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 bool CMBitMap::allocate(ReservedSpace heap_rs) { 105 _bmStartWord = (HeapWord*)(heap_rs.base()); 106 _bmWordSize = heap_rs.size()/HeapWordSize; // heap_rs.size() is in bytes 107 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 108 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 109 if (!brs.is_reserved()) { 110 warning("ConcurrentMark marking bit map allocation failure"); 111 return false; 112 } 113 MemTracker::record_virtual_memory_type((address)brs.base(), mtGC); 114 // For now we'll just commit all of the bit map up front. 115 // Later on we'll try to be more parsimonious with swap. 116 if (!_virtual_space.initialize(brs, brs.size())) { 117 warning("ConcurrentMark marking bit map backing store failure"); 118 return false; 119 } 120 assert(_virtual_space.committed_size() == brs.size(), 121 "didn't reserve backing store for all of concurrent marking bit map?"); 122 _bm.set_map((uintptr_t*)_virtual_space.low()); 123 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 124 _bmWordSize, "inconsistency in bit map sizing"); 125 _bm.set_size(_bmWordSize >> _shifter); 126 return true; 127 } 128 129 void CMBitMap::clearAll() { 130 _bm.clear(); 131 return; 132 } 133 134 void CMBitMap::markRange(MemRegion mr) { 135 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 136 assert(!mr.is_empty(), "unexpected empty region"); 137 assert((offsetToHeapWord(heapWordToOffset(mr.end())) == 138 ((HeapWord *) mr.end())), 139 "markRange memory region end is not card aligned"); 140 // convert address range into offset range 141 _bm.at_put_range(heapWordToOffset(mr.start()), 142 heapWordToOffset(mr.end()), true); 143 } 144 145 void CMBitMap::clearRange(MemRegion mr) { 146 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 147 assert(!mr.is_empty(), "unexpected empty region"); 148 // convert address range into offset range 149 _bm.at_put_range(heapWordToOffset(mr.start()), 150 heapWordToOffset(mr.end()), false); 151 } 152 153 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr, 154 HeapWord* end_addr) { 155 HeapWord* start = getNextMarkedWordAddress(addr); 156 start = MIN2(start, end_addr); 157 HeapWord* end = getNextUnmarkedWordAddress(start); 158 end = MIN2(end, end_addr); 159 assert(start <= end, "Consistency check"); 160 MemRegion mr(start, end); 161 if (!mr.is_empty()) { 162 clearRange(mr); 163 } 164 return mr; 165 } 166 167 CMMarkStack::CMMarkStack(ConcurrentMark* cm) : 168 _base(NULL), _cm(cm) 169 #ifdef ASSERT 170 , _drain_in_progress(false) 171 , _drain_in_progress_yields(false) 172 #endif 173 {} 174 175 bool CMMarkStack::allocate(size_t capacity) { 176 // allocate a stack of the requisite depth 177 ReservedSpace rs(ReservedSpace::allocation_align_size_up(capacity * sizeof(oop))); 178 if (!rs.is_reserved()) { 179 warning("ConcurrentMark MarkStack allocation failure"); 180 return false; 181 } 182 MemTracker::record_virtual_memory_type((address)rs.base(), mtGC); 183 if (!_virtual_space.initialize(rs, rs.size())) { 184 warning("ConcurrentMark MarkStack backing store failure"); 185 // Release the virtual memory reserved for the marking stack 186 rs.release(); 187 return false; 188 } 189 assert(_virtual_space.committed_size() == rs.size(), 190 "Didn't reserve backing store for all of ConcurrentMark stack?"); 191 _base = (oop*) _virtual_space.low(); 192 setEmpty(); 193 _capacity = (jint) capacity; 194 _saved_index = -1; 195 _should_expand = false; 196 NOT_PRODUCT(_max_depth = 0); 197 return true; 198 } 199 200 void CMMarkStack::expand() { 201 // Called, during remark, if we've overflown the marking stack during marking. 202 assert(isEmpty(), "stack should been emptied while handling overflow"); 203 assert(_capacity <= (jint) MarkStackSizeMax, "stack bigger than permitted"); 204 // Clear expansion flag 205 _should_expand = false; 206 if (_capacity == (jint) MarkStackSizeMax) { 207 if (PrintGCDetails && Verbose) { 208 gclog_or_tty->print_cr(" (benign) Can't expand marking stack capacity, at max size limit"); 209 } 210 return; 211 } 212 // Double capacity if possible 213 jint new_capacity = MIN2(_capacity*2, (jint) MarkStackSizeMax); 214 // Do not give up existing stack until we have managed to 215 // get the double capacity that we desired. 216 ReservedSpace rs(ReservedSpace::allocation_align_size_up(new_capacity * 217 sizeof(oop))); 218 if (rs.is_reserved()) { 219 // Release the backing store associated with old stack 220 _virtual_space.release(); 221 // Reinitialize virtual space for new stack 222 if (!_virtual_space.initialize(rs, rs.size())) { 223 fatal("Not enough swap for expanded marking stack capacity"); 224 } 225 _base = (oop*)(_virtual_space.low()); 226 _index = 0; 227 _capacity = new_capacity; 228 } else { 229 if (PrintGCDetails && Verbose) { 230 // Failed to double capacity, continue; 231 gclog_or_tty->print(" (benign) Failed to expand marking stack capacity from " 232 SIZE_FORMAT"K to " SIZE_FORMAT"K", 233 _capacity / K, new_capacity / K); 234 } 235 } 236 } 237 238 void CMMarkStack::set_should_expand() { 239 // If we're resetting the marking state because of an 240 // marking stack overflow, record that we should, if 241 // possible, expand the stack. 242 _should_expand = _cm->has_overflown(); 243 } 244 245 CMMarkStack::~CMMarkStack() { 246 if (_base != NULL) { 247 _base = NULL; 248 _virtual_space.release(); 249 } 250 } 251 252 void CMMarkStack::par_push(oop ptr) { 253 while (true) { 254 if (isFull()) { 255 _overflow = true; 256 return; 257 } 258 // Otherwise... 259 jint index = _index; 260 jint next_index = index+1; 261 jint res = Atomic::cmpxchg(next_index, &_index, index); 262 if (res == index) { 263 _base[index] = ptr; 264 // Note that we don't maintain this atomically. We could, but it 265 // doesn't seem necessary. 266 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 267 return; 268 } 269 // Otherwise, we need to try again. 270 } 271 } 272 273 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) { 274 while (true) { 275 if (isFull()) { 276 _overflow = true; 277 return; 278 } 279 // Otherwise... 280 jint index = _index; 281 jint next_index = index + n; 282 if (next_index > _capacity) { 283 _overflow = true; 284 return; 285 } 286 jint res = Atomic::cmpxchg(next_index, &_index, index); 287 if (res == index) { 288 for (int i = 0; i < n; i++) { 289 int ind = index + i; 290 assert(ind < _capacity, "By overflow test above."); 291 _base[ind] = ptr_arr[i]; 292 } 293 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 294 return; 295 } 296 // Otherwise, we need to try again. 297 } 298 } 299 300 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { 301 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 302 jint start = _index; 303 jint next_index = start + n; 304 if (next_index > _capacity) { 305 _overflow = true; 306 return; 307 } 308 // Otherwise. 309 _index = next_index; 310 for (int i = 0; i < n; i++) { 311 int ind = start + i; 312 assert(ind < _capacity, "By overflow test above."); 313 _base[ind] = ptr_arr[i]; 314 } 315 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); 316 } 317 318 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { 319 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 320 jint index = _index; 321 if (index == 0) { 322 *n = 0; 323 return false; 324 } else { 325 int k = MIN2(max, index); 326 jint new_ind = index - k; 327 for (int j = 0; j < k; j++) { 328 ptr_arr[j] = _base[new_ind + j]; 329 } 330 _index = new_ind; 331 *n = k; 332 return true; 333 } 334 } 335 336 template<class OopClosureClass> 337 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) { 338 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after 339 || SafepointSynchronize::is_at_safepoint(), 340 "Drain recursion must be yield-safe."); 341 bool res = true; 342 debug_only(_drain_in_progress = true); 343 debug_only(_drain_in_progress_yields = yield_after); 344 while (!isEmpty()) { 345 oop newOop = pop(); 346 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop"); 347 assert(newOop->is_oop(), "Expected an oop"); 348 assert(bm == NULL || bm->isMarked((HeapWord*)newOop), 349 "only grey objects on this stack"); 350 newOop->oop_iterate(cl); 351 if (yield_after && _cm->do_yield_check()) { 352 res = false; 353 break; 354 } 355 } 356 debug_only(_drain_in_progress = false); 357 return res; 358 } 359 360 void CMMarkStack::note_start_of_gc() { 361 assert(_saved_index == -1, 362 "note_start_of_gc()/end_of_gc() bracketed incorrectly"); 363 _saved_index = _index; 364 } 365 366 void CMMarkStack::note_end_of_gc() { 367 // This is intentionally a guarantee, instead of an assert. If we 368 // accidentally add something to the mark stack during GC, it 369 // will be a correctness issue so it's better if we crash. we'll 370 // only check this once per GC anyway, so it won't be a performance 371 // issue in any way. 372 guarantee(_saved_index == _index, 373 err_msg("saved index: %d index: %d", _saved_index, _index)); 374 _saved_index = -1; 375 } 376 377 void CMMarkStack::oops_do(OopClosure* f) { 378 assert(_saved_index == _index, 379 err_msg("saved index: %d index: %d", _saved_index, _index)); 380 for (int i = 0; i < _index; i += 1) { 381 f->do_oop(&_base[i]); 382 } 383 } 384 385 bool ConcurrentMark::not_yet_marked(oop obj) const { 386 return _g1h->is_obj_ill(obj); 387 } 388 389 CMRootRegions::CMRootRegions() : 390 _young_list(NULL), _cm(NULL), _scan_in_progress(false), 391 _should_abort(false), _next_survivor(NULL) { } 392 393 void CMRootRegions::init(G1CollectedHeap* g1h, ConcurrentMark* cm) { 394 _young_list = g1h->young_list(); 395 _cm = cm; 396 } 397 398 void CMRootRegions::prepare_for_scan() { 399 assert(!scan_in_progress(), "pre-condition"); 400 401 // Currently, only survivors can be root regions. 402 assert(_next_survivor == NULL, "pre-condition"); 403 _next_survivor = _young_list->first_survivor_region(); 404 _scan_in_progress = (_next_survivor != NULL); 405 _should_abort = false; 406 } 407 408 HeapRegion* CMRootRegions::claim_next() { 409 if (_should_abort) { 410 // If someone has set the should_abort flag, we return NULL to 411 // force the caller to bail out of their loop. 412 return NULL; 413 } 414 415 // Currently, only survivors can be root regions. 416 HeapRegion* res = _next_survivor; 417 if (res != NULL) { 418 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 419 // Read it again in case it changed while we were waiting for the lock. 420 res = _next_survivor; 421 if (res != NULL) { 422 if (res == _young_list->last_survivor_region()) { 423 // We just claimed the last survivor so store NULL to indicate 424 // that we're done. 425 _next_survivor = NULL; 426 } else { 427 _next_survivor = res->get_next_young_region(); 428 } 429 } else { 430 // Someone else claimed the last survivor while we were trying 431 // to take the lock so nothing else to do. 432 } 433 } 434 assert(res == NULL || res->is_survivor(), "post-condition"); 435 436 return res; 437 } 438 439 void CMRootRegions::scan_finished() { 440 assert(scan_in_progress(), "pre-condition"); 441 442 // Currently, only survivors can be root regions. 443 if (!_should_abort) { 444 assert(_next_survivor == NULL, "we should have claimed all survivors"); 445 } 446 _next_survivor = NULL; 447 448 { 449 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 450 _scan_in_progress = false; 451 RootRegionScan_lock->notify_all(); 452 } 453 } 454 455 bool CMRootRegions::wait_until_scan_finished() { 456 if (!scan_in_progress()) return false; 457 458 { 459 MutexLockerEx x(RootRegionScan_lock, Mutex::_no_safepoint_check_flag); 460 while (scan_in_progress()) { 461 RootRegionScan_lock->wait(Mutex::_no_safepoint_check_flag); 462 } 463 } 464 return true; 465 } 466 467 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 468 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 469 #endif // _MSC_VER 470 471 uint ConcurrentMark::scale_parallel_threads(uint n_par_threads) { 472 return MAX2((n_par_threads + 2) / 4, 1U); 473 } 474 475 ConcurrentMark::ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs) : 476 _g1h(g1h), 477 _markBitMap1(MinObjAlignment - 1), 478 _markBitMap2(MinObjAlignment - 1), 479 480 _parallel_marking_threads(0), 481 _max_parallel_marking_threads(0), 482 _sleep_factor(0.0), 483 _marking_task_overhead(1.0), 484 _cleanup_sleep_factor(0.0), 485 _cleanup_task_overhead(1.0), 486 _cleanup_list("Cleanup List"), 487 _region_bm((BitMap::idx_t)(g1h->max_regions()), false /* in_resource_area*/), 488 _card_bm((heap_rs.size() + CardTableModRefBS::card_size - 1) >> 489 CardTableModRefBS::card_shift, 490 false /* in_resource_area*/), 491 492 _prevMarkBitMap(&_markBitMap1), 493 _nextMarkBitMap(&_markBitMap2), 494 495 _markStack(this), 496 // _finger set in set_non_marking_state 497 498 _max_worker_id(MAX2((uint)ParallelGCThreads, 1U)), 499 // _active_tasks set in set_non_marking_state 500 // _tasks set inside the constructor 501 _task_queues(new CMTaskQueueSet((int) _max_worker_id)), 502 _terminator(ParallelTaskTerminator((int) _max_worker_id, _task_queues)), 503 504 _has_overflown(false), 505 _concurrent(false), 506 _has_aborted(false), 507 _restart_for_overflow(false), 508 _concurrent_marking_in_progress(false), 509 510 // _verbose_level set below 511 512 _init_times(), 513 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), 514 _cleanup_times(), 515 _total_counting_time(0.0), 516 _total_rs_scrub_time(0.0), 517 518 _parallel_workers(NULL), 519 520 _count_card_bitmaps(NULL), 521 _count_marked_bytes(NULL), 522 _completed_initialization(false) { 523 CMVerboseLevel verbose_level = (CMVerboseLevel) G1MarkingVerboseLevel; 524 if (verbose_level < no_verbose) { 525 verbose_level = no_verbose; 526 } 527 if (verbose_level > high_verbose) { 528 verbose_level = high_verbose; 529 } 530 _verbose_level = verbose_level; 531 532 if (verbose_low()) { 533 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", " 534 "heap end = "PTR_FORMAT, _heap_start, _heap_end); 535 } 536 537 if (!_markBitMap1.allocate(heap_rs)) { 538 warning("Failed to allocate first CM bit map"); 539 return; 540 } 541 if (!_markBitMap2.allocate(heap_rs)) { 542 warning("Failed to allocate second CM bit map"); 543 return; 544 } 545 546 // Create & start a ConcurrentMark thread. 547 _cmThread = new ConcurrentMarkThread(this); 548 assert(cmThread() != NULL, "CM Thread should have been created"); 549 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); 550 551 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 552 assert(_markBitMap1.covers(heap_rs), "_markBitMap1 inconsistency"); 553 assert(_markBitMap2.covers(heap_rs), "_markBitMap2 inconsistency"); 554 555 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 556 satb_qs.set_buffer_size(G1SATBBufferSize); 557 558 _root_regions.init(_g1h, this); 559 560 if (ConcGCThreads > ParallelGCThreads) { 561 warning("Can't have more ConcGCThreads (" UINT32_FORMAT ") " 562 "than ParallelGCThreads (" UINT32_FORMAT ").", 563 ConcGCThreads, ParallelGCThreads); 564 return; 565 } 566 if (ParallelGCThreads == 0) { 567 // if we are not running with any parallel GC threads we will not 568 // spawn any marking threads either 569 _parallel_marking_threads = 0; 570 _max_parallel_marking_threads = 0; 571 _sleep_factor = 0.0; 572 _marking_task_overhead = 1.0; 573 } else { 574 if (!FLAG_IS_DEFAULT(ConcGCThreads) && ConcGCThreads > 0) { 575 // Note: ConcGCThreads has precedence over G1MarkingOverheadPercent 576 // if both are set 577 _sleep_factor = 0.0; 578 _marking_task_overhead = 1.0; 579 } else if (G1MarkingOverheadPercent > 0) { 580 // We will calculate the number of parallel marking threads based 581 // on a target overhead with respect to the soft real-time goal 582 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; 583 double overall_cm_overhead = 584 (double) MaxGCPauseMillis * marking_overhead / 585 (double) GCPauseIntervalMillis; 586 double cpu_ratio = 1.0 / (double) os::processor_count(); 587 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); 588 double marking_task_overhead = 589 overall_cm_overhead / marking_thread_num * 590 (double) os::processor_count(); 591 double sleep_factor = 592 (1.0 - marking_task_overhead) / marking_task_overhead; 593 594 FLAG_SET_ERGO(uintx, ConcGCThreads, (uint) marking_thread_num); 595 _sleep_factor = sleep_factor; 596 _marking_task_overhead = marking_task_overhead; 597 } else { 598 // Calculate the number of parallel marking threads by scaling 599 // the number of parallel GC threads. 