1 /* 2 * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/symbolTable.hpp" 27 #include "gc_implementation/g1/concurrentMark.inline.hpp" 28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 30 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 31 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 32 #include "gc_implementation/g1/g1Log.hpp" 33 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 34 #include "gc_implementation/g1/g1RemSet.hpp" 35 #include "gc_implementation/g1/heapRegion.inline.hpp" 36 #include "gc_implementation/g1/heapRegionRemSet.hpp" 37 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 38 #include "gc_implementation/shared/vmGCOperations.hpp" 39 #include "memory/genOopClosures.inline.hpp" 40 #include "memory/referencePolicy.hpp" 41 #include "memory/resourceArea.hpp" 42 #include "oops/oop.inline.hpp" 43 #include "runtime/handles.inline.hpp" 44 #include "runtime/java.hpp" 45 #include "services/memTracker.hpp" 46 47 // Concurrent marking bit map wrapper 48 49 CMBitMapRO::CMBitMapRO(int shifter) : 50 _bm(), 51 _shifter(shifter) { 52 _bmStartWord = 0; 53 _bmWordSize = 0; 54 } 55 56 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr, 57 HeapWord* limit) const { 58 // First we must round addr *up* to a possible object boundary. 59 addr = (HeapWord*)align_size_up((intptr_t)addr, 60 HeapWordSize << _shifter); 61 size_t addrOffset = heapWordToOffset(addr); 62 if (limit == NULL) { 63 limit = _bmStartWord + _bmWordSize; 64 } 65 size_t limitOffset = heapWordToOffset(limit); 66 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); 67 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 68 assert(nextAddr >= addr, "get_next_one postcondition"); 69 assert(nextAddr == limit || isMarked(nextAddr), 70 "get_next_one postcondition"); 71 return nextAddr; 72 } 73 74 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr, 75 HeapWord* limit) const { 76 size_t addrOffset = heapWordToOffset(addr); 77 if (limit == NULL) { 78 limit = _bmStartWord + _bmWordSize; 79 } 80 size_t limitOffset = heapWordToOffset(limit); 81 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset); 82 HeapWord* nextAddr = offsetToHeapWord(nextOffset); 83 assert(nextAddr >= addr, "get_next_one postcondition"); 84 assert(nextAddr == limit || !isMarked(nextAddr), 85 "get_next_one postcondition"); 86 return nextAddr; 87 } 88 89 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const { 90 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check"); 91 return (int) (diff >> _shifter); 92 } 93 94 #ifndef PRODUCT 95 bool CMBitMapRO::covers(ReservedSpace heap_rs) const { 96 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 97 assert(((size_t)_bm.size() * ((size_t)1 << _shifter)) == _bmWordSize, 98 "size inconsistency"); 99 return _bmStartWord == (HeapWord*)(heap_rs.base()) && 100 _bmWordSize == heap_rs.size()>>LogHeapWordSize; 101 } 102 #endif 103 104 void CMBitMapRO::print_on_error(outputStream* st, const char* prefix) const { 105 _bm.print_on_error(st, prefix); 106 } 107 108 bool CMBitMap::allocate(ReservedSpace heap_rs) { 109 _bmStartWord = (HeapWord*)(heap_rs.base()); 110 _bmWordSize = heap_rs.size()/HeapWordSize; // heap_rs.size() is in bytes 111 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 112 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 113 if (!brs.is_reserved()) { 114 warning("ConcurrentMark marking bit map allocation failure"); 115 return false; 116 } 117 MemTracker::record_virtual_memory_type((address)brs.base(), mtGC); 118 // For now we'll just commit all of the bit map up front. 119 // Later on we'll try to be more parsimonious with swap. 120 if (!_virtual_space.initialize(brs, brs.size())) { 121 warning("ConcurrentMark marking bit map backing store failure"); 122 return false; 123 } 124 assert(_virtual_space.committed_size() == brs.size(), 125 "didn't reserve backing store for all of concurrent marking bit map?"); 126 _bm.set_map((uintptr_t*)_virtual_space.low()); 127 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 128 _bmWordSize, "inconsistency in bit map sizing"); 129 _bm.set_size(_bmWordSize >> _shifter); 130 return true; 131 } 132 133 void CMBitMap::clearAll() { 134 _bm.clear(); 135 return; 136 } 137 138 void CMBitMap::markRange(MemRegion mr) { 139 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 140 assert(!mr.is_empty(), "unexpected empty region"); 141 assert((offsetToHeapWord(heapWordToOffset(mr.end())) == 142 ((HeapWord *) mr.end())), 143 "markRange memory region end is not card aligned"); 144 // convert address range into offset range 145 _bm.at_put_range(heapWordToOffset(mr.start()), 146 heapWordToOffset(mr.end()), true); 147 } 148 149 void CMBitMap::clearRange(MemRegion mr) { 150 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 151 assert(!mr.is_empty(), "unexpected empty region"); 152 // convert address range into offset range 153 _bm.at_put_range(heapWordToOffset(mr.start()), 154 heapWordToOffset(mr.end()), false); 155 } 156 157 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr, 158 HeapWord* end_addr) { 159 HeapWord* start = getNextMarkedWordAddress(addr); 160 start = MIN2(start, end_addr); 161 HeapWord* end = getNextUnmarkedWordAddress(start); 162 end = MIN2(end, end_addr); 163 assert(start <= end, "Consistency check"); 164 MemRegion mr(start, end); 165 if (!mr.is_empty()) { 166 clearRange(mr); 167 } 168 return mr; 169 } 170 171 CMMarkStack::CMMarkStack(ConcurrentMark* cm) : 172 _base(NULL), _cm(cm) 173 #ifdef ASSERT 174 , _drain_in_progress(false) 175 , _drain_in_progress_yields(false) 176 #endif 177 {} 178 179 bool CMMarkStack::allocate(size_t capacity) { 180 // allocate a stack of the requisite depth 181 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop))); 182 if (!rs.is_reserved()) { 183 warning("ConcurrentMark MarkStack allocation failure"); 184 return false; 185 } 186 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 187 if (!_virtual_space.initialize(rs, rs.size())) { 188 warning("ConcurrentMark MarkStack backing store failure"); 189 // Release the virtual memory reserved for the marking stack 190 rs.release(); 191 return false; 192 } 193 assert(_virtual_space.committed_size() == rs.size(), 194 "Didn't reserve backing store for all of ConcurrentMark stack?"); 195 _base = (oop*) _virtual_space.low(); 196 setEmpty(); 197 _capacity = (jint) capacity; 198 _saved_index = -1; 199 _should_expand = false; 200 NOT_PRODUCT(_max_depth = 0); 201 return true; 202 } 203 204 void CMMarkStack::expand() { 205 // Called, during remark, if we've overflown the marking stack during marking. 206 assert(isEmpty(), "stack should been emptied while handling overflow"); 207 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted"); 208 // Clear expansion flag 209 _should_expand = false; 210 if (_capacity == (jint) MarkStackSizeMax) { 211 if (PrintGCDetails && Verbose) { 212 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit"); 213 } 214 return; 215 } 216 // Double capacity if possible 217 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax); 218 // Do not give up existing stack until we have managed to 219 // get the double capacity that we desired. 220 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity * 221 sizeof(oop))); 222 if (rs.is_reserved()) { 223 // Release the backing store associated with old stack 224 _virtual_space.release(); 225 // Reinitialize virtual space for new stack 226 if (!_virtual_space.initialize(rs, rs.size())) { 227 fatal("Not enough swap for expanded marking stack capacity"); 228 } 229 _base = (oop*)(_virtual_space.low()); 230 _index = 0; 231 _capacity = new_capacity; 232 } else { 233 if (PrintGCDetails && Verbose) { 234 // Failed to double capacity, continue; 235 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from " 236 SIZE_FORMAT"K to " SIZE_FORMAT"K", 237 _capacity / K, new_capacity / K); 238 } 239 } 240 } 241 242 void CMMarkStack::set_should_expand() { 243 // If we're resetting the marking state because of an 244 // marking stack overflow, record that we should, if 245 // possible, expand the stack. 246 _should_expand = _cm->has_overflown(); 247 } 248 249 CMMarkStack::~CMMarkStack() { 250 if (_base != NULL) { 251 _base = NULL; 252 _virtual_space.release(); 253 } 254 } 255 256 void CMMarkStack::par_push(oop ptr) { 257 while (true) { 258 if (isFull()) { 259 _overflow = true; 260 return; 261 } 262 // Otherwise... 263 jint index = _index; 264 jint next_index = index+1; 265 jint res = Atomic::cmpxchg(next_index, &_index, index); 266 if (res == index) { 267 _base[index] = ptr; 268 // Note that we don't maintain this atomically. We could, but it 269 // doesn't seem necessary. 270 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 271 return; 272 } 273 // Otherwise, we need to try again. 274 } 275 } 276 277 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) { 278 while (true) { 279 if (isFull()) { 280 _overflow = true; 281 return; 282 } 283 // Otherwise... 284 jint index = _index; 285 jint next_index = index + n; 286 if (next_index > _capacity) { 287 _overflow = true; 288 return; 289 } 290 jint res = Atomic::cmpxchg(next_index, &_index, index); 291 if (res == index) { 292 for (int i = 0; i < n; i++) { 293 int ind = index + i; 294 assert(ind < _capacity, "By overflow test above."); 295 _base[ind] = ptr_arr[i]; 296 } 297 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 298 return; 299 } 300 // Otherwise, we need to try again. 301 } 302 } 303 304 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { 305 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 306 jint start = _index; 307 jint next_index = start + n; 308 if (next_index > _capacity) { 309 _overflow = true; 310 return; 311 } 312 // Otherwise. 313 _index = next_index; 314 for (int i = 0; i < n; i++) { 315 int ind = start + i; 316 assert(ind < _capacity, "By overflow test above."); 317 _base[ind] = ptr_arr[i]; 318 } 319 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 320 } 321 322 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { 323 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 324 jint index = _index; 325 if (index == 0) { 326 *n = 0; 327 return false; 328 } else { 329 int k = MIN2(max, index); 330 jint new_ind = index - k; 331 for (int j = 0; j < k; j++) { 332 ptr_arr[j] = _base[new_ind + j]; 333 } 334 _index = new_ind; 335 *n = k; 336 return true; 337 } 338 } 339 340 template<class OopClosureClass> 341 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) { 342 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after 343 || SafepointSynchronize::is_at_safepoint(), 344 "Drain recursion must be yield-safe."); 345 bool res = true; 346 debug_only(_drain_in_progress = true); 347 debug_only(_drain_in_progress_yields = yield_after); 348 while (!isEmpty()) { 349 oop newOop = pop(); 350 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop"); 351 assert(newOop->is_oop(), "Expected an oop"); 352 assert(bm == NULL || bm->isMarked((HeapWord*)newOop), 353 "only grey objects on this stack"); 354 newOop->oop_iterate(cl); 355 if (yield_after && _cm->do_yield_check()) { 356 res = false; 357 break; 358 } 359 } 360 debug_only(_drain_in_progress = false); 361 return res; 362 } 363 364 void CMMarkStack::note_start_of_gc() { 365 assert(_saved_index == -1, 366 "note_start_of_gc()/end_of_gc() bracketed incorrectly"); 367 _saved_index = _index; 368 } 369 370 void CMMarkStack::note_end_of_gc() { 371 // This is intentionally a guarantee, instead of an assert. If we 372 // accidentally add something to the mark stack during GC, it 373 // will be a correctness issue so it's better if we crash. we'll 374 // only check this once per GC anyway, so it won't be a performance 375 // issue in any way. 376 guarantee(_saved_index == _index, 377 err_msg("saved index: %d index: %d", _saved_index, _index)); 378 _saved_index = -1; 379 } 380 381 void CMMarkStack::oops_do(OopClosure* f) { 382 assert(_saved_index == _index, 383 err_msg("saved index: %d index: %d", _saved_index, _index)); 384 for (int i = 0; i < _index; i += 1) { 385 f->do_oop(&_base[i]); 386 } 387 } 388 389 bool ConcurrentMark::not_yet_marked(oop obj) const { 390 return _g1h->is_obj_ill(obj); 391 } 392 393 CMRootRegions::CMRootRegions() : 394 _young_list(NULL), _cm(NULL), _scan_in_progress(false), 395 _should_abort(false), _next_survivor(NULL) { } 396 397 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) { 398 _young_list = g1h->young_list(); 399 _cm = cm; 400 } 401 402 void CMRootRegions::prepare_for_scan() { 403 assert(!scan_in_progress(), "pre-condition"); 404 405 // Currently, only survivors can be root regions. 406 assert(_next_survivor == NULL, "pre-condition"); 407 _next_survivor = _young_list->first_survivor_region(); 408 _scan_in_progress = (_next_survivor != NULL); 409 _should_abort = false; 410 } 411 412 HeapRegion* CMRootRegions::claim_next() { 413 if (_should_abort) { 414 // If someone has set the should_abort flag, we return NULL to 415 // force the caller to bail out of their loop. 416 return NULL; 417 } 418 419 // Currently, only survivors can be root regions. 420 HeapRegion* res = _next_survivor; 421 if (res != NULL) { 422 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 423 // Read it again in case it changed while we were waiting for the lock. 424 res = _next_survivor; 425 if (res != NULL) { 426 if (res == _young_list->last_survivor_region()) { 427 // We just claimed the last survivor so store NULL to indicate 428 // that we're done. 429 _next_survivor = NULL; 430 } else { 431 _next_survivor = res->get_next_young_region(); 432 } 433 } else { 434 // Someone else claimed the last survivor while we were trying 435 // to take the lock so nothing else to do. 436 } 437 } 438 assert(res == NULL || res->is_survivor(), "post-condition"); 439 440 return res; 441 } 442 443 void CMRootRegions::scan_finished() { 444 assert(scan_in_progress(), "pre-condition"); 445 446 // Currently, only survivors can be root regions. 447 if (!_should_abort) { 448 assert(_next_survivor == NULL, "we should have claimed all survivors"); 449 } 450 _next_survivor = NULL; 451 452 { 453 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 454 _scan_in_progress = false; 455 RootRegionScan_lock->notify_all(); 456 } 457 } 458 459 bool CMRootRegions::wait_until_scan_finished() { 460 if (!scan_in_progress()) return false; 461 462 { 463 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 464 while (scan_in_progress()) { 465 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 466 } 467 } 468 return true; 469 } 470 471 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 472 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 473 #endif // _MSC_VER 474 475 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 476 return MAX2((n_par_threads + 2) / 4, 1U); 477 } 478 479 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs) : 480 _g1h(g1h), 481 _markBitMap1(MinObjAlignment - 1), 482 _markBitMap2(MinObjAlignment - 1), 483 484 _parallel_marking_threads(0), 485 _max_parallel_marking_threads(0), 486 _sleep_factor(0.0), 487 _marking_task_overhead(1.0), 488 _cleanup_sleep_factor(0.0), 489 _cleanup_task_overhead(1.0), 490 _cleanup_list("Cleanup List"), 491 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/), 492 _card_bm((heap_rs.size() + CardTableModRefBS::card_size - 1) >> 493 CardTableModRefBS::card_shift, 494 false /* in_resource_area*/), 495 496 _prevMarkBitMap(&_markBitMap1), 497 _nextMarkBitMap(&_markBitMap2), 498 499 _markStack(this), 500 // _finger set in set_non_marking_state 501 502 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)), 503 // _active_tasks set in set_non_marking_state 504 // _tasks set inside the constructor 505 _task_queues(new CMTaskQueueSet((int) _max_worker_id)), 506 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)), 507 508 _has_overflown(false), 509 _concurrent(false), 510 _has_aborted(false), 511 _restart_for_overflow(false), 512 _concurrent_marking_in_progress(false), 513 514 // _verbose_level set below 515 516 _init_times(), 517 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 518 _cleanup_times(), 519 _total_counting_time(0.0), 520 _total_rs_scrub_time(0.0), 521 522 _parallel_workers(NULL), 523 524 _count_card_bitmaps(NULL), 525 _count_marked_bytes(NULL), 526 _completed_initialization(false) { 527 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel; 528 if (verbose_level < no_verbose) { 529 verbose_level = no_verbose; 530 } 531 if (verbose_level > high_verbose) { 532 verbose_level = high_verbose; 533 } 534 _verbose_level = verbose_level; 535 536 if (verbose_low()) { 537 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", " 538 "heap end = "PTR_FORMAT, _heap_start, _heap_end); 539 } 540 541 if (!_markBitMap1.allocate(heap_rs)) { 542 warning("Failed to allocate first CM bit map"); 543 return; 544 } 545 if (!_markBitMap2.allocate(heap_rs)) { 546 warning("Failed to allocate second CM bit map"); 547 return; 548 } 549 550 // Create & start a ConcurrentMark thread. 551 _cmThread = new ConcurrentMarkThread(this); 552 assert(cmThread() != NULL, "CM Thread should have been created"); 553 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 554 555 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 556 assert(_markBitMap1.covers(heap_rs), "_markBitMap1 inconsistency"); 557 assert(_markBitMap2.covers(heap_rs), "_markBitMap2 inconsistency"); 558 559 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 560 satb_qs.set_buffer_size(G1SATBBufferSize); 561 562 _root_regions.init(_g1h, this); 563 564 if (ConcGCThreads > ParallelGCThreads) { 565 warning("Can't have more ConcGCThreads (" UINT32_FORMAT ") " 566 "than ParallelGCThreads (" UINT32_FORMAT ").", 567 ConcGCThreads, ParallelGCThreads); 568 return; 569 } 570 if (ParallelGCThreads == 0) { 571 // if we are not running with any parallel GC threads we will not 572 // spawn any marking threads either 573 _parallel_marking_threads = 0; 574 _max_parallel_marking_threads = 0; 575 _sleep_factor = 0.0; 576 _marking_task_overhead = 1.0; 577 } else { 578 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 579 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 580 // if both are set 581 _sleep_factor = 0.0; 582 _marking_task_overhead = 1.0; 583 } else if (G1MarkingOverheadPercent > 0) { 584 // We will calculate the number of parallel marking threads based 585 // on a target overhead with respect to the soft real-time goal 586 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 587 double overall_cm_overhead = 588 (double) MaxGCPauseMillis * marking_overhead / 589 (double) GCPauseIntervalMillis; 590 double cpu_ratio = 1.0 / (double) os::processor_count(); 591 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 592 double marking_task_overhead = 593 overall_cm_overhead / marking_thread_num * 594 (double) os::processor_count(); 595 double sleep_factor = 596 (1.0 - marking_task_overhead) / marking_task_overhead; 597 598 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num); 599 _sleep_factor = sleep_factor; 600 _marking_task_overhead = marking_task_overhead; 601 } else { 602 // Calculate the number of parallel marking threads by scaling 603 // the number of parallel GC threads. 604 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads); 605 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num); 606 _sleep_factor = 0.0; 607 _marking_task_overhead = 1.0; 608 } 609 610 assert(ConcGCThreads > 0, "Should have been set"); 611 _parallel_marking_threads = (uint) ConcGCThreads; 612 _max_parallel_marking_threads = _parallel_marking_threads; 613 614 if (parallel_marking_threads() > 1) { 615 _cleanup_task_overhead = 1.0; 616 } else { 617 _cleanup_task_overhead = marking_task_overhead(); 618 } 619 _cleanup_sleep_factor = 620 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead(); 621 622 #if 0 623 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads()); 624 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead()); 625 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor()); 626 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead()); 627 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor()); 628 #endif 629 630 guarantee(parallel_marking_threads() > 0, "peace of mind"); 631 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads", 632 _max_parallel_marking_threads, false, true); 633 if (_parallel_workers == NULL) { 634 vm_exit_during_initialization("Failed necessary allocation."); 635 } else { 636 _parallel_workers->initialize_workers(); 637 } 638 } 639 640 if (FLAG_IS_DEFAULT(MarkStackSize)) { 641 uintx mark_stack_size = 642 MIN2(MarkStackSizeMax, 643 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE))); 644 // Verify that the calculated value for MarkStackSize is in range. 645 // It would be nice to use the private utility routine from Arguments. 