this to false before the next 4480 // safepoint. This way we ensure that the marking phase 4481 // doesn't observe any more heap expansions. 4482 _cm->clear_concurrent_marking_in_progress(); 4483 } 4484 } 4485 4486 // We can now guarantee that the global stack is empty, since 4487 // all other tasks have finished. We separated the guarantees so 4488 // that, if a condition is false, we can immediately find out 4489 // which one. 4490 guarantee(_cm->out_of_regions(), "only way to reach here"); 4491 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 4492 guarantee(_task_queue->size() == 0, "only way to reach here"); 4493 guarantee(!_cm->has_overflown(), "only way to reach here"); 4494 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); 4495 4496 if (_cm->verbose_low()) { 4497 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id); 4498 } 4499 } else { 4500 // Apparently there's more work to do. Let's abort this task. It 4501 // will restart it and we can hopefully find more things to do. 4502 4503 if (_cm->verbose_low()) { 4504 gclog_or_tty->print_cr("[%u] apparently there is more work to do", 4505 _worker_id); 4506 } 4507 4508 set_has_aborted(); 4509 statsOnly( ++_aborted_termination ); 4510 } 4511 } 4512 4513 // Mainly for debugging purposes to make sure that a pointer to the 4514 // closure which was statically allocated in this frame doesn't 4515 // escape it by accident. 4516 set_cm_oop_closure(NULL); 4517 double end_time_ms = os::elapsedVTime() * 1000.0; 4518 double elapsed_time_ms = end_time_ms - _start_time_ms; 4519 // Update the step history. 4520 _step_times_ms.add(elapsed_time_ms); 4521 4522 if (has_aborted()) { 4523 // The task was aborted for some reason. 4524 4525 statsOnly( ++_aborted ); 4526 4527 if (_has_timed_out) { 4528 double diff_ms = elapsed_time_ms - _time_target_ms; 4529 // Keep statistics of how well we did with respect to hitting 4530 // our target only if we actually timed out (if we aborted for 4531 // other reasons, then the results might get skewed). 4532 _marking_step_diffs_ms.add(diff_ms); 4533 } 4534 4535 if (_cm->has_overflown()) { 4536 // This is the interesting one. We aborted because a global 4537 // overflow was raised. This means we have to restart the 4538 // marking phase and start iterating over regions. However, in 4539 // order to do this we have to make sure that all tasks stop 4540 // what they are doing and re-initialize in a safe manner. We 4541 // will achieve this with the use of two barrier sync points. 4542 4543 if (_cm->verbose_low()) { 4544 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id); 4545 } 4546 4547 if (!is_serial) { 4548 // We only need to enter the sync barrier if being called 4549 // from a parallel context 4550 _cm->enter_first_sync_barrier(_worker_id); 4551 4552 // When we exit this sync barrier we know that all tasks have 4553 // stopped doing marking work. So, it's now safe to 4554 // re-initialize our data structures. At the end of this method, 4555 // task 0 will clear the global data structures. 4556 } 4557 4558 statsOnly( ++_aborted_overflow ); 4559 4560 // We clear the local state of this task... 4561 clear_region_fields(); 4562 4563 if (!is_serial) { 4564 // ...and enter the second barrier. 4565 _cm->enter_second_sync_barrier(_worker_id); 4566 } 4567 // At this point, if we're during the concurrent phase of 4568 // marking, everything has been re-initialized and we're 4569 // ready to restart. 4570 } 4571 4572 if (_cm->verbose_low()) { 4573 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, " 4574 "elapsed = %1.2lfms <<<<<<<<<<", 4575 _worker_id, _time_target_ms, elapsed_time_ms); 4576 if (_cm->has_aborted()) { 4577 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========", 4578 _worker_id); 4579 } 4580 } 4581 } else { 4582 if (_cm->verbose_low()) { 4583 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, " 4584 "elapsed = %1.2lfms <<<<<<<<<<", 4585 _worker_id, _time_target_ms, elapsed_time_ms); 4586 } 4587 } 4588 4589 _claimed = false; 4590 } 4591 4592 CMTask::CMTask(uint worker_id, 4593 ConcurrentMark* cm, 4594 