600 uint marking_thread_num = scale_parallel_threads((uint) ParallelGCThreads); 601 FLAG_SET_ERGO(uintx, ConcGCThreads, marking_thread_num); 602 _sleep_factor = 0.0; 603 _marking_task_overhead = 1.0; 604 } 605 606 assert(ConcGCThreads > 0, "Should have been set"); 607 _parallel_marking_threads = (uint) ConcGCThreads; 608 _max_parallel_marking_threads = _parallel_marking_threads; 609 610 if (parallel_marking_threads() > 1) { 611 _cleanup_task_overhead = 1.0; 612 } else { 613 _cleanup_task_overhead = marking_task_overhead(); 614 } 615 _cleanup_sleep_factor = 616 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead(); 617 618 #if 0 619 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads()); 620 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead()); 621 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor()); 622 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead()); 623 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor()); 624 #endif 625 626 guarantee(parallel_marking_threads() > 0, "peace of mind"); 627 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads", 628 _max_parallel_marking_threads, false, true); 629 if (_parallel_workers == NULL) { 630 vm_exit_during_initialization("Failed necessary allocation."); 631 } else { 632 _parallel_workers->initialize_workers(); 633 } 634 } 635 636 if (FLAG_IS_DEFAULT(MarkStackSize)) { 637 uintx mark_stack_size = 638 MIN2(MarkStackSizeMax, 639 MAX2(MarkStackSize, (uintx) (parallel_marking_threads() * TASKQUEUE_SIZE))); 640 // Verify that the calculated value for MarkStackSize is in range. 641 // It would be nice to use the private utility routine from Arguments. 642 if (!(mark_stack_size >= 1 && mark_stack_size <= MarkStackSizeMax)) { 643 warning("Invalid value calculated for MarkStackSize (" UINTX_FORMAT "): " 644 "must be between " UINTX_FORMAT " and " UINTX_FORMAT, 645 mark_stack_size, 1, MarkStackSizeMax); 646 return; 647 } 648 FLAG_SET_ERGO(uintx, MarkStackSize, mark_stack_size); 649 } else { 650 // Verify MarkStackSize is in range. 651 if (FLAG_IS_CMDLINE(MarkStackSize)) { 652 if (FLAG_IS_DEFAULT(MarkStackSizeMax)) { 653 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 654 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT "): " 655 "must be between " UINTX_FORMAT " and " UINTX_FORMAT, 656 MarkStackSize, 1, MarkStackSizeMax); 657 return; 658 } 659 } else if (FLAG_IS_CMDLINE(MarkStackSizeMax)) { 660 if (!(MarkStackSize >= 1 && MarkStackSize <= MarkStackSizeMax)) { 661 warning("Invalid value specified for MarkStackSize (" UINTX_FORMAT ")" 662 " or for MarkStackSizeMax (" UINTX_FORMAT ")", 663 MarkStackSize, MarkStackSizeMax); 664 return; 665 } 666 } 667 } 668 } 669 670 if (!_markStack.allocate(MarkStackSize)) { 671 warning("Failed to allocate CM marking stack"); 672 return; 673 } 674 675 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_worker_id, mtGC); 676 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_worker_id, mtGC); 677 678 _count_card_bitmaps = NEW_C_HEAP_ARRAY(BitMap, _max_worker_id, mtGC); 679 _count_marked_bytes = NEW_C_HEAP_ARRAY(size_t*, _max_worker_id, mtGC); 680 681 BitMap::idx_t card_bm_size = _card_bm.size(); 682 683 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail 684 _active_tasks = _max_worker_id; 685 686 size_t max_regions = (size_t) _g1h->max_regions(); 687 for (uint i = 0; i < _max_worker_id; ++i) { 688 CMTaskQueue* task_queue = new CMTaskQueue(); 689 task_queue->initialize(); 690 _task_queues->register_queue(i, task_queue); 691 692 _count_card_bitmaps[i] = BitMap(card_bm_size, false); 693 _count_marked_bytes[i] = NEW_C_HEAP_ARRAY(size_t, max_regions, mtGC); 694 695 _tasks[i] = new CMTask(i, this, 696 _count_marked_bytes[i], 697 &_count_card_bitmaps[i], 698 task_queue, _task_queues); 699 700 _accum_task_vtime[i] = 0.0; 701 } 702 703 // Calculate the card number for the bottom of the heap. Used 704 // in biasing indexes into the accounting card bitmaps. 705 _heap_bottom_card_num = 706 intptr_t(uintptr_t(_g1h->reserved_region().start()) >> 707 CardTableModRefBS::card_shift); 708 709 // Clear all the liveness counting data 710 clear_all_count_data(); 711 712 // so that the call below can read a sensible value 713 _heap_start = (HeapWord*) heap_rs.base(); 714 set_non_marking_state(); 715 _completed_initialization = true; 716 } 717 718 void ConcurrentMark::update_g1_committed(bool force) { 719 // If concurrent marking is not in progress, then we do not need to 720 // update _heap_end. 721 if (!concurrent_marking_in_progress() && !force) return; 722 723 MemRegion committed = _g1h->g1_committed(); 724 assert(committed.start() == _heap_start, "start shouldn't change"); 725 HeapWord* new_end = committed.end(); 726 if (new_end > _heap_end) { 727 // The heap has been expanded. 728 729 _heap_end = new_end; 730 } 731 // Notice that the heap can also shrink. However, this only happens 732 // during a Full GC (at least currently) and the entire marking 733 // phase will bail out and the task will not be restarted. So, let's 734 // do nothing. 735 } 736 737 void ConcurrentMark::reset() { 738 // Starting values for these two. This should be called in a STW 739 // phase. CM will be notified of any future g1_committed expansions 740 // will be at the end of evacuation pauses, when tasks are 741 // inactive. 742 MemRegion committed = _g1h->g1_committed(); 743 _heap_start = committed.start(); 744 _heap_end = committed.end(); 745 746 // Separated the asserts so that we know which one fires. 747 assert(_heap_start != NULL, "heap bounds should look ok"); 748 assert(_heap_end != NULL, "heap bounds should look ok"); 749 assert(_heap_start < _heap_end, "heap bounds should look ok"); 750 751 // Reset all the marking data structures and any necessary flags 752 reset_marking_state(); 753 754 if (verbose_low()) { 755 gclog_or_tty->print_cr("[global] resetting"); 756 } 757 758 // We do reset all of them, since different phases will use 759 // different number of active threads. So, it's easiest to have all 760 // of them ready. 761 for (uint i = 0; i < _max_worker_id; ++i) { 762 _tasks[i]->reset(_nextMarkBitMap); 763 } 764 765 // we need this to make sure that the flag is on during the evac 766 // pause with initial mark piggy-backed 767 set_concurrent_marking_in_progress(); 768 } 769 770 771 void ConcurrentMark::reset_marking_state(bool clear_overflow) { 772 _markStack.set_should_expand(); 773 _markStack.setEmpty(); // Also clears the _markStack overflow flag 774 if (clear_overflow) { 775 clear_has_overflown(); 776 } else { 777 assert(has_overflown(), "pre-condition"); 778 } 779 _finger = _heap_start; 780 781 for (uint i = 0; i < _max_worker_id; ++i) { 782 CMTaskQueue* queue = _task_queues->queue(i); 783 queue->set_empty(); 784 } 785 } 786 787 void ConcurrentMark::set_phase(uint active_tasks, bool concurrent) { 788 assert(active_tasks <= _max_worker_id, "we should not have more"); 789 790 _active_tasks = active_tasks; 791 // Need to update the three data structures below according to the 792 // number of active threads for this phase. 793 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); 794 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); 795 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); 796 797 _concurrent = concurrent; 798 // We propagate this to all tasks, not just the active ones. 799 for (uint i = 0; i < _max_worker_id; ++i) 800 _tasks[i]->set_concurrent(concurrent); 801 802 if (concurrent) { 803 set_concurrent_marking_in_progress(); 804 } else { 805 // We currently assume that the concurrent flag has been set to 806 // false before we start remark. At this point we should also be 807 // in a STW phase. 808 assert(!concurrent_marking_in_progress(), "invariant"); 809 assert(_finger == _heap_end, "only way to get here"); 810 update_g1_committed(true); 811 } 812 } 813 814 void ConcurrentMark::set_non_marking_state() { 815 // We set the global marking state to some default values when we're 816 // not doing marking. 817 reset_marking_state(); 818 _active_tasks = 0; 819 clear_concurrent_marking_in_progress(); 820 } 821 822 ConcurrentMark::~ConcurrentMark() { 823 // The ConcurrentMark instance is never freed. 824 ShouldNotReachHere(); 825 } 826 827 void ConcurrentMark::clearNextBitmap() { 828 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 829 G1CollectorPolicy* g1p = g1h->g1_policy(); 830 831 // Make sure that the concurrent mark thread looks to still be in 832 // the current cycle. 833 guarantee(cmThread()->during_cycle(), "invariant"); 834 835 // We are finishing up the current cycle by clearing the next 836 // marking bitmap and getting it ready for the next cycle. During 837 // this time no other cycle can start. So, let's make sure that this 838 // is the case. 839 guarantee(!g1h->mark_in_progress(), "invariant"); 840 841 // clear the mark bitmap (no grey objects to start with). 842 // We need to do this in chunks and offer to yield in between 843 // each chunk. 844 HeapWord* start = _nextMarkBitMap->startWord(); 845 HeapWord* end = _nextMarkBitMap->endWord(); 846 HeapWord* cur = start; 847 size_t chunkSize = M; 848 while (cur < end) { 849 HeapWord* next = cur + chunkSize; 850 if (next > end) { 851 next = end; 852 } 853 MemRegion mr(cur,next); 854 _nextMarkBitMap->clearRange(mr); 855 cur = next; 856 do_yield_check(); 857 858 // Repeat the asserts from above. We'll do them as asserts here to 859 // minimize their overhead on the product. However, we'll have 860 // them as guarantees at the beginning / end of the bitmap 861 // clearing to get some checking in the product. 862 assert(cmThread()->during_cycle(), "invariant"); 863 assert(!g1h->mark_in_progress(), "invariant"); 864 } 865 866 // Clear the liveness counting data 867 clear_all_count_data(); 868 869 // Repeat the asserts from above. 870 guarantee(cmThread()->during_cycle(), "invariant"); 871 guarantee(!g1h->mark_in_progress(), "invariant"); 872 } 873 874 class NoteStartOfMarkHRClosure: public HeapRegionClosure { 875 public: 876 bool doHeapRegion(HeapRegion* r) { 877 if (!r->continuesHumongous()) { 878 r->note_start_of_marking(); 879 } 880 return false; 881 } 882 }; 883 884 void ConcurrentMark::checkpointRootsInitialPre() { 885 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 886 G1CollectorPolicy* g1p = g1h->g1_policy(); 887 888 _has_aborted = false; 889 890 #ifndef PRODUCT 891 if (G1PrintReachableAtInitialMark) { 892 print_reachable("at-cycle-start", 893 VerifyOption_G1UsePrevMarking, true /* all */); 894 } 895 #endif 896 897 // Initialise marking structures. This has to be done in a STW phase. 898 reset(); 899 900 // For each region note start of marking. 901 NoteStartOfMarkHRClosure startcl; 902 g1h->heap_region_iterate(&startcl); 903 } 904 905 906 void ConcurrentMark::checkpointRootsInitialPost() { 907 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 908 909 // If we force an overflow during remark, the remark operation will 910 // actually abort and we'll restart concurrent marking. If we always 911 // force an oveflow during remark we'll never actually complete the 912 // marking phase. So, we initilize this here, at the start of the 913 // cycle, so that at the remaining overflow number will decrease at 914 // every remark and we'll eventually not need to cause one. 915 force_overflow_stw()->init(); 916 917 // Start Concurrent Marking weak-reference discovery. 918 ReferenceProcessor* rp = g1h->ref_processor_cm(); 919 // enable ("weak") refs discovery 920 rp->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 921 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle 922 923 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 924 // This is the start of the marking cycle, we're expected all 925 // threads to have SATB queues with active set to false. 926 satb_mq_set.set_active_all_threads(true, /* new active value */ 927 false /* expected_active */); 928 929 _root_regions.prepare_for_scan(); 930 931 // update_g1_committed() will be called at the end of an evac pause 932 // when marking is on. So, it's also called at the end of the 933 // initial-mark pause to update the heap end, if the heap expands 934 // during it. No need to call it here. 935 } 936 937 /* 938 * Notice that in the next two methods, we actually leave the STS 939 * during the barrier sync and join it immediately afterwards. If we 940 * do not do this, the following deadlock can occur: one thread could 941 * be in the barrier sync code, waiting for the other thread to also 942 * sync up, whereas another one could be trying to yield, while also 943 * waiting for the other threads to sync up too. 944 * 945 * Note, however, that this code is also used during remark and in 946 * this case we should not attempt to leave / enter the STS, otherwise 947 * we'll either hit an asseert (debug / fastdebug) or deadlock 948 * (product). So we should only leave / enter the STS if we are 949 * operating concurrently. 950 * 951 * Because the thread that does the sync barrier has left the STS, it 952 * is possible to be suspended for a Full GC or an evacuation pause 953 * could occur. This is actually safe, since the entering the sync 954 * barrier is one of the last things do_marking_step() does, and it 955 * doesn't manipulate any data structures afterwards. 956 */ 957 958 void ConcurrentMark::enter_first_sync_barrier(uint worker_id) { 959 if (verbose_low()) { 960 gclog_or_tty->print_cr("[%u] entering first barrier", worker_id); 961 } 962 963 if (concurrent()) { 964 ConcurrentGCThread::stsLeave(); 965 } 966 _first_overflow_barrier_sync.enter(); 967 if (concurrent()) { 968 ConcurrentGCThread::stsJoin(); 969 } 970 // at this point everyone should have synced up and not be doing any 971 // more work 972 973 if (verbose_low()) { 974 gclog_or_tty->print_cr("[%u] leaving first barrier", worker_id); 975 } 976 977 // let the task associated with with worker 0 do this 978 if (worker_id == 0) { 979 // task 0 is responsible for clearing the global data structures 980 // We should be here because of an overflow. During STW we should 981 // not clear the overflow flag since we rely on it being true when 982 // we exit this method to abort the pause and restart concurent 983 // marking. 984 reset_marking_state(concurrent() /* clear_overflow */); 985 force_overflow()->update(); 986 987 if (G1Log::fine()) { 988 gclog_or_tty->date_stamp(PrintGCDateStamps); 989 gclog_or_tty->stamp(PrintGCTimeStamps); 990 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]"); 991 } 992 } 993 994 // after this, each task should reset its own data structures then 995 // then go into the second barrier 996 } 997 998 void ConcurrentMark::enter_second_sync_barrier(uint worker_id) { 999 if (verbose_low()) { 1000 gclog_or_tty->print_cr("[%u] entering second barrier", worker_id); 1001 } 1002 1003 if (concurrent()) { 1004 ConcurrentGCThread::stsLeave(); 1005 } 1006 _second_overflow_barrier_sync.enter(); 1007 if (concurrent()) { 1008 ConcurrentGCThread::stsJoin(); 1009 } 1010 // at this point everything should be re-initialised and ready to go 1011 1012 if (verbose_low()) { 1013 gclog_or_tty->print_cr("[%u] leaving second barrier", worker_id); 1014 } 1015 } 1016 1017 #ifndef PRODUCT 1018 void ForceOverflowSettings::init() { 1019 _num_remaining = G1ConcMarkForceOverflow; 1020 _force = false; 1021 update(); 1022 } 1023 1024 void ForceOverflowSettings::update() { 1025 if (_num_remaining > 0) { 1026 _num_remaining -= 1; 1027 _force = true; 1028 } else { 1029 _force = false; 1030 } 1031 } 1032 1033 bool ForceOverflowSettings::should_force() { 1034 if (_force) { 1035 _force = false; 1036 return true; 1037 } else { 1038 return false; 1039 } 1040 } 1041 #endif // !PRODUCT 1042 1043 class CMConcurrentMarkingTask: public AbstractGangTask { 1044 private: 1045 ConcurrentMark* _cm; 1046 ConcurrentMarkThread* _cmt; 1047 1048 public: 1049 void work(uint worker_id) { 1050 assert(Thread::current()->is_ConcurrentGC_thread(), 1051 "this should only be done by a conc GC thread"); 1052 ResourceMark rm; 1053 1054 double start_vtime = os::elapsedVTime(); 1055 1056 ConcurrentGCThread::stsJoin(); 1057 1058 assert(worker_id < _cm->active_tasks(), "invariant"); 1059 CMTask* the_task = _cm->task(worker_id); 1060 the_task->record_start_time(); 1061 if (!