646 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 647 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): " 648 "must be between " UINTX_FORMAT " and " UINTX_FORMAT, 649 mark_stack_size, 1, MarkStackSizeMax); 650 return; 651 } 652 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size); 653 } else { 654 // Verify MarkStackSize is in range. 655 if (FLAG_IS_CMDLINE(MarkStackSize)) { 656 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 657 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 658 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): " 659 "must be between " UINTX_FORMAT " and " UINTX_FORMAT, 660 MarkStackSize, 1, MarkStackSizeMax); 661 return; 662 } 663 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 664 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 665 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")" 666 " or for MarkStackSizeMax (" UINTX_FORMAT ")", 667 MarkStackSize, MarkStackSizeMax); 668 return; 669 } 670 } 671 } 672 } 673 674 if (!_markStack.allocate(MarkStackSize)) { 675 warning("Failed to allocate CM marking stack"); 676 return; 677 } 678 679 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC); 680 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC); 681 682 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC); 683 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC); 684 685 BitMap::idx_t card_bm_size = _card_bm.size(); 686 687 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 688 _active_tasks = _max_worker_id; 689 690 size_t max_regions = (size_t) _g1h->max_regions(); 691 for (uint i = 0; i < _max_worker_id; ++i) { 692 CMTaskQueue* task_queue = new CMTaskQueue(); 693 task_queue->initialize(); 694 _task_queues->register_queue(i, task_queue); 695 696 _count_card_bitmaps[i] = BitMap(card_bm_size, false); 697 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC); 698 699 _tasks[i] = new CMTask(i, this, 700 _count_marked_bytes[i], 701 &_count_card_bitmaps[i], 702 task_queue, _task_queues); 703 704 _accum_task_vtime[i] = 0.0; 705 } 706 707 // Calculate the card number for the bottom of the heap. Used 708 // in biasing indexes into the accounting card bitmaps. 709 _heap_bottom_card_num = 710 intptr_t(uintptr_t(_g1h->reserved_region().start()) >> 711 CardTableModRefBS::card_shift); 712 713 // Clear all the liveness counting data 714 clear_all_count_data(); 715 716 // so that the call below can read a sensible value 717 _heap_start = (HeapWord*) heap_rs.base(); 718 set_non_marking_state(); 719 _completed_initialization = true; 720 } 721 722 void ConcurrentMark::update_g1_committed(bool force) { 723 // If concurrent marking is not in progress, then we do not need to 724 // update _heap_end. 725 if (!concurrent_marking_in_progress() && !force) return; 726 727 MemRegion committed = _g1h->g1_committed(); 728 assert(committed.start() == _heap_start, "start shouldn't change"); 729 HeapWord* new_end = committed.end(); 730 if (new_end > _heap_end) { 731 // The heap has been expanded. 732 733 _heap_end = new_end; 734 } 735 // Notice that the heap can also shrink. However, this only happens 736 // during a Full GC (at least currently) and the entire marking 737 // phase will bail out and the task will not be restarted. So, let's 738 // do nothing. 739 } 740 741 void ConcurrentMark::reset() { 742 // Starting values for these two. This should be called in a STW 743 // phase. CM will be notified of any future g1_committed expansions 744 // will be at the end of evacuation pauses, when tasks are 745 // inactive. 746 MemRegion committed = _g1h->g1_committed(); 747 _heap_start = committed.start(); 748 _heap_end = committed.end(); 749 750 // Separated the asserts so that we know which one fires. 751 assert(_heap_start != NULL, "heap bounds should look ok"); 752 assert(_heap_end != NULL, "heap bounds should look ok"); 753 assert(_heap_start < _heap_end, "heap bounds should look ok"); 754 755 // Reset all the marking data structures and any necessary flags 756 reset_marking_state(); 757 758 if (verbose_low()) { 759 gclog_or_tty->print_cr("[global] resetting"); 760 } 761 762 // We do reset all of them, since different phases will use 763 // different number of active threads. So, it's easiest to have all 764 // of them ready. 765 for (uint i = 0; i < _max_worker_id; ++i) { 766 _tasks[i]->reset(_nextMarkBitMap); 767 } 768 769 // we need this to make sure that the flag is on during the evac 770 // pause with initial mark piggy-backed 771 set_concurrent_marking_in_progress(); 772 } 773 774 775 void ConcurrentMark::reset_marking_state(bool clear_overflow) { 776 _markStack.set_should_expand(); 777 _markStack.setEmpty(); // Also clears the _markStack overflow flag 778 if (clear_overflow) { 779 clear_has_overflown(); 780 } else { 781 assert(has_overflown(), "pre-condition"); 782 } 783 _finger = _heap_start; 784 785 for (uint i = 0; i < _max_worker_id; ++i) { 786 CMTaskQueue* queue = _task_queues->queue(i); 787 queue->set_empty(); 788 } 789 } 790 791 void ConcurrentMark::set_concurrency(uint active_tasks) { 792 assert(active_tasks <= _max_worker_id, "we should not have more"); 793 794 _active_tasks = active_tasks; 795 // Need to update the three data structures below according to the 796 // number of active threads for this phase. 797 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 798 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 799 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 800 } 801 802 void ConcurrentMark::set_concurrency_and_phase(uint active_tasks, bool concurrent) { 803 set_concurrency(active_tasks); 804 805 _concurrent = concurrent; 806 // We propagate this to all tasks, not just the active ones. 807 for (uint i = 0; i < _max_worker_id; ++i) 808 _tasks[i]->set_concurrent(concurrent); 809 810 if (concurrent) { 811 set_concurrent_marking_in_progress(); 812 } else { 813 // We currently assume that the concurrent flag has been set to 814 // false before we start remark. At this point we should also be 815 // in a STW phase. 816 assert(!concurrent_marking_in_progress(), "invariant"); 817 assert(_finger == _heap_end, 818 err_msg("only way to get here: _finger: "PTR_FORMAT", _heap_end: "PTR_FORMAT, 819 _finger, _heap_end)); 820 update_g1_committed(true); 821 } 822 } 823 824 void ConcurrentMark::set_non_marking_state() { 825 // We set the global marking state to some default values when we're 826 // not doing marking. 827 reset_marking_state(); 828 _active_tasks = 0; 829 clear_concurrent_marking_in_progress(); 830 } 831 832 ConcurrentMark::~ConcurrentMark() { 833 // The ConcurrentMark instance is never freed. 834 ShouldNotReachHere(); 835 } 836 837 void ConcurrentMark::clearNextBitmap() { 838 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 839 G1CollectorPolicy* g1p = g1h->g1_policy(); 840 841 // Make sure that the concurrent mark thread looks to still be in 842 // the current cycle. 843 guarantee(cmThread()->during_cycle(), "invariant"); 844 845 // We are finishing up the current cycle by clearing the next 846 // marking bitmap and getting it ready for the next cycle. During 847 // this time no other cycle can start. So, let's make sure that this 848 // is the case. 849 guarantee(!g1h->mark_in_progress(), "invariant"); 850 851 // clear the mark bitmap (no grey objects to start with). 852 // We need to do this in chunks and offer to yield in between 853 // each chunk. 854 HeapWord* start = _nextMarkBitMap->startWord(); 855 HeapWord* end = _nextMarkBitMap->endWord(); 856 HeapWord* cur = start; 857 size_t chunkSize = M; 858 while (cur < end) { 859 HeapWord* next = cur + chunkSize; 860 if (next > end) { 861 next = end; 862 } 863 MemRegion mr(cur,next); 864 _nextMarkBitMap->clearRange(mr); 865 cur = next; 866 do_yield_check(); 867 868 // Repeat the asserts from above. We'll do them as asserts here to 869 // minimize their overhead on the product. However, we'll have 870 // them as guarantees at the beginning / end of the bitmap 871 // clearing to get some checking in the product. 872 assert(cmThread()->during_cycle(), "invariant"); 873 assert(!g1h->mark_in_progress(), "invariant"); 874 } 875 876 // Clear the liveness counting data 877 clear_all_count_data(); 878 879 // Repeat the asserts from above. 880 guarantee(cmThread()->during_cycle(), "invariant"); 881 guarantee(!g1h->mark_in_progress(), "invariant"); 882 } 883 884 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 885 public: 886 bool doHeapRegion(HeapRegion* r) { 887 if (!r->continuesHumongous()) { 888 r->note_start_of_marking(); 889 } 890 return false; 891 } 892 }; 893 894 void ConcurrentMark::checkpointRootsInitialPre() { 895 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 896 G1CollectorPolicy* g1p = g1h->g1_policy(); 897 898 _has_aborted = false; 899 900 #ifndef PRODUCT 901 if (G1PrintReachableAtInitialMark) { 902 print_reachable("at-cycle-start", 903 VerifyOption_G1UsePrevMarking, true /* all */); 904 } 905 #endif 906 907 // Initialise marking structures. This has to be done in a STW phase. 908 reset(); 909 910 // For each region note start of marking. 911 NoteStartOfMarkHRClosure startcl; 912 g1h->heap_region_iterate(&startcl); 913 } 914 915 916 void ConcurrentMark::checkpointRootsInitialPost() { 917 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 918 919 // If we force an overflow during remark, the remark operation will 920 // actually abort and we'll restart concurrent marking. If we always 921 // force an oveflow during remark we'll never actually complete the 922 // marking phase. So, we initilize this here, at the start of the 923 // cycle, so that at the remaining overflow number will decrease at 924 // every remark and we'll eventually not need to cause one. 925 force_overflow_stw()->init(); 926 927 // Start Concurrent Marking weak-reference discovery. 928 ReferenceProcessor* rp = g1h->ref_processor_cm(); 929 // enable ("weak") refs discovery 930 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 931 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 932 933 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 934 // This is the start of the marking cycle, we're expected all 935 // threads to have SATB queues with active set to false. 936 satb_mq_set.set_active_all_threads(true, /* new active value */ 937 false /* expected_active */); 938 939 _root_regions.prepare_for_scan(); 940 941 // update_g1_committed() will be called at the end of an evac pause 942 // when marking is on. So, it's also called at the end of the 943 // initial-mark pause to update the heap end, if the heap expands 944 // during it. No need to call it here. 945 } 946 947 /* 948 * Notice that in the next two methods, we actually leave the STS 949 * during the barrier sync and join it immediately afterwards. If we 950 * do not do this, the following deadlock can occur: one thread could 951 * be in the barrier sync code, waiting for the other thread to also 952 * sync up, whereas another one could be trying to yield, while also 953 * waiting for the other threads to sync up too. 954 * 955 * Note, however, that this code is also used during remark and in 956 * this case we should not attempt to leave / enter the STS, otherwise 957 * we'll either hit an asseert (debug / fastdebug) or deadlock 958 * (product). So we should only leave / enter the STS if we are 959 * operating concurrently. 960 * 961 * Because the thread that does the sync barrier has left the STS, it 962 * is possible to be suspended for a Full GC or an evacuation pause 963 * could occur. This is actually safe, since the entering the sync 964 * barrier is one of the last things do_marking_step() does, and it 965 * doesn't manipulate any data structures afterwards. 966 */ 967 968 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 969 if (verbose_low()) { 970 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id); 971 } 972 973 if (concurrent()) { 974 ConcurrentGCThread::stsLeave(); 975 } 976 _first_overflow_barrier_sync.enter(); 977 if (concurrent()) { 978 ConcurrentGCThread::stsJoin(); 979 } 980 // at this point everyone should have synced up and not be doing any 981 // more work 982 983 if (verbose_low()) { 984 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id); 985 } 986 987 // If we're executing the concurrent phase of marking, reset the marking 988 // state; otherwise the marking state is reset after reference processing, 989 // during the remark pause. 990 // If we reset here as a result of an overflow during the remark we will 991 // see assertion failures from any subsequent set_concurrency_and_phase() 992 // calls. 993 if (concurrent()) { 994 // let the task associated with with worker 0 do this 995 if (worker_id == 0) { 996 // task 0 is responsible for clearing the global data structures 997 // We should be here because of an overflow. During STW we should 998 // not clear the overflow flag since we rely on it being true when 999 // we exit this method to abort the pause and restart concurent 1000 // marking. 1001 reset_marking_state(true /* clear_overflow */); 1002 force_overflow()->update(); 1003 1004 if (G1Log::fine()) { 1005 gclog_or_tty->date_stamp(PrintGCDateStamps); 1006 gclog_or_tty->stamp(PrintGCTimeStamps); 1007 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]"); 1008 } 1009 } 1010 } 1011 1012 // after this, each task should reset its own data structures then 1013 // then go into the second barrier 1014 } 1015 1016 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 1017 if (verbose_low()) { 1018 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id); 1019 } 1020 1021 if (concurrent()) { 1022 ConcurrentGCThread::stsLeave(); 1023 } 1024 _second_overflow_barrier_sync.enter(); 1025 if (concurrent()) { 1026 ConcurrentGCThread::stsJoin(); 1027 } 1028 // at this point everything should be re-initialized and ready to go 1029 1030 if (verbose_low()) { 1031 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id); 1032 } 1033 } 1034 1035 #ifndef PRODUCT 1036 void ForceOverflowSettings::init() { 1037 _num_remaining = G1ConcMarkForceOverflow; 1038 _force = false; 1039 update(); 1040 } 1041 1042 void ForceOverflowSettings::update() { 1043 if (_num_remaining > 0) { 1044 _num_remaining -= 1; 1045 _force = true; 1046 } else { 1047 _force = false; 1048 } 1049 } 1050 1051 bool ForceOverflowSettings::should_force() { 1052 if (_force) { 1053 _force = false; 1054 return true; 1055 } else { 1056 return false; 1057 } 1058 } 1059 #endif // !PRODUCT 1060 1061 class CMConcurrentMarkingTask: public AbstractGangTask { 1062 private: 1063 ConcurrentMark* _cm; 1064 ConcurrentMarkThread* _cmt; 1065 1066 public: 1067 void work(uint worker_id) { 1068 assert(Thread::current()->is_ConcurrentGC_thread(), 1069 "this should only be done by a conc GC thread"); 1070 ResourceMark rm; 1071 1072 double start_vtime = os::elapsedVTime(); 1073 1074 ConcurrentGCThread::stsJoin(); 1075 1076 assert(worker_id < _cm->active_tasks(), "invariant"); 1077 CMTask* the_task = _cm->task(worker_id); 1078 the_task->record_start_time(); 1079 if (!_cm->has_aborted()) { 1080 do { 1081 double start_vtime_sec = os::elapsedVTime(); 1082 double start_time_sec = os::elapsedTime(); 1083 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1084 1085 the_task->do_marking_step(mark_step_duration_ms, 1086 true /* do_termination */, 1087 false /* is_serial*/); 1088 1089 double end_time_sec = os::elapsedTime(); 1090 double end_vtime_sec = os::elapsedVTime(); 1091 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 1092 double elapsed_time_sec = end_time_sec - start_time_sec; 1093 _cm->clear_has_overflown(); 1094 1095 bool ret = _cm->do_yield_check(worker_id); 1096 1097 jlong sleep_time_ms; 1098 if (!_cm->has_aborted() && the_task->has_aborted()) { 1099 sleep_time_ms = 1100 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 1101 ConcurrentGCThread::stsLeave(); 1102 os::sleep(Thread::current(), sleep_time_ms, false); 1103 ConcurrentGCThread::stsJoin(); 1104 } 1105 double end_time2_sec = os::elapsedTime(); 1106 double elapsed_time2_sec = end_time2_sec - start_time_sec; 1107 1108 #if 0 1109 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, " 1110 "overhead %1.4lf", 1111 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms, 1112 the_task->conc_overhead(os::elapsedTime()) * 8.0); 1113 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms", 1114 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0); 1115 #endif 1116 } while (!_cm->has_aborted() && the_task->has_aborted()); 1117 } 1118 the_task->record_end_time(); 1119 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 1120 1121 ConcurrentGCThread::stsLeave(); 1122 1123 double end_vtime = os::elapsedVTime(); 1124 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 1125 } 1126 1127 CMConcurrentMarkingTask(ConcurrentMark* cm, 1128 ConcurrentMarkThread* cmt) : 1129 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 1130 1131 ~CMConcurrentMarkingTask() { } 1132 }; 1133 1134 // Calculates the number of active workers for a concurrent 1135 // phase. 1136 uint ConcurrentMark::calc_parallel_marking_threads() { 1137 if (G1CollectedHeap::use_parallel_gc_threads()) { 1138 uint n_conc_workers = 0; 1139 if (!UseDynamicNumberOfGCThreads || 1140 (!FLAG_IS_DEFAULT(ConcGCThreads) && 1141 !ForceDynamicNumberOfGCThreads)) { 1142 n_conc_workers = max_parallel_marking_threads(); 1143 } else { 1144 n_conc_workers = 1145 AdaptiveSizePolicy::calc_default_active_workers( 1146 max_parallel_marking_threads(), 1147 1, /* Minimum workers */ 1148 parallel_marking_threads(), 1149 Threads::number_of_non_daemon_threads()); 1150 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 1151 // that scaling has already gone into "_max_parallel_marking_threads". 1152 } 1153 assert(n_conc_workers > 0, "Always need at least 1"); 1154 return n_conc_workers; 1155 } 1156 // If we are not running with any parallel GC threads we will not 1157 // have spawned any marking threads either. Hence the number of 1158 // concurrent workers should be 0. 1159 return 0; 1160 } 1161 1162 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) { 1163 // Currently, only survivors can be root regions. 1164 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 1165 G1RootRegionScanClosure cl(_g1h, this, worker_id); 1166 1167 const uintx interval = PrefetchScanIntervalInBytes; 1168 HeapWord* curr = hr->bottom(); 1169 const HeapWord* end = hr->top(); 1170 while (curr < end) { 1171 Prefetch::read(curr, interval); 1172 oop obj = oop(curr); 1173 int size = obj->oop_iterate(&cl); 1174 assert(size == obj->size(), "sanity"); 1175 curr += size; 1176 } 1177 } 1178 1179 class CMRootRegionScanTask : public AbstractGangTask { 1180 private: 1181 ConcurrentMark* _cm; 1182 1183 public: 1184 CMRootRegionScanTask(ConcurrentMark* cm) : 1185 AbstractGangTask("Root Region Scan"), _cm(cm) { } 1186 1187 void work(uint worker_id) { 1188 assert(Thread::current()->is_ConcurrentGC_thread(), 1189 "this should only be done by a conc GC thread"); 1190 1191 CMRootRegions* root_regions = _cm->root_regions(); 1192 HeapRegion* hr = root_regions->claim_next(); 1193 while (hr != NULL) { 1194 _cm->scanRootRegion(hr, worker_id); 1195 hr = root_regions->claim_next(); 1196 } 1197 } 1198 }; 1199 1200 void ConcurrentMark::scanRootRegions() { 1201 // scan_in_progress() will have been set to true only if there was 1202 // at least one root region to scan. So, if it's false, we 1203 // should not attempt to do any further work. 1204 if (root_regions()->scan_in_progress()) { 1205 _parallel_marking_threads = calc_parallel_marking_threads(); 1206 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1207 "Maximum number of marking threads exceeded"); 1208 uint active_workers = MAX2(1U, parallel_marking_threads()); 1209 1210 CMRootRegionScanTask task(this); 1211 if (use_parallel_marking_threads()) { 1212 _parallel_workers->set_active_workers((int) active_workers); 1213 _parallel_workers->run_task(&task); 1214 } else { 1215 task.work(0); 1216 } 1217 1218 // It's possible that has_aborted() is true here without actually 1219 // aborting the survivor scan earlier. This is OK as it's 1220 // mainly used for sanity checking. 1221 root_regions()->scan_finished(); 1222 } 1223 } 1224 1225 void ConcurrentMark::markFromRoots() { 1226 // we might be tempted to assert that: 1227 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1228 // "inconsistent argument?"); 1229 // However that wouldn't be right, because it's possible that 1230 // a safepoint is indeed in progress as a younger generation 1231 // stop-the-world GC happens even as we mark in this generation. 