size_t* marked_bytes, 4595 BitMap* card_bm, 4596 CMTaskQueue* task_queue, 4597 CMTaskQueueSet* task_queues) 4598 : _g1h(G1CollectedHeap::heap()), 4599 _worker_id(worker_id), _cm(cm), 4600 _claimed(false), 4601 _nextMarkBitMap(NULL), _hash_seed(17), 4602 _task_queue(task_queue), 4603 _task_queues(task_queues), 4604 _cm_oop_closure(NULL), 4605 _marked_bytes_array(marked_bytes), 4606 _card_bm(card_bm) { 4607 guarantee(task_queue != NULL, "invariant"); 4608 guarantee(task_queues != NULL, "invariant"); 4609 4610 statsOnly( _clock_due_to_scanning = 0; 4611 _clock_due_to_marking = 0 ); 4612 4613 _marking_step_diffs_ms.add(0.5); 4614 } 4615 4616 // These are formatting macros that are used below to ensure 4617 // consistent formatting. The *_H_* versions are used to format the 4618 // header for a particular value and they should be kept consistent 4619 // with the corresponding macro. Also note that most of the macros add 4620 // the necessary white space (as a prefix) which makes them a bit 4621 // easier to compose. 4622 4623 // All the output lines are prefixed with this string to be able to 4624 // identify them easily in a large log file. 4625 #define G1PPRL_LINE_PREFIX "###" 4626 4627 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT 4628 #ifdef _LP64 4629 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 4630 #else // _LP64 4631 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 4632 #endif // _LP64 4633 4634 // For per-region info 4635 #define G1PPRL_TYPE_FORMAT " %-4s" 4636 #define G1PPRL_TYPE_H_FORMAT " %4s" 4637 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9) 4638 #define G1PPRL_BYTE_H_FORMAT " %9s" 4639 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 4640 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 4641 4642 // For summary info 4643 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT 4644 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT 4645 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB" 4646 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%" 4647 4648 G1PrintRegionLivenessInfoClosure:: 4649 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name) 4650 : _out(out), 4651 _total_used_bytes(0), _total_capacity_bytes(0), 4652 _total_prev_live_bytes(0), _total_next_live_bytes(0), 4653 _hum_used_bytes(0), _hum_capacity_bytes(0), 4654 _hum_prev_live_bytes(0), _hum_next_live_bytes(0), 4655 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) { 4656 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 4657 MemRegion g1_reserved = g1h->g1_reserved(); 4658 double now = os::elapsedTime(); 4659 4660 // Print the header of the output. 4661 _out->cr(); 4662 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 4663 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP" 4664 G1PPRL_SUM_ADDR_FORMAT("reserved") 4665 G1PPRL_SUM_BYTE_FORMAT("region-size"), 4666 p2i(g1_reserved.start()), p2i(g1_reserved.end()), 4667 HeapRegion::GrainBytes); 4668 _out->print_cr(G1PPRL_LINE_PREFIX); 4669 _out->print_cr(G1PPRL_LINE_PREFIX 4670 G1PPRL_TYPE_H_FORMAT 4671 G1PPRL_ADDR_BASE_H_FORMAT 4672 G1PPRL_BYTE_H_FORMAT 4673 G1PPRL_BYTE_H_FORMAT 4674 G1PPRL_BYTE_H_FORMAT 4675 G1PPRL_DOUBLE_H_FORMAT 4676 G1PPRL_BYTE_H_FORMAT 4677 G1PPRL_BYTE_H_FORMAT, 4678 "type", "address-range", 4679 "used", "prev-live", "next-live", "gc-eff", 4680 "remset", "code-roots"); 4681 _out->print_cr(G1PPRL_LINE_PREFIX 4682 G1PPRL_TYPE_H_FORMAT 4683 G1PPRL_ADDR_BASE_H_FORMAT 4684 G1PPRL_BYTE_H_FORMAT 4685 G1PPRL_BYTE_H_FORMAT 4686 G1PPRL_BYTE_H_FORMAT 4687 G1PPRL_DOUBLE_H_FORMAT 4688 G1PPRL_BYTE_H_FORMAT 4689 G1PPRL_BYTE_H_FORMAT, 4690 "", "", 4691 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", 4692 "(bytes)", "(bytes)"); 4693 } 4694 4695 // It takes as a parameter a reference to one of the _hum_* fields, it 4696 // deduces the corresponding value for a region in a humongous region 4697 // series (either the region size, or what's left if the _hum_* field 4698 // is < the region size), and updates the _hum_* field accordingly. 