_cm->has_aborted()) { 1062 do { 1063 double start_vtime_sec = os::elapsedVTime(); 1064 double start_time_sec = os::elapsedTime(); 1065 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 1066 1067 the_task->do_marking_step(mark_step_duration_ms, 1068 true /* do_stealing */, 1069 true /* do_termination */); 1070 1071 double end_time_sec = os::elapsedTime(); 1072 double end_vtime_sec = os::elapsedVTime(); 1073 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; 1074 double elapsed_time_sec = end_time_sec - start_time_sec; 1075 _cm->clear_has_overflown(); 1076 1077 bool ret = _cm->do_yield_check(worker_id); 1078 1079 jlong sleep_time_ms; 1080 if (!_cm->has_aborted() && the_task->has_aborted()) { 1081 sleep_time_ms = 1082 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); 1083 ConcurrentGCThread::stsLeave(); 1084 os::sleep(Thread::current(), sleep_time_ms, false); 1085 ConcurrentGCThread::stsJoin(); 1086 } 1087 double end_time2_sec = os::elapsedTime(); 1088 double elapsed_time2_sec = end_time2_sec - start_time_sec; 1089 1090 #if 0 1091 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, " 1092 "overhead %1.4lf", 1093 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms, 1094 the_task->conc_overhead(os::elapsedTime()) * 8.0); 1095 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms", 1096 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0); 1097 #endif 1098 } while (!_cm->has_aborted() && the_task->has_aborted()); 1099 } 1100 the_task->record_end_time(); 1101 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); 1102 1103 ConcurrentGCThread::stsLeave(); 1104 1105 double end_vtime = os::elapsedVTime(); 1106 _cm->update_accum_task_vtime(worker_id, end_vtime - start_vtime); 1107 } 1108 1109 CMConcurrentMarkingTask(ConcurrentMark* cm, 1110 ConcurrentMarkThread* cmt) : 1111 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } 1112 1113 ~CMConcurrentMarkingTask() { } 1114 }; 1115 1116 // Calculates the number of active workers for a concurrent 1117 // phase. 1118 uint ConcurrentMark::calc_parallel_marking_threads() { 1119 if (G1CollectedHeap::use_parallel_gc_threads()) { 1120 uint n_conc_workers = 0; 1121 if (!UseDynamicNumberOfGCThreads || 1122 (!FLAG_IS_DEFAULT(ConcGCThreads) && 1123 !ForceDynamicNumberOfGCThreads)) { 1124 n_conc_workers = max_parallel_marking_threads(); 1125 } else { 1126 n_conc_workers = 1127 AdaptiveSizePolicy::calc_default_active_workers( 1128 max_parallel_marking_threads(), 1129 1, /* Minimum workers */ 1130 parallel_marking_threads(), 1131 Threads::number_of_non_daemon_threads()); 1132 // Don't scale down "n_conc_workers" by scale_parallel_threads() because 1133 // that scaling has already gone into "_max_parallel_marking_threads". 1134 } 1135 assert(n_conc_workers > 0, "Always need at least 1"); 1136 return n_conc_workers; 1137 } 1138 // If we are not running with any parallel GC threads we will not 1139 // have spawned any marking threads either. Hence the number of 1140 // concurrent workers should be 0. 1141 return 0; 1142 } 1143 1144 void ConcurrentMark::scanRootRegion(HeapRegion* hr, uint worker_id) { 1145 // Currently, only survivors can be root regions. 1146 assert(hr->next_top_at_mark_start() == hr->bottom(), "invariant"); 1147 G1RootRegionScanClosure cl(_g1h, this, worker_id); 1148 1149 const uintx interval = PrefetchScanIntervalInBytes; 1150 HeapWord* curr = hr->bottom(); 1151 const HeapWord* end = hr->top(); 1152 while (curr < end) { 1153 Prefetch::read(curr, interval); 1154 oop obj = oop(curr); 1155 int size = obj->oop_iterate(&cl); 1156 assert(size == obj->size(), "sanity"); 1157 curr += size; 1158 } 1159 } 1160 1161 class CMRootRegionScanTask : public AbstractGangTask { 1162 private: 1163 ConcurrentMark* _cm; 1164 1165 public: 1166 CMRootRegionScanTask(ConcurrentMark* cm) : 1167 AbstractGangTask("Root Region Scan"), _cm(cm) { } 1168 1169 void work(uint worker_id) { 1170 assert(Thread::current()->is_ConcurrentGC_thread(), 1171 "this should only be done by a conc GC thread"); 1172 1173 CMRootRegions* root_regions = _cm->root_regions(); 1174 HeapRegion* hr = root_regions->claim_next(); 1175 while (hr != NULL) { 1176 _cm->scanRootRegion(hr, worker_id); 1177 hr = root_regions->claim_next(); 1178 } 1179 } 1180 }; 1181 1182 void ConcurrentMark::scanRootRegions() { 1183 // scan_in_progress() will have been set to true only if there was 1184 // at least one root region to scan. So, if it's false, we 1185 // should not attempt to do any further work. 1186 if (root_regions()->scan_in_progress()) { 1187 _parallel_marking_threads = calc_parallel_marking_threads(); 1188 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1189 "Maximum number of marking threads exceeded"); 1190 uint active_workers = MAX2(1U, parallel_marking_threads()); 1191 1192 CMRootRegionScanTask task(this); 1193 if (use_parallel_marking_threads()) { 1194 _parallel_workers->set_active_workers((int) active_workers); 1195 _parallel_workers->run_task(&task); 1196 } else { 1197 task.work(0); 1198 } 1199 1200 // It's possible that has_aborted() is true here without actually 1201 // aborting the survivor scan earlier. This is OK as it's 1202 // mainly used for sanity checking. 1203 root_regions()->scan_finished(); 1204 } 1205 } 1206 1207 void ConcurrentMark::markFromRoots() { 1208 // we might be tempted to assert that: 1209 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), 1210 // "inconsistent argument?"); 1211 // However that wouldn't be right, because it's possible that 1212 // a safepoint is indeed in progress as a younger generation 1213 // stop-the-world GC happens even as we mark in this generation. 1214 1215 _restart_for_overflow = false; 1216 force_overflow_conc()->init(); 1217 1218 // _g1h has _n_par_threads 1219 _parallel_marking_threads = calc_parallel_marking_threads(); 1220 assert(parallel_marking_threads() <= max_parallel_marking_threads(), 1221 "Maximum number of marking threads exceeded"); 1222 1223 uint active_workers = MAX2(1U, parallel_marking_threads()); 1224 1225 // Parallel task terminator is set in "set_phase()" 1226 set_phase(active_workers, true /* concurrent */); 1227 1228 CMConcurrentMarkingTask markingTask(this, cmThread()); 1229 if (use_parallel_marking_threads()) { 1230 _parallel_workers->set_active_workers((int)active_workers); 1231 // Don't set _n_par_threads because it affects MT in proceess_strong_roots() 1232 // and the decisions on that MT processing is made elsewhere. 1233 assert(_parallel_workers->active_workers() > 0, "Should have been set"); 1234 _parallel_workers->run_task(&markingTask); 1235 } else { 1236 markingTask.work(0); 1237 } 1238 print_stats(); 1239 } 1240 1241 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { 1242 // world is stopped at this checkpoint 1243 assert(SafepointSynchronize::is_at_safepoint(), 1244 "world should be stopped"); 1245 1246 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1247 1248 // If a full collection has happened, we shouldn't do this. 1249 if (has_aborted()) { 1250 g1h->set_marking_complete(); // So bitmap clearing isn't confused 1251 return; 1252 } 1253 1254 SvcGCMarker sgcm(SvcGCMarker::OTHER); 1255 1256 if (VerifyDuringGC) { 1257 HandleMark hm; // handle scope 1258 gclog_or_tty->print(" VerifyDuringGC:(before)"); 1259 Universe::heap()->prepare_for_verify(); 1260 Universe::verify(/* silent */ false, 1261 /* option */ VerifyOption_G1UsePrevMarking); 1262 } 1263 1264 G1CollectorPolicy* g1p = g1h->g1_policy(); 1265 g1p->record_concurrent_mark_remark_start(); 1266 1267 double start = os::elapsedTime(); 1268 1269 checkpointRootsFinalWork(); 1270 1271 double mark_work_end = os::elapsedTime(); 1272 1273 weakRefsWork(clear_all_soft_refs); 1274 1275 if (has_overflown()) { 1276 // Oops. We overflowed. Restart concurrent marking. 1277 _restart_for_overflow = true; 1278 if (G1TraceMarkStackOverflow) { 1279 gclog_or_tty->print_cr("\nRemark led to restart for overflow."); 1280 } 1281 1282 // Verify the heap w.r.t. the previous marking bitmap. 1283 if (VerifyDuringGC) { 1284 HandleMark hm; // handle scope 1285 gclog_or_tty->print(" VerifyDuringGC:(overflow)"); 1286 Universe::heap()->prepare_for_verify(); 1287 Universe::verify(/* silent */ false, 1288 /* option */ VerifyOption_G1UsePrevMarking); 1289 } 1290 1291 // Clear the marking state because we will be restarting 1292 // marking due to overflowing the global mark stack. 1293 reset_marking_state(); 1294 } else { 1295 // Aggregate the per-task counting data that we have accumulated 1296 // while marking. 1297 aggregate_count_data(); 1298 1299 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 1300 // We're done with marking. 1301 // This is the end of the marking cycle, we're expected all 1302 // threads to have SATB queues with active set to true. 1303 satb_mq_set.set_active_all_threads(false, /* new active value */ 1304 true /* expected_active */); 1305 1306 if (VerifyDuringGC) { 1307 HandleMark hm; // handle scope 1308 gclog_or_tty->print(" VerifyDuringGC:(after)"); 1309 Universe::heap()->prepare_for_verify(); 1310 Universe::verify(/* silent */ false, 1311 /* option */ VerifyOption_G1UseNextMarking); 1312 } 1313 assert(!restart_for_overflow(), "sanity"); 1314 // Completely reset the marking state since marking completed 1315 set_non_marking_state(); 1316 } 1317 1318 // Expand the marking stack, if we have to and if we can. 1319 if (_markStack.should_expand()) { 1320 _markStack.expand(); 1321 } 1322 1323 #if VERIFY_OBJS_PROCESSED 1324 _scan_obj_cl.objs_processed = 0; 1325 ThreadLocalObjQueue::objs_enqueued = 0; 1326 #endif 1327 1328 // Statistics 1329 double now = os::elapsedTime(); 1330 _remark_mark_times.add((mark_work_end - start) * 1000.0); 1331 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); 1332 _remark_times.add((now - start) * 1000.0); 1333 1334 g1p->record_concurrent_mark_remark_end(); 1335 } 1336 1337 // Base class of the closures that finalize and verify the 1338 // liveness counting data. 1339 class CMCountDataClosureBase: public HeapRegionClosure { 1340 protected: 1341 G1CollectedHeap* _g1h; 1342 ConcurrentMark* _cm; 1343 CardTableModRefBS* _ct_bs; 1344 1345 BitMap* _region_bm; 1346 BitMap* _card_bm; 1347 1348 // Takes a region that's not empty (i.e., it has at least one 1349 // live object in it and sets its corresponding bit on the region 1350 // bitmap to 1. If the region is "starts humongous" it will also set 1351 // to 1 the bits on the region bitmap that correspond to its 1352 // associated "continues humongous" regions. 1353 void set_bit_for_region(HeapRegion* hr) { 1354 assert(!hr->continuesHumongous(), "should have filtered those out"); 1355 1356 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index(); 1357 if (!hr->startsHumongous()) { 1358 // Normal (non-humongous) case: just set the bit. 1359 _region_bm->par_at_put(index, true); 1360 } else { 1361 // Starts humongous case: calculate how many regions are part of 1362 // this humongous region and then set the bit range. 1363 BitMap::idx_t end_index = (BitMap::idx_t) hr->last_hc_index(); 1364 _region_bm->par_at_put_range(index, end_index, true); 1365 } 1366 } 1367 1368 public: 1369 CMCountDataClosureBase(G1CollectedHeap* g1h, 1370 BitMap* region_bm, BitMap* card_bm): 1371 _g1h(g1h), _cm(g1h->concurrent_mark()), 1372 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 1373 _region_bm(region_bm), _card_bm(card_bm) { } 1374 }; 1375 1376 // Closure that calculates the # live objects per region. Used 1377 // for verification purposes during the cleanup pause. 1378 class CalcLiveObjectsClosure: public CMCountDataClosureBase { 1379 CMBitMapRO* _bm; 1380 size_t _region_marked_bytes; 1381 1382 public: 1383 CalcLiveObjectsClosure(CMBitMapRO *bm, G1CollectedHeap* g1h, 1384 BitMap* region_bm, BitMap* card_bm) : 1385 CMCountDataClosureBase(g1h, region_bm, card_bm), 1386 _bm(bm), _region_marked_bytes(0) { } 1387 1388 bool doHeapRegion(HeapRegion* hr) { 1389 1390 if (hr->continuesHumongous()) { 1391 // We will ignore these here and process them when their 1392 // associated "starts humongous" region is processed (see 1393 // set_bit_for_heap_region()). Note that we cannot rely on their 1394 // associated "starts humongous" region to have their bit set to 1395 // 1 since, due to the region chunking in the parallel region 1396 // iteration, a "continues humongous" region might be visited 1397 // before its associated "starts humongous". 1398 return false; 1399 } 1400 1401 HeapWord* ntams = hr->next_top_at_mark_start(); 1402 HeapWord* start = hr->bottom(); 1403 1404 assert(start <= hr->end() && start <= ntams && ntams <= hr->end(), 1405 err_msg("Preconditions not met - " 1406 "start: "PTR_FORMAT", ntams: "PTR_FORMAT", end: "PTR_FORMAT, 1407 start, ntams, hr->end())); 1408 1409 // Find the first marked object at or after "start". 1410 start = _bm->getNextMarkedWordAddress(start, ntams); 1411 1412 size_t marked_bytes = 0; 1413 1414 while (start < ntams) { 1415 oop obj = oop(start); 1416 int obj_sz = obj->size(); 1417 HeapWord* obj_end = start + obj_sz; 1418 1419 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 1420 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(obj_end); 1421 1422 // Note: if we're looking at the last region in heap - obj_end 1423 // could be actually just beyond the end of the heap; end_idx 1424 // will then correspond to a (non-existent) card that is also 1425 // just beyond the heap. 1426 if (_g1h->is_in_g1_reserved(obj_end) && !_ct_bs->is_card_aligned(obj_end)) { 1427 // end of object is not card aligned - increment to cover 1428 // all the cards spanned by the object 1429 end_idx += 1; 1430 } 1431 1432 // Set the bits in the card BM for the cards spanned by this object. 1433 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1434 1435 // Add the size of this object to the number of marked bytes. 1436 marked_bytes += (size_t)obj_sz * HeapWordSize; 1437 1438 // Find the next marked object after this one. 1439 start = _bm->getNextMarkedWordAddress(obj_end, ntams); 1440 } 1441 1442 // Mark the allocated-since-marking portion... 1443 HeapWord* top = hr->top(); 1444 if (ntams < top) { 1445 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1446 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1447 1448 // Note: if we're looking at the last region in heap - top 1449 // could be actually just beyond the end of the heap; end_idx 1450 // will then correspond to a (non-existent) card that is also 1451 // just beyond the heap. 1452 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1453 // end of object is not card aligned - increment to cover 1454 // all the cards spanned by the object 1455 end_idx += 1; 1456 } 1457 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1458 1459 // This definitely means the region has live objects. 1460 set_bit_for_region(hr); 1461 } 1462 1463 // Update the live region bitmap. 1464 if (marked_bytes > 0) { 1465 set_bit_for_region(hr); 1466 } 1467 1468 // Set the marked bytes for the current region so that 1469 // it can be queried by a calling verificiation routine 1470 _region_marked_bytes = marked_bytes; 1471 1472 return false; 1473 } 1474 1475 size_t region_marked_bytes() const { return _region_marked_bytes; } 1476 }; 1477 1478 // Heap region closure used for verifying the counting data 1479 // that was accumulated concurrently and aggregated during 1480 // the remark pause. This closure is applied to the heap 1481 // regions during the STW cleanup pause. 1482 1483 class VerifyLiveObjectDataHRClosure: public HeapRegionClosure { 1484 G1CollectedHeap* _g1h; 1485 ConcurrentMark* _cm; 1486 CalcLiveObjectsClosure _calc_cl; 1487 BitMap* _region_bm; // Region BM to be verified 1488 BitMap* _card_bm; // Card BM to be verified 1489 bool _verbose; // verbose output? 