1232 1233 _restart_for_overflow = false; 1234 force_overflow_conc()->init(); 1235 1236 // _g1h has _n_par_threads 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 1241 uint active_workers = MAX2(1U, parallel_marking_threads()); 1242 1243 // Parallel task terminator is set in "set_concurrency_and_phase()" 1244 set_concurrency_and_phase(active_workers, true /* concurrent */); 1245 1246 CMConcurrentMarkingTask markingTask(this, cmThread()); 1247 if (use_parallel_marking_threads()) { 1248 _parallel_workers->set_active_workers((int)active_workers); 1249 // Don't set _n_par_threads because it affects MT in proceess_strong_roots() 1250 // and the decisions on that MT processing is made elsewhere. 1251 assert(_parallel_workers->active_workers() > 0, "Should have been set"); 1252 _parallel_workers->run_task(&markingTask); 1253 } else { 1254 markingTask.work(0); 1255 } 1256 print_stats(); 1257 } 1258 1259 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1260 // world is stopped at this checkpoint 1261 assert(SafepointSynchronize::is_at_safepoint(), 1262 "world should be stopped"); 1263 1264 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1265 1266 // If a full collection has happened, we shouldn't do this. 1267 if (has_aborted()) { 1268 g1h->set_marking_complete(); // So bitmap clearing isn't confused 1269 return; 1270 } 1271 1272 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1273 1274 if (VerifyDuringGC) { 1275 HandleMark hm; // handle scope 1276 Universe::heap()->prepare_for_verify(); 1277 Universe::verify(VerifyOption_G1UsePrevMarking, 1278 " VerifyDuringGC:(before)"); 1279 } 1280 1281 G1CollectorPolicy* g1p = g1h->g1_policy(); 1282 g1p->record_concurrent_mark_remark_start(); 1283 1284 double start = os::elapsedTime(); 1285 1286 checkpointRootsFinalWork(); 1287 1288 double mark_work_end = os::elapsedTime(); 1289 1290 weakRefsWork(clear_all_soft_refs); 1291 1292 if (has_overflown()) { 1293 // Oops. We overflowed. Restart concurrent marking. 1294 _restart_for_overflow = true; 1295 if (G1TraceMarkStackOverflow) { 1296 gclog_or_tty->print_cr("\nRemark led to restart for overflow."); 1297 } 1298 1299 // Verify the heap w.r.t. the previous marking bitmap. 1300 if (VerifyDuringGC) { 1301 HandleMark hm; // handle scope 1302 Universe::heap()->prepare_for_verify(); 1303 Universe::verify(VerifyOption_G1UsePrevMarking, 1304 " VerifyDuringGC:(overflow)"); 1305 } 1306 1307 // Clear the marking state because we will be restarting 1308 // marking due to overflowing the global mark stack. 1309 reset_marking_state(); 1310 } else { 1311 // Aggregate the per-task counting data that we have accumulated 1312 // while marking. 1313 aggregate_count_data(); 1314 1315 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1316 // We're done with marking. 1317 // This is the end of the marking cycle, we're expected all 1318 // threads to have SATB queues with active set to true. 1319 satb_mq_set.set_active_all_threads(false, /* new active value */ 1320 true /* expected_active */); 1321 1322 if (VerifyDuringGC) { 1323 HandleMark hm; // handle scope 1324 Universe::heap()->prepare_for_verify(); 1325 Universe::verify(VerifyOption_G1UseNextMarking, 1326 " VerifyDuringGC:(after)"); 1327 } 1328 assert(!restart_for_overflow(), "sanity"); 1329 // Completely reset the marking state since marking completed 1330 set_non_marking_state(); 1331 } 1332 1333 // Expand the marking stack, if we have to and if we can. 1334 if (_markStack.should_expand()) { 1335 _markStack.expand(); 1336 } 1337 1338 // Statistics 1339 double now = os::elapsedTime(); 1340 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1341 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1342 _remark_times.add((now - start) * 1000.0); 1343 1344 g1p->record_concurrent_mark_remark_end(); 1345 } 1346 1347 // Base class of the closures that finalize and verify the 1348 // liveness counting data. 1349 class CMCountDataClosureBase: public HeapRegionClosure { 1350 protected: 1351 G1CollectedHeap* _g1h; 1352 ConcurrentMark* _cm; 1353 CardTableModRefBS* _ct_bs; 1354 1355 BitMap* _region_bm; 1356 BitMap* _card_bm; 1357 1358 // Takes a region that's not empty (i.e., it has at least one 1359 // live object in it and sets its corresponding bit on the region 1360 // bitmap to 1. If the region is "starts humongous" it will also set 1361 // to 1 the bits on the region bitmap that correspond to its 1362 // associated "continues humongous" regions. 1363 void set_bit_for_region(HeapRegion* hr) { 1364 assert(!hr->continuesHumongous(), "should have filtered those out"); 1365 1366 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index(); 1367 if (!hr->startsHumongous()) { 1368 // Normal (non-humongous) case: just set the bit. 1369 _region_bm->par_at_put(index, true); 1370 } else { 1371 // Starts humongous case: calculate how many regions are part of 1372 // this humongous region and then set the bit range. 1373 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index(); 1374 _region_bm->par_at_put_range(index, end_index, true); 1375 } 1376 } 1377 1378 public: 1379 CMCountDataClosureBase(G1CollectedHeap* g1h, 1380 BitMap* region_bm, BitMap* card_bm): 1381 _g1h(g1h), _cm(g1h->concurrent_mark()), 1382 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 1383 _region_bm(region_bm), _card_bm(card_bm) { } 1384 }; 1385 1386 // Closure that calculates the # live objects per region. Used 1387 // for verification purposes during the cleanup pause. 1388 class CalcLiveObjectsClosure: public CMCountDataClosureBase { 1389 CMBitMapRO* _bm; 1390 size_t _region_marked_bytes; 1391 1392 public: 1393 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h, 1394 BitMap* region_bm, BitMap* card_bm) : 1395 CMCountDataClosureBase(g1h, region_bm, card_bm), 1396 _bm(bm), _region_marked_bytes(0) { } 1397 1398 bool doHeapRegion(HeapRegion* hr) { 1399 1400 if (hr->continuesHumongous()) { 1401 // We will ignore these here and process them when their 1402 // associated "starts humongous" region is processed (see 1403 // set_bit_for_heap_region()). Note that we cannot rely on their 1404 // associated "starts humongous" region to have their bit set to 1405 // 1 since, due to the region chunking in the parallel region 1406 // iteration, a "continues humongous" region might be visited 1407 // before its associated "starts humongous". 1408 return false; 1409 } 1410 1411 HeapWord* ntams = hr->next_top_at_mark_start(); 1412 HeapWord* start = hr->bottom(); 1413 1414 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(), 1415 err_msg("Preconditions not met - " 1416 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT, 1417 start, ntams, hr->end())); 1418 1419 // Find the first marked object at or after "start". 1420 start = _bm->getNextMarkedWordAddress(start, ntams); 1421 1422 size_t marked_bytes = 0; 1423 1424 while (start < ntams) { 1425 oop obj = oop(start); 1426 int obj_sz = obj->size(); 1427 HeapWord* obj_end = start + obj_sz; 1428 1429 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 1430 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end); 1431 1432 // Note: if we're looking at the last region in heap - obj_end 1433 // could be actually just beyond the end of the heap; end_idx 1434 // will then correspond to a (non-existent) card that is also 1435 // just beyond the heap. 1436 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) { 1437 // end of object is not card aligned - increment to cover 1438 // all the cards spanned by the object 1439 end_idx += 1; 1440 } 1441 1442 // Set the bits in the card BM for the cards spanned by this object. 1443 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1444 1445 // Add the size of this object to the number of marked bytes. 1446 marked_bytes += (size_t)obj_sz * HeapWordSize; 1447 1448 // Find the next marked object after this one. 1449 start = _bm->getNextMarkedWordAddress(obj_end, ntams); 1450 } 1451 1452 // Mark the allocated-since-marking portion... 1453 HeapWord* top = hr->top(); 1454 if (ntams < top) { 1455 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1456 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1457 1458 // Note: if we're looking at the last region in heap - top 1459 // could be actually just beyond the end of the heap; end_idx 1460 // will then correspond to a (non-existent) card that is also 1461 // just beyond the heap. 1462 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1463 // end of object is not card aligned - increment to cover 1464 // all the cards spanned by the object 1465 end_idx += 1; 1466 } 1467 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1468 1469 // This definitely means the region has live objects. 1470 set_bit_for_region(hr); 1471 } 1472 1473 // Update the live region bitmap. 1474 if (marked_bytes > 0) { 1475 set_bit_for_region(hr); 1476 } 1477 1478 // Set the marked bytes for the current region so that 1479 // it can be queried by a calling verificiation routine 1480 _region_marked_bytes = marked_bytes; 1481 1482 return false; 1483 } 1484 1485 size_t region_marked_bytes() const { return _region_marked_bytes; } 1486 }; 1487 1488 // Heap region closure used for verifying the counting data 1489 // that was accumulated concurrently and aggregated during 1490 // the remark pause. This closure is applied to the heap 1491 // regions during the STW cleanup pause. 1492 1493 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure { 1494 G1CollectedHeap* _g1h; 1495 ConcurrentMark* _cm; 1496 CalcLiveObjectsClosure _calc_cl; 1497 BitMap* _region_bm; // Region BM to be verified 1498 BitMap* _card_bm; // Card BM to be verified 1499 bool _verbose; // verbose output? 1500 1501 BitMap* _exp_region_bm; // Expected Region BM values 1502 BitMap* _exp_card_bm; // Expected card BM values 1503 1504 int _failures; 1505 1506 public: 1507 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h, 1508 BitMap* region_bm, 1509 BitMap* card_bm, 1510 BitMap* exp_region_bm, 1511 BitMap* exp_card_bm, 1512 bool verbose) : 1513 _g1h(g1h), _cm(g1h->concurrent_mark()), 1514 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm), 1515 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose), 1516 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm), 1517 _failures(0) { } 1518 1519 int failures() const { return _failures; } 1520 1521 bool doHeapRegion(HeapRegion* hr) { 1522 if (hr->continuesHumongous()) { 1523 // We will ignore these here and process them when their 1524 // associated "starts humongous" region is processed (see 1525 // set_bit_for_heap_region()). Note that we cannot rely on their 1526 // associated "starts humongous" region to have their bit set to 1527 // 1 since, due to the region chunking in the parallel region 1528 // iteration, a "continues humongous" region might be visited 1529 // before its associated "starts humongous". 1530 return false; 1531 } 1532 1533 int failures = 0; 1534 1535 // Call the CalcLiveObjectsClosure to walk the marking bitmap for 1536 // this region and set the corresponding bits in the expected region 1537 // and card bitmaps. 1538 bool res = _calc_cl.doHeapRegion(hr); 1539 assert(res == false, "should be continuing"); 1540 1541 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL), 1542 Mutex::_no_safepoint_check_flag); 1543 1544 // Verify the marked bytes for this region. 1545 size_t exp_marked_bytes = _calc_cl.region_marked_bytes(); 1546 size_t act_marked_bytes = hr->next_marked_bytes(); 1547 1548 // We're not OK if expected marked bytes > actual marked bytes. It means 1549 // we have missed accounting some objects during the actual marking. 1550 if (exp_marked_bytes > act_marked_bytes) { 1551 if (_verbose) { 1552 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: " 1553 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT, 1554 hr->hrs_index(), exp_marked_bytes, act_marked_bytes); 1555 } 1556 failures += 1; 1557 } 1558 1559 // Verify the bit, for this region, in the actual and expected 1560 // (which was just calculated) region bit maps. 1561 // We're not OK if the bit in the calculated expected region 1562 // bitmap is set and the bit in the actual region bitmap is not. 1563 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index(); 1564 1565 bool expected = _exp_region_bm->at(index); 1566 bool actual = _region_bm->at(index); 1567 if (expected && !actual) { 1568 if (_verbose) { 1569 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: " 1570 "expected: %s, actual: %s", 1571 hr->hrs_index(), 1572 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1573 } 1574 failures += 1; 1575 } 1576 1577 // Verify that the card bit maps for the cards spanned by the current 1578 // region match. We have an error if we have a set bit in the expected 1579 // bit map and the corresponding bit in the actual bitmap is not set. 1580 1581 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom()); 1582 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top()); 1583 1584 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) { 1585 expected = _exp_card_bm->at(i); 1586 actual = _card_bm->at(i); 1587 1588 if (expected && !actual) { 1589 if (_verbose) { 1590 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": " 1591 "expected: %s, actual: %s", 1592 hr->hrs_index(), i, 1593 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1594 } 1595 failures += 1; 1596 } 1597 } 1598 1599 if (failures > 0 && _verbose) { 1600 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", " 1601 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT, 1602 HR_FORMAT_PARAMS(hr), hr->next_top_at_mark_start(), 1603 _calc_cl.region_marked_bytes(), hr->next_marked_bytes()); 1604 } 1605 1606 _failures += failures; 1607 1608 // We could stop iteration over the heap when we 1609 // find the first violating region by returning true. 1610 return false; 1611 } 1612 }; 1613 1614 1615 class G1ParVerifyFinalCountTask: public AbstractGangTask { 1616 protected: 1617 G1CollectedHeap* _g1h; 1618 ConcurrentMark* _cm; 1619 BitMap* _actual_region_bm; 1620 BitMap* _actual_card_bm; 1621 1622 uint _n_workers; 1623 1624 BitMap* _expected_region_bm; 1625 BitMap* _expected_card_bm; 1626 1627 int _failures; 1628 bool _verbose; 1629 1630 public: 1631 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h, 1632 BitMap* region_bm, BitMap* card_bm, 1633 BitMap* expected_region_bm, BitMap* expected_card_bm) 1634 : AbstractGangTask("G1 verify final counting"), 1635 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1636 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1637 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm), 1638 _failures(0), _verbose(false), 1639 _n_workers(0) { 1640 assert(VerifyDuringGC, "don't call this otherwise"); 1641 1642 // Use the value already set as the number of active threads 1643 // in the call to run_task(). 1644 if (G1CollectedHeap::use_parallel_gc_threads()) { 1645 assert( _g1h->workers()->active_workers() > 0, 1646 "Should have been previously set"); 1647 _n_workers = _g1h->workers()->active_workers(); 1648 } else { 1649 _n_workers = 1; 1650 } 1651 1652 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity"); 1653 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity"); 1654 1655 _verbose = _cm->verbose_medium(); 1656 } 1657 1658 void work(uint worker_id) { 1659 assert(worker_id < _n_workers, "invariant"); 1660 1661 VerifyLiveObjectDataHRClosure verify_cl(_g1h, 1662 _actual_region_bm, _actual_card_bm, 1663 _expected_region_bm, 1664 _expected_card_bm, 1665 _verbose); 1666 1667 if (G1CollectedHeap::use_parallel_gc_threads()) { 1668 _g1h->heap_region_par_iterate_chunked(&verify_cl, 1669 worker_id, 1670 _n_workers, 1671 HeapRegion::VerifyCountClaimValue); 1672 } else { 1673 _g1h->heap_region_iterate(&verify_cl); 1674 } 1675 1676 Atomic::add(verify_cl.failures(), &_failures); 1677 } 1678 1679 int failures() const { return _failures; } 1680 }; 1681 1682 // Closure that finalizes the liveness counting data. 1683 // Used during the cleanup pause. 1684 // Sets the bits corresponding to the interval [NTAMS, top] 1685 // (which contains the implicitly live objects) in the 1686 // card liveness bitmap. Also sets the bit for each region, 1687 // containing live data, in the region liveness bitmap. 1688 1689 class FinalCountDataUpdateClosure: public CMCountDataClosureBase { 1690 public: 1691 FinalCountDataUpdateClosure(G1CollectedHeap* g1h, 1692 BitMap* region_bm, 1693 BitMap* card_bm) : 1694 CMCountDataClosureBase(g1h, region_bm, card_bm) { } 1695 1696 bool doHeapRegion(HeapRegion* hr) { 1697 1698 if (hr->continuesHumongous()) { 1699 // We will ignore these here and process them when their 1700 // associated "starts humongous" region is processed (see 1701 // set_bit_for_heap_region()). Note that we cannot rely on their 1702 // associated "starts humongous" region to have their bit set to 1703 // 1 since, due to the region chunking in the parallel region 1704 // iteration, a "continues humongous" region might be visited 1705 // before its associated "starts humongous". 1706 return false; 1707 } 1708 1709 HeapWord* ntams = hr->next_top_at_mark_start(); 1710 HeapWord* top = hr->top(); 1711 1712 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions."); 1713 1714 // Mark the allocated-since-marking portion... 1715 if (ntams < top) { 1716 // This definitely means the region has live objects. 1717 set_bit_for_region(hr); 1718 1719 // Now set the bits in the card bitmap for [ntams, top) 1720 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1721 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1722 1723 // Note: if we're looking at the last region in heap - top 1724 // could be actually just beyond the end of the heap; end_idx 1725 // will then correspond to a (non-existent) card that is also 1726 // just beyond the heap. 1727 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1728 // end of object is not card aligned - increment to cover 1729 // all the cards spanned by the object 1730 end_idx += 1; 1731 } 1732 1733 assert(end_idx <= _card_bm->size(), 1734 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1735 end_idx, _card_bm->size())); 1736 assert(start_idx < _card_bm->size(), 1737 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1738 start_idx, _card_bm->size())); 1739 1740 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1741 } 1742 1743 // Set the bit for the region if it contains live data 1744 if (hr->next_marked_bytes() > 0) { 1745 set_bit_for_region(hr); 1746 } 1747 1748 return false; 1749 } 1750 }; 1751 1752 class G1ParFinalCountTask: public AbstractGangTask { 1753 protected: 1754 G1CollectedHeap* _g1h; 1755 ConcurrentMark* _cm; 1756 BitMap* _actual_region_bm; 1757 BitMap* _actual_card_bm; 1758 1759 uint _n_workers; 1760 1761 public: 1762 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm) 1763 : AbstractGangTask("G1 final counting"), 1764 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1765 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1766 _n_workers(0) { 1767 // Use the value already set as the number of active threads 1768 // in the call to run_task(). 