4699 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) { 4700 size_t bytes = 0; 4701 // The > 0 check is to deal with the prev and next live bytes which 4702 // could be 0. 4703 if (*hum_bytes > 0) { 4704 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes); 4705 *hum_bytes -= bytes; 4706 } 4707 return bytes; 4708 } 4709 4710 // It deduces the values for a region in a humongous region series 4711 // from the _hum_* fields and updates those accordingly. It assumes 4712 // that that _hum_* fields have already been set up from the "starts 4713 // humongous" region and we visit the regions in address order. 4714 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes, 4715 size_t* capacity_bytes, 4716 size_t* prev_live_bytes, 4717 size_t* next_live_bytes) { 4718 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition"); 4719 *used_bytes = get_hum_bytes(&_hum_used_bytes); 4720 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes); 4721 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes); 4722 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes); 4723 } 4724 4725 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) { 4726 const char* type = r->get_type_str(); 4727 HeapWord* bottom = r->bottom(); 4728 HeapWord* end = r->end(); 4729 size_t capacity_bytes = r->capacity(); 4730 size_t used_bytes = r->used(); 4731 size_t prev_live_bytes = r->live_bytes(); 4732 size_t next_live_bytes = r->next_live_bytes(); 4733 double gc_eff = r->gc_efficiency(); 4734 size_t remset_bytes = r->rem_set()->mem_size(); 4735 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size(); 4736 4737 if (r->is_starts_humongous()) { 4738 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 && 4739 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0, 4740 "they should have been zeroed after the last time we used them"); 4741 // Set up the _hum_* fields. 4742 _hum_capacity_bytes = capacity_bytes; 4743 _hum_used_bytes = used_bytes; 4744 _hum_prev_live_bytes = prev_live_bytes; 4745 _hum_next_live_bytes = next_live_bytes; 4746 get_hum_bytes(&used_bytes, &capacity_bytes, 4747 &prev_live_bytes, &next_live_bytes); 4748 end = bottom + HeapRegion::GrainWords; 4749 } else if (r->is_continues_humongous()) { 4750 get_hum_bytes(&used_bytes, &capacity_bytes, 4751 &prev_live_bytes, &next_live_bytes); 4752 assert(end == bottom + HeapRegion::GrainWords, "invariant"); 4753 } 4754 4755 _total_used_bytes += used_bytes; 4756 _total_capacity_bytes += capacity_bytes; 4757 _total_prev_live_bytes += prev_live_bytes; 4758 _total_next_live_bytes += next_live_bytes; 4759 _total_remset_bytes += remset_bytes; 4760 _total_strong_code_roots_bytes += strong_code_roots_bytes; 4761 4762 // Print a line for this particular region. 4763 _out->print_cr(G1PPRL_LINE_PREFIX 4764 G1PPRL_TYPE_FORMAT 4765 G1PPRL_ADDR_BASE_FORMAT 4766 G1PPRL_BYTE_FORMAT 4767 G1PPRL_BYTE_FORMAT 4768 G1PPRL_BYTE_FORMAT 4769 G1PPRL_DOUBLE_FORMAT 4770 G1PPRL_BYTE_FORMAT 4771 G1PPRL_BYTE_FORMAT, 4772 type, p2i(bottom), p2i(end), 4773 used_bytes, prev_live_bytes, next_live_bytes, gc_eff, 4774 remset_bytes, strong_code_roots_bytes); 4775 4776 return false; 4777 } 4778 4779 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 4780 // add static memory usages to remembered set sizes 4781 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size(); 4782 // Print the footer of the output. 4783 _out->print_cr(G1PPRL_LINE_PREFIX); 4784 _out->print_cr(G1PPRL_LINE_PREFIX 4785 " SUMMARY" 4786 G1PPRL_SUM_MB_FORMAT("capacity") 4787 G1PPRL_SUM_MB_PERC_FORMAT("used") 4788 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 4789 G1PPRL_SUM_MB_PERC_FORMAT("next-live") 4790 G1PPRL_SUM_MB_FORMAT("remset") 4791 G1PPRL_SUM_MB_FORMAT("code-roots"), 4792 bytes_to_mb(_total_capacity_bytes), 4793 bytes_to_mb(_total_used_bytes), 4794 perc(_total_used_bytes, _total_capacity_bytes), 4795 bytes_to_mb(_total_prev_live_bytes), 4796 perc(_total_prev_live_bytes, _total_capacity_bytes), 4797 bytes_to_mb(_total_next_live_bytes), 4798 perc(_total_next_live_bytes, _total_capacity_bytes), 4799 bytes_to_mb(_total_remset_bytes), 4800 bytes_to_mb(_total_strong_code_roots_bytes)); 4801 _out->cr(); 4802 }