1490 1491 BitMap* _exp_region_bm; // Expected Region BM values 1492 BitMap* _exp_card_bm; // Expected card BM values 1493 1494 int _failures; 1495 1496 public: 1497 VerifyLiveObjectDataHRClosure(G1CollectedHeap* g1h, 1498 BitMap* region_bm, 1499 BitMap* card_bm, 1500 BitMap* exp_region_bm, 1501 BitMap* exp_card_bm, 1502 bool verbose) : 1503 _g1h(g1h), _cm(g1h->concurrent_mark()), 1504 _calc_cl(_cm->nextMarkBitMap(), g1h, exp_region_bm, exp_card_bm), 1505 _region_bm(region_bm), _card_bm(card_bm), _verbose(verbose), 1506 _exp_region_bm(exp_region_bm), _exp_card_bm(exp_card_bm), 1507 _failures(0) { } 1508 1509 int failures() const { return _failures; } 1510 1511 bool doHeapRegion(HeapRegion* hr) { 1512 if (hr->continuesHumongous()) { 1513 // We will ignore these here and process them when their 1514 // associated "starts humongous" region is processed (see 1515 // set_bit_for_heap_region()). Note that we cannot rely on their 1516 // associated "starts humongous" region to have their bit set to 1517 // 1 since, due to the region chunking in the parallel region 1518 // iteration, a "continues humongous" region might be visited 1519 // before its associated "starts humongous". 1520 return false; 1521 } 1522 1523 int failures = 0; 1524 1525 // Call the CalcLiveObjectsClosure to walk the marking bitmap for 1526 // this region and set the corresponding bits in the expected region 1527 // and card bitmaps. 1528 bool res = _calc_cl.doHeapRegion(hr); 1529 assert(res == false, "should be continuing"); 1530 1531 MutexLockerEx x((_verbose ? ParGCRareEvent_lock : NULL), 1532 Mutex::_no_safepoint_check_flag); 1533 1534 // Verify the marked bytes for this region. 1535 size_t exp_marked_bytes = _calc_cl.region_marked_bytes(); 1536 size_t act_marked_bytes = hr->next_marked_bytes(); 1537 1538 // We're not OK if expected marked bytes > actual marked bytes. It means 1539 // we have missed accounting some objects during the actual marking. 1540 if (exp_marked_bytes > act_marked_bytes) { 1541 if (_verbose) { 1542 gclog_or_tty->print_cr("Region %u: marked bytes mismatch: " 1543 "expected: " SIZE_FORMAT ", actual: " SIZE_FORMAT, 1544 hr->hrs_index(), exp_marked_bytes, act_marked_bytes); 1545 } 1546 failures += 1; 1547 } 1548 1549 // Verify the bit, for this region, in the actual and expected 1550 // (which was just calculated) region bit maps. 1551 // We're not OK if the bit in the calculated expected region 1552 // bitmap is set and the bit in the actual region bitmap is not. 1553 BitMap::idx_t index = (BitMap::idx_t) hr->hrs_index(); 1554 1555 bool expected = _exp_region_bm->at(index); 1556 bool actual = _region_bm->at(index); 1557 if (expected && !actual) { 1558 if (_verbose) { 1559 gclog_or_tty->print_cr("Region %u: region bitmap mismatch: " 1560 "expected: %s, actual: %s", 1561 hr->hrs_index(), 1562 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1563 } 1564 failures += 1; 1565 } 1566 1567 // Verify that the card bit maps for the cards spanned by the current 1568 // region match. We have an error if we have a set bit in the expected 1569 // bit map and the corresponding bit in the actual bitmap is not set. 1570 1571 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(hr->bottom()); 1572 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(hr->top()); 1573 1574 for (BitMap::idx_t i = start_idx; i < end_idx; i+=1) { 1575 expected = _exp_card_bm->at(i); 1576 actual = _card_bm->at(i); 1577 1578 if (expected && !actual) { 1579 if (_verbose) { 1580 gclog_or_tty->print_cr("Region %u: card bitmap mismatch at " SIZE_FORMAT ": " 1581 "expected: %s, actual: %s", 1582 hr->hrs_index(), i, 1583 BOOL_TO_STR(expected), BOOL_TO_STR(actual)); 1584 } 1585 failures += 1; 1586 } 1587 } 1588 1589 if (failures > 0 && _verbose) { 1590 gclog_or_tty->print_cr("Region " HR_FORMAT ", ntams: " PTR_FORMAT ", " 1591 "marked_bytes: calc/actual " SIZE_FORMAT "/" SIZE_FORMAT, 1592 HR_FORMAT_PARAMS(hr), hr->next_top_at_mark_start(), 1593 _calc_cl.region_marked_bytes(), hr->next_marked_bytes()); 1594 } 1595 1596 _failures += failures; 1597 1598 // We could stop iteration over the heap when we 1599 // find the first violating region by returning true. 1600 return false; 1601 } 1602 }; 1603 1604 1605 class G1ParVerifyFinalCountTask: public AbstractGangTask { 1606 protected: 1607 G1CollectedHeap* _g1h; 1608 ConcurrentMark* _cm; 1609 BitMap* _actual_region_bm; 1610 BitMap* _actual_card_bm; 1611 1612 uint _n_workers; 1613 1614 BitMap* _expected_region_bm; 1615 BitMap* _expected_card_bm; 1616 1617 int _failures; 1618 bool _verbose; 1619 1620 public: 1621 G1ParVerifyFinalCountTask(G1CollectedHeap* g1h, 1622 BitMap* region_bm, BitMap* card_bm, 1623 BitMap* expected_region_bm, BitMap* expected_card_bm) 1624 : AbstractGangTask("G1 verify final counting"), 1625 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1626 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1627 _expected_region_bm(expected_region_bm), _expected_card_bm(expected_card_bm), 1628 _failures(0), _verbose(false), 1629 _n_workers(0) { 1630 assert(VerifyDuringGC, "don't call this otherwise"); 1631 1632 // Use the value already set as the number of active threads 1633 // in the call to run_task(). 1634 if (G1CollectedHeap::use_parallel_gc_threads()) { 1635 assert( _g1h->workers()->active_workers() > 0, 1636 "Should have been previously set"); 1637 _n_workers = _g1h->workers()->active_workers(); 1638 } else { 1639 _n_workers = 1; 1640 } 1641 1642 assert(_expected_card_bm->size() == _actual_card_bm->size(), "sanity"); 1643 assert(_expected_region_bm->size() == _actual_region_bm->size(), "sanity"); 1644 1645 _verbose = _cm->verbose_medium(); 1646 } 1647 1648 void work(uint worker_id) { 1649 assert(worker_id < _n_workers, "invariant"); 1650 1651 VerifyLiveObjectDataHRClosure verify_cl(_g1h, 1652 _actual_region_bm, _actual_card_bm, 1653 _expected_region_bm, 1654 _expected_card_bm, 1655 _verbose); 1656 1657 if (G1CollectedHeap::use_parallel_gc_threads()) { 1658 _g1h->heap_region_par_iterate_chunked(&verify_cl, 1659 worker_id, 1660 _n_workers, 1661 HeapRegion::VerifyCountClaimValue); 1662 } else { 1663 _g1h->heap_region_iterate(&verify_cl); 1664 } 1665 1666 Atomic::add(verify_cl.failures(), &_failures); 1667 } 1668 1669 int failures() const { return _failures; } 1670 }; 1671 1672 // Closure that finalizes the liveness counting data. 1673 // Used during the cleanup pause. 1674 // Sets the bits corresponding to the interval [NTAMS, top] 1675 // (which contains the implicitly live objects) in the 1676 // card liveness bitmap. Also sets the bit for each region, 1677 // containing live data, in the region liveness bitmap. 1678 1679 class FinalCountDataUpdateClosure: public CMCountDataClosureBase { 1680 public: 1681 FinalCountDataUpdateClosure(G1CollectedHeap* g1h, 1682 BitMap* region_bm, 1683 BitMap* card_bm) : 1684 CMCountDataClosureBase(g1h, region_bm, card_bm) { } 1685 1686 bool doHeapRegion(HeapRegion* hr) { 1687 1688 if (hr->continuesHumongous()) { 1689 // We will ignore these here and process them when their 1690 // associated "starts humongous" region is processed (see 1691 // set_bit_for_heap_region()). Note that we cannot rely on their 1692 // associated "starts humongous" region to have their bit set to 1693 // 1 since, due to the region chunking in the parallel region 1694 // iteration, a "continues humongous" region might be visited 1695 // before its associated "starts humongous". 1696 return false; 1697 } 1698 1699 HeapWord* ntams = hr->next_top_at_mark_start(); 1700 HeapWord* top = hr->top(); 1701 1702 assert(hr->bottom() <= ntams && ntams <= hr->end(), "Preconditions."); 1703 1704 // Mark the allocated-since-marking portion... 1705 if (ntams < top) { 1706 // This definitely means the region has live objects. 1707 set_bit_for_region(hr); 1708 1709 // Now set the bits in the card bitmap for [ntams, top) 1710 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(ntams); 1711 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(top); 1712 1713 // Note: if we're looking at the last region in heap - top 1714 // could be actually just beyond the end of the heap; end_idx 1715 // will then correspond to a (non-existent) card that is also 1716 // just beyond the heap. 1717 if (_g1h->is_in_g1_reserved(top) && !_ct_bs->is_card_aligned(top)) { 1718 // end of object is not card aligned - increment to cover 1719 // all the cards spanned by the object 1720 end_idx += 1; 1721 } 1722 1723 assert(end_idx <= _card_bm->size(), 1724 err_msg("oob: end_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1725 end_idx, _card_bm->size())); 1726 assert(start_idx < _card_bm->size(), 1727 err_msg("oob: start_idx= "SIZE_FORMAT", bitmap size= "SIZE_FORMAT, 1728 start_idx, _card_bm->size())); 1729 1730 _cm->set_card_bitmap_range(_card_bm, start_idx, end_idx, true /* is_par */); 1731 } 1732 1733 // Set the bit for the region if it contains live data 1734 if (hr->next_marked_bytes() > 0) { 1735 set_bit_for_region(hr); 1736 } 1737 1738 return false; 1739 } 1740 }; 1741 1742 class G1ParFinalCountTask: public AbstractGangTask { 1743 protected: 1744 G1CollectedHeap* _g1h; 1745 ConcurrentMark* _cm; 1746 BitMap* _actual_region_bm; 1747 BitMap* _actual_card_bm; 1748 1749 uint _n_workers; 1750 1751 public: 1752 G1ParFinalCountTask(G1CollectedHeap* g1h, BitMap* region_bm, BitMap* card_bm) 1753 : AbstractGangTask("G1 final counting"), 1754 _g1h(g1h), _cm(_g1h->concurrent_mark()), 1755 _actual_region_bm(region_bm), _actual_card_bm(card_bm), 1756 _n_workers(0) { 1757 // Use the value already set as the number of active threads 1758 // in the call to run_task(). 1759 if (G1CollectedHeap::use_parallel_gc_threads()) { 1760 assert( _g1h->workers()->active_workers() > 0, 1761 "Should have been previously set"); 1762 _n_workers = _g1h->workers()->active_workers(); 1763 } else { 1764 _n_workers = 1; 1765 } 1766 } 1767 1768 void work(uint worker_id) { 1769 assert(worker_id < _n_workers, "invariant"); 1770 1771 FinalCountDataUpdateClosure final_update_cl(_g1h, 1772 _actual_region_bm, 1773 _actual_card_bm); 1774 1775 if (G1CollectedHeap::use_parallel_gc_threads()) { 1776 _g1h->heap_region_par_iterate_chunked(&final_update_cl, 1777 worker_id, 1778 _n_workers, 1779 HeapRegion::FinalCountClaimValue); 1780 } else { 1781 _g1h->heap_region_iterate(&final_update_cl); 1782 } 1783 } 1784 }; 1785 1786 class G1ParNoteEndTask; 1787 1788 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { 1789 G1CollectedHeap* _g1; 1790 int _worker_num; 1791 size_t _max_live_bytes; 1792 uint _regions_claimed; 1793 size_t _freed_bytes; 1794 FreeRegionList* _local_cleanup_list; 1795 OldRegionSet* _old_proxy_set; 1796 HumongousRegionSet* _humongous_proxy_set; 1797 HRRSCleanupTask* _hrrs_cleanup_task; 1798 double _claimed_region_time; 1799 double _max_region_time; 1800 1801 public: 1802 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, 1803 int worker_num, 1804 FreeRegionList* local_cleanup_list, 1805 OldRegionSet* old_proxy_set, 1806 HumongousRegionSet* humongous_proxy_set, 1807 HRRSCleanupTask* hrrs_cleanup_task) : 1808 _g1(g1), _worker_num(worker_num), 1809 _max_live_bytes(0), _regions_claimed(0), 1810 _freed_bytes(0), 1811 _claimed_region_time(0.0), _max_region_time(0.0), 1812 _local_cleanup_list(local_cleanup_list), 1813 _old_proxy_set(old_proxy_set), 1814 _humongous_proxy_set(humongous_proxy_set), 1815 _hrrs_cleanup_task(hrrs_cleanup_task) { } 1816 1817 size_t freed_bytes() { return _freed_bytes; } 1818 1819 bool doHeapRegion(HeapRegion *hr) { 1820 if (hr->continuesHumongous()) { 1821 return false; 1822 } 1823 // We use a claim value of zero here because all regions 1824 // were claimed with value 1 in the FinalCount task. 1825 _g1->reset_gc_time_stamps(hr); 1826 double start = os::elapsedTime(); 1827 _regions_claimed++; 1828 hr->note_end_of_marking(); 1829 _max_live_bytes += hr->max_live_bytes(); 1830 _g1->free_region_if_empty(hr, 1831 &_freed_bytes, 1832 _local_cleanup_list, 1833 _old_proxy_set, 1834 _humongous_proxy_set, 1835 _hrrs_cleanup_task, 1836 true /* par */); 1837 double region_time = (os::elapsedTime() - start); 1838 _claimed_region_time += region_time; 1839 if (region_time > _max_region_time) { 1840 _max_region_time = region_time; 1841 } 1842 return false; 1843 } 1844 1845 size_t max_live_bytes() { return _max_live_bytes; } 1846 uint regions_claimed() { return _regions_claimed; } 1847 double claimed_region_time_sec() { return _claimed_region_time; } 1848 double max_region_time_sec() { return _max_region_time; } 1849 }; 1850 1851 class G1ParNoteEndTask: public AbstractGangTask { 1852 friend class G1NoteEndOfConcMarkClosure; 1853 1854 protected: 1855 G1CollectedHeap* _g1h; 1856 size_t _max_live_bytes; 1857 size_t _freed_bytes; 1858 FreeRegionList* _cleanup_list; 1859 1860 public: 1861 G1ParNoteEndTask(G1CollectedHeap* g1h, 1862 FreeRegionList* cleanup_list) : 1863 AbstractGangTask("G1 note end"), _g1h(g1h), 1864 _max_live_bytes(0), _freed_bytes(0), _cleanup_list(cleanup_list) { } 1865 1866 void work(uint worker_id) { 1867 double start = os::elapsedTime(); 1868 FreeRegionList local_cleanup_list("Local Cleanup List"); 1869 OldRegionSet old_proxy_set("Local Cleanup Old Proxy Set"); 1870 HumongousRegionSet humongous_proxy_set("Local Cleanup Humongous Proxy Set"); 1871 HRRSCleanupTask hrrs_cleanup_task; 1872 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, worker_id, &local_cleanup_list, 1873 &old_proxy_set, 1874 &humongous_proxy_set, 1875 &hrrs_cleanup_task); 1876 if (G1CollectedHeap::use_parallel_gc_threads()) { 1877 _g1h->heap_region_par_iterate_chunked(&g1_note_end, worker_id, 1878 _g1h->workers()->active_workers(), 1879 HeapRegion::NoteEndClaimValue); 1880 } else { 1881 _g1h->heap_region_iterate(&g1_note_end); 1882 } 1883 assert(g1_note_end.complete(), "Shouldn't have yielded!"); 1884 1885 // Now update the lists 1886 _g1h->update_sets_after_freeing_regions(g1_note_end.freed_bytes(), 1887 NULL /* free_list */, 1888 &old_proxy_set, 1889 &humongous_proxy_set, 1890 true /* par */); 1891 { 1892 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 1893 _max_live_bytes += g1_note_end.max_live_bytes(); 1894 _freed_bytes += g1_note_end.freed_bytes(); 1895 1896 // If we iterate over the global cleanup list at the end of 1897 // cleanup to do this printing we will not guarantee to only 1898 // generate output for the newly-reclaimed regions (the list 1899 // might not be empty at the beginning of cleanup; we might 1900 // still be working on its previous contents). So we do the 1901 // printing here, before we append the new regions to the global 1902 // cleanup list. 1903 1904 G1HRPrinter* hr_printer = _g1h->hr_printer(); 1905 if (hr_printer->is_active()) { 1906 HeapRegionLinkedListIterator iter(&local_cleanup_list); 1907 while (iter.more_available()) { 1908 HeapRegion* hr = iter.get_next(); 1909 hr_printer->cleanup(hr); 1910 } 1911 } 1912 1913 _cleanup_list->add_as_tail(&local_cleanup_list); 1914 assert(local_cleanup_list.is_empty(), "post-condition"); 1915 1916 HeapRegionRemSet::finish_cleanup_task(&hrrs_cleanup_task); 1917 } 1918 } 1919 size_t max_live_bytes() { return _max_live_bytes; } 1920 size_t freed_bytes() { return _freed_bytes; } 1921 }; 1922 1923 class G1ParScrubRemSetTask: public AbstractGangTask { 1924 protected: 1925 G1RemSet* _g1rs; 1926 BitMap* _region_bm; 1927 BitMap* _card_bm; 1928 public: 1929 G1ParScrubRemSetTask(G1CollectedHeap* g1h, 1930 BitMap* region_bm, BitMap* card_bm) : 1931 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), 1932 _region_bm(region_bm), _card_bm(card_bm) { } 1933 1934 void work(uint worker_id) { 1935 if (G1CollectedHeap::use_parallel_gc_threads()) { 1936 _g1rs->scrub_par(_region_bm, _card_bm, worker_id, 1937 HeapRegion::ScrubRemSetClaimValue); 1938 } else { 1939 _g1rs->scrub(_region_bm, _card_bm); 1940 } 1941 } 1942 1943 }; 1944 1945 void ConcurrentMark::cleanup() { 1946 // world is stopped at this checkpoint 1947 assert(SafepointSynchronize::is_at_safepoint(), 1948 "world should be stopped"); 1949 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 1950 1951 // If a full collection has happened, we shouldn't do this. 1952 if (has_aborted()) { 1953 g1h->set_marking_complete(); // So bitmap clearing isn't confused 1954 return; 1955 } 1956 1957 HRSPhaseSetter x(HRSPhaseCleanup); 1958 g1h->verify_region_sets_optional(); 1959 1960 if (VerifyDuringGC) { 1961 HandleMark hm; // handle scope 1962 gclog_or_tty->print(" VerifyDuringGC:(before)"); 1963 Universe::heap()->prepare_for_verify(); 1964 Universe::verify(/* silent */ false, 1965 /* option */ VerifyOption_G1UsePrevMarking); 1966 } 1967 1968 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); 1969 g1p->record_concurrent_mark_cleanup_start(); 1970 1971 double start = os::elapsedTime(); 1972 1973 HeapRegionRemSet::reset_for_cleanup_tasks(); 1974 1975 uint n_workers; 1976 1977 // Do counting once more with the world stopped for good measure. 