1769 if (G1CollectedHeap::use_parallel_gc_threads()) { 1770 assert( _g1h->workers()->active_workers() > 0, 1771 "Should have been previously set"); 1772 _n_workers = _g1h->workers()->active_workers(); 1773 } else { 1774 _n_workers = 1; 1775 } 1776 } 1777 1778 void work(uint worker_id) { 1779 assert(worker_id < _n_workers, "invariant"); 1780 1781 FinalCountDataUpdateClosure final_update_cl(_g1h, 1782 _actual_region_bm, 1783 _actual_card_bm); 1784 1785 if (G1CollectedHeap::use_parallel_gc_threads()) { 1786 _g1h->heap_region_par_iterate_chunked(&final_update_cl, 1787 worker_id, 1788 _n_workers, 1789 HeapRegion::FinalCountClaimValue); 1790 } else { 1791 _g1h->heap_region_iterate(&final_update_cl); 1792 } 1793 } 1794 }; 1795 1796 class G1ParNoteEndTask; 1797 1798 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1799 G1CollectedHeap* _g1; 1800 int _worker_num; 1801 size_t _max_live_bytes; 1802 uint _regions_claimed; 1803 size_t _freed_bytes; 1804 FreeRegionList* _local_cleanup_list; 1805 OldRegionSet* _old_proxy_set; 1806 HumongousRegionSet* _humongous_proxy_set; 1807 HRRSCleanupTask* _hrrs_cleanup_task; 1808 double _claimed_region_time; 1809 double _max_region_time; 1810 1811 public: 1812 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1813 int worker_num, 1814 FreeRegionList* local_cleanup_list, 1815 OldRegionSet* old_proxy_set, 1816 HumongousRegionSet* humongous_proxy_set, 1817 HRRSCleanupTask* hrrs_cleanup_task) : 1818 _g1(g1), _worker_num(worker_num), 1819 _max_live_bytes(0), _regions_claimed(0), 1820 _freed_bytes(0), 1821 _claimed_region_time(0.0), _max_region_time(0.0), 1822 _local_cleanup_list(local_cleanup_list), 1823 _old_proxy_set(old_proxy_set), 1824 _humongous_proxy_set(humongous_proxy_set), 1825 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1826 1827 size_t freed_bytes() { return _freed_bytes; } 1828 1829 bool doHeapRegion(HeapRegion *hr) { 1830 if (hr->continuesHumongous()) { 1831 return false; 1832 } 1833 // We use a claim value of zero here because all regions 1834 // were claimed with value 1 in the FinalCount task. 1835 _g1->reset_gc_time_stamps(hr); 1836 double start = os::elapsedTime(); 1837 _regions_claimed++; 1838 hr->note_end_of_marking(); 1839 _max_live_bytes += hr->max_live_bytes(); 1840 _g1->free_region_if_empty(hr, 1841 &_freed_bytes, 1842 _local_cleanup_list, 1843 _old_proxy_set, 1844 _humongous_proxy_set, 1845 _hrrs_cleanup_task, 1846 true /* par */); 1847 double region_time = (os::elapsedTime() - start); 1848 _claimed_region_time += region_time; 1849 if (region_time > _max_region_time) { 1850 _max_region_time = region_time; 1851 } 1852 return false; 1853 } 1854 1855 size_t max_live_bytes() { return _max_live_bytes; } 1856 uint regions_claimed() { return _regions_claimed; } 1857 double claimed_region_time_sec() { return _claimed_region_time; } 1858 double max_region_time_sec() { return _max_region_time; } 1859 }; 1860 1861 class G1ParNoteEndTask: public AbstractGangTask { 1862 friend class G1NoteEndOfConcMarkClosure; 1863 1864 protected: 1865 G1CollectedHeap* _g1h; 1866 size_t _max_live_bytes; 1867 size_t _freed_bytes; 1868 FreeRegionList* _cleanup_list; 1869 1870 public: 1871 G1ParNoteEndTask(G1CollectedHeap* g1h, 1872 FreeRegionList* cleanup_list) : 1873 AbstractGangTask("G1 note end"), _g1h(g1h), 1874 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { } 1875 1876 void work(uint worker_id) { 1877 double start = os::elapsedTime(); 1878 FreeRegionList local_cleanup_list("Local Cleanup List"); 1879 OldRegionSet old_proxy_set("Local Cleanup Old Proxy Set"); 1880 HumongousRegionSet humongous_proxy_set("Local Cleanup Humongous Proxy Set"); 1881 HRRSCleanupTask hrrs_cleanup_task; 1882 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, worker_id, &local_cleanup_list, 1883 &old_proxy_set, 1884 &humongous_proxy_set, 1885 &hrrs_cleanup_task); 1886 if (G1CollectedHeap::use_parallel_gc_threads()) { 1887 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id, 1888 _g1h->workers()->active_workers(), 1889 HeapRegion::NoteEndClaimValue); 1890 } else { 1891 _g1h->heap_region_iterate(&g1_note_end); 1892 } 1893 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1894 1895 // Now update the lists 1896 _g1h->update_sets_after_freeing_regions(g1_note_end.freed_bytes(), 1897 NULL /* free_list */, 1898 &old_proxy_set, 1899 &humongous_proxy_set, 1900 true /* par */); 1901 { 1902 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1903 _max_live_bytes += g1_note_end.max_live_bytes(); 1904 _freed_bytes += g1_note_end.freed_bytes(); 1905 1906 // If we iterate over the global cleanup list at the end of 1907 // cleanup to do this printing we will not guarantee to only 1908 // generate output for the newly-reclaimed regions (the list 1909 // might not be empty at the beginning of cleanup; we might 1910 // still be working on its previous contents). So we do the 1911 // printing here, before we append the new regions to the global 1912 // cleanup list. 1913 1914 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1915 if (hr_printer->is_active()) { 1916 HeapRegionLinkedListIterator iter(&local_cleanup_list); 1917 while (iter.more_available()) { 1918 HeapRegion* hr = iter.get_next(); 1919 hr_printer->cleanup(hr); 1920 } 1921 } 1922 1923 _cleanup_list->add_as_tail(&local_cleanup_list); 1924 assert(local_cleanup_list.is_empty(), "post-condition"); 1925 1926 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1927 } 1928 } 1929 size_t max_live_bytes() { return _max_live_bytes; } 1930 size_t freed_bytes() { return _freed_bytes; } 1931 }; 1932 1933 class G1ParScrubRemSetTask: public AbstractGangTask { 1934 protected: 1935 G1RemSet* _g1rs; 1936 BitMap* _region_bm; 1937 BitMap* _card_bm; 1938 public: 1939 G1ParScrubRemSetTask(G1CollectedHeap* g1h, 1940 BitMap* region_bm, BitMap* card_bm) : 1941 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), 1942 _region_bm(region_bm), _card_bm(card_bm) { } 1943 1944 void work(uint worker_id) { 1945 if (G1CollectedHeap::use_parallel_gc_threads()) { 1946 _g1rs->scrub_par(_region_bm, _card_bm, worker_id, 1947 HeapRegion::ScrubRemSetClaimValue); 1948 } else { 1949 _g1rs->scrub(_region_bm, _card_bm); 1950 } 1951 } 1952 1953 }; 1954 1955 void ConcurrentMark::cleanup() { 1956 // world is stopped at this checkpoint 1957 assert(SafepointSynchronize::is_at_safepoint(), 1958 "world should be stopped"); 1959 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1960 1961 // If a full collection has happened, we shouldn't do this. 1962 if (has_aborted()) { 1963 g1h->set_marking_complete(); // So bitmap clearing isn't confused 1964 return; 1965 } 1966 1967 HRSPhaseSetter x(HRSPhaseCleanup); 1968 g1h->verify_region_sets_optional(); 1969 1970 if (VerifyDuringGC) { 1971 HandleMark hm; // handle scope 1972 Universe::heap()->prepare_for_verify(); 1973 Universe::verify(VerifyOption_G1UsePrevMarking, 1974 " VerifyDuringGC:(before)"); 1975 } 1976 1977 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); 1978 g1p->record_concurrent_mark_cleanup_start(); 1979 1980 double start = os::elapsedTime(); 1981 1982 HeapRegionRemSet::reset_for_cleanup_tasks(); 1983 1984 uint n_workers; 1985 1986 // Do counting once more with the world stopped for good measure. 1987 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm); 1988 1989 if (G1CollectedHeap::use_parallel_gc_threads()) { 1990 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 1991 "sanity check"); 1992 1993 g1h->set_par_threads(); 1994 n_workers = g1h->n_par_threads(); 1995 assert(g1h->n_par_threads() == n_workers, 1996 "Should not have been reset"); 1997 g1h->workers()->run_task(&g1_par_count_task); 1998 // Done with the parallel phase so reset to 0. 1999 g1h->set_par_threads(0); 2000 2001 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue), 2002 "sanity check"); 2003 } else { 2004 n_workers = 1; 2005 g1_par_count_task.work(0); 2006 } 2007 2008 if (VerifyDuringGC) { 2009 // Verify that the counting data accumulated during marking matches 2010 // that calculated by walking the marking bitmap. 2011 2012 // Bitmaps to hold expected values 2013 BitMap expected_region_bm(_region_bm.size(), false); 2014 BitMap expected_card_bm(_card_bm.size(), false); 2015 2016 G1ParVerifyFinalCountTask g1_par_verify_task(g1h, 2017 &_region_bm, 2018 &_card_bm, 2019 &expected_region_bm, 2020 &expected_card_bm); 2021 2022 if (G1CollectedHeap::use_parallel_gc_threads()) { 2023 g1h->set_par_threads((int)n_workers); 2024 g1h->workers()->run_task(&g1_par_verify_task); 2025 // Done with the parallel phase so reset to 0. 2026 g1h->set_par_threads(0); 2027 2028 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue), 2029 "sanity check"); 2030 } else { 2031 g1_par_verify_task.work(0); 2032 } 2033 2034 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures"); 2035 } 2036 2037 size_t start_used_bytes = g1h->used(); 2038 g1h->set_marking_complete(); 2039 2040 double count_end = os::elapsedTime(); 2041 double this_final_counting_time = (count_end - start); 2042 _total_counting_time += this_final_counting_time; 2043 2044 if (G1PrintRegionLivenessInfo) { 2045 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking"); 2046 _g1h->heap_region_iterate(&cl); 2047 } 2048 2049 // Install newly created mark bitMap as "prev". 2050 swapMarkBitMaps(); 2051 2052 g1h->reset_gc_time_stamp(); 2053 2054 // Note end of marking in all heap regions. 2055 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list); 2056 if (G1CollectedHeap::use_parallel_gc_threads()) { 2057 g1h->set_par_threads((int)n_workers); 2058 g1h->workers()->run_task(&g1_par_note_end_task); 2059 g1h->set_par_threads(0); 2060 2061 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue), 2062 "sanity check"); 2063 } else { 2064 g1_par_note_end_task.work(0); 2065 } 2066 g1h->check_gc_time_stamps(); 2067 2068 if (!cleanup_list_is_empty()) { 2069 // The cleanup list is not empty, so we'll have to process it 2070 // concurrently. Notify anyone else that might be wanting free 2071 // regions that there will be more free regions coming soon. 2072 g1h->set_free_regions_coming(); 2073 } 2074 2075 // call below, since it affects the metric by which we sort the heap 2076 // regions. 2077 if (G1ScrubRemSets) { 2078 double rs_scrub_start = os::elapsedTime(); 2079 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm); 2080 if (G1CollectedHeap::use_parallel_gc_threads()) { 2081 g1h->set_par_threads((int)n_workers); 2082 g1h->workers()->run_task(&g1_par_scrub_rs_task); 2083 g1h->set_par_threads(0); 2084 2085 assert(g1h->check_heap_region_claim_values( 2086 HeapRegion::ScrubRemSetClaimValue), 2087 "sanity check"); 2088 } else { 2089 g1_par_scrub_rs_task.work(0); 2090 } 2091 2092 double rs_scrub_end = os::elapsedTime(); 2093 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); 2094 _total_rs_scrub_time += this_rs_scrub_time; 2095 } 2096 2097 // this will also free any regions totally full of garbage objects, 2098 // and sort the regions. 2099 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers); 2100 2101 // Statistics. 2102 double end = os::elapsedTime(); 2103 _cleanup_times.add((end - start) * 1000.0); 2104 2105 if (G1Log::fine()) { 2106 g1h->print_size_transition(gclog_or_tty, 2107 start_used_bytes, 2108 g1h->used(), 2109 g1h->capacity()); 2110 } 2111 2112 // Clean up will have freed any regions completely full of garbage. 2113 // Update the soft reference policy with the new heap occupancy. 2114 Universe::update_heap_info_at_gc(); 2115 2116 // We need to make this be a "collection" so any collection pause that 2117 // races with it goes around and waits for completeCleanup to finish. 2118 g1h->increment_total_collections(); 2119 2120 // We reclaimed old regions so we should calculate the sizes to make 2121 // sure we update the old gen/space data. 2122 g1h->g1mm()->update_sizes(); 2123 2124 if (VerifyDuringGC) { 2125 HandleMark hm; // handle scope 2126 Universe::heap()->prepare_for_verify(); 2127 Universe::verify(VerifyOption_G1UsePrevMarking, 2128 " VerifyDuringGC:(after)"); 2129 } 2130 2131 g1h->verify_region_sets_optional(); 2132 } 2133 2134 void ConcurrentMark::completeCleanup() { 2135 if (has_aborted()) return; 2136 2137 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2138 2139 _cleanup_list.verify_optional(); 2140 FreeRegionList tmp_free_list("Tmp Free List"); 2141 2142 if (G1ConcRegionFreeingVerbose) { 2143 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2144 "cleanup list has %u entries", 2145 _cleanup_list.length()); 2146 } 2147 2148 // Noone else should be accessing the _cleanup_list at this point, 2149 // so it's not necessary to take any locks 2150 while (!_cleanup_list.is_empty()) { 2151 HeapRegion* hr = _cleanup_list.remove_head(); 2152 assert(hr != NULL, "the list was not empty"); 2153 hr->par_clear(); 2154 tmp_free_list.add_as_tail(hr); 2155 2156 // Instead of adding one region at a time to the secondary_free_list, 2157 // we accumulate them in the local list and move them a few at a 2158 // time. This also cuts down on the number of notify_all() calls 2159 // we do during this process. We'll also append the local list when 2160 // _cleanup_list is empty (which means we just removed the last 2161 // region from the _cleanup_list). 2162 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 2163 _cleanup_list.is_empty()) { 2164 if (G1ConcRegionFreeingVerbose) { 2165 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2166 "appending %u entries to the secondary_free_list, " 2167 "cleanup list still has %u entries", 2168 tmp_free_list.length(), 2169 _cleanup_list.length()); 2170 } 2171 2172 { 2173 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 2174 g1h->secondary_free_list_add_as_tail(&tmp_free_list); 2175 SecondaryFreeList_lock->notify_all(); 2176 } 2177 2178 if (G1StressConcRegionFreeing) { 2179 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 2180 os::sleep(Thread::current(), (jlong) 1, false); 2181 } 2182 } 2183 } 2184 } 2185 assert(tmp_free_list.is_empty(), "post-condition"); 2186 } 2187 2188 // Supporting Object and Oop closures for reference discovery 2189 // and processing in during marking 2190 2191 bool G1CMIsAliveClosure::do_object_b(oop obj) { 2192 HeapWord* addr = (HeapWord*)obj; 2193 return addr != NULL && 2194 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 2195 } 2196 2197 // 'Keep Alive' oop closure used by both serial parallel reference processing. 2198 // Uses the CMTask associated with a worker thread (for serial reference 2199 // processing the CMTask for worker 0 is used) to preserve (mark) and 2200 // trace referent objects. 2201 // 2202 // Using the CMTask and embedded local queues avoids having the worker 2203 // threads operating on the global mark stack. This reduces the risk 2204 // of overflowing the stack - which we would rather avoid at this late 2205 // state. Also using the tasks' local queues removes the potential 2206 // of the workers interfering with each other that could occur if 2207 // operating on the global stack. 2208 2209 class G1CMKeepAliveAndDrainClosure: public OopClosure { 2210 ConcurrentMark* _cm; 2211 CMTask* _task; 2212 int _ref_counter_limit; 2213 int _ref_counter; 2214 bool _is_serial; 2215 public: 2216 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) : 2217 _cm(cm), _task(task), _is_serial(is_serial), 2218 _ref_counter_limit(G1RefProcDrainInterval) { 2219 assert(_ref_counter_limit > 0, "sanity"); 2220 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 2221 _ref_counter = _ref_counter_limit; 2222 } 2223 2224 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 2225 virtual void do_oop( oop* p) { do_oop_work(p); } 2226 2227 template <class T> void do_oop_work(T* p) { 2228 if (!_cm->has_overflown()) { 2229 oop obj = oopDesc::load_decode_heap_oop(p); 2230 if (_cm->verbose_high()) { 2231 gclog_or_tty->print_cr("\t[%u] we're looking at location " 2232 "*"PTR_FORMAT" = "PTR_FORMAT, 2233 _task->worker_id(), p, (void*) obj); 2234 } 2235 2236 _task->deal_with_reference(obj); 2237 _ref_counter--; 2238 2239 if (_ref_counter == 0) { 2240 // We have dealt with _ref_counter_limit references, pushing them 2241 // and objects reachable from them on to the local stack (and 2242 // possibly the global stack). Call CMTask::do_marking_step() to 2243 // process these entries. 2244 // 2245 // We call CMTask::do_marking_step() in a loop, which we'll exit if 2246 // there's nothing more to do (i.e. we're done with the entries that 2247 // were pushed as a result of the CMTask::deal_with_reference() calls 2248 // above) or we overflow. 2249 // 2250 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2251 // flag while there may still be some work to do. (See the comment at 2252 // the beginning of CMTask::do_marking_step() for those conditions - 2253 // one of which is reaching the specified time target.) It is only 2254 // when CMTask::do_marking_step() returns without setting the 2255 // has_aborted() flag that the marking step has completed. 2256 do { 2257 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 2258 _task->do_marking_step(mark_step_duration_ms, 2259 false /* do_termination */, 2260 _is_serial); 2261 } while (_task->has_aborted() && !_cm->has_overflown()); 2262 _ref_counter = _ref_counter_limit; 2263 } 2264 } else { 2265 if (_cm->verbose_high()) { 2266 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id()); 2267 } 2268 } 2269 } 2270 }; 2271 2272 // 'Drain' oop closure used by both serial and parallel reference processing. 2273 // Uses the CMTask associated with a given worker thread (for serial 2274 // reference processing the CMtask for worker 0 is used). Calls the 2275 // do_marking_step routine, with an unbelievably large timeout value, 2276 // to drain the marking data structures of the remaining entries 2277 // added by the 'keep alive' oop closure above. 2278 2279 class G1CMDrainMarkingStackClosure: public VoidClosure { 2280 ConcurrentMark* _cm; 2281 CMTask* _task; 2282 bool _is_serial; 2283 public: 2284 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_serial) : 2285 _cm(cm), _task(task), _is_serial(is_serial) { 2286 assert(!_is_serial || _task->worker_id() == 0, "only task 0 for serial code"); 2287 } 2288 2289 void do_void() { 2290 do { 2291 if (_cm->verbose_high()) { 2292 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - serial: %s", 2293 _task->worker_id(), BOOL_TO_STR(_is_serial)); 2294 } 2295 2296 // We call CMTask::do_marking_step() to completely drain the local 2297 // and global marking stacks of entries pushed by the 'keep alive' 2298 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 2299 // 2300 // CMTask::do_marking_step() is called in a loop, which we'll exit 2301 // if there's nothing more to do (i.e. we'completely drained the 2302 // entries that were pushed as a a result of applying the 'keep alive' 2303 // closure to the entries on the discovered ref lists) or we overflow 2304 // the global marking stack. 2305 // 2306 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2307 // flag while there may still be some work to do. (See the comment at 2308 // the beginning of CMTask::do_marking_step() for those conditions - 2309 // one of which is reaching the specified time target.) It is only 2310 // when CMTask::do_marking_step() returns without setting the 2311 // has_aborted() flag that the marking step has completed. 2312 2313 _task->do_marking_step(1000000000.0 /* something very large */, 2314 true /* do_termination */, 2315 _is_serial); 2316 } while (_task->has_aborted() && !_cm->has_overflown()); 2317 } 2318 }; 2319 2320 // Implementation of AbstractRefProcTaskExecutor for parallel 2321 // reference processing at the end of G1 concurrent marking 2322 2323 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2324 private: 2325 G1CollectedHeap* _g1h; 2326 ConcurrentMark* _cm; 2327 WorkGang* _workers; 2328 int _active_workers; 2329 2330 public: 2331 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 2332 ConcurrentMark* cm, 2333 WorkGang* workers, 2334 int n_workers) : 2335 _g1h(g1h), _cm(cm), 2336 _workers(workers), _active_workers(n_workers) { } 2337 2338 // Executes the given task using concurrent marking worker threads. 2339 virtual void execute(ProcessTask& task); 2340 virtual void execute(EnqueueTask& task); 2341 }; 2342 2343 class G1CMRefProcTaskProxy: public AbstractGangTask { 2344 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2345 ProcessTask& _proc_task; 2346 G1CollectedHeap* _g1h; 2347 ConcurrentMark* _cm; 2348 2349 public: 2350 G1CMRefProcTaskProxy(ProcessTask& proc_task, 2351 G1CollectedHeap* g1h, 2352 ConcurrentMark* cm) : 2353 AbstractGangTask("Process reference objects in parallel"), 2354 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 2355 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 2356 assert(rp->processing_is_mt(), "shouldn't be here otherwise"); 2357 } 2358 2359 virtual void work(uint worker_id) { 2360 CMTask* task = _cm->task(worker_id); 2361 G1CMIsAliveClosure g1_is_alive(_g1h); 2362 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, task, false /* is_serial */); 2363 G1CMDrainMarkingStackClosure g1_par_drain(_cm, task, false /* is_serial */); 2364 2365 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 2366 } 2367 }; 2368 2369 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 2370 assert(_workers != NULL, "Need parallel worker threads."); 2371 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2372 2373 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 2374 2375 // We need to reset the concurrency level before each 2376 // proxy task execution, so that the termination protocol 2377 // and overflow handling in CMTask::do_marking_step() knows 2378 // how many workers to wait for. 2379 _cm->set_concurrency(_active_workers); 2380 _g1h->set_par_threads(_active_workers); 2381 _workers->run_task(&proc_task_proxy); 2382 _g1h->set_par_threads(0); 2383 } 2384 2385 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 2386 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 2387 EnqueueTask& _enq_task; 2388 2389 public: 2390 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 2391 AbstractGangTask("Enqueue reference objects in parallel"), 2392 _enq_task(enq_task) { } 2393 2394 virtual void work(uint worker_id) { 2395 _enq_task.work(worker_id); 2396 } 2397 }; 2398 2399 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 2400 assert(_workers != NULL, "Need parallel worker threads."); 2401 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2402 2403 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 2404 2405 // Not strictly necessary but... 2406 // 2407 // We need to reset the concurrency level before each 2408 // proxy task execution, so that the termination protocol 2409 // and overflow handling in CMTask::do_marking_step() knows 2410 // how many workers to wait for. 2411 _cm->set_concurrency(_active_workers); 2412 _g1h->set_par_threads(_active_workers); 2413 _workers->run_task(&enq_task_proxy); 2414 _g1h->set_par_threads(0); 2415 } 2416 2417 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 2418 if (has_overflown()) { 2419 // Skip processing the discovered references if we have 2420 // overflown the global marking stack. Reference objects 2421 // only get discovered once so it is OK to not 2422 // de-populate the discovered reference lists. We could have, 2423 // but the only benefit would be that, when marking restarts, 2424 // less reference objects are discovered. 