1978 G1ParFinalCountTask g1_par_count_task(g1h, &_region_bm, &_card_bm); 1979 1980 if (G1CollectedHeap::use_parallel_gc_threads()) { 1981 assert(g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 1982 "sanity check"); 1983 1984 g1h->set_par_threads(); 1985 n_workers = g1h->n_par_threads(); 1986 assert(g1h->n_par_threads() == n_workers, 1987 "Should not have been reset"); 1988 g1h->workers()->run_task(&g1_par_count_task); 1989 // Done with the parallel phase so reset to 0. 1990 g1h->set_par_threads(0); 1991 1992 assert(g1h->check_heap_region_claim_values(HeapRegion::FinalCountClaimValue), 1993 "sanity check"); 1994 } else { 1995 n_workers = 1; 1996 g1_par_count_task.work(0); 1997 } 1998 1999 if (VerifyDuringGC) { 2000 // Verify that the counting data accumulated during marking matches 2001 // that calculated by walking the marking bitmap. 2002 2003 // Bitmaps to hold expected values 2004 BitMap expected_region_bm(_region_bm.size(), false); 2005 BitMap expected_card_bm(_card_bm.size(), false); 2006 2007 G1ParVerifyFinalCountTask g1_par_verify_task(g1h, 2008 &_region_bm, 2009 &_card_bm, 2010 &expected_region_bm, 2011 &expected_card_bm); 2012 2013 if (G1CollectedHeap::use_parallel_gc_threads()) { 2014 g1h->set_par_threads((int)n_workers); 2015 g1h->workers()->run_task(&g1_par_verify_task); 2016 // Done with the parallel phase so reset to 0. 2017 g1h->set_par_threads(0); 2018 2019 assert(g1h->check_heap_region_claim_values(HeapRegion::VerifyCountClaimValue), 2020 "sanity check"); 2021 } else { 2022 g1_par_verify_task.work(0); 2023 } 2024 2025 guarantee(g1_par_verify_task.failures() == 0, "Unexpected accounting failures"); 2026 } 2027 2028 size_t start_used_bytes = g1h->used(); 2029 g1h->set_marking_complete(); 2030 2031 double count_end = os::elapsedTime(); 2032 double this_final_counting_time = (count_end - start); 2033 _total_counting_time += this_final_counting_time; 2034 2035 if (G1PrintRegionLivenessInfo) { 2036 G1PrintRegionLivenessInfoClosure cl(gclog_or_tty, "Post-Marking"); 2037 _g1h->heap_region_iterate(&cl); 2038 } 2039 2040 // Install newly created mark bitMap as "prev". 2041 swapMarkBitMaps(); 2042 2043 g1h->reset_gc_time_stamp(); 2044 2045 // Note end of marking in all heap regions. 2046 G1ParNoteEndTask g1_par_note_end_task(g1h, &_cleanup_list); 2047 if (G1CollectedHeap::use_parallel_gc_threads()) { 2048 g1h->set_par_threads((int)n_workers); 2049 g1h->workers()->run_task(&g1_par_note_end_task); 2050 g1h->set_par_threads(0); 2051 2052 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue), 2053 "sanity check"); 2054 } else { 2055 g1_par_note_end_task.work(0); 2056 } 2057 g1h->check_gc_time_stamps(); 2058 2059 if (!cleanup_list_is_empty()) { 2060 // The cleanup list is not empty, so we'll have to process it 2061 // concurrently. Notify anyone else that might be wanting free 2062 // regions that there will be more free regions coming soon. 2063 g1h->set_free_regions_coming(); 2064 } 2065 2066 // call below, since it affects the metric by which we sort the heap 2067 // regions. 2068 if (G1ScrubRemSets) { 2069 double rs_scrub_start = os::elapsedTime(); 2070 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm); 2071 if (G1CollectedHeap::use_parallel_gc_threads()) { 2072 g1h->set_par_threads((int)n_workers); 2073 g1h->workers()->run_task(&g1_par_scrub_rs_task); 2074 g1h->set_par_threads(0); 2075 2076 assert(g1h->check_heap_region_claim_values( 2077 HeapRegion::ScrubRemSetClaimValue), 2078 "sanity check"); 2079 } else { 2080 g1_par_scrub_rs_task.work(0); 2081 } 2082 2083 double rs_scrub_end = os::elapsedTime(); 2084 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); 2085 _total_rs_scrub_time += this_rs_scrub_time; 2086 } 2087 2088 // this will also free any regions totally full of garbage objects, 2089 // and sort the regions. 2090 g1h->g1_policy()->record_concurrent_mark_cleanup_end((int)n_workers); 2091 2092 // Statistics. 2093 double end = os::elapsedTime(); 2094 _cleanup_times.add((end - start) * 1000.0); 2095 2096 if (G1Log::fine()) { 2097 g1h->print_size_transition(gclog_or_tty, 2098 start_used_bytes, 2099 g1h->used(), 2100 g1h->capacity()); 2101 } 2102 2103 // Clean up will have freed any regions completely full of garbage. 2104 // Update the soft reference policy with the new heap occupancy. 2105 Universe::update_heap_info_at_gc(); 2106 2107 // We need to make this be a "collection" so any collection pause that 2108 // races with it goes around and waits for completeCleanup to finish. 2109 g1h->increment_total_collections(); 2110 2111 // We reclaimed old regions so we should calculate the sizes to make 2112 // sure we update the old gen/space data. 2113 g1h->g1mm()->update_sizes(); 2114 2115 if (VerifyDuringGC) { 2116 HandleMark hm; // handle scope 2117 gclog_or_tty->print(" VerifyDuringGC:(after)"); 2118 Universe::heap()->prepare_for_verify(); 2119 Universe::verify(/* silent */ false, 2120 /* option */ VerifyOption_G1UsePrevMarking); 2121 } 2122 2123 g1h->verify_region_sets_optional(); 2124 } 2125 2126 void ConcurrentMark::completeCleanup() { 2127 if (has_aborted()) return; 2128 2129 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2130 2131 _cleanup_list.verify_optional(); 2132 FreeRegionList tmp_free_list("Tmp Free List"); 2133 2134 if (G1ConcRegionFreeingVerbose) { 2135 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2136 "cleanup list has %u entries", 2137 _cleanup_list.length()); 2138 } 2139 2140 // Noone else should be accessing the _cleanup_list at this point, 2141 // so it's not necessary to take any locks 2142 while (!_cleanup_list.is_empty()) { 2143 HeapRegion* hr = _cleanup_list.remove_head(); 2144 assert(hr != NULL, "the list was not empty"); 2145 hr->par_clear(); 2146 tmp_free_list.add_as_tail(hr); 2147 2148 // Instead of adding one region at a time to the secondary_free_list, 2149 // we accumulate them in the local list and move them a few at a 2150 // time. This also cuts down on the number of notify_all() calls 2151 // we do during this process. We'll also append the local list when 2152 // _cleanup_list is empty (which means we just removed the last 2153 // region from the _cleanup_list). 2154 if ((tmp_free_list.length() % G1SecondaryFreeListAppendLength == 0) || 2155 _cleanup_list.is_empty()) { 2156 if (G1ConcRegionFreeingVerbose) { 2157 gclog_or_tty->print_cr("G1ConcRegionFreeing [complete cleanup] : " 2158 "appending %u entries to the secondary_free_list, " 2159 "cleanup list still has %u entries", 2160 tmp_free_list.length(), 2161 _cleanup_list.length()); 2162 } 2163 2164 { 2165 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 2166 g1h->secondary_free_list_add_as_tail(&tmp_free_list); 2167 SecondaryFreeList_lock->notify_all(); 2168 } 2169 2170 if (G1StressConcRegionFreeing) { 2171 for (uintx i = 0; i < G1StressConcRegionFreeingDelayMillis; ++i) { 2172 os::sleep(Thread::current(), (jlong) 1, false); 2173 } 2174 } 2175 } 2176 } 2177 assert(tmp_free_list.is_empty(), "post-condition"); 2178 } 2179 2180 // Supporting Object and Oop closures for reference discovery 2181 // and processing in during marking 2182 2183 bool G1CMIsAliveClosure::do_object_b(oop obj) { 2184 HeapWord* addr = (HeapWord*)obj; 2185 return addr != NULL && 2186 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); 2187 } 2188 2189 // 'Keep Alive' oop closure used by both serial parallel reference processing. 2190 // Uses the CMTask associated with a worker thread (for serial reference 2191 // processing the CMTask for worker 0 is used) to preserve (mark) and 2192 // trace referent objects. 2193 // 2194 // Using the CMTask and embedded local queues avoids having the worker 2195 // threads operating on the global mark stack. This reduces the risk 2196 // of overflowing the stack - which we would rather avoid at this late 2197 // state. Also using the tasks' local queues removes the potential 2198 // of the workers interfering with each other that could occur if 2199 // operating on the global stack. 2200 2201 class G1CMKeepAliveAndDrainClosure: public OopClosure { 2202 ConcurrentMark* _cm; 2203 CMTask* _task; 2204 int _ref_counter_limit; 2205 int _ref_counter; 2206 public: 2207 G1CMKeepAliveAndDrainClosure(ConcurrentMark* cm, CMTask* task) : 2208 _cm(cm), _task(task), _ref_counter_limit(G1RefProcDrainInterval) { 2209 assert(_ref_counter_limit > 0, "sanity"); 2210 _ref_counter = _ref_counter_limit; 2211 } 2212 2213 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 2214 virtual void do_oop( oop* p) { do_oop_work(p); } 2215 2216 template <class T> void do_oop_work(T* p) { 2217 if (!_cm->has_overflown()) { 2218 oop obj = oopDesc::load_decode_heap_oop(p); 2219 if (_cm->verbose_high()) { 2220 gclog_or_tty->print_cr("\t[%u] we're looking at location " 2221 "*"PTR_FORMAT" = "PTR_FORMAT, 2222 _task->worker_id(), p, (void*) obj); 2223 } 2224 2225 _task->deal_with_reference(obj); 2226 _ref_counter--; 2227 2228 if (_ref_counter == 0) { 2229 // We have dealt with _ref_counter_limit references, pushing them 2230 // and objects reachable from them on to the local stack (and 2231 // possibly the global stack). Call CMTask::do_marking_step() to 2232 // process these entries. 2233 // 2234 // We call CMTask::do_marking_step() in a loop, which we'll exit if 2235 // there's nothing more to do (i.e. we're done with the entries that 2236 // were pushed as a result of the CMTask::deal_with_reference() calls 2237 // above) or we overflow. 2238 // 2239 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2240 // flag while there may still be some work to do. (See the comment at 2241 // the beginning of CMTask::do_marking_step() for those conditions - 2242 // one of which is reaching the specified time target.) It is only 2243 // when CMTask::do_marking_step() returns without setting the 2244 // has_aborted() flag that the marking step has completed. 2245 do { 2246 double mark_step_duration_ms = G1ConcMarkStepDurationMillis; 2247 _task->do_marking_step(mark_step_duration_ms, 2248 false /* do_stealing */, 2249 false /* do_termination */); 2250 } while (_task->has_aborted() && !_cm->has_overflown()); 2251 _ref_counter = _ref_counter_limit; 2252 } 2253 } else { 2254 if (_cm->verbose_high()) { 2255 gclog_or_tty->print_cr("\t[%u] CM Overflow", _task->worker_id()); 2256 } 2257 } 2258 } 2259 }; 2260 2261 // 'Drain' oop closure used by both serial and parallel reference processing. 2262 // Uses the CMTask associated with a given worker thread (for serial 2263 // reference processing the CMtask for worker 0 is used). Calls the 2264 // do_marking_step routine, with an unbelievably large timeout value, 2265 // to drain the marking data structures of the remaining entries 2266 // added by the 'keep alive' oop closure above. 2267 2268 class G1CMDrainMarkingStackClosure: public VoidClosure { 2269 ConcurrentMark* _cm; 2270 CMTask* _task; 2271 bool _do_stealing; 2272 bool _do_termination; 2273 public: 2274 G1CMDrainMarkingStackClosure(ConcurrentMark* cm, CMTask* task, bool is_par) : 2275 _cm(cm), _task(task) { 2276 assert(is_par || _task->worker_id() == 0, 2277 "Only task for worker 0 should be used if ref processing is single threaded"); 2278 // We only allow stealing and only enter the termination protocol 2279 // in CMTask::do_marking_step() if this closure is being instantiated 2280 // for parallel reference processing. 2281 _do_stealing = _do_termination = is_par; 2282 } 2283 2284 void do_void() { 2285 do { 2286 if (_cm->verbose_high()) { 2287 gclog_or_tty->print_cr("\t[%u] Drain: Calling do_marking_step - " 2288 "stealing: %s, termination: %s", 2289 _task->worker_id(), 2290 BOOL_TO_STR(_do_stealing), 2291 BOOL_TO_STR(_do_termination)); 2292 } 2293 2294 // We call CMTask::do_marking_step() to completely drain the local 2295 // and global marking stacks of entries pushed by the 'keep alive' 2296 // oop closure (an instance of G1CMKeepAliveAndDrainClosure above). 2297 // 2298 // CMTask::do_marking_step() is called in a loop, which we'll exit 2299 // if there's nothing more to do (i.e. we'completely drained the 2300 // entries that were pushed as a a result of applying the 'keep alive' 2301 // closure to the entries on the discovered ref lists) or we overflow 2302 // the global marking stack. 2303 // 2304 // Note: CMTask::do_marking_step() can set the CMTask::has_aborted() 2305 // flag while there may still be some work to do. (See the comment at 2306 // the beginning of CMTask::do_marking_step() for those conditions - 2307 // one of which is reaching the specified time target.) It is only 2308 // when CMTask::do_marking_step() returns without setting the 2309 // has_aborted() flag that the marking step has completed. 2310 2311 _task->do_marking_step(1000000000.0 /* something very large */, 2312 _do_stealing, 2313 _do_termination); 2314 } while (_task->has_aborted() && !_cm->has_overflown()); 2315 } 2316 }; 2317 2318 // Implementation of AbstractRefProcTaskExecutor for parallel 2319 // reference processing at the end of G1 concurrent marking 2320 2321 class G1CMRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 2322 private: 2323 G1CollectedHeap* _g1h; 2324 ConcurrentMark* _cm; 2325 WorkGang* _workers; 2326 int _active_workers; 2327 2328 public: 2329 G1CMRefProcTaskExecutor(G1CollectedHeap* g1h, 2330 ConcurrentMark* cm, 2331 WorkGang* workers, 2332 int n_workers) : 2333 _g1h(g1h), _cm(cm), 2334 _workers(workers), _active_workers(n_workers) { } 2335 2336 // Executes the given task using concurrent marking worker threads. 2337 virtual void execute(ProcessTask& task); 2338 virtual void execute(EnqueueTask& task); 2339 }; 2340 2341 class G1CMRefProcTaskProxy: public AbstractGangTask { 2342 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 2343 ProcessTask& _proc_task; 2344 G1CollectedHeap* _g1h; 2345 ConcurrentMark* _cm; 2346 bool _processing_is_mt; 2347 2348 public: 2349 G1CMRefProcTaskProxy(ProcessTask& proc_task, 2350 G1CollectedHeap* g1h, 2351 ConcurrentMark* cm) : 2352 AbstractGangTask("Process reference objects in parallel"), 2353 _proc_task(proc_task), _g1h(g1h), _cm(cm) { 2354 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 2355 _processing_is_mt = rp->processing_is_mt(); 2356 } 2357 2358 virtual void work(uint worker_id) { 2359 CMTask* marking_task = _cm->task(worker_id); 2360 G1CMIsAliveClosure g1_is_alive(_g1h); 2361 G1CMKeepAliveAndDrainClosure g1_par_keep_alive(_cm, marking_task); 2362 G1CMDrainMarkingStackClosure g1_par_drain(_cm, marking_task, _processing_is_mt); 2363 2364 _proc_task.work(worker_id, g1_is_alive, g1_par_keep_alive, g1_par_drain); 2365 } 2366 }; 2367 2368 void G1CMRefProcTaskExecutor::execute(ProcessTask& proc_task) { 2369 assert(_workers != NULL, "Need parallel worker threads."); 2370 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2371 2372 G1CMRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _cm); 2373 2374 // We need to reset the phase for each task execution so that 2375 // the termination protocol of CMTask::do_marking_step works. 2376 _cm->set_phase(_active_workers, false /* concurrent */); 2377 _g1h->set_par_threads(_active_workers); 2378 _workers->run_task(&proc_task_proxy); 2379 _g1h->set_par_threads(0); 2380 } 2381 2382 class G1CMRefEnqueueTaskProxy: public AbstractGangTask { 2383 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 2384 EnqueueTask& _enq_task; 2385 2386 public: 2387 G1CMRefEnqueueTaskProxy(EnqueueTask& enq_task) : 2388 AbstractGangTask("Enqueue reference objects in parallel"), 2389 _enq_task(enq_task) { } 2390 2391 virtual void work(uint worker_id) { 2392 _enq_task.work(worker_id); 2393 } 2394 }; 2395 2396 void G1CMRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 2397 assert(_workers != NULL, "Need parallel worker threads."); 2398 assert(_g1h->ref_processor_cm()->processing_is_mt(), "processing is not MT"); 2399 2400 G1CMRefEnqueueTaskProxy enq_task_proxy(enq_task); 2401 2402 _g1h->set_par_threads(_active_workers); 2403 _workers->run_task(&enq_task_proxy); 2404 _g1h->set_par_threads(0); 2405 } 2406 2407 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { 2408 if (has_overflown()) { 2409 // If we have overflown the marking stack then just return 2410 // without processing the discovered references. We will be 2411 // restarting marking because of the overflow and any 2412 // currently discovered reference will stay discovered. 2413 // They will be processed when the remark task successfully 2414 // completes. 2415 return; 2416 } 2417 2418 ResourceMark rm; 2419 HandleMark hm; 2420 2421 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2422 2423 // Is alive closure. 2424 G1CMIsAliveClosure g1_is_alive(g1h); 2425 2426 // Inner scope to exclude the cleaning of the string and symbol 2427 // tables from the displayed time. 2428 { 2429 if (G1Log::finer()) { 2430 gclog_or_tty->put(' '); 2431 } 2432 TraceTime t("GC ref-proc", G1Log::finer(), false, gclog_or_tty); 2433 2434 ReferenceProcessor* rp = g1h->ref_processor_cm(); 2435 2436 // See the comment in G1CollectedHeap::ref_processing_init() 2437 // about how reference processing currently works in G1. 2438 2439 // Set the soft reference policy 2440 rp->setup_policy(clear_all_soft_refs); 2441 assert(_markStack.isEmpty(), "mark stack should be empty"); 2442 2443 // Non-MT instances 'Keep Alive' and 'Complete GC' oop closures. 2444 G1CMKeepAliveAndDrainClosure g1_keep_alive(this, task(0)); 2445 G1CMDrainMarkingStackClosure g1_drain_mark_stack(this, task(0), false); 2446 2447 // We need at least one active thread. If reference processing is 2448 // not multi-threaded we use the current (ConcurrentMarkThread) thread, 2449 // otherwise we use the work gang from the G1CollectedHeap and we 2450 // utilize all the worker threads we can. 2451 uint active_workers = (rp->processing_is_mt() && g1h->workers() != NULL 2452 ? g1h->workers()->active_workers() 2453 : 1U); 2454 2455 active_workers = MAX2(MIN2(active_workers, _max_worker_id), 1U); 2456 2457 G1CMRefProcTaskExecutor par_task_executor(g1h, this, 2458 g1h->workers(), active_workers); 2459 2460 AbstractRefProcTaskExecutor* executor = (rp->processing_is_mt() 2461 ? &par_task_executor 2462 : NULL); 2463 2464 // Set the degree of MT processing here. If the discovery was done MT, 2465 // the number of threads involved during discovery could differ from 2466 // the number of active workers. This is OK as long as the discovered 2467 // Reference lists are balanced (see balance_all_queues() and balance_queues()). 2468 rp->set_active_mt_degree(active_workers); 2469 2470 // Process the weak references. 2471 rp->process_discovered_references(&g1_is_alive, 2472 &g1_keep_alive, 2473 &g1_drain_mark_stack, 2474 executor); 2475 2476 // The do_oop work routines of the keep_alive and drain_marking_stack 2477 // oop closures will set the has_overflown flag if we overflow the 2478 // global marking stack. 2479 2480 assert(_markStack.overflow() || _markStack.isEmpty(), 2481 "mark stack should be empty (unless it overflowed)"); 2482 if (_markStack.overflow()) { 2483 // This should have been done already when we tried to push an 2484 // entry on to the global mark stack. But let's do it again. 2485 set_has_overflown(); 2486 } 2487 2488 assert(rp->num_q() == active_workers, "why not"); 2489 2490 rp->enqueue_discovered_references(executor); 2491 2492 rp->verify_no_references_recorded(); 2493 assert(!rp->discovery_enabled(), "Post condition"); 2494 } 2495 2496 // Now clean up stale oops in StringTable 2497 StringTable::unlink(&g1_is_alive); 2498 // Clean up unreferenced symbols in symbol table. 2499 SymbolTable::unlink(); 2500 } 2501 2502 void ConcurrentMark::swapMarkBitMaps() { 2503 CMBitMapRO* temp = _prevMarkBitMap; 2504 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; 2505 _nextMarkBitMap = (CMBitMap*) temp; 2506 } 2507 2508 class CMRemarkTask: public AbstractGangTask { 2509 private: 2510 ConcurrentMark *_cm; 2511 2512 public: 2513 void work(uint worker_id) { 2514 // Since all available tasks are actually started, we should 2515 // only proceed if we're supposed to be actived. 2516 if (worker_id < _cm->active_tasks()) { 2517 CMTask* task = _cm->task(worker_id); 2518 task->record_start_time(); 2519 do { 2520 task->do_marking_step(1000000000.0 /* something very large */, 2521 true /* do_stealing */, 2522 true /* do_termination */); 2523 } while (task->has_aborted() && !_cm->has_overflown()); 2524 // If we overflow, then we do not want to restart. We instead 2525 // want to abort remark and do concurrent marking again. 2526 task->record_end_time(); 2527 } 2528 } 2529 2530 CMRemarkTask(ConcurrentMark* cm, int active_workers) : 2531 AbstractGangTask("Par Remark"), _cm(cm) { 2532 _cm->terminator()->reset_for_reuse(active_workers); 2533 } 2534 }; 2535 2536 void ConcurrentMark::checkpointRootsFinalWork() { 2537 ResourceMark rm; 2538 HandleMark hm; 2539 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 2540 2541 g1h->ensure_parsability(false); 2542 2543 if (G1CollectedHeap::use_parallel_gc_threads()) { 2544 G1CollectedHeap::StrongRootsScope srs(g1h); 2545 // this is remark, so we'll use up all active threads 2546 uint active_workers = g1h->workers()->active_workers(); 2547 if (active_workers == 0) { 2548 assert(active_workers > 0, "Should have been set earlier"); 2549 active_workers = (uint) ParallelGCThreads; 2550 g1h->workers()->set_active_workers(active_workers); 2551 } 2552 set_phase(active_workers, false /* concurrent */); 2553 // Leave _parallel_marking_threads at it's 2554 // value originally calculated in the ConcurrentMark 2555 // constructor and pass values of the active workers 2556 // through the gang in the task. 2557 2558 CMRemarkTask remarkTask(this, active_workers); 2559 g1h->set_par_threads(active_workers); 2560 g1h->workers()->run_task(&remarkTask); 2561 g1h->set_par_threads(0); 2562 } else { 2563 G1CollectedHeap::StrongRootsScope srs(g1h); 2564 // this is remark, so we'll use up all available threads 2565 uint active_workers = 1; 2566 set_phase(active_workers, false /* concurrent */); 2567 2568 CMRemarkTask remarkTask(this, active_workers); 2569 // We will start all available threads, even if we decide that the 2570 // active_workers will be fewer. The extra ones will just bail out 2571 // immediately. 2572 remarkTask.work(0); 2573 } 2574 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 2575 guarantee(has_overflown() || 2576 satb_mq_set.completed_buffers_num() == 0, 2577 err_msg("Invariant: has_overflown = %s, num buffers = %d", 2578 BOOL_TO_STR(has_overflown()), 2579 satb_mq_set.completed_buffers_num())); 2580 2581 print_stats(); 2582 2583 #if VERIFY_OBJS_PROCESSED 2584 if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) { 2585 gclog_or_tty->print_cr("Processed = %d, enqueued = %d.", 2586 _scan_obj_cl.objs_processed, 2587 ThreadLocalObjQueue::objs_enqueued); 2588 guarantee(_scan_obj_cl.objs_processed == 2589 ThreadLocalObjQueue::objs_enqueued, 2590 "Different number of objs processed and enqueued."); 2591 } 2592 #endif 2593 } 2594 2595 #ifndef PRODUCT 2596 2597 class PrintReachableOopClosure: public OopClosure { 2598 private: 2599 G1CollectedHeap* _g1h; 2600 outputStream* _out; 2601 VerifyOption _vo; 2602 bool _all; 2603 2604 public: 2605 PrintReachableOopClosure(outputStream* out, 2606 VerifyOption vo, 2607 bool all) : 2608 _g1h(G1CollectedHeap::heap()), 2609 _out(out), _vo(vo), _all(all) { } 2610 2611 void do_oop(narrowOop* p) { do_oop_work(p); } 2612 void do_oop( oop* p) { do_oop_work(p); } 2613 2614 template <class T> void do_oop_work(T* p) { 2615 oop obj = oopDesc::load_decode_heap_oop(p); 2616 const char* str = NULL; 2617 const char* str2 = ""; 2618 2619 if (obj == NULL) { 2620 str = ""; 2621 } else if (!_g1h->is_in_g1_reserved(obj)) { 2622 str = " O"; 2623 } else { 2624 HeapRegion* hr = _g1h->heap_region_containing(obj); 2625 guarantee(hr != NULL, "invariant"); 2626 bool over_tams = _g1h->allocated_since_marking(obj, hr, _vo); 2627 bool marked = _g1h->is_marked(obj, _vo); 2628 2629 if (over_tams) { 2630 str = " >"; 2631 if (marked) { 2632 str2 = " AND MARKED"; 2633 } 2634 } else if (marked) { 2635 str = " M"; 2636 } else { 2637 str = " NOT"; 2638 } 2639 } 2640 2641 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s", 2642 p, (void*) obj, str, str2); 2643 } 2644 }; 2645 2646 class PrintReachableObjectClosure : public ObjectClosure { 2647 private: 2648 G1CollectedHeap* _g1h; 2649 outputStream* _out; 2650 VerifyOption _vo; 2651 bool _all; 2652 HeapRegion* _hr; 2653 2654 public: 2655 PrintReachableObjectClosure(outputStream* out, 2656 VerifyOption vo, 2657 bool all, 2658 HeapRegion* hr) : 2659 _g1h(G1CollectedHeap::heap()), 2660 _out(out), _vo(vo), _all(all), _hr(hr) { } 2661 2662 void do_object(oop o) { 2663 bool over_tams = _g1h->allocated_since_marking(o, _hr, _vo); 2664 bool marked = _g1h->is_marked(o, _vo); 2665 bool print_it = _all || over_tams || marked; 2666 2667 if (print_it) { 2668 _out->print_cr(" "PTR_FORMAT"%s", 2669 o, (over_tams) ? " >" : (marked) ? " M" : ""); 2670 PrintReachableOopClosure oopCl(_out, _vo, _all); 2671 o->oop_iterate_no_header(&oopCl); 2672 } 2673 } 2674 }; 2675 2676 class PrintReachableRegionClosure : public HeapRegionClosure { 2677 private: 2678 G1CollectedHeap* _g1h; 2679 outputStream* _out; 2680 VerifyOption _vo; 2681 bool _all; 2682 2683 public: 2684 bool doHeapRegion(HeapRegion* hr) { 2685 HeapWord* b = hr->bottom(); 2686 HeapWord* e = hr->end(); 2687 HeapWord* t = hr->top(); 2688 HeapWord* p = _g1h->top_at_mark_start(hr, _vo); 2689 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" " 2690 "TAMS: "PTR_FORMAT, b, e, t, p); 2691 _out->cr(); 2692 2693 HeapWord* from = b; 2694 HeapWord* to = t; 2695 2696 if (to > from) { 2697 _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to); 2698 _out->cr(); 2699 PrintReachableObjectClosure ocl(_out, _vo, _all, hr); 2700 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl); 2701 _out->cr(); 2702 } 2703 2704 return false; 2705 } 2706 2707 PrintReachableRegionClosure(outputStream* out, 2708 VerifyOption vo, 2709 bool all) : 2710 _g1h(G1CollectedHeap::heap()), _out(out), _vo(vo), _all(all) { } 2711 }; 2712 2713 void ConcurrentMark::print_reachable(const char* str, 2714 VerifyOption vo, 2715 bool all) { 2716 gclog_or_tty->cr(); 2717 gclog_or_tty->print_cr("== Doing heap dump... "); 2718 2719 if (G1PrintReachableBaseFile == NULL) { 2720 gclog_or_tty->print_cr(" #### error: no base file defined"); 2721 return; 2722 } 2723 2724 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) > 2725 (JVM_MAXPATHLEN - 1)) { 2726 gclog_or_tty->print_cr(" #### error: file name too long"); 2727 return; 2728 } 2729 2730 char file_name[JVM_MAXPATHLEN]; 2731 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str); 2732 gclog_or_tty->print_cr(" dumping to file %s", file_name); 2733 2734 fileStream fout(file_name); 2735 if (!fout.is_open()) { 2736 gclog_or_tty->print_cr(" #### error: could not open file"); 2737 return; 2738 } 2739 2740 outputStream* out = &fout; 2741 out->print_cr("-- USING %s", _g1h->top_at_mark_start_str(vo)); 2742 out->cr(); 2743 2744 out->print_cr("--- ITERATING OVER REGIONS"); 2745 out->cr(); 2746 PrintReachableRegionClosure rcl(out, vo, all); 2747 _g1h->heap_region_iterate(&rcl); 2748 out->cr(); 2749 2750 gclog_or_tty->print_cr(" done"); 2751 gclog_or_tty->flush(); 2752 } 2753 2754 #endif // PRODUCT 2755 2756 void ConcurrentMark::clearRangePrevBitmap(MemRegion mr) { 2757 // Note we are overriding the read-only view of the prev map here, via 2758 // the cast. 2759 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); 2760 } 2761 2762 void ConcurrentMark::clearRangeNextBitmap(MemRegion mr) { 2763 _nextMarkBitMap->clearRange(mr); 2764 } 2765 2766 void ConcurrentMark::clearRangeBothBitmaps(MemRegion mr) { 2767 clearRangePrevBitmap(mr); 2768 clearRangeNextBitmap(mr); 2769 } 2770 2771 HeapRegion* 2772 ConcurrentMark::claim_region(uint worker_id) { 2773 // "checkpoint" the finger 2774 HeapWord* finger = _finger; 2775 2776 // _heap_end will not change underneath our feet; it only changes at 2777 // yield points. 2778 while (finger < _heap_end) { 2779 assert(_g1h->is_in_g1_reserved(finger), "invariant"); 2780 2781 // Note on how this code handles humongous regions. In the 2782 // normal case the finger will reach the start of a "starts 2783 // humongous" (SH) region. Its end will either be the end of the 2784 // last "continues humongous" (CH) region in the sequence, or the 2785 // standard end of the SH region (if the SH is the only region in 2786 // the sequence). That way claim_region() will skip over the CH 2787 // regions. However, there is a subtle race between a CM thread 2788 // executing this method and a mutator thread doing a humongous 2789 // object allocation. The two are not mutually exclusive as the CM 2790 // thread does not need to hold the Heap_lock when it gets 2791 // here. So there is a chance that claim_region() will come across 2792 // a free region that's in the progress of becoming a SH or a CH 2793 // region. In the former case, it will either 2794 // a) Miss the update to the region's end, in which case it will 2795 // visit every subsequent CH region, will find their bitmaps 2796 // empty, and do nothing, or 2797 // b) Will observe the update of the region's end (in which case 2798 // it will skip the subsequent CH regions). 2799 // If it comes across a region that suddenly becomes CH, the 2800 // scenario will be similar to b). So, the race between 2801 // claim_region() and a humongous object allocation might force us 2802 // to do a bit of unnecessary work (due to some unnecessary bitmap 2803 // iterations) but it should not introduce and correctness issues. 2804 HeapRegion* curr_region = _g1h->heap_region_containing_raw(finger); 2805 HeapWord* bottom = curr_region->bottom(); 2806 HeapWord* end = curr_region->end(); 2807 HeapWord* limit = curr_region->next_top_at_mark_start(); 2808 2809 if (verbose_low()) { 2810 gclog_or_tty->print_cr("[%u] curr_region = "PTR_FORMAT" " 2811 "["PTR_FORMAT", "PTR_FORMAT"), " 2812 "limit = "PTR_FORMAT, 2813 worker_id, curr_region, bottom, end, limit); 2814 } 2815 2816 // Is the gap between reading the finger and doing the CAS too long? 2817 HeapWord* res = (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); 2818 if (res == finger) { 2819 // we succeeded 2820 2821 // notice that _finger == end cannot be guaranteed here since, 2822 // someone else might have moved the finger even further 2823 assert(_finger >= end, "the finger should have moved forward"); 2824 2825 if (verbose_low()) { 2826 gclog_or_tty->print_cr("[%u] we were successful with region = " 2827 PTR_FORMAT, worker_id, curr_region); 2828 } 2829 2830 if (limit > bottom) { 2831 if (verbose_low()) { 2832 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is not empty, " 2833 "returning it ", worker_id, curr_region); 2834 } 2835 return curr_region; 2836 } else { 2837 assert(limit == bottom, 2838 "the region limit should be at bottom"); 2839 if (verbose_low()) { 2840 gclog_or_tty->print_cr("[%u] region "PTR_FORMAT" is empty, " 2841 "returning NULL", worker_id, curr_region); 2842 } 2843 // we return NULL and the caller should try calling 2844 // claim_region() again. 2845 return NULL; 2846 } 2847 } else { 2848 assert(_finger > finger, "the finger should have moved forward"); 2849 if (verbose_low()) { 2850 gclog_or_tty->print_cr("[%u] somebody else moved the finger, " 2851 "global finger = "PTR_FORMAT", " 2852 "our finger = "PTR_FORMAT, 2853 worker_id, _finger, finger); 2854 } 2855 2856 // read it again 2857 finger = _finger; 2858 } 2859 } 2860 2861 return NULL; 2862 } 2863 2864 #ifndef PRODUCT 2865 enum VerifyNoCSetOopsPhase { 2866 VerifyNoCSetOopsStack, 2867 VerifyNoCSetOopsQueues, 2868 VerifyNoCSetOopsSATBCompleted, 2869 VerifyNoCSetOopsSATBThread 2870 }; 2871 2872 class VerifyNoCSetOopsClosure : public OopClosure, public ObjectClosure { 2873 private: 2874 G1CollectedHeap* _g1h; 2875 VerifyNoCSetOopsPhase _phase; 2876 int _info; 2877 2878 const char* phase_str() { 2879 switch (_phase) { 2880 case VerifyNoCSetOopsStack: return "Stack"; 2881 case VerifyNoCSetOopsQueues: return "Queue"; 2882 case VerifyNoCSetOopsSATBCompleted: return "Completed SATB Buffers"; 2883 case VerifyNoCSetOopsSATBThread: return "Thread SATB Buffers"; 2884 default: ShouldNotReachHere(); 2885 } 2886 return NULL; 2887 } 2888 2889 void do_object_work(oop obj) { 2890 guarantee(!