2425 return; 2426 } 2427 2428 ResourceMark rm; 2429 HandleMark hm; 2430 2431 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2432 2433 // Is alive closure. 2434 G1CMIsAliveClosure g1_is_alive(g1h); 2435 2436 // Inner scope to exclude the cleaning of the string and symbol 2437 // tables from the displayed time. 2438 { 2439 if (G1Log::finer()) { 2440 gclog_or_tty->put(' '); 2441 } 2442 TraceTime t("GC ref-proc", G1Log::finer(), false, gclog_or_tty); 2443 2444 ReferenceProcessor* rp = g1h->ref_processor_cm(); 2445 2446 // See the comment in G1CollectedHeap::ref_processing_init() 2447 // about how reference processing currently works in G1. 2448 2449 // Set the soft reference policy 2450 rp->setup_policy(clear_all_soft_refs); 2451 assert(_markStack.isEmpty(), "mark stack should be empty"); 2452 2453 // Instances of the 'Keep Alive' and 'Complete GC' closures used 2454 // in serial reference processing. Note these closures are also 2455 // used for serially processing (by the the current thread) the 2456 // JNI references during parallel reference processing. 2457 // 2458 // These closures do not need to synchronize with the worker 2459 // threads involved in parallel reference processing as these 2460 // instances are executed serially by the current thread (e.g. 2461 // reference processing is not multi-threaded and is thus 2462 // performed by the current thread instead of a gang worker). 2463 // 2464 // The gang tasks involved in parallel reference procssing create 2465 // their own instances of these closures, which do their own 2466 // synchronization among themselves. 2467 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0), true /* is_serial */); 2468 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), true /* is_serial */); 2469 2470 // We need at least one active thread. If reference processing 2471 // is not multi-threaded we use the current (VMThread) thread, 2472 // otherwise we use the work gang from the G1CollectedHeap and 2473 // we utilize all the worker threads we can. 2474 bool processing_is_mt = rp->processing_is_mt() && g1h->workers() != NULL; 2475 uint active_workers = (processing_is_mt ? g1h->workers()->active_workers() : 1U); 2476 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U); 2477 2478 // Parallel processing task executor. 2479 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 2480 g1h->workers(), active_workers); 2481 AbstractRefProcTaskExecutor* executor = (processing_is_mt ? &par_task_executor : NULL); 2482 2483 // Set the concurrency level. The phase was already set prior to 2484 // executing the remark task. 2485 set_concurrency(active_workers); 2486 2487 // Set the degree of MT processing here. If the discovery was done MT, 2488 // the number of threads involved during discovery could differ from 2489 // the number of active workers. This is OK as long as the discovered 2490 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 2491 rp->set_active_mt_degree(active_workers); 2492 2493 // Process the weak references. 2494 rp->process_discovered_references(&g1_is_alive, 2495 &g1_keep_alive, 2496 &g1_drain_mark_stack, 2497 executor); 2498 2499 // The do_oop work routines of the keep_alive and drain_marking_stack 2500 // oop closures will set the has_overflown flag if we overflow the 2501 // global marking stack. 2502 2503 assert(_markStack.overflow() || _markStack.isEmpty(), 2504 "mark stack should be empty (unless it overflowed)"); 2505 2506 if (_markStack.overflow()) { 2507 // This should have been done already when we tried to push an 2508 // entry on to the global mark stack. But let's do it again. 2509 set_has_overflown(); 2510 } 2511 2512 assert(rp->num_q() == active_workers, "why not"); 2513 2514 rp->enqueue_discovered_references(executor); 2515 2516 rp->verify_no_references_recorded(); 2517 assert(!rp->discovery_enabled(), "Post condition"); 2518 } 2519 2520 // Now clean up stale oops in StringTable 2521 StringTable::unlink(&g1_is_alive); 2522 // Clean up unreferenced symbols in symbol table. 2523 SymbolTable::unlink(); 2524 } 2525 2526 void ConcurrentMark::swapMarkBitMaps() { 2527 CMBitMapRO* temp = _prevMarkBitMap; 2528 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; 2529 _nextMarkBitMap = (CMBitMap*) temp; 2530 } 2531 2532 class CMRemarkTask: public AbstractGangTask { 2533 private: 2534 ConcurrentMark* _cm; 2535 bool _is_serial; 2536 public: 2537 void work(uint worker_id) { 2538 // Since all available tasks are actually started, we should 2539 // only proceed if we're supposed to be actived. 2540 if (worker_id < _cm->active_tasks()) { 2541 CMTask* task = _cm->task(worker_id); 2542 task->record_start_time(); 2543 do { 2544 task->do_marking_step(1000000000.0 /* something very large */, 2545 true /* do_termination */, 2546 _is_serial); 2547 } while (task->has_aborted() && !_cm->has_overflown()); 2548 // If we overflow, then we do not want to restart. We instead 2549 // want to abort remark and do concurrent marking again. 2550 task->record_end_time(); 2551 } 2552 } 2553 2554 CMRemarkTask(ConcurrentMark* cm, int active_workers, bool is_serial) : 2555 AbstractGangTask("Par Remark"), _cm(cm), _is_serial(is_serial) { 2556 _cm->terminator()->reset_for_reuse(active_workers); 2557 } 2558 }; 2559 2560 void ConcurrentMark::checkpointRootsFinalWork() { 2561 ResourceMark rm; 2562 HandleMark hm; 2563 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2564 2565 g1h->ensure_parsability(false); 2566 2567 if (G1CollectedHeap::use_parallel_gc_threads()) { 2568 G1CollectedHeap::StrongRootsScope srs(g1h); 2569 // this is remark, so we'll use up all active threads 2570 uint active_workers = g1h->workers()->active_workers(); 2571 if (active_workers == 0) { 2572 assert(active_workers > 0, "Should have been set earlier"); 2573 active_workers = (uint) ParallelGCThreads; 2574 g1h->workers()->set_active_workers(active_workers); 2575 } 2576 set_concurrency_and_phase(active_workers, false /* concurrent */); 2577 // Leave _parallel_marking_threads at it's 2578 // value originally calculated in the ConcurrentMark 2579 // constructor and pass values of the active workers 2580 // through the gang in the task. 2581 2582 CMRemarkTask remarkTask(this, active_workers, false /* is_serial */); 2583 // We will start all available threads, even if we decide that the 2584 // active_workers will be fewer. The extra ones will just bail out 2585 // immediately. 2586 g1h->set_par_threads(active_workers); 2587 g1h->workers()->run_task(&remarkTask); 2588 g1h->set_par_threads(0); 2589 } else { 2590 G1CollectedHeap::StrongRootsScope srs(g1h); 2591 uint active_workers = 1; 2592 set_concurrency_and_phase(active_workers, false /* concurrent */); 2593 2594 // Note - if there's no work gang then the VMThread will be 2595 // the thread to execute the remark - serially. We have 2596 // to pass true for the is_serial parameter so that 2597 // CMTask::do_marking_step() doesn't enter the sync 2598 // barriers in the event of an overflow. Doing so will 2599 // cause an assert that the current thread is not a 2600 // concurrent GC thread. 2601 CMRemarkTask remarkTask(this, active_workers, true /* is_serial*/); 2602 remarkTask.work(0); 2603 } 2604 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2605 guarantee(has_overflown() || 2606 satb_mq_set.completed_buffers_num() == 0, 2607 err_msg("Invariant: has_overflown = %s, num buffers = %d", 2608 BOOL_TO_STR(has_overflown()), 2609 satb_mq_set.completed_buffers_num())); 2610 2611 print_stats(); 2612 } 2613 2614 #ifndef PRODUCT 2615 2616 class PrintReachableOopClosure: public OopClosure { 2617 private: 2618 G1CollectedHeap* _g1h; 2619 outputStream* _out; 2620 VerifyOption _vo; 2621 bool _all; 2622 2623 public: 2624 PrintReachableOopClosure(outputStream* out, 2625 VerifyOption vo, 2626 bool all) : 2627 _g1h(G1CollectedHeap::heap()), 2628 _out(out), _vo(vo), _all(all) { } 2629 2630 void do_oop(narrowOop* p) { do_oop_work(p); } 2631 void do_oop( oop* p) { do_oop_work(p); } 2632 2633 template <class T> void do_oop_work(T* p) { 2634 oop obj = oopDesc::load_decode_heap_oop(p); 2635 const char* str = NULL; 2636 const char* str2 = ""; 2637 2638 if (obj == NULL) { 2639 str = ""; 2640 } else if (!_g1h->is_in_g1_reserved(obj)) { 2641 str = " O"; 2642 } else { 2643 HeapRegion* hr = _g1h->heap_region_containing(obj); 2644 guarantee(hr != NULL, "invariant"); 2645 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo); 2646 bool marked = _g1h->is_marked(obj, _vo); 2647 2648 if (over_tams) { 2649 str = " >"; 2650 if (marked) { 2651 str2 = " AND MARKED"; 2652 } 2653 } else if (marked) { 2654 str = " M"; 2655 } else { 2656 str = " NOT"; 2657 } 2658 } 2659 2660 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s", 2661 p, (void*) obj, str, str2); 2662 } 2663 }; 2664 2665 class PrintReachableObjectClosure : public ObjectClosure { 2666 private: 2667 G1CollectedHeap* _g1h; 2668 outputStream* _out; 2669 VerifyOption _vo; 2670 bool _all; 2671 HeapRegion* _hr; 2672 2673 public: 2674 PrintReachableObjectClosure(outputStream* out, 2675 VerifyOption vo, 2676 bool all, 2677 HeapRegion* hr) : 2678 _g1h(G1CollectedHeap::heap()), 2679 _out(out), _vo(vo), _all(all), _hr(hr) { } 2680 2681 void do_object(oop o) { 2682 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo); 2683 bool marked = _g1h->is_marked(o, _vo); 2684 bool print_it = _all || over_tams || marked; 2685 2686 if (print_it) { 2687 _out->print_cr(" "PTR_FORMAT"%s", 2688 o, (over_tams) ? " >" : (marked) ? " M" : ""); 2689 PrintReachableOopClosure oopCl(_out, _vo, _all); 2690 o->oop_iterate_no_header(&oopCl); 2691 } 2692 } 2693 }; 2694 2695 class PrintReachableRegionClosure : public HeapRegionClosure { 2696 private: 2697 G1CollectedHeap* _g1h; 2698 outputStream* _out; 2699 VerifyOption _vo; 2700 bool _all; 2701 2702 public: 2703 bool doHeapRegion(HeapRegion* hr) { 2704 HeapWord* b = hr->bottom(); 2705 HeapWord* e = hr->end(); 2706 HeapWord* t = hr->top(); 2707 HeapWord* p = _g1h->top_at_mark_start(hr, _vo); 2708 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" " 2709 "TAMS: "PTR_FORMAT, b, e, t, p); 2710 _out->cr(); 2711 2712 HeapWord* from = b; 2713 HeapWord* to = t; 2714 2715 if (to > from) { 2716 _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to); 2717 _out->cr(); 2718 PrintReachableObjectClosure ocl(_out, _vo, _all, hr); 2719 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl); 2720 _out->cr(); 2721 } 2722 2723 return false; 2724 } 2725 2726 PrintReachableRegionClosure(outputStream* out, 2727 VerifyOption vo, 2728 bool all) : 2729 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { } 2730 }; 2731 2732 void ConcurrentMark::print_reachable(const char* str, 2733 VerifyOption vo, 2734 bool all) { 2735 gclog_or_tty->cr(); 2736 gclog_or_tty->print_cr("== Doing heap dump... "); 2737 2738 if (G1PrintReachableBaseFile == NULL) { 2739 gclog_or_tty->print_cr(" #### error: no base file defined"); 2740 return; 2741 } 2742 2743 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) > 2744 (JVM_MAXPATHLEN - 1)) { 2745 gclog_or_tty->print_cr(" #### error: file name too long"); 2746 return; 2747 } 2748 2749 char file_name[JVM_MAXPATHLEN]; 2750 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str); 2751 gclog_or_tty->print_cr(" dumping to file %s", file_name); 2752 2753 fileStream fout(file_name); 2754 if (!fout.is_open()) { 2755 gclog_or_tty->print_cr(" #### error: could not open file"); 2756 return; 2757 } 2758 2759 outputStream* out = &fout; 2760 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo)); 2761 out->cr(); 2762 2763 out->print_cr("--- ITERATING OVER REGIONS"); 2764 out->cr(); 2765 PrintReachableRegionClosure rcl(out, vo, all); 2766 _g1h->heap_region_iterate(&rcl); 2767 out->cr(); 2768 2769 gclog_or_tty->print_cr(" done"); 2770 gclog_or_tty->flush(); 2771 } 2772 2773 #endif // PRODUCT 2774 2775 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 2776 // Note we are overriding the read-only view of the prev map here, via 2777 // the cast. 2778 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); 2779 } 2780 2781 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) { 2782 _nextMarkBitMap->clearRange(mr); 2783 } 2784 2785 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) { 2786 clearRangePrevBitmap(mr); 2787 clearRangeNextBitmap(mr); 2788 } 2789 2790 HeapRegion* 2791 ConcurrentMark::claim_region(uint worker_id) { 2792 // "checkpoint" the finger 2793 HeapWord* finger = _finger; 2794 2795 // _heap_end will not change underneath our feet; it only changes at 2796 // yield points. 2797 while (finger < _heap_end) { 2798 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 2799 2800 // Note on how this code handles humongous regions. In the 2801 // normal case the finger will reach the start of a "starts 2802 // humongous" (SH) region. Its end will either be the end of the 2803 // last "continues humongous" (CH) region in the sequence, or the 2804 // standard end of the SH region (if the SH is the only region in 2805 // the sequence). That way claim_region() will skip over the CH 2806 // regions. However, there is a subtle race between a CM thread 2807 // executing this method and a mutator thread doing a humongous 2808 // object allocation. The two are not mutually exclusive as the CM 2809 // thread does not need to hold the Heap_lock when it gets 2810 // here. So there is a chance that claim_region() will come across 2811 // a free region that's in the progress of becoming a SH or a CH 2812 // region. In the former case, it will either 2813 // a) Miss the update to the region's end, in which case it will 2814 // visit every subsequent CH region, will find their bitmaps 2815 // empty, and do nothing, or 2816 // b) Will observe the update of the region's end (in which case 2817 // it will skip the subsequent CH regions). 2818 // If it comes across a region that suddenly becomes CH, the 2819 // scenario will be similar to b). So, the race between 2820 // claim_region() and a humongous object allocation might force us 2821 // to do a bit of unnecessary work (due to some unnecessary bitmap 2822 // iterations) but it should not introduce and correctness issues. 2823 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger); 2824 HeapWord* bottom = curr_region->bottom(); 2825 HeapWord* end = curr_region->end(); 2826 HeapWord* limit = curr_region->next_top_at_mark_start(); 2827 2828 if (verbose_low()) { 2829 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" " 2830 "["PTR_FORMAT", "PTR_FORMAT"), " 2831 "limit = "PTR_FORMAT, 2832 worker_id, curr_region, bottom, end, limit); 2833 } 2834 2835 // Is the gap between reading the finger and doing the CAS too long? 2836 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 2837 if (res == finger) { 2838 // we succeeded 2839 2840 // notice that _finger == end cannot be guaranteed here since, 2841 // someone else might have moved the finger even further 2842 assert(_finger >= end, "the finger should have moved forward"); 2843 2844 if (verbose_low()) { 2845 gclog_or_tty->print_cr("[%u] we were successful with region = " 2846 PTR_FORMAT, worker_id, curr_region); 2847 } 2848 2849 if (limit > bottom) { 2850 if (verbose_low()) { 2851 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, " 2852 "returning it ", worker_id, curr_region); 2853 } 2854 return curr_region; 2855 } else { 2856 assert(limit == bottom, 2857 "the region limit should be at bottom"); 2858 if (verbose_low()) { 2859 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, " 2860 "returning NULL", worker_id, curr_region); 2861 } 2862 // we return NULL and the caller should try calling 2863 // claim_region() again. 2864 return NULL; 2865 } 2866 } else { 2867 assert(_finger > finger, "the finger should have moved forward"); 2868 if (verbose_low()) { 2869 gclog_or_tty->print_cr("[%u] somebody else moved the finger, " 2870 "global finger = "PTR_FORMAT", " 2871 "our finger = "PTR_FORMAT, 2872 worker_id, _finger, finger); 2873 } 2874 2875 // read it again 2876 finger = _finger; 2877 } 2878 } 2879 2880 return NULL; 2881 } 2882 2883 #ifndef PRODUCT 2884 enum VerifyNoCSetOopsPhase { 2885 VerifyNoCSetOopsStack, 2886 VerifyNoCSetOopsQueues, 2887 VerifyNoCSetOopsSATBCompleted, 2888 VerifyNoCSetOopsSATBThread 2889 }; 2890 2891 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure { 2892 private: 2893 G1CollectedHeap* _g1h; 2894 VerifyNoCSetOopsPhase _phase; 2895 int _info; 2896 2897 const char* phase_str() { 2898 switch (_phase) { 2899 case VerifyNoCSetOopsStack: return "Stack"; 2900 case VerifyNoCSetOopsQueues: return "Queue"; 2901 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers"; 2902 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers"; 2903 default: ShouldNotReachHere(); 2904 } 2905 return NULL; 2906 } 2907 2908 void do_object_work(oop obj) { 2909 guarantee(!_g1h->obj_in_cs(obj), 2910 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d", 2911 (void*) obj, phase_str(), _info)); 2912 } 2913 2914 public: 2915 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { } 2916 2917 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) { 2918 _phase = phase; 2919 _info = info; 2920 } 2921 2922 virtual void do_oop(oop* p) { 2923 oop obj = oopDesc::load_decode_heap_oop(p); 2924 do_object_work(obj); 2925 } 2926 2927 virtual void do_oop(narrowOop* p) { 2928 // We should not come across narrow oops while scanning marking 2929 // stacks and SATB buffers. 2930 ShouldNotReachHere(); 2931 } 2932 2933 virtual void do_object(oop obj) { 2934 do_object_work(obj); 2935 } 2936 }; 2937 2938 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks, 2939 bool verify_enqueued_buffers, 2940 bool verify_thread_buffers, 2941 bool verify_fingers) { 2942 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 2943 if (!G1CollectedHeap::heap()->mark_in_progress()) { 2944 return; 2945 } 2946 2947 VerifyNoCSetOopsClosure cl; 2948 2949 if (verify_stacks) { 2950 // Verify entries on the global mark stack 2951 cl.set_phase(VerifyNoCSetOopsStack); 2952 _markStack.oops_do(&cl); 2953 2954 // Verify entries on the task queues 2955 for (uint i = 0; i < _max_worker_id; i += 1) { 2956 cl.set_phase(VerifyNoCSetOopsQueues, i); 2957 CMTaskQueue* queue = _task_queues->queue(i); 2958 queue->oops_do(&cl); 2959 } 2960 } 2961 2962 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 2963 2964 // Verify entries on the enqueued SATB buffers 2965 if (verify_enqueued_buffers) { 2966 cl.set_phase(VerifyNoCSetOopsSATBCompleted); 2967 satb_qs.iterate_completed_buffers_read_only(&cl); 2968 } 2969 2970 // Verify entries on the per-thread SATB buffers 2971 if (verify_thread_buffers) { 2972 cl.set_phase(VerifyNoCSetOopsSATBThread); 2973 satb_qs.iterate_thread_buffers_read_only(&cl); 2974 } 2975 2976 if (verify_fingers) { 2977 // Verify the global finger 2978 HeapWord* global_finger = finger(); 2979 if (global_finger != NULL && global_finger < _heap_end) { 2980 // The global finger always points to a heap region boundary. We 2981 // use heap_region_containing_raw() to get the containing region 2982 // given that the global finger could be pointing to a free region 2983 // which subsequently becomes continues humongous. If that 2984 // happens, heap_region_containing() will return the bottom of the 2985 // corresponding starts humongous region and the check below will 2986 // not hold any more. 2987 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger); 2988 guarantee(global_finger == global_hr->bottom(), 2989 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT, 2990 global_finger, HR_FORMAT_PARAMS(global_hr))); 2991 } 2992 2993 // Verify the task fingers 2994 assert(parallel_marking_threads() <= _max_worker_id, "sanity"); 2995 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) { 2996 CMTask* task = _tasks[i]; 2997 HeapWord* task_finger = task->finger(); 2998 if (task_finger != NULL && task_finger < _heap_end) { 2999 // See above note on the global finger verification. 