_g1h->obj_in_cs(obj), 2891 err_msg("obj: "PTR_FORMAT" in CSet, phase: %s, info: %d", 2892 (void*) obj, phase_str(), _info)); 2893 } 2894 2895 public: 2896 VerifyNoCSetOopsClosure() : _g1h(G1CollectedHeap::heap()) { } 2897 2898 void set_phase(VerifyNoCSetOopsPhase phase, int info = -1) { 2899 _phase = phase; 2900 _info = info; 2901 } 2902 2903 virtual void do_oop(oop* p) { 2904 oop obj = oopDesc::load_decode_heap_oop(p); 2905 do_object_work(obj); 2906 } 2907 2908 virtual void do_oop(narrowOop* p) { 2909 // We should not come across narrow oops while scanning marking 2910 // stacks and SATB buffers. 2911 ShouldNotReachHere(); 2912 } 2913 2914 virtual void do_object(oop obj) { 2915 do_object_work(obj); 2916 } 2917 }; 2918 2919 void ConcurrentMark::verify_no_cset_oops(bool verify_stacks, 2920 bool verify_enqueued_buffers, 2921 bool verify_thread_buffers, 2922 bool verify_fingers) { 2923 assert(SafepointSynchronize::is_at_safepoint(), "should be at a safepoint"); 2924 if (!G1CollectedHeap::heap()->mark_in_progress()) { 2925 return; 2926 } 2927 2928 VerifyNoCSetOopsClosure cl; 2929 2930 if (verify_stacks) { 2931 // Verify entries on the global mark stack 2932 cl.set_phase(VerifyNoCSetOopsStack); 2933 _markStack.oops_do(&cl); 2934 2935 // Verify entries on the task queues 2936 for (uint i = 0; i < _max_worker_id; i += 1) { 2937 cl.set_phase(VerifyNoCSetOopsQueues, i); 2938 CMTaskQueue* queue = _task_queues->queue(i); 2939 queue->oops_do(&cl); 2940 } 2941 } 2942 2943 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); 2944 2945 // Verify entries on the enqueued SATB buffers 2946 if (verify_enqueued_buffers) { 2947 cl.set_phase(VerifyNoCSetOopsSATBCompleted); 2948 satb_qs.iterate_completed_buffers_read_only(&cl); 2949 } 2950 2951 // Verify entries on the per-thread SATB buffers 2952 if (verify_thread_buffers) { 2953 cl.set_phase(VerifyNoCSetOopsSATBThread); 2954 satb_qs.iterate_thread_buffers_read_only(&cl); 2955 } 2956 2957 if (verify_fingers) { 2958 // Verify the global finger 2959 HeapWord* global_finger = finger(); 2960 if (global_finger != NULL && global_finger < _heap_end) { 2961 // The global finger always points to a heap region boundary. We 2962 // use heap_region_containing_raw() to get the containing region 2963 // given that the global finger could be pointing to a free region 2964 // which subsequently becomes continues humongous. If that 2965 // happens, heap_region_containing() will return the bottom of the 2966 // corresponding starts humongous region and the check below will 2967 // not hold any more. 2968 HeapRegion* global_hr = _g1h->heap_region_containing_raw(global_finger); 2969 guarantee(global_finger == global_hr->bottom(), 2970 err_msg("global finger: "PTR_FORMAT" region: "HR_FORMAT, 2971 global_finger, HR_FORMAT_PARAMS(global_hr))); 2972 } 2973 2974 // Verify the task fingers 2975 assert(parallel_marking_threads() <= _max_worker_id, "sanity"); 2976 for (int i = 0; i < (int) parallel_marking_threads(); i += 1) { 2977 CMTask* task = _tasks[i]; 2978 HeapWord* task_finger = task->finger(); 2979 if (task_finger != NULL && task_finger < _heap_end) { 2980 // See above note on the global finger verification. 2981 HeapRegion* task_hr = _g1h->heap_region_containing_raw(task_finger); 2982 guarantee(task_finger == task_hr->bottom() || 2983 !task_hr->in_collection_set(), 2984 err_msg("task finger: "PTR_FORMAT" region: "HR_FORMAT, 2985 task_finger, HR_FORMAT_PARAMS(task_hr))); 2986 } 2987 } 2988 } 2989 } 2990 #endif // PRODUCT 2991 2992 // Aggregate the counting data that was constructed concurrently 2993 // with marking. 2994 class AggregateCountDataHRClosure: public HeapRegionClosure { 2995 G1CollectedHeap* _g1h; 2996 ConcurrentMark* _cm; 2997 CardTableModRefBS* _ct_bs; 2998 BitMap* _cm_card_bm; 2999 uint _max_worker_id; 3000 3001 public: 3002 AggregateCountDataHRClosure(G1CollectedHeap* g1h, 3003 BitMap* cm_card_bm, 3004 uint max_worker_id) : 3005 _g1h(g1h), _cm(g1h->concurrent_mark()), 3006 _ct_bs((CardTableModRefBS*) (g1h->barrier_set())), 3007 _cm_card_bm(cm_card_bm), _max_worker_id(max_worker_id) { } 3008 3009 bool doHeapRegion(HeapRegion* hr) { 3010 if (hr->continuesHumongous()) { 3011 // We will ignore these here and process them when their 3012 // associated "starts humongous" region is processed. 3013 // Note that we cannot rely on their associated 3014 // "starts humongous" region to have their bit set to 1 3015 // since, due to the region chunking in the parallel region 3016 // iteration, a "continues humongous" region might be visited 3017 // before its associated "starts humongous". 3018 return false; 3019 } 3020 3021 HeapWord* start = hr->bottom(); 3022 HeapWord* limit = hr->next_top_at_mark_start(); 3023 HeapWord* end = hr->end(); 3024 3025 assert(start <= limit && limit <= hr->top() && hr->top() <= hr->end(), 3026 err_msg("Preconditions not met - " 3027 "start: "PTR_FORMAT", limit: "PTR_FORMAT", " 3028 "top: "PTR_FORMAT", end: "PTR_FORMAT, 3029 start, limit, hr->top(), hr->end())); 3030 3031 assert(hr->next_marked_bytes() == 0, "Precondition"); 3032 3033 if (start == limit) { 3034 // NTAMS of this region has not been set so nothing to do. 3035 return false; 3036 } 3037 3038 // 'start' should be in the heap. 3039 assert(_g1h->is_in_g1_reserved(start) && _ct_bs->is_card_aligned(start), "sanity"); 3040 // 'end' *may* be just beyone the end of the heap (if hr is the last region) 3041 assert(!_g1h->is_in_g1_reserved(end) || _ct_bs->is_card_aligned(end), "sanity"); 3042 3043 BitMap::idx_t start_idx = _cm->card_bitmap_index_for(start); 3044 BitMap::idx_t limit_idx = _cm->card_bitmap_index_for(limit); 3045 BitMap::idx_t end_idx = _cm->card_bitmap_index_for(end); 3046 3047 // If ntams is not card aligned then we bump card bitmap index 3048 // for limit so that we get the all the cards spanned by 3049 // the object ending at ntams. 3050 // Note: if this is the last region in the heap then ntams 3051 // could be actually just beyond the end of the the heap; 3052 // limit_idx will then correspond to a (non-existent) card 3053 // that is also outside the heap. 3054 if (_g1h->is_in_g1_reserved(limit) && !_ct_bs->is_card_aligned(limit)) { 3055 limit_idx += 1; 3056 } 3057 3058 assert(limit_idx <= end_idx, "or else use atomics"); 3059 3060 // Aggregate the "stripe" in the count data associated with hr. 3061 uint hrs_index = hr->hrs_index(); 3062 size_t marked_bytes = 0; 3063 3064 for (uint i = 0; i < _max_worker_id; i += 1) { 3065 size_t* marked_bytes_array = _cm->count_marked_bytes_array_for(i); 3066 BitMap* task_card_bm = _cm->count_card_bitmap_for(i); 3067 3068 // Fetch the marked_bytes in this region for task i and 3069 // add it to the running total for this region. 3070 marked_bytes += marked_bytes_array[hrs_index]; 3071 3072 // Now union the bitmaps[0,max_worker_id)[start_idx..limit_idx) 3073 // into the global card bitmap. 3074 BitMap::idx_t scan_idx = task_card_bm->get_next_one_offset(start_idx, limit_idx); 3075 3076 while (scan_idx < limit_idx) { 3077 assert(task_card_bm->at(scan_idx) == true, "should be"); 3078 _cm_card_bm->set_bit(scan_idx); 3079 assert(_cm_card_bm->at(scan_idx) == true, "should be"); 3080 3081 // BitMap::get_next_one_offset() can handle the case when 3082 // its left_offset parameter is greater than its right_offset 3083 // parameter. It does, however, have an early exit if 3084 // left_offset == right_offset. So let's limit the value 3085 // passed in for left offset here. 3086 BitMap::idx_t next_idx = MIN2(scan_idx + 1, limit_idx); 3087 scan_idx = task_card_bm->get_next_one_offset(next_idx, limit_idx); 3088 } 3089 } 3090 3091 // Update the marked bytes for this region. 3092 hr->add_to_marked_bytes(marked_bytes); 3093 3094 // Next heap region 3095 return false; 3096 } 3097 }; 3098 3099 class G1AggregateCountDataTask: public AbstractGangTask { 3100 protected: 3101 G1CollectedHeap* _g1h; 3102 ConcurrentMark* _cm; 3103 BitMap* _cm_card_bm; 3104 uint _max_worker_id; 3105 int _active_workers; 3106 3107 public: 3108 G1AggregateCountDataTask(G1CollectedHeap* g1h, 3109 ConcurrentMark* cm, 3110 BitMap* cm_card_bm, 3111 uint max_worker_id, 3112 int n_workers) : 3113 AbstractGangTask("Count Aggregation"), 3114 _g1h(g1h), _cm(cm), _cm_card_bm(cm_card_bm), 3115 _max_worker_id(max_worker_id), 3116 _active_workers(n_workers) { } 3117 3118 void work(uint worker_id) { 3119 AggregateCountDataHRClosure cl(_g1h, _cm_card_bm, _max_worker_id); 3120 3121 if (G1CollectedHeap::use_parallel_gc_threads()) { 3122 _g1h->heap_region_par_iterate_chunked(&cl, worker_id, 3123 _active_workers, 3124 HeapRegion::AggregateCountClaimValue); 3125 } else { 3126 _g1h->heap_region_iterate(&cl); 3127 } 3128 } 3129 }; 3130 3131 3132 void ConcurrentMark::aggregate_count_data() { 3133 int n_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3134 _g1h->workers()->active_workers() : 3135 1); 3136 3137 G1AggregateCountDataTask g1_par_agg_task(_g1h, this, &_card_bm, 3138 _max_worker_id, n_workers); 3139 3140 if (G1CollectedHeap::use_parallel_gc_threads()) { 3141 assert(_g1h->check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3142 "sanity check"); 3143 _g1h->set_par_threads(n_workers); 3144 _g1h->workers()->run_task(&g1_par_agg_task); 3145 _g1h->set_par_threads(0); 3146 3147 assert(_g1h->check_heap_region_claim_values(HeapRegion::AggregateCountClaimValue), 3148 "sanity check"); 3149 _g1h->reset_heap_region_claim_values(); 3150 } else { 3151 g1_par_agg_task.work(0); 3152 } 3153 } 3154 3155 // Clear the per-worker arrays used to store the per-region counting data 3156 void ConcurrentMark::clear_all_count_data() { 3157 // Clear the global card bitmap - it will be filled during 3158 // liveness count aggregation (during remark) and the 3159 // final counting task. 3160 _card_bm.clear(); 3161 3162 // Clear the global region bitmap - it will be filled as part 3163 // of the final counting task. 3164 _region_bm.clear(); 3165 3166 uint max_regions = _g1h->max_regions(); 3167 assert(_max_worker_id > 0, "uninitialized"); 3168 3169 for (uint i = 0; i < _max_worker_id; i += 1) { 3170 BitMap* task_card_bm = count_card_bitmap_for(i); 3171 size_t* marked_bytes_array = count_marked_bytes_array_for(i); 3172 3173 assert(task_card_bm->size() == _card_bm.size(), "size mismatch"); 3174 assert(marked_bytes_array != NULL, "uninitialized"); 3175 3176 memset(marked_bytes_array, 0, (size_t) max_regions * sizeof(size_t)); 3177 task_card_bm->clear(); 3178 } 3179 } 3180 3181 void ConcurrentMark::print_stats() { 3182 if (verbose_stats()) { 3183 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3184 for (size_t i = 0; i < _active_tasks; ++i) { 3185 _tasks[i]->print_stats(); 3186 gclog_or_tty->print_cr("---------------------------------------------------------------------"); 3187 } 3188 } 3189 } 3190 3191 // abandon current marking iteration due to a Full GC 3192 void ConcurrentMark::abort() { 3193 // Clear all marks to force marking thread to do nothing 3194 _nextMarkBitMap->clearAll(); 3195 // Clear the liveness counting data 3196 clear_all_count_data(); 3197 // Empty mark stack 3198 reset_marking_state(); 3199 for (uint i = 0; i < _max_worker_id; ++i) { 3200 _tasks[i]->clear_region_fields(); 3201 } 3202 _has_aborted = true; 3203 3204 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3205 satb_mq_set.abandon_partial_marking(); 3206 // This can be called either during or outside marking, we'll read 3207 // the expected_active value from the SATB queue set. 3208 satb_mq_set.set_active_all_threads( 3209 false, /* new active value */ 3210 satb_mq_set.is_active() /* expected_active */); 3211 } 3212 3213 static void print_ms_time_info(const char* prefix, const char* name, 3214 NumberSeq& ns) { 3215 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", 3216 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); 3217 if (ns.num() > 0) { 3218 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]", 3219 prefix, ns.sd(), ns.maximum()); 3220 } 3221 } 3222 3223 void ConcurrentMark::print_summary_info() { 3224 gclog_or_tty->print_cr(" Concurrent marking:"); 3225 print_ms_time_info(" ", "init marks", _init_times); 3226 print_ms_time_info(" ", "remarks", _remark_times); 3227 { 3228 print_ms_time_info(" ", "final marks", _remark_mark_times); 3229 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); 3230 3231 } 3232 print_ms_time_info(" ", "cleanups", _cleanup_times); 3233 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).", 3234 _total_counting_time, 3235 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / 3236 (double)_cleanup_times.num() 3237 : 0.0)); 3238 if (G1ScrubRemSets) { 3239 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).", 3240 _total_rs_scrub_time, 3241 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / 3242 (double)_cleanup_times.num() 3243 : 0.0)); 3244 } 3245 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.", 3246 (_init_times.sum() + _remark_times.sum() + 3247 _cleanup_times.sum())/1000.0); 3248 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s " 3249 "(%8.2f s marking).", 3250 cmThread()->vtime_accum(), 3251 cmThread()->vtime_mark_accum()); 3252 } 3253 3254 void ConcurrentMark::print_worker_threads_on(outputStream* st) const { 3255 if (use_parallel_marking_threads()) { 3256 _parallel_workers->print_worker_threads_on(st); 3257 } 3258 } 3259 3260 // We take a break if someone is trying to stop the world. 3261 bool ConcurrentMark::do_yield_check(uint worker_id) { 3262 if (should_yield()) { 3263 if (worker_id == 0) { 3264 _g1h->g1_policy()->record_concurrent_pause(); 3265 } 3266 cmThread()->yield(); 3267 return true; 3268 } else { 3269 return false; 3270 } 3271 } 3272 3273 bool ConcurrentMark::should_yield() { 3274 return cmThread()->should_yield(); 3275 } 3276 3277 bool ConcurrentMark::containing_card_is_marked(void* p) { 3278 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1); 3279 return _card_bm.at(offset >> CardTableModRefBS::card_shift); 3280 } 3281 3282 bool ConcurrentMark::containing_cards_are_marked(void* start, 3283 void* last) { 3284 return containing_card_is_marked(start) && 3285 containing_card_is_marked(last); 3286 } 3287 3288 #ifndef PRODUCT 3289 // for debugging purposes 3290 void ConcurrentMark::print_finger() { 3291 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT, 3292 _heap_start, _heap_end, _finger); 3293 for (uint i = 0; i < _max_worker_id; ++i) { 3294 gclog_or_tty->print(" %u: "PTR_FORMAT, i, _tasks[i]->finger()); 3295 } 3296 gclog_or_tty->print_cr(""); 3297 } 3298 #endif 3299 3300 void CMTask::scan_object(oop obj) { 3301 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant"); 3302 3303 if (_cm->verbose_high()) { 3304 gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT, 3305 _worker_id, (void*) obj); 3306 } 3307 3308 size_t obj_size = obj->size(); 3309 _words_scanned += obj_size; 3310 3311 obj->oop_iterate(_cm_oop_closure); 3312 statsOnly( ++_objs_scanned ); 3313 check_limits(); 3314 } 3315 3316 // Closure for iteration over bitmaps 3317 class CMBitMapClosure : public BitMapClosure { 3318 private: 3319 // the bitmap that is being iterated over 3320 CMBitMap* _nextMarkBitMap; 3321 ConcurrentMark* _cm; 3322 CMTask* _task; 3323 3324 public: 3325 CMBitMapClosure(CMTask *task, ConcurrentMark* cm, CMBitMap* nextMarkBitMap) : 3326 _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } 3327 3328 bool do_bit(size_t offset) { 3329 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); 3330 assert(_nextMarkBitMap->isMarked(addr), "invariant"); 3331 assert( addr < _cm->finger(), "invariant"); 3332 3333 statsOnly( _task->increase_objs_found_on_bitmap() ); 3334 assert(addr >= _task->finger(), "invariant"); 3335 3336 // We move that task's local finger along. 3337 _task->move_finger_to(addr); 3338 3339 _task->scan_object(oop(addr)); 3340 // we only partially drain the local queue and global stack 3341 _task->drain_local_queue(true); 3342 _task->drain_global_stack(true); 3343 3344 // if the has_aborted flag has been raised, we need to bail out of 3345 // the iteration 3346 return !_task->has_aborted(); 3347 } 3348 }; 3349 3350 // Closure for iterating over objects, currently only used for 3351 // processing SATB buffers. 