3000 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger); 3001 guarantee(task_finger == task_hr->bottom() || 3002 !task_hr->in_collection_set(), 3003 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT, 3004 task_finger, HR_FORMAT_PARAMS(task_hr))); 3005 } 3006 } 3007 } 3008 } 3009 #endif // PRODUCT 3010 3011 // Aggregate the counting data that was constructed concurrently 3012 // with marking. 3013 class AggregateCountDataHRClosure: public HeapRegionClosure { 3014 G1CollectedHeap* _g1h; 3015 ConcurrentMark* _cm; 3016 CardTableModRefBS* _ct_bs; 3017 BitMap* _cm_card_bm; 3018 uint _max_worker_id; 3019 3020 public: 3021 AggregateCountDataHRClosure(G1CollectedHeap* g1h, 3022 BitMap* cm_card_bm, 3023 uint max_worker_id) : 3024 _g1h(g1h), _cm(g1h->concurrent_mark()), 3025 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 3026 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { } 3027 3028 bool doHeapRegion(HeapRegion* hr) { 3029 if (hr->continuesHumongous()) { 3030 // We will ignore these here and process them when their 3031 // associated "starts humongous" region is processed. 3032 // Note that we cannot rely on their associated 3033 // "starts humongous" region to have their bit set to 1 3034 // since, due to the region chunking in the parallel region 3035 // iteration, a "continues humongous" region might be visited 3036 // before its associated "starts humongous". 3037 return false; 3038 } 3039 3040 HeapWord* start = hr->bottom(); 3041 HeapWord* limit = hr->next_top_at_mark_start(); 3042 HeapWord* end = hr->end(); 3043 3044 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(), 3045 err_msg("Preconditions not met - " 3046 "start: "PTR_FORMAT", limit: "PTR_FORMAT", " 3047 "top: "PTR_FORMAT", end: "PTR_FORMAT, 3048 start, limit, hr->top(), hr->end())); 3049 3050 assert(hr->next_marked_bytes() == 0, "Precondition"); 3051 3052 if (start == limit) { 3053 // NTAMS of this region has not been set so nothing to do. 3054 return false; 3055 } 3056 3057 // 'start' should be in the heap. 3058 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity"); 3059 // 'end' *may* be just beyone the end of the heap (if hr is the last region) 3060 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity"); 3061 3062 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 3063 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit); 3064 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end); 3065 3066 // If ntams is not card aligned then we bump card bitmap index 3067 // for limit so that we get the all the cards spanned by 3068 // the object ending at ntams. 3069 // Note: if this is the last region in the heap then ntams 3070 // could be actually just beyond the end of the the heap; 3071 // limit_idx will then correspond to a (non-existent) card 3072 // that is also outside the heap. 3073 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) { 3074 limit_idx += 1; 3075 } 3076 3077 assert(limit_idx <= end_idx, "or else use atomics"); 3078 3079 // Aggregate the "stripe" in the count data associated with hr. 3080 uint hrs_index = hr->hrs_index(); 3081 size_t marked_bytes = 0; 3082 3083 for (uint i = 0; i < _max_worker_id; i += 1) { 3084 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i); 3085 BitMap* task_card_bm = _cm->count_card_bitmap_for(i); 3086 3087 // Fetch the marked_bytes in this region for task i and 3088 // add it to the running total for this region. 3089 marked_bytes += marked_bytes_array[hrs_index]; 3090 3091 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx) 3092 // into the global card bitmap. 3093 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx); 3094 3095 while (scan_idx < limit_idx) { 3096 assert(task_card_bm->at(scan_idx) == true, "should be"); 3097 _cm_card_bm->set_bit(scan_idx); 3098 assert(_cm_card_bm->at(scan_idx) == true, "should be"); 3099 3100 // BitMap::get_next_one_offset() can handle the case when 3101 // its left_offset parameter is greater than its right_offset 3102 // parameter. It does, however, have an early exit if 3103 // left_offset == right_offset. So let's limit the value 3104 // passed in for left offset here. 3105 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx); 3106 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx); 3107 } 3108 } 3109 3110 // Update the marked bytes for this region. 3111 hr->add_to_marked_bytes(marked_bytes); 3112 3113 // Next heap region 3114 return false; 3115 } 3116 }; 3117 3118 class G1AggregateCountDataTask: public AbstractGangTask { 3119 protected: 3120 G1CollectedHeap* _g1h; 3121 ConcurrentMark* _cm; 3122 BitMap* _cm_card_bm; 3123 uint _max_worker_id; 3124 int _active_workers; 3125 3126 public: 3127 G1AggregateCountDataTask(G1CollectedHeap* g1h, 3128 ConcurrentMark* cm, 3129 BitMap* cm_card_bm, 3130 uint max_worker_id, 3131 int n_workers) : 3132 AbstractGangTask("Count Aggregation"), 3133 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm), 3134 _max_worker_id(max_worker_id), 3135 _active_workers(n_workers) { } 3136 3137 void work(uint worker_id) { 3138 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id); 3139 3140 if (G1CollectedHeap::use_parallel_gc_threads()) { 3141 _g1h->heap_region_par_iterate_chunked(&cl, worker_id, 3142 _active_workers, 3143 HeapRegion::AggregateCountClaimValue); 3144 } else { 3145 _g1h->heap_region_iterate(&cl); 3146 } 3147 } 3148 }; 3149 3150 3151 void ConcurrentMark::aggregate_count_data() { 3152 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3153 _g1h->workers()->active_workers() : 3154 1); 3155 3156 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm, 3157 _max_worker_id, n_workers); 3158 3159 if (G1CollectedHeap::use_parallel_gc_threads()) { 3160 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3161 "sanity check"); 3162 _g1h->set_par_threads(n_workers); 3163 _g1h->workers()->run_task(&g1_par_agg_task); 3164 _g1h->set_par_threads(0); 3165 3166 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue), 3167 "sanity check"); 3168 _g1h->reset_heap_region_claim_values(); 3169 } else { 3170 g1_par_agg_task.work(0); 3171 } 3172 } 3173 3174 // Clear the per-worker arrays used to store the per-region counting data 3175 void ConcurrentMark::clear_all_count_data() { 3176 // Clear the global card bitmap - it will be filled during 3177 // liveness count aggregation (during remark) and the 3178 // final counting task. 3179 _card_bm.clear(); 3180 3181 // Clear the global region bitmap - it will be filled as part 3182 // of the final counting task. 3183 _region_bm.clear(); 3184 3185 uint max_regions = _g1h->max_regions(); 3186 assert(_max_worker_id > 0, "uninitialized"); 3187 3188 for (uint i = 0; i < _max_worker_id; i += 1) { 3189 BitMap* task_card_bm = count_card_bitmap_for(i); 3190 size_t* marked_bytes_array = count_marked_bytes_array_for(i); 3191 3192 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 3193 assert(marked_bytes_array != NULL, "uninitialized"); 3194 3195 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t)); 3196 task_card_bm->clear(); 3197 } 3198 } 3199 3200 void ConcurrentMark::print_stats() { 3201 if (verbose_stats()) { 3202 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3203 for (size_t i = 0; i < _active_tasks; ++i) { 3204 _tasks[i]->print_stats(); 3205 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3206 } 3207 } 3208 } 3209 3210 // abandon current marking iteration due to a Full GC 3211 void ConcurrentMark::abort() { 3212 // Clear all marks to force marking thread to do nothing 3213 _nextMarkBitMap->clearAll(); 3214 // Clear the liveness counting data 3215 clear_all_count_data(); 3216 // Empty mark stack 3217 reset_marking_state(); 3218 for (uint i = 0; i < _max_worker_id; ++i) { 3219 _tasks[i]->clear_region_fields(); 3220 } 3221 _has_aborted = true; 3222 3223 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3224 satb_mq_set.abandon_partial_marking(); 3225 // This can be called either during or outside marking, we'll read 3226 // the expected_active value from the SATB queue set. 3227 satb_mq_set.set_active_all_threads( 3228 false, /* new active value */ 3229 satb_mq_set.is_active() /* expected_active */); 3230 } 3231 3232 static void print_ms_time_info(const char* prefix, const char* name, 3233 NumberSeq& ns) { 3234 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 3235 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 3236 if (ns.num() > 0) { 3237 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]", 3238 prefix, ns.sd(), ns.maximum()); 3239 } 3240 } 3241 3242 void ConcurrentMark::print_summary_info() { 3243 gclog_or_tty->print_cr(" Concurrent marking:"); 3244 print_ms_time_info(" ", "init marks", _init_times); 3245 print_ms_time_info(" ", "remarks", _remark_times); 3246 { 3247 print_ms_time_info(" ", "final marks", _remark_mark_times); 3248 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 3249 3250 } 3251 print_ms_time_info(" ", "cleanups", _cleanup_times); 3252 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).", 3253 _total_counting_time, 3254 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / 3255 (double)_cleanup_times.num() 3256 : 0.0)); 3257 if (G1ScrubRemSets) { 3258 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 3259 _total_rs_scrub_time, 3260 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / 3261 (double)_cleanup_times.num() 3262 : 0.0)); 3263 } 3264 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.", 3265 (_init_times.sum() + _remark_times.sum() + 3266 _cleanup_times.sum())/1000.0); 3267 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s " 3268 "(%8.2f s marking).", 3269 cmThread()->vtime_accum(), 3270 cmThread()->vtime_mark_accum()); 3271 } 3272 3273 void ConcurrentMark::print_worker_threads_on(outputStream* st) const { 3274 if (use_parallel_marking_threads()) { 3275 _parallel_workers->print_worker_threads_on(st); 3276 } 3277 } 3278 3279 void ConcurrentMark::print_on_error(outputStream* st) const { 3280 st->print_cr("Marking Bits (Prev, Next): (CMBitMap*) " PTR_FORMAT ", (CMBitMap*) " PTR_FORMAT, 3281 _prevMarkBitMap, _nextMarkBitMap); 3282 _prevMarkBitMap->print_on_error(st, " Prev Bits: "); 3283 _nextMarkBitMap->print_on_error(st, " Next Bits: "); 3284 } 3285 3286 // We take a break if someone is trying to stop the world. 3287 bool ConcurrentMark::do_yield_check(uint worker_id) { 3288 if (should_yield()) { 3289 if (worker_id == 0) { 3290 _g1h->g1_policy()->record_concurrent_pause(); 3291 } 3292 cmThread()->yield(); 3293 return true; 3294 } else { 3295 return false; 3296 } 3297 } 3298 3299 bool ConcurrentMark::should_yield() { 3300 return cmThread()->should_yield(); 3301 } 3302 3303 bool ConcurrentMark::containing_card_is_marked(void* p) { 3304 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1); 3305 return _card_bm.at(offset >> CardTableModRefBS::card_shift); 3306 } 3307 3308 bool ConcurrentMark::containing_cards_are_marked(void* start, 3309 void* last) { 3310 return containing_card_is_marked(start) && 3311 containing_card_is_marked(last); 3312 } 3313 3314 #ifndef PRODUCT 3315 // for debugging purposes 3316 void ConcurrentMark::print_finger() { 3317 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT, 3318 _heap_start, _heap_end, _finger); 3319 for (uint i = 0; i < _max_worker_id; ++i) { 3320 gclog_or_tty->print(" %u: "PTR_FORMAT, i, _tasks[i]->finger()); 3321 } 3322 gclog_or_tty->print_cr(""); 3323 } 3324 #endif 3325 3326 void CMTask::scan_object(oop obj) { 3327 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant"); 3328 3329 if (_cm->verbose_high()) { 3330 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT, 3331 _worker_id, (void*) obj); 3332 } 3333 3334 size_t obj_size = obj->size(); 3335 _words_scanned += obj_size; 3336 3337 obj->oop_iterate(_cm_oop_closure); 3338 statsOnly( ++_objs_scanned ); 3339 check_limits(); 3340 } 3341 3342 // Closure for iteration over bitmaps 3343 class CMBitMapClosure : public BitMapClosure { 3344 private: 3345 // the bitmap that is being iterated over 3346 CMBitMap* _nextMarkBitMap; 3347 ConcurrentMark* _cm; 3348 CMTask* _task; 3349 3350 public: 3351 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) : 3352 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } 3353 3354 bool do_bit(size_t offset) { 3355 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); 3356 assert(_nextMarkBitMap->isMarked(addr), "invariant"); 3357 assert( addr < _cm->finger(), "invariant"); 3358 3359 statsOnly( _task->increase_objs_found_on_bitmap() ); 3360 assert(addr >= _task->finger(), "invariant"); 3361 3362 // We move that task's local finger along. 3363 _task->move_finger_to(addr); 3364 3365 _task->scan_object(oop(addr)); 3366 // we only partially drain the local queue and global stack 3367 _task->drain_local_queue(true); 3368 _task->drain_global_stack(true); 3369 3370 // if the has_aborted flag has been raised, we need to bail out of 3371 // the iteration 3372 return !_task->has_aborted(); 3373 } 3374 }; 3375 3376 // Closure for iterating over objects, currently only used for 3377 // processing SATB buffers. 3378 class CMObjectClosure : public ObjectClosure { 3379 private: 3380 CMTask* _task; 3381 3382 public: 3383 void do_object(oop obj) { 3384 _task->deal_with_reference(obj); 3385 } 3386 3387 CMObjectClosure(CMTask* task) : _task(task) { } 3388 }; 3389 3390 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 3391 ConcurrentMark* cm, 3392 CMTask* task) 3393 : _g1h(g1h), _cm(cm), _task(task) { 3394 assert(_ref_processor == NULL, "should be initialized to NULL"); 3395 3396 if (G1UseConcMarkReferenceProcessing) { 3397 _ref_processor = g1h->ref_processor_cm(); 3398 assert(_ref_processor != NULL, "should not be NULL"); 3399 } 3400 } 3401 3402 void CMTask::setup_for_region(HeapRegion* hr) { 3403 // Separated the asserts so that we know which one fires. 3404 assert(hr != NULL, 3405 "claim_region() should have filtered out continues humongous regions"); 3406 assert(!hr->continuesHumongous(), 3407 "claim_region() should have filtered out continues humongous regions"); 3408 3409 if (_cm->verbose_low()) { 3410 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT, 3411 _worker_id, hr); 3412 } 3413 3414 _curr_region = hr; 3415 _finger = hr->bottom(); 3416 update_region_limit(); 3417 } 3418 3419 void CMTask::update_region_limit() { 3420 HeapRegion* hr = _curr_region; 3421 HeapWord* bottom = hr->bottom(); 3422 HeapWord* limit = hr->next_top_at_mark_start(); 3423 3424 if (limit == bottom) { 3425 if (_cm->verbose_low()) { 3426 gclog_or_tty->print_cr("[%u] found an empty region " 3427 "["PTR_FORMAT", "PTR_FORMAT")", 3428 _worker_id, bottom, limit); 3429 } 3430 // The region was collected underneath our feet. 3431 // We set the finger to bottom to ensure that the bitmap 3432 // iteration that will follow this will not do anything. 3433 // (this is not a condition that holds when we set the region up, 3434 // as the region is not supposed to be empty in the first place) 3435 _finger = bottom; 3436 } else if (limit >= _region_limit) { 3437 assert(limit >= _finger, "peace of mind"); 3438 } else { 3439 assert(limit < _region_limit, "only way to get here"); 3440 // This can happen under some pretty unusual circumstances. An 3441 // evacuation pause empties the region underneath our feet (NTAMS 3442 // at bottom). We then do some allocation in the region (NTAMS 3443 // stays at bottom), followed by the region being used as a GC 3444 // alloc region (NTAMS will move to top() and the objects 3445 // originally below it will be grayed). All objects now marked in 3446 // the region are explicitly grayed, if below the global finger, 3447 // and we do not need in fact to scan anything else. So, we simply 3448 // set _finger to be limit to ensure that the bitmap iteration 3449 // doesn't do anything. 3450 _finger = limit; 3451 } 3452 3453 _region_limit = limit; 3454 } 3455 3456 void CMTask::giveup_current_region() { 3457 assert(_curr_region != NULL, "invariant"); 3458 if (_cm->verbose_low()) { 3459 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT, 3460 _worker_id, _curr_region); 3461 } 3462 clear_region_fields(); 3463 } 3464 3465 void CMTask::clear_region_fields() { 3466 // Values for these three fields that indicate that we're not 3467 // holding on to a region. 3468 _curr_region = NULL; 3469 _finger = NULL; 3470 _region_limit = NULL; 3471 } 3472 3473 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 3474 if (cm_oop_closure == NULL) { 3475 assert(_cm_oop_closure != NULL, "invariant"); 3476 } else { 3477 assert(_cm_oop_closure == NULL, "invariant"); 3478 } 3479 _cm_oop_closure = cm_oop_closure; 3480 } 3481 3482 void CMTask::reset(CMBitMap* nextMarkBitMap) { 3483 guarantee(nextMarkBitMap != NULL, "invariant"); 3484 3485 if (_cm->verbose_low()) { 3486 gclog_or_tty->print_cr("[%u] resetting", _worker_id); 3487 } 3488 3489 _nextMarkBitMap = nextMarkBitMap; 3490 clear_region_fields(); 3491 3492 _calls = 0; 3493 _elapsed_time_ms = 0.0; 3494 _termination_time_ms = 0.0; 3495 _termination_start_time_ms = 0.0; 3496 3497 #if _MARKING_STATS_ 3498 _local_pushes = 0; 3499 _local_pops = 0; 3500 _local_max_size = 0; 3501 _objs_scanned = 0; 3502 _global_pushes = 0; 3503 _global_pops = 0; 3504 _global_max_size = 0; 3505 _global_transfers_to = 0; 3506 _global_transfers_from = 0; 3507 _regions_claimed = 0; 3508 _objs_found_on_bitmap = 0; 3509 _satb_buffers_processed = 0; 3510 _steal_attempts = 0; 3511 _steals = 0; 3512 _aborted = 0; 3513 _aborted_overflow = 0; 3514 _aborted_cm_aborted = 0; 3515 _aborted_yield = 0; 3516 _aborted_timed_out = 0; 3517 _aborted_satb = 0; 3518 _aborted_termination = 0; 3519 #endif // _MARKING_STATS_ 3520 } 3521 3522 bool CMTask::should_exit_termination() { 3523 regular_clock_call(); 3524 // This is called when we are in the termination protocol. We should 3525 // quit if, for some reason, this task wants to abort or the global 3526 // stack is not empty (this means that we can get work from it). 3527 return !_cm->mark_stack_empty() || has_aborted(); 3528 } 3529 3530 void CMTask::reached_limit() { 3531 assert(_words_scanned >= _words_scanned_limit || 3532 _refs_reached >= _refs_reached_limit , 3533 "shouldn't have been called otherwise"); 3534 regular_clock_call(); 3535 } 3536 3537 void CMTask::regular_clock_call() { 3538 if (has_aborted()) return; 3539 3540 // First, we need to recalculate the words scanned and refs reached 3541 // limits for the next clock call. 3542 recalculate_limits(); 3543 3544 // During the regular clock call we do the following 3545 3546 // (1) If an overflow has been flagged, then we abort. 3547 if (_cm->has_overflown()) { 3548 set_has_aborted(); 3549 return; 3550 } 3551 3552 // If we are not concurrent (i.e. we're doing remark) we don't need 3553 // to check anything else. The other steps are only needed during 3554 // the concurrent marking phase. 3555 if (!concurrent()) return; 3556 3557 // (2) If marking has been aborted for Full GC, then we also abort. 3558 if (_cm->has_aborted()) { 3559 set_has_aborted(); 3560 statsOnly( ++_aborted_cm_aborted ); 3561 return; 3562 } 3563 3564 double curr_time_ms = os::elapsedVTime() * 1000.0; 3565 3566 // (3) If marking stats are enabled, then we update the step history. 3567 #if _MARKING_STATS_ 3568 if (_words_scanned >= _words_scanned_limit) { 3569 ++_clock_due_to_scanning; 3570 } 3571 if (_refs_reached >= _refs_reached_limit) { 3572 ++_clock_due_to_marking; 3573 } 3574 3575 double last_interval_ms = curr_time_ms - _interval_start_time_ms; 3576 _interval_start_time_ms = curr_time_ms; 3577 _all_clock_intervals_ms.add(last_interval_ms); 3578 3579 if (_cm->verbose_medium()) { 3580 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, " 3581 "scanned = %d%s, refs reached = %d%s", 3582 _worker_id, last_interval_ms, 3583 _words_scanned, 3584 (_words_scanned >= _words_scanned_limit) ? " (*)" : "", 3585 _refs_reached, 3586 (_refs_reached >= _refs_reached_limit) ? " (*)" : ""); 3587 } 3588 #endif // _MARKING_STATS_ 3589 3590 // (4) We check whether we should yield. If we have to, then we abort. 3591 if (_cm->should_yield()) { 3592 // We should yield. To do this we abort the task. The caller is 3593 // responsible for yielding. 3594 set_has_aborted(); 3595 statsOnly( ++_aborted_yield ); 3596 return; 3597 } 3598 3599 // (5) We check whether we've reached our time quota. If we have, 3600 // then we abort. 3601 double elapsed_time_ms = curr_time_ms - _start_time_ms; 3602 if (elapsed_time_ms > _time_target_ms) { 3603 set_has_aborted(); 3604 _has_timed_out = true; 3605 statsOnly( ++_aborted_timed_out ); 3606 return; 3607 } 3608 3609 // (6) Finally, we check whether there are enough completed STAB 3610 // buffers available for processing. If there are, we abort. 3611 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3612 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 3613 if (_cm->verbose_low()) { 3614 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers", 3615 _worker_id); 3616 } 3617 // we do need to process SATB buffers, we'll abort and restart 3618 // the marking task to do so 3619 set_has_aborted(); 3620 statsOnly( ++_aborted_satb ); 3621 return; 3622 } 3623 } 3624 3625 void CMTask::recalculate_limits() { 3626 _real_words_scanned_limit = _words_scanned + words_scanned_period; 3627 _words_scanned_limit = _real_words_scanned_limit; 3628 3629 _real_refs_reached_limit = _refs_reached + refs_reached_period; 3630 _refs_reached_limit = _real_refs_reached_limit; 3631 } 3632 3633 void CMTask::decrease_limits() { 3634 // This is called when we believe that we're going to do an infrequent 3635 // operation which will increase the per byte scanned cost (i.e. move 3636 // entries to/from the global stack). It basically tries to decrease the 3637 // scanning limit so that the clock is called earlier. 3638 3639 if (_cm->verbose_medium()) { 3640 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id); 3641 } 3642 3643 _words_scanned_limit = _real_words_scanned_limit - 3644 3 * words_scanned_period / 4; 3645 _refs_reached_limit = _real_refs_reached_limit - 3646 3 * refs_reached_period / 4; 3647 } 3648 3649 void CMTask::move_entries_to_global_stack() { 3650 // local array where we'll store the entries that will be popped 3651 // from the local queue 3652 oop buffer[global_stack_transfer_size]; 3653 3654 int n = 0; 3655 oop obj; 3656 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { 3657 buffer[n] = obj; 3658 ++n; 3659 } 3660 3661 if (n > 0) { 3662 // we popped at least one entry from the local queue 3663 3664 statsOnly( ++_global_transfers_to; _local_pops += n ); 3665 3666 if (!_cm->mark_stack_push(buffer, n)) { 3667 if (_cm->verbose_low()) { 3668 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow", 3669 _worker_id); 3670 } 3671 set_has_aborted(); 3672 } else { 3673 // the transfer was successful 3674 3675 if (_cm->verbose_medium()) { 3676 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack", 3677 _worker_id, n); 3678 } 3679 statsOnly( int tmp_size = _cm->mark_stack_size(); 3680 if (tmp_size > _global_max_size) { 3681 _global_max_size = tmp_size; 3682 } 3683 _global_pushes += n ); 3684 } 3685 } 3686 3687 // this operation was quite expensive, so decrease the limits 3688 decrease_limits(); 3689 } 3690 3691 void CMTask::get_entries_from_global_stack() { 3692 // local array where we'll store the entries that will be popped 3693 // from the global stack. 3694 oop buffer[global_stack_transfer_size]; 3695 int n; 3696 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); 3697 assert(n <= global_stack_transfer_size, 3698 "we should not pop more than the given limit"); 3699 if (n > 0) { 3700 // yes, we did actually pop at least one entry 3701 3702 statsOnly( ++_global_transfers_from; _global_pops += n ); 3703 if (_cm->verbose_medium()) { 3704 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack", 3705 _worker_id, n); 3706 } 3707 for (int i = 0; i < n; ++i) { 3708 bool success = _task_queue->push(buffer[i]); 3709 // We only call this when the local queue is empty or under a 3710 // given target limit. So, we do not expect this push to fail. 3711 assert(success, "invariant"); 3712 } 3713 3714 statsOnly( int tmp_size = _task_queue->size(); 3715 if (tmp_size > _local_max_size) { 3716 _local_max_size = tmp_size; 3717 } 3718 _local_pushes += n ); 3719 } 3720 3721 // this operation was quite expensive, so decrease the limits 3722 decrease_limits(); 3723 } 3724 3725 void CMTask::drain_local_queue(bool partially) { 3726 if (has_aborted()) return; 3727 3728 // Decide what the target size is, depending whether we're going to 3729 // drain it partially (so that other tasks can steal if they run out 3730 // of things to do) or totally (at the very end). 3731 size_t target_size; 3732 if (partially) { 3733 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 3734 } else { 3735 target_size = 0; 3736 } 3737 3738 if (_task_queue->size() > target_size) { 3739 if (_cm->verbose_high()) { 3740 gclog_or_tty->print_cr("[%u] draining local queue, target size = %d", 3741 _worker_id, target_size); 3742 } 3743 3744 oop obj; 3745 bool ret = _task_queue->pop_local(obj); 3746 while (ret) { 3747 statsOnly( ++_local_pops ); 3748 3749 if (_cm->verbose_high()) { 3750 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id, 3751 (void*) obj); 3752 } 3753 3754 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); 3755 assert(!_g1h->is_on_master_free_list( 3756 _g1h->heap_region_containing((HeapWord*) obj)), "invariant"); 3757 3758 scan_object(obj); 3759 3760 if (_task_queue->size() <= target_size || has_aborted()) { 3761 ret = false; 3762 } else { 3763 ret = _task_queue->pop_local(obj); 3764 } 3765 } 3766 3767 if (_cm->verbose_high()) { 3768 gclog_or_tty->print_cr("[%u] drained local queue, size = %d", 3769 _worker_id, _task_queue->size()); 3770 } 3771 } 3772 } 3773 3774 void CMTask::drain_global_stack(bool partially) { 3775 if (has_aborted()) return; 3776 3777 // We have a policy to drain the local queue before we attempt to 3778 // drain the global stack. 3779 assert(partially || _task_queue->size() == 0, "invariant"); 3780 3781 // Decide what the target size is, depending whether we're going to 3782 // drain it partially (so that other tasks can steal if they run out 3783 // of things to do) or totally (at the very end). Notice that, 3784 // because we move entries from the global stack in chunks or 3785 // because another task might be doing the same, we might in fact 3786 // drop below the target. But, this is not a problem. 3787 size_t target_size; 3788 if (partially) { 3789 target_size = _cm->partial_mark_stack_size_target(); 3790 } else { 3791 target_size = 0; 3792 } 3793 3794 if (_cm->mark_stack_size() > target_size) { 3795 if (_cm->verbose_low()) { 3796 gclog_or_tty->print_cr("[%u] draining global_stack, target size %d", 3797 _worker_id, target_size); 3798 } 3799 3800 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 3801 get_entries_from_global_stack(); 3802 drain_local_queue(partially); 3803 } 3804 3805 if (_cm->verbose_low()) { 3806 gclog_or_tty->print_cr("[%u] drained global stack, size = %d", 3807 _worker_id, _cm->mark_stack_size()); 3808 } 3809 } 3810 } 3811 3812 // SATB Queue has several assumptions on whether to call the par or 3813 // non-par versions of the methods. this is why some of the code is 3814 // replicated. We should really get rid of the single-threaded version 3815 // of the code to simplify things. 3816 void CMTask::drain_satb_buffers() { 3817 if (has_aborted()) return; 3818 3819 // We set this so that the regular clock knows that we're in the 3820 // middle of draining buffers and doesn't set the abort flag when it 3821 // notices that SATB buffers are available for draining. It'd be 3822 // very counter productive if it did that. :-) 3823 _draining_satb_buffers = true; 3824 3825 CMObjectClosure oc(this); 3826 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3827 if (G1CollectedHeap::use_parallel_gc_threads()) { 3828 satb_mq_set.set_par_closure(_worker_id, &oc); 3829 } else { 3830 satb_mq_set.set_closure(&oc); 3831 } 3832 3833 // This keeps claiming and applying the closure to completed buffers 3834 // until we run out of buffers or we need to abort. 3835 if (G1CollectedHeap::use_parallel_gc_threads()) { 3836 while (!has_aborted() && 3837 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) { 3838 if (_cm->verbose_medium()) { 3839 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id); 3840 } 3841 statsOnly( ++_satb_buffers_processed ); 3842 regular_clock_call(); 3843 } 3844 } else { 3845 while (!has_aborted() && 3846 satb_mq_set.apply_closure_to_completed_buffer()) { 3847 if (_cm->verbose_medium()) { 3848 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id); 3849 } 3850 statsOnly( ++_satb_buffers_processed ); 3851 regular_clock_call(); 3852 } 3853 } 3854 3855 if (!concurrent() && !has_aborted()) { 3856 // We should only do this during remark. 3857 if (G1CollectedHeap::use_parallel_gc_threads()) { 3858 satb_mq_set.par_iterate_closure_all_threads(_worker_id); 3859 } else { 3860 satb_mq_set.iterate_closure_all_threads(); 3861 } 3862 } 3863 3864 _draining_satb_buffers = false; 3865 3866 assert(has_aborted() || 3867 concurrent() || 3868 satb_mq_set.completed_buffers_num() == 0, "invariant"); 3869 3870 if (G1CollectedHeap::use_parallel_gc_threads()) { 3871 satb_mq_set.set_par_closure(_worker_id, NULL); 3872 } else { 3873 satb_mq_set.set_closure(NULL); 3874 } 3875 3876 // again, this was a potentially expensive operation, decrease the 3877 // limits to get the regular clock call early 3878 decrease_limits(); 3879 } 3880 3881 void CMTask::print_stats() { 3882 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d", 3883 _worker_id, _calls); 3884 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 3885 _elapsed_time_ms, _termination_time_ms); 3886 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 3887 _step_times_ms.num(), _step_times_ms.avg(), 3888 _step_times_ms.sd()); 3889 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 3890 _step_times_ms.maximum(), _step_times_ms.sum()); 3891 3892 #if _MARKING_STATS_ 3893 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 3894 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(), 3895 _all_clock_intervals_ms.sd()); 3896 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 3897 _all_clock_intervals_ms.maximum(), 3898 _all_clock_intervals_ms.sum()); 3899 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d", 3900 _clock_due_to_scanning, _clock_due_to_marking); 3901 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d", 3902 _objs_scanned, _objs_found_on_bitmap); 3903 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d", 3904 _local_pushes, _local_pops, _local_max_size); 3905 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d", 3906 _global_pushes, _global_pops, _global_max_size); 3907 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d", 3908 _global_transfers_to,_global_transfers_from); 3909 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed); 3910 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed); 3911 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d", 3912 _steal_attempts, _steals); 3913 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted); 3914 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d", 3915 _aborted_overflow, _aborted_cm_aborted, _aborted_yield); 3916 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d", 3917 _aborted_timed_out, _aborted_satb, _aborted_termination); 3918 #endif // _MARKING_STATS_ 3919 } 3920 3921 /***************************************************************************** 3922 3923 The do_marking_step(time_target_ms, ...) method is the building 3924 block of the parallel marking framework. It can be called in parallel 3925 with other invocations of do_marking_step() on different tasks 3926 (but only one per task, obviously) and concurrently with the 3927 mutator threads, or during remark, hence it eliminates the need 3928 for two versions of the code. When called during remark, it will 3929 pick up from where the task left off during the concurrent marking 3930 phase. Interestingly, tasks are also claimable during evacuation 3931 pauses too, since do_marking_step() ensures that it aborts before 3932 it needs to yield. 3933 3934 The data structures that it uses to do marking work are the 3935 following: 3936 3937 (1) Marking Bitmap. If there are gray objects that appear only 3938 on the bitmap (this happens either when dealing with an overflow 3939 or when the initial marking phase has simply marked the roots 3940 and didn't push them on the stack), then tasks claim heap 3941 regions whose bitmap they then scan to find gray objects. A 3942 global finger indicates where the end of the last claimed region 3943 is. A local finger indicates how far into the region a task has 3944 scanned. The two fingers are used to determine how to gray an 3945 object (i.e. whether simply marking it is OK, as it will be 3946 visited by a task in the future, or whether it needs to be also 3947 pushed on a stack). 3948 3949 (2) Local Queue. The local queue of the task which is accessed 3950 reasonably efficiently by the task. Other tasks can steal from 3951 it when they run out of work. Throughout the marking phase, a 3952 task attempts to keep its local queue short but not totally 3953 empty, so that entries are available for stealing by other 3954 tasks. Only when there is no more work, a task will totally 3955 drain its local queue. 3956 3957 (3) Global Mark Stack. This handles local queue overflow. During 3958 marking only sets of entries are moved between it and the local 3959 queues, as access to it requires a mutex and more fine-grain 3960 interaction with it which might cause contention. If it 3961 overflows, then the marking phase should restart and iterate 3962 over the bitmap to identify gray objects. Throughout the marking 3963 phase, tasks attempt to keep the global mark stack at a small 3964 length but not totally empty, so that entries are available for 3965 popping by other tasks. Only when there is no more work, tasks 3966 will totally drain the global mark stack. 3967 3968 (4) SATB Buffer Queue. This is where completed SATB buffers are 3969 made available. Buffers are regularly removed from this queue 3970 and scanned for roots, so that the queue doesn't get too 3971 long. During remark, all completed buffers are processed, as 3972 well as the filled in parts of any uncompleted buffers. 3973 3974 The do_marking_step() method tries to abort when the time target 3975 has been reached. There are a few other cases when the 3976 do_marking_step() method also aborts: 3977 3978 (1) When the marking phase has been aborted (after a Full GC). 3979 3980 (2) When a global overflow (on the global stack) has been 3981 triggered. Before the task aborts, it will actually sync up with 3982 the other tasks to ensure that all the marking data structures 3983 (local queues, stacks, fingers etc.) are re-initialized so that 3984 when do_marking_step() completes, the marking phase can 3985 immediately restart. 3986 3987 (3) When enough completed SATB buffers are available. The 3988 do_marking_step() method only tries to drain SATB buffers right 3989 at the beginning. So, if enough buffers are available, the 3990 marking step aborts and the SATB buffers are processed at 3991 the beginning of the next invocation. 3992 3993 (4) To yield. when we have to yield then we abort and yield 3994 right at the end of do_marking_step(). This saves us from a lot 3995 of hassle as, by yielding we might allow a Full GC. If this 3996 happens then objects will be compacted underneath our feet, the 3997 heap might shrink, etc. We save checking for this by just 3998 aborting and doing the yield right at the end. 3999 4000 From the above it follows that the do_marking_step() method should 4001 be called in a loop (or, otherwise, regularly) until it completes. 4002 4003 If a marking step completes without its has_aborted() flag being 4004 true, it means it has completed the current marking phase (and 4005 also all other marking tasks have done so and have all synced up). 4006 4007 A method called regular_clock_call() is invoked "regularly" (in 4008 sub ms intervals) throughout marking. It is this clock method that 4009 checks all the abort conditions which were mentioned above and 4010 decides when the task should abort. A work-based scheme is used to 4011 trigger this clock method: when the number of object words the 4012 marking phase has scanned or the number of references the marking 4013 phase has visited reach a given limit. Additional invocations to 4014 the method clock have been planted in a few other strategic places 4015 too. The initial reason for the clock method was to avoid calling 4016 vtime too regularly, as it is quite expensive. So, once it was in 4017 place, it was natural to piggy-back all the other conditions on it 4018 too and not constantly check them throughout the code. 4019 4020 If do_termination is true then do_marking_step will enter its 4021 termination protocol. 4022 4023 The value of is_serial must be true when do_marking_step is being 4024 called serially (i.e. by the VMThread) and do_marking_step should 4025 skip any synchronization in the termination and overflow code. 4026 Examples include the serial remark code and the serial reference 4027 processing closures. 4028 4029 The value of is_serial must be false when do_marking_step is 4030 being called by any of the worker threads in a work gang. 4031 Examples include the concurrent marking code (CMMarkingTask), 4032 the MT remark code, and the MT reference processing closures. 4033 4034 *****************************************************************************/ 4035 4036 void CMTask::do_marking_step(double time_target_ms, 4037 bool do_termination, 4038 bool is_serial) { 4039 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 4040 assert(concurrent() == _cm->concurrent(), "they should be the same"); 4041 4042 G1CollectorPolicy* g1_policy = _g1h->g1_policy(); 4043 assert(_task_queues != NULL, "invariant"); 4044 assert(_task_queue != NULL, "invariant"); 4045 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant"); 4046 4047 assert(!_claimed, 4048 "only one thread should claim this task at any one time"); 4049 4050 // OK, this doesn't safeguard again all possible scenarios, as it is 4051 // possible for two threads to set the _claimed flag at the same 4052 // time. But it is only for debugging purposes anyway and it will 4053 // catch most problems. 4054 _claimed = true; 4055 4056 _start_time_ms = os::elapsedVTime() * 1000.0; 4057 statsOnly( _interval_start_time_ms = _start_time_ms ); 4058 4059 // If do_stealing is true then do_marking_step will attempt to 4060 // steal work from the other CMTasks. It only makes sense to 4061 // enable stealing when the termination protocol is enabled 4062 // and do_marking_step() is not being called serially. 4063 bool do_stealing = do_termination && !is_serial; 4064 4065 double diff_prediction_ms = 4066 g1_policy->get_new_prediction(&_marking_step_diffs_ms); 4067 _time_target_ms = time_target_ms - diff_prediction_ms; 4068 4069 // set up the variables that are used in the work-based scheme to 4070 // call the regular clock method 4071 _words_scanned = 0; 4072 _refs_reached = 0; 4073 recalculate_limits(); 4074 4075 // clear all flags 4076 clear_has_aborted(); 4077 _has_timed_out = false; 4078 _draining_satb_buffers = false; 4079 4080 ++_calls; 4081 4082 if (_cm->verbose_low()) { 4083 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, " 4084 "target = %1.2lfms >>>>>>>>>>", 4085 _worker_id, _calls, _time_target_ms); 4086 } 4087 4088 // Set up the bitmap and oop closures. Anything that uses them is 4089 // eventually called from this method, so it is OK to allocate these 4090 // statically. 4091 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); 4092 G1CMOopClosure cm_oop_closure(_g1h, _cm, this); 4093 set_cm_oop_closure(&cm_oop_closure); 4094 4095 if (_cm->has_overflown()) { 4096 // This can happen if the mark stack overflows during a GC pause 4097 // and this task, after a yield point, restarts. We have to abort 4098 // as we need to get into the overflow protocol which happens 4099 // right at the end of this task. 4100 set_has_aborted(); 4101 } 4102 4103 // First drain any available SATB buffers. After this, we will not 4104 // look at SATB buffers before the next invocation of this method. 4105 // If enough completed SATB buffers are queued up, the regular clock 4106 // will abort this task so that it restarts. 4107 drain_satb_buffers(); 4108 // ...then partially drain the local queue and the global stack 4109 drain_local_queue(true); 4110 drain_global_stack(true); 4111 4112 do { 4113 if (!has_aborted() && _curr_region != NULL) { 4114 // This means that we're already holding on to a region. 4115 assert(_finger != NULL, "if region is not NULL, then the finger " 4116 "should not be NULL either"); 4117 4118 // We might have restarted this task after an evacuation pause 4119 // which might have evacuated the region we're holding on to 4120 // underneath our feet. Let's read its limit again to make sure 4121 // that we do not iterate over a region of the heap that 4122 // contains garbage (update_region_limit() will also move 4123 // _finger to the start of the region if it is found empty). 4124 update_region_limit(); 4125 // We will start from _finger not from the start of the region, 4126 // as we might be restarting this task after aborting half-way 4127 // through scanning this region. In this case, _finger points to 4128 // the address where we last found a marked object. If this is a 4129 // fresh region, _finger points to start(). 4130 MemRegion mr = MemRegion(_finger, _region_limit); 4131 4132 if (_cm->verbose_low()) { 4133 gclog_or_tty->print_cr("[%u] we're scanning part " 4134 "["PTR_FORMAT", "PTR_FORMAT") " 4135 "of region "HR_FORMAT, 4136 _worker_id, _finger, _region_limit, 4137 HR_FORMAT_PARAMS(_curr_region)); 4138 } 4139 4140 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(), 4141 "humongous regions should go around loop once only"); 4142 4143 // Some special cases: 4144 // If the memory region is empty, we can just give up the region. 4145 // If the current region is humongous then we only need to check 4146 // the bitmap for the bit associated with the start of the object, 4147 // scan the object if it's live, and give up the region. 4148 // Otherwise, let's iterate over the bitmap of the part of the region 4149 // that is left. 4150 // If the iteration is successful, give up the region. 4151 if (mr.is_empty()) { 4152 giveup_current_region(); 4153 regular_clock_call(); 4154 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) { 4155 if (_nextMarkBitMap->isMarked(mr.start())) { 4156 // The object is marked - apply the closure 4157 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start()); 4158 bitmap_closure.do_bit(offset); 4159 } 4160 // Even if this task aborted while scanning the humongous object 4161 // we can (and should) give up the current region. 4162 giveup_current_region(); 4163 regular_clock_call(); 4164 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) { 4165 giveup_current_region(); 4166 regular_clock_call(); 4167 } else { 4168 assert(has_aborted(), "currently the only way to do so"); 4169 // The only way to abort the bitmap iteration is to return 4170 // false from the do_bit() method. However, inside the 4171 // do_bit() method we move the _finger to point to the 4172 // object currently being looked at. So, if we bail out, we 4173 // have definitely set _finger to something non-null. 4174 assert(_finger != NULL, "invariant"); 4175 4176 // Region iteration was actually aborted. So now _finger 4177 // points to the address of the object we last scanned. If we 4178 // leave it there, when we restart this task, we will rescan 4179 // the object. It is easy to avoid this. We move the finger by 4180 // enough to point to the next possible object header (the 4181 // bitmap knows by how much we need to move it as it knows its 4182 // granularity). 4183 assert(_finger < _region_limit, "invariant"); 4184 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger); 4185 // Check if bitmap iteration was aborted while scanning the last object 4186 if (new_finger >= _region_limit) { 4187 giveup_current_region(); 4188 } else { 4189 move_finger_to(new_finger); 4190 } 4191 } 4192 } 4193 // At this point we have either completed iterating over the 4194 // region we were holding on to, or we have aborted. 4195 4196 // We then partially drain the local queue and the global stack. 4197 // (Do we really need this?) 4198 drain_local_queue(true); 4199 drain_global_stack(true); 4200 4201 // Read the note on the claim_region() method on why it might 4202 // return NULL with potentially more regions available for 4203 // claiming and why we have to check out_of_regions() to determine 4204 // whether we're done or not. 4205 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 4206 // We are going to try to claim a new region. We should have 4207 // given up on the previous one. 4208 // Separated the asserts so that we know which one fires. 4209 assert(_curr_region == NULL, "invariant"); 4210 assert(_finger == NULL, "invariant"); 4211 assert(_region_limit == NULL, "invariant"); 4212 if (_cm->verbose_low()) { 4213 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id); 4214 } 4215 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 4216 if (claimed_region != NULL) { 4217 // Yes, we managed to claim one 4218 statsOnly( ++_regions_claimed ); 4219 4220 if (_cm->verbose_low()) { 4221 gclog_or_tty->print_cr("[%u] we successfully claimed " 4222 "region "PTR_FORMAT, 4223 _worker_id, claimed_region); 4224 } 4225 4226 setup_for_region(claimed_region); 4227 assert(_curr_region == claimed_region, "invariant"); 4228 } 4229 // It is important to call the regular clock here. It might take 4230 // a while to claim a region if, for example, we hit a large 4231 // block of empty regions. So we need to call the regular clock 4232 // method once round the loop to make sure it's called 4233 // frequently enough. 4234 regular_clock_call(); 4235 } 4236 4237 if (!has_aborted() && _curr_region == NULL) { 4238 assert(_cm->out_of_regions(), 4239 "at this point we should be out of regions"); 4240 } 4241 } while ( _curr_region != NULL && !has_aborted()); 4242 4243 if (!has_aborted()) { 4244 // We cannot check whether the global stack is empty, since other 4245 // tasks might be pushing objects to it concurrently. 4246 assert(_cm->out_of_regions(), 4247 "at this point we should be out of regions"); 4248 4249 if (_cm->verbose_low()) { 4250 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id); 4251 } 4252 4253 // Try to reduce the number of available SATB buffers so that 4254 // remark has less work to do. 4255 drain_satb_buffers(); 4256 } 4257 4258 // Since we've done everything else, we can now totally drain the 4259 // local queue and global stack. 4260 drain_local_queue(false); 4261 drain_global_stack(false); 4262 4263 // Attempt at work stealing from other task's queues. 4264 if (do_stealing && !has_aborted()) { 4265 // We have not aborted. This means that we have finished all that 4266 // we could. Let's try to do some stealing... 4267 4268 // We cannot check whether the global stack is empty, since other 4269 // tasks might be pushing objects to it concurrently. 4270 assert(_cm->out_of_regions() && _task_queue->size() == 0, 4271 "only way to reach here"); 4272 4273 if (_cm->verbose_low()) { 4274 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id); 4275 } 4276 4277 while (!has_aborted()) { 4278 oop obj; 4279 statsOnly( ++_steal_attempts ); 4280 4281 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) { 4282 if (_cm->verbose_medium()) { 4283 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully", 4284 _worker_id, (void*) obj); 4285 } 4286 4287 statsOnly( ++_steals ); 4288 4289 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), 4290 "any stolen object should be marked"); 4291 scan_object(obj); 4292 4293 // And since we're towards the end, let's totally drain the 4294 // local queue and global stack. 4295 drain_local_queue(false); 4296 drain_global_stack(false); 4297 } else { 4298 break; 4299 } 4300 } 4301 } 4302 4303 // If we are about to wrap up and go into termination, check if we 4304 // should raise the overflow flag. 4305 if (do_termination && !has_aborted()) { 4306 if (_cm->force_overflow()->should_force()) { 4307 _cm->set_has_overflown(); 4308 regular_clock_call(); 4309 } 4310 } 4311 4312 // We still haven't aborted. Now, let's try to get into the 4313 // termination protocol. 4314 if (do_termination && !has_aborted()) { 4315 // We cannot check whether the global stack is empty, since other 4316 // tasks might be concurrently pushing objects on it. 4317 // Separated the asserts so that we know which one fires. 4318 assert(_cm->out_of_regions(), "only way to reach here"); 4319 assert(_task_queue->size() == 0, "only way to reach here"); 4320 4321 if (_cm->verbose_low()) { 4322 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id); 4323 } 4324 4325 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 4326 4327 // The CMTask class also extends the TerminatorTerminator class, 4328 // hence its should_exit_termination() method will also decide 4329 // whether to exit the termination protocol or not. 4330 bool finished = (is_serial || 4331 _cm->terminator()->offer_termination(this)); 4332 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 4333 _termination_time_ms += 4334 termination_end_time_ms - _termination_start_time_ms; 4335 4336 if (finished) { 4337 // We're all done. 4338 4339 if (_worker_id == 0) { 4340 // let's allow task 0 to do this 4341 if (concurrent()) { 4342 assert(_cm->concurrent_marking_in_progress(), "invariant"); 4343 // we need to set this to false before the next 4344 // safepoint. This way we ensure that the marking phase 4345 // doesn't observe any more heap expansions. 4346 _cm->clear_concurrent_marking_in_progress(); 4347 } 4348 } 4349 4350 // We can now guarantee that the global stack is empty, since 4351 // all other tasks have finished. We separated the guarantees so 4352 // that, if a condition is false, we can immediately find out 4353 // which one. 4354 guarantee(_cm->out_of_regions(), "only way to reach here"); 4355 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 4356 guarantee(_task_queue->size() == 0, "only way to reach here"); 4357 guarantee(!_cm->has_overflown(), "only way to reach here"); 4358 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); 4359 4360 if (_cm->verbose_low()) { 4361 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id); 4362 } 4363 } else { 4364 // Apparently there's more work to do. Let's abort this task. It 4365 // will restart it and we can hopefully find more things to do. 4366 4367 if (_cm->verbose_low()) { 4368 gclog_or_tty->print_cr("[%u] apparently there is more work to do", 4369 _worker_id); 4370 } 4371 4372 set_has_aborted(); 4373 statsOnly( ++_aborted_termination ); 4374 } 4375 } 4376 4377 // Mainly for debugging purposes to make sure that a pointer to the 4378 // closure which was statically allocated in this frame doesn't 4379 // escape it by accident. 4380 set_cm_oop_closure(NULL); 4381 double end_time_ms = os::elapsedVTime() * 1000.0; 4382 double elapsed_time_ms = end_time_ms - _start_time_ms; 4383 // Update the step history. 4384 _step_times_ms.add(elapsed_time_ms); 4385 4386 if (has_aborted()) { 4387 // The task was aborted for some reason. 4388 4389 statsOnly( ++_aborted ); 4390 4391 if (_has_timed_out) { 4392 double diff_ms = elapsed_time_ms - _time_target_ms; 4393 // Keep statistics of how well we did with respect to hitting 4394 // our target only if we actually timed out (if we aborted for 4395 // other reasons, then the results might get skewed). 4396 _marking_step_diffs_ms.add(diff_ms); 4397 } 4398 4399 if (_cm->has_overflown()) { 4400 // This is the interesting one. We aborted because a global 4401 // overflow was raised. This means we have to restart the 4402 // marking phase and start iterating over regions. However, in 4403 // order to do this we have to make sure that all tasks stop 4404 // what they are doing and re-initialise in a safe manner. We 4405 // will achieve this with the use of two barrier sync points. 4406 4407 if (_cm->verbose_low()) { 4408 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id); 4409 } 4410 4411 if (!is_serial) { 4412 // We only need to enter the sync barrier if being called 4413 // from a parallel context 4414 _cm->enter_first_sync_barrier(_worker_id); 4415 4416 // When we exit this sync barrier we know that all tasks have 4417 // stopped doing marking work. So, it's now safe to 4418 // re-initialise our data structures. At the end of this method, 4419 // task 0 will clear the global data structures. 4420 } 4421 4422 statsOnly( ++_aborted_overflow ); 4423 4424 // We clear the local state of this task... 4425 clear_region_fields(); 4426 4427 if (!is_serial) { 4428 // ...and enter the second barrier. 4429 _cm->enter_second_sync_barrier(_worker_id); 4430 } 4431 // At this point, if we're during the concurrent phase of 4432 // marking, everything has been re-initialized and we're 4433 // ready to restart. 4434 } 4435 4436 if (_cm->verbose_low()) { 4437 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, " 4438 "elapsed = %1.2lfms <<<<<<<<<<", 4439 _worker_id, _time_target_ms, elapsed_time_ms); 4440 if (_cm->has_aborted()) { 4441 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========", 4442 _worker_id); 4443 } 4444 } 4445 } else { 4446 if (_cm->verbose_low()) { 4447 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, " 4448 "elapsed = %1.2lfms <<<<<<<<<<", 4449 _worker_id, _time_target_ms, elapsed_time_ms); 4450 } 4451 } 4452 4453 _claimed = false; 4454 } 4455 4456 CMTask::CMTask(uint worker_id, 4457 ConcurrentMark* cm, 4458 size_t* marked_bytes, 4459 BitMap* card_bm, 4460 CMTaskQueue* task_queue, 4461 CMTaskQueueSet* task_queues) 4462 : _g1h(G1CollectedHeap::heap()), 4463 _worker_id(worker_id), _cm(cm), 4464 _claimed(false), 4465 _nextMarkBitMap(NULL), _hash_seed(17), 4466 _task_queue(task_queue), 4467 _task_queues(task_queues), 4468 _cm_oop_closure(NULL), 4469 _marked_bytes_array(marked_bytes), 4470 _card_bm(card_bm) { 4471 guarantee(task_queue != NULL, "invariant"); 4472 guarantee(task_queues != NULL, "invariant"); 4473 4474 statsOnly( _clock_due_to_scanning = 0; 4475 _clock_due_to_marking = 0 ); 4476 4477 _marking_step_diffs_ms.add(0.5); 4478 } 4479 4480 // These are formatting macros that are used below to ensure 4481 // consistent formatting. The *_H_* versions are used to format the 4482 // header for a particular value and they should be kept consistent 4483 // with the corresponding macro. Also note that most of the macros add 4484 // the necessary white space (as a prefix) which makes them a bit 4485 // easier to compose. 4486 4487 // All the output lines are prefixed with this string to be able to 4488 // identify them easily in a large log file. 4489 #define G1PPRL_LINE_PREFIX "###" 4490 4491 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT 4492 #ifdef _LP64 4493 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 4494 #else // _LP64 4495 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 4496 #endif // _LP64 4497 4498 // For per-region info 4499 #define G1PPRL_TYPE_FORMAT " %-4s" 4500 #define G1PPRL_TYPE_H_FORMAT " %4s" 4501 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9) 4502 #define G1PPRL_BYTE_H_FORMAT " %9s" 4503 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 4504 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 4505 4506 // For summary info 4507 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT 4508 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT 4509 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB" 4510 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%" 4511 4512 G1PrintRegionLivenessInfoClosure:: 4513 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name) 4514 : _out(out), 4515 _total_used_bytes(0), _total_capacity_bytes(0), 4516 _total_prev_live_bytes(0), _total_next_live_bytes(0), 4517 _hum_used_bytes(0), _hum_capacity_bytes(0), 4518 _hum_prev_live_bytes(0), _hum_next_live_bytes(0), 4519 _total_remset_bytes(0), _total_strong_code_roots_bytes(0) { 4520 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 4521 MemRegion g1_committed = g1h->g1_committed(); 4522 MemRegion g1_reserved = g1h->g1_reserved(); 4523 double now = os::elapsedTime(); 4524 4525 // Print the header of the output. 4526 _out->cr(); 4527 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 4528 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP" 4529 G1PPRL_SUM_ADDR_FORMAT("committed") 4530 G1PPRL_SUM_ADDR_FORMAT("reserved") 4531 G1PPRL_SUM_BYTE_FORMAT("region-size"), 4532 g1_committed.start(), g1_committed.end(), 4533 g1_reserved.start(), g1_reserved.end(), 4534 HeapRegion::GrainBytes); 4535 _out->print_cr(G1PPRL_LINE_PREFIX); 4536 _out->print_cr(G1PPRL_LINE_PREFIX 4537 G1PPRL_TYPE_H_FORMAT 4538 G1PPRL_ADDR_BASE_H_FORMAT 4539 G1PPRL_BYTE_H_FORMAT 4540 G1PPRL_BYTE_H_FORMAT 4541 G1PPRL_BYTE_H_FORMAT 4542 G1PPRL_DOUBLE_H_FORMAT 4543 G1PPRL_BYTE_H_FORMAT 4544 G1PPRL_BYTE_H_FORMAT, 4545 "type", "address-range", 4546 "used", "prev-live", "next-live", "gc-eff", 4547 "remset", "code-roots"); 4548 _out->print_cr(G1PPRL_LINE_PREFIX 4549 G1PPRL_TYPE_H_FORMAT 4550 G1PPRL_ADDR_BASE_H_FORMAT 4551 G1PPRL_BYTE_H_FORMAT 4552 G1PPRL_BYTE_H_FORMAT 4553 G1PPRL_BYTE_H_FORMAT 4554 G1PPRL_DOUBLE_H_FORMAT 4555 G1PPRL_BYTE_H_FORMAT 4556 G1PPRL_BYTE_H_FORMAT, 4557 "", "", 4558 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)", 4559 "(bytes)", "(bytes)"); 4560 } 4561 4562 // It takes as a parameter a reference to one of the _hum_* fields, it 4563 // deduces the corresponding value for a region in a humongous region 4564 // series (either the region size, or what's left if the _hum_* field 4565 // is < the region size), and updates the _hum_* field accordingly. 4566 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) { 4567 size_t bytes = 0; 4568 // The > 0 check is to deal with the prev and next live bytes which 4569 // could be 0. 4570 if (*hum_bytes > 0) { 4571 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes); 4572 *hum_bytes -= bytes; 4573 } 4574 return bytes; 4575 } 4576 4577 // It deduces the values for a region in a humongous region series 4578 // from the _hum_* fields and updates those accordingly. It assumes 4579 // that that _hum_* fields have already been set up from the "starts 4580 // humongous" region and we visit the regions in address order. 4581 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes, 4582 size_t* capacity_bytes, 4583 size_t* prev_live_bytes, 4584 size_t* next_live_bytes) { 4585 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition"); 4586 *used_bytes = get_hum_bytes(&_hum_used_bytes); 4587 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes); 4588 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes); 4589 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes); 4590 } 4591 4592 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) { 4593 const char* type = ""; 4594 HeapWord* bottom = r->bottom(); 4595 HeapWord* end = r->end(); 4596 size_t capacity_bytes = r->capacity(); 4597 size_t used_bytes = r->used(); 4598 size_t prev_live_bytes = r->live_bytes(); 4599 size_t next_live_bytes = r->next_live_bytes(); 4600 double gc_eff = r->gc_efficiency(); 4601 size_t remset_bytes = r->rem_set()->mem_size(); 4602 size_t strong_code_roots_bytes = r->rem_set()->strong_code_roots_mem_size(); 4603 4604 if (r->used() == 0) { 4605 type = "FREE"; 4606 } else if (r->is_survivor()) { 4607 type = "SURV"; 4608 } else if (r->is_young()) { 4609 type = "EDEN"; 4610 } else if (r->startsHumongous()) { 4611 type = "HUMS"; 4612 4613 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 && 4614 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0, 4615 "they should have been zeroed after the last time we used them"); 4616 // Set up the _hum_* fields. 4617 _hum_capacity_bytes = capacity_bytes; 4618 _hum_used_bytes = used_bytes; 4619 _hum_prev_live_bytes = prev_live_bytes; 4620 _hum_next_live_bytes = next_live_bytes; 4621 get_hum_bytes(&used_bytes, &capacity_bytes, 4622 &prev_live_bytes, &next_live_bytes); 4623 end = bottom + HeapRegion::GrainWords; 4624 } else if (r->continuesHumongous()) { 4625 type = "HUMC"; 4626 get_hum_bytes(&used_bytes, &capacity_bytes, 4627 &prev_live_bytes, &next_live_bytes); 4628 assert(end == bottom + HeapRegion::GrainWords, "invariant"); 4629 } else { 4630 type = "OLD"; 4631 } 4632 4633 _total_used_bytes += used_bytes; 4634 _total_capacity_bytes += capacity_bytes; 4635 _total_prev_live_bytes += prev_live_bytes; 4636 _total_next_live_bytes += next_live_bytes; 4637 _total_remset_bytes += remset_bytes; 4638 _total_strong_code_roots_bytes += strong_code_roots_bytes; 4639 4640 // Print a line for this particular region. 4641 _out->print_cr(G1PPRL_LINE_PREFIX 4642 G1PPRL_TYPE_FORMAT 4643 G1PPRL_ADDR_BASE_FORMAT 4644 G1PPRL_BYTE_FORMAT 4645 G1PPRL_BYTE_FORMAT 4646 G1PPRL_BYTE_FORMAT 4647 G1PPRL_DOUBLE_FORMAT 4648 G1PPRL_BYTE_FORMAT 4649 G1PPRL_BYTE_FORMAT, 4650 type, bottom, end, 4651 used_bytes, prev_live_bytes, next_live_bytes, gc_eff, 4652 remset_bytes, strong_code_roots_bytes); 4653 4654 return false; 4655 } 4656 4657 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 4658 // add static memory usages to remembered set sizes 4659 _total_remset_bytes += HeapRegionRemSet::fl_mem_size() + HeapRegionRemSet::static_mem_size(); 4660 // Print the footer of the output. 4661 _out->print_cr(G1PPRL_LINE_PREFIX); 4662 _out->print_cr(G1PPRL_LINE_PREFIX 4663 " SUMMARY" 4664 G1PPRL_SUM_MB_FORMAT("capacity") 4665 G1PPRL_SUM_MB_PERC_FORMAT("used") 4666 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 4667 G1PPRL_SUM_MB_PERC_FORMAT("next-live") 4668 G1PPRL_SUM_MB_FORMAT("remset") 4669 G1PPRL_SUM_MB_FORMAT("code-roots"), 4670 bytes_to_mb(_total_capacity_bytes), 4671 bytes_to_mb(_total_used_bytes), 4672 perc(_total_used_bytes, _total_capacity_bytes), 4673 bytes_to_mb(_total_prev_live_bytes), 4674 perc(_total_prev_live_bytes, _total_capacity_bytes), 4675 bytes_to_mb(_total_next_live_bytes), 4676 perc(_total_next_live_bytes, _total_capacity_bytes), 4677 bytes_to_mb(_total_remset_bytes), 4678 bytes_to_mb(_total_strong_code_roots_bytes)); 4679 _out->cr(); 4680 }