3352 class CMObjectClosure : public ObjectClosure { 3353 private: 3354 CMTask* _task; 3355 3356 public: 3357 void do_object(oop obj) { 3358 _task->deal_with_reference(obj); 3359 } 3360 3361 CMObjectClosure(CMTask* task) : _task(task) { } 3362 }; 3363 3364 G1CMOopClosure::G1CMOopClosure(G1CollectedHeap* g1h, 3365 ConcurrentMark* cm, 3366 CMTask* task) 3367 : _g1h(g1h), _cm(cm), _task(task) { 3368 assert(_ref_processor == NULL, "should be initialized to NULL"); 3369 3370 if (G1UseConcMarkReferenceProcessing) { 3371 _ref_processor = g1h->ref_processor_cm(); 3372 assert(_ref_processor != NULL, "should not be NULL"); 3373 } 3374 } 3375 3376 void CMTask::setup_for_region(HeapRegion* hr) { 3377 // Separated the asserts so that we know which one fires. 3378 assert(hr != NULL, 3379 "claim_region() should have filtered out continues humongous regions"); 3380 assert(!hr->continuesHumongous(), 3381 "claim_region() should have filtered out continues humongous regions"); 3382 3383 if (_cm->verbose_low()) { 3384 gclog_or_tty->print_cr("[%u] setting up for region "PTR_FORMAT, 3385 _worker_id, hr); 3386 } 3387 3388 _curr_region = hr; 3389 _finger = hr->bottom(); 3390 update_region_limit(); 3391 } 3392 3393 void CMTask::update_region_limit() { 3394 HeapRegion* hr = _curr_region; 3395 HeapWord* bottom = hr->bottom(); 3396 HeapWord* limit = hr->next_top_at_mark_start(); 3397 3398 if (limit == bottom) { 3399 if (_cm->verbose_low()) { 3400 gclog_or_tty->print_cr("[%u] found an empty region " 3401 "["PTR_FORMAT", "PTR_FORMAT")", 3402 _worker_id, bottom, limit); 3403 } 3404 // The region was collected underneath our feet. 3405 // We set the finger to bottom to ensure that the bitmap 3406 // iteration that will follow this will not do anything. 3407 // (this is not a condition that holds when we set the region up, 3408 // as the region is not supposed to be empty in the first place) 3409 _finger = bottom; 3410 } else if (limit >= _region_limit) { 3411 assert(limit >= _finger, "peace of mind"); 3412 } else { 3413 assert(limit < _region_limit, "only way to get here"); 3414 // This can happen under some pretty unusual circumstances. An 3415 // evacuation pause empties the region underneath our feet (NTAMS 3416 // at bottom). We then do some allocation in the region (NTAMS 3417 // stays at bottom), followed by the region being used as a GC 3418 // alloc region (NTAMS will move to top() and the objects 3419 // originally below it will be grayed). All objects now marked in 3420 // the region are explicitly grayed, if below the global finger, 3421 // and we do not need in fact to scan anything else. So, we simply 3422 // set _finger to be limit to ensure that the bitmap iteration 3423 // doesn't do anything. 3424 _finger = limit; 3425 } 3426 3427 _region_limit = limit; 3428 } 3429 3430 void CMTask::giveup_current_region() { 3431 assert(_curr_region != NULL, "invariant"); 3432 if (_cm->verbose_low()) { 3433 gclog_or_tty->print_cr("[%u] giving up region "PTR_FORMAT, 3434 _worker_id, _curr_region); 3435 } 3436 clear_region_fields(); 3437 } 3438 3439 void CMTask::clear_region_fields() { 3440 // Values for these three fields that indicate that we're not 3441 // holding on to a region. 3442 _curr_region = NULL; 3443 _finger = NULL; 3444 _region_limit = NULL; 3445 } 3446 3447 void CMTask::set_cm_oop_closure(G1CMOopClosure* cm_oop_closure) { 3448 if (cm_oop_closure == NULL) { 3449 assert(_cm_oop_closure != NULL, "invariant"); 3450 } else { 3451 assert(_cm_oop_closure == NULL, "invariant"); 3452 } 3453 _cm_oop_closure = cm_oop_closure; 3454 } 3455 3456 void CMTask::reset(CMBitMap* nextMarkBitMap) { 3457 guarantee(nextMarkBitMap != NULL, "invariant"); 3458 3459 if (_cm->verbose_low()) { 3460 gclog_or_tty->print_cr("[%u] resetting", _worker_id); 3461 } 3462 3463 _nextMarkBitMap = nextMarkBitMap; 3464 clear_region_fields(); 3465 3466 _calls = 0; 3467 _elapsed_time_ms = 0.0; 3468 _termination_time_ms = 0.0; 3469 _termination_start_time_ms = 0.0; 3470 3471 #if _MARKING_STATS_ 3472 _local_pushes = 0; 3473 _local_pops = 0; 3474 _local_max_size = 0; 3475 _objs_scanned = 0; 3476 _global_pushes = 0; 3477 _global_pops = 0; 3478 _global_max_size = 0; 3479 _global_transfers_to = 0; 3480 _global_transfers_from = 0; 3481 _regions_claimed = 0; 3482 _objs_found_on_bitmap = 0; 3483 _satb_buffers_processed = 0; 3484 _steal_attempts = 0; 3485 _steals = 0; 3486 _aborted = 0; 3487 _aborted_overflow = 0; 3488 _aborted_cm_aborted = 0; 3489 _aborted_yield = 0; 3490 _aborted_timed_out = 0; 3491 _aborted_satb = 0; 3492 _aborted_termination = 0; 3493 #endif // _MARKING_STATS_ 3494 } 3495 3496 bool CMTask::should_exit_termination() { 3497 regular_clock_call(); 3498 // This is called when we are in the termination protocol. We should 3499 // quit if, for some reason, this task wants to abort or the global 3500 // stack is not empty (this means that we can get work from it). 3501 return !_cm->mark_stack_empty() || has_aborted(); 3502 } 3503 3504 void CMTask::reached_limit() { 3505 assert(_words_scanned >= _words_scanned_limit || 3506 _refs_reached >= _refs_reached_limit , 3507 "shouldn't have been called otherwise"); 3508 regular_clock_call(); 3509 } 3510 3511 void CMTask::regular_clock_call() { 3512 if (has_aborted()) return; 3513 3514 // First, we need to recalculate the words scanned and refs reached 3515 // limits for the next clock call. 3516 recalculate_limits(); 3517 3518 // During the regular clock call we do the following 3519 3520 // (1) If an overflow has been flagged, then we abort. 3521 if (_cm->has_overflown()) { 3522 set_has_aborted(); 3523 return; 3524 } 3525 3526 // If we are not concurrent (i.e. we're doing remark) we don't need 3527 // to check anything else. The other steps are only needed during 3528 // the concurrent marking phase. 3529 if (!concurrent()) return; 3530 3531 // (2) If marking has been aborted for Full GC, then we also abort. 3532 if (_cm->has_aborted()) { 3533 set_has_aborted(); 3534 statsOnly( ++_aborted_cm_aborted ); 3535 return; 3536 } 3537 3538 double curr_time_ms = os::elapsedVTime() * 1000.0; 3539 3540 // (3) If marking stats are enabled, then we update the step history. 3541 #if _MARKING_STATS_ 3542 if (_words_scanned >= _words_scanned_limit) { 3543 ++_clock_due_to_scanning; 3544 } 3545 if (_refs_reached >= _refs_reached_limit) { 3546 ++_clock_due_to_marking; 3547 } 3548 3549 double last_interval_ms = curr_time_ms - _interval_start_time_ms; 3550 _interval_start_time_ms = curr_time_ms; 3551 _all_clock_intervals_ms.add(last_interval_ms); 3552 3553 if (_cm->verbose_medium()) { 3554 gclog_or_tty->print_cr("[%u] regular clock, interval = %1.2lfms, " 3555 "scanned = %d%s, refs reached = %d%s", 3556 _worker_id, last_interval_ms, 3557 _words_scanned, 3558 (_words_scanned >= _words_scanned_limit) ? " (*)" : "", 3559 _refs_reached, 3560 (_refs_reached >= _refs_reached_limit) ? " (*)" : ""); 3561 } 3562 #endif // _MARKING_STATS_ 3563 3564 // (4) We check whether we should yield. If we have to, then we abort. 3565 if (_cm->should_yield()) { 3566 // We should yield. To do this we abort the task. The caller is 3567 // responsible for yielding. 3568 set_has_aborted(); 3569 statsOnly( ++_aborted_yield ); 3570 return; 3571 } 3572 3573 // (5) We check whether we've reached our time quota. If we have, 3574 // then we abort. 3575 double elapsed_time_ms = curr_time_ms - _start_time_ms; 3576 if (elapsed_time_ms > _time_target_ms) { 3577 set_has_aborted(); 3578 _has_timed_out = true; 3579 statsOnly( ++_aborted_timed_out ); 3580 return; 3581 } 3582 3583 // (6) Finally, we check whether there are enough completed STAB 3584 // buffers available for processing. If there are, we abort. 3585 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3586 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { 3587 if (_cm->verbose_low()) { 3588 gclog_or_tty->print_cr("[%u] aborting to deal with pending SATB buffers", 3589 _worker_id); 3590 } 3591 // we do need to process SATB buffers, we'll abort and restart 3592 // the marking task to do so 3593 set_has_aborted(); 3594 statsOnly( ++_aborted_satb ); 3595 return; 3596 } 3597 } 3598 3599 void CMTask::recalculate_limits() { 3600 _real_words_scanned_limit = _words_scanned + words_scanned_period; 3601 _words_scanned_limit = _real_words_scanned_limit; 3602 3603 _real_refs_reached_limit = _refs_reached + refs_reached_period; 3604 _refs_reached_limit = _real_refs_reached_limit; 3605 } 3606 3607 void CMTask::decrease_limits() { 3608 // This is called when we believe that we're going to do an infrequent 3609 // operation which will increase the per byte scanned cost (i.e. move 3610 // entries to/from the global stack). It basically tries to decrease the 3611 // scanning limit so that the clock is called earlier. 3612 3613 if (_cm->verbose_medium()) { 3614 gclog_or_tty->print_cr("[%u] decreasing limits", _worker_id); 3615 } 3616 3617 _words_scanned_limit = _real_words_scanned_limit - 3618 3 * words_scanned_period / 4; 3619 _refs_reached_limit = _real_refs_reached_limit - 3620 3 * refs_reached_period / 4; 3621 } 3622 3623 void CMTask::move_entries_to_global_stack() { 3624 // local array where we'll store the entries that will be popped 3625 // from the local queue 3626 oop buffer[global_stack_transfer_size]; 3627 3628 int n = 0; 3629 oop obj; 3630 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { 3631 buffer[n] = obj; 3632 ++n; 3633 } 3634 3635 if (n > 0) { 3636 // we popped at least one entry from the local queue 3637 3638 statsOnly( ++_global_transfers_to; _local_pops += n ); 3639 3640 if (!_cm->mark_stack_push(buffer, n)) { 3641 if (_cm->verbose_low()) { 3642 gclog_or_tty->print_cr("[%u] aborting due to global stack overflow", 3643 _worker_id); 3644 } 3645 set_has_aborted(); 3646 } else { 3647 // the transfer was successful 3648 3649 if (_cm->verbose_medium()) { 3650 gclog_or_tty->print_cr("[%u] pushed %d entries to the global stack", 3651 _worker_id, n); 3652 } 3653 statsOnly( int tmp_size = _cm->mark_stack_size(); 3654 if (tmp_size > _global_max_size) { 3655 _global_max_size = tmp_size; 3656 } 3657 _global_pushes += n ); 3658 } 3659 } 3660 3661 // this operation was quite expensive, so decrease the limits 3662 decrease_limits(); 3663 } 3664 3665 void CMTask::get_entries_from_global_stack() { 3666 // local array where we'll store the entries that will be popped 3667 // from the global stack. 3668 oop buffer[global_stack_transfer_size]; 3669 int n; 3670 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); 3671 assert(n <= global_stack_transfer_size, 3672 "we should not pop more than the given limit"); 3673 if (n > 0) { 3674 // yes, we did actually pop at least one entry 3675 3676 statsOnly( ++_global_transfers_from; _global_pops += n ); 3677 if (_cm->verbose_medium()) { 3678 gclog_or_tty->print_cr("[%u] popped %d entries from the global stack", 3679 _worker_id, n); 3680 } 3681 for (int i = 0; i < n; ++i) { 3682 bool success = _task_queue->push(buffer[i]); 3683 // We only call this when the local queue is empty or under a 3684 // given target limit. So, we do not expect this push to fail. 3685 assert(success, "invariant"); 3686 } 3687 3688 statsOnly( int tmp_size = _task_queue->size(); 3689 if (tmp_size > _local_max_size) { 3690 _local_max_size = tmp_size; 3691 } 3692 _local_pushes += n ); 3693 } 3694 3695 // this operation was quite expensive, so decrease the limits 3696 decrease_limits(); 3697 } 3698 3699 void CMTask::drain_local_queue(bool partially) { 3700 if (has_aborted()) return; 3701 3702 // Decide what the target size is, depending whether we're going to 3703 // drain it partially (so that other tasks can steal if they run out 3704 // of things to do) or totally (at the very end). 3705 size_t target_size; 3706 if (partially) { 3707 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); 3708 } else { 3709 target_size = 0; 3710 } 3711 3712 if (_task_queue->size() > target_size) { 3713 if (_cm->verbose_high()) { 3714 gclog_or_tty->print_cr("[%u] draining local queue, target size = %d", 3715 _worker_id, target_size); 3716 } 3717 3718 oop obj; 3719 bool ret = _task_queue->pop_local(obj); 3720 while (ret) { 3721 statsOnly( ++_local_pops ); 3722 3723 if (_cm->verbose_high()) { 3724 gclog_or_tty->print_cr("[%u] popped "PTR_FORMAT, _worker_id, 3725 (void*) obj); 3726 } 3727 3728 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); 3729 assert(!_g1h->is_on_master_free_list( 3730 _g1h->heap_region_containing((HeapWord*) obj)), "invariant"); 3731 3732 scan_object(obj); 3733 3734 if (_task_queue->size() <= target_size || has_aborted()) { 3735 ret = false; 3736 } else { 3737 ret = _task_queue->pop_local(obj); 3738 } 3739 } 3740 3741 if (_cm->verbose_high()) { 3742 gclog_or_tty->print_cr("[%u] drained local queue, size = %d", 3743 _worker_id, _task_queue->size()); 3744 } 3745 } 3746 } 3747 3748 void CMTask::drain_global_stack(bool partially) { 3749 if (has_aborted()) return; 3750 3751 // We have a policy to drain the local queue before we attempt to 3752 // drain the global stack. 3753 assert(partially || _task_queue->size() == 0, "invariant"); 3754 3755 // Decide what the target size is, depending whether we're going to 3756 // drain it partially (so that other tasks can steal if they run out 3757 // of things to do) or totally (at the very end). Notice that, 3758 // because we move entries from the global stack in chunks or 3759 // because another task might be doing the same, we might in fact 3760 // drop below the target. But, this is not a problem. 3761 size_t target_size; 3762 if (partially) { 3763 target_size = _cm->partial_mark_stack_size_target(); 3764 } else { 3765 target_size = 0; 3766 } 3767 3768 if (_cm->mark_stack_size() > target_size) { 3769 if (_cm->verbose_low()) { 3770 gclog_or_tty->print_cr("[%u] draining global_stack, target size %d", 3771 _worker_id, target_size); 3772 } 3773 3774 while (!has_aborted() && _cm->mark_stack_size() > target_size) { 3775 get_entries_from_global_stack(); 3776 drain_local_queue(partially); 3777 } 3778 3779 if (_cm->verbose_low()) { 3780 gclog_or_tty->print_cr("[%u] drained global stack, size = %d", 3781 _worker_id, _cm->mark_stack_size()); 3782 } 3783 } 3784 } 3785 3786 // SATB Queue has several assumptions on whether to call the par or 3787 // non-par versions of the methods. this is why some of the code is 3788 // replicated. We should really get rid of the single-threaded version 3789 // of the code to simplify things. 3790 void CMTask::drain_satb_buffers() { 3791 if (has_aborted()) return; 3792 3793 // We set this so that the regular clock knows that we're in the 3794 // middle of draining buffers and doesn't set the abort flag when it 3795 // notices that SATB buffers are available for draining. It'd be 3796 // very counter productive if it did that. :-) 3797 _draining_satb_buffers = true; 3798 3799 CMObjectClosure oc(this); 3800 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); 3801 if (G1CollectedHeap::use_parallel_gc_threads()) { 3802 satb_mq_set.set_par_closure(_worker_id, &oc); 3803 } else { 3804 satb_mq_set.set_closure(&oc); 3805 } 3806 3807 // This keeps claiming and applying the closure to completed buffers 3808 // until we run out of buffers or we need to abort. 3809 if (G1CollectedHeap::use_parallel_gc_threads()) { 3810 while (!has_aborted() && 3811 satb_mq_set.par_apply_closure_to_completed_buffer(_worker_id)) { 3812 if (_cm->verbose_medium()) { 3813 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id); 3814 } 3815 statsOnly( ++_satb_buffers_processed ); 3816 regular_clock_call(); 3817 } 3818 } else { 3819 while (!has_aborted() && 3820 satb_mq_set.apply_closure_to_completed_buffer()) { 3821 if (_cm->verbose_medium()) { 3822 gclog_or_tty->print_cr("[%u] processed an SATB buffer", _worker_id); 3823 } 3824 statsOnly( ++_satb_buffers_processed ); 3825 regular_clock_call(); 3826 } 3827 } 3828 3829 if (!concurrent() && !has_aborted()) { 3830 // We should only do this during remark. 3831 if (G1CollectedHeap::use_parallel_gc_threads()) { 3832 satb_mq_set.par_iterate_closure_all_threads(_worker_id); 3833 } else { 3834 satb_mq_set.iterate_closure_all_threads(); 3835 } 3836 } 3837 3838 _draining_satb_buffers = false; 3839 3840 assert(has_aborted() || 3841 concurrent() || 3842 satb_mq_set.completed_buffers_num() == 0, "invariant"); 3843 3844 if (G1CollectedHeap::use_parallel_gc_threads()) { 3845 satb_mq_set.set_par_closure(_worker_id, NULL); 3846 } else { 3847 satb_mq_set.set_closure(NULL); 3848 } 3849 3850 // again, this was a potentially expensive operation, decrease the 3851 // limits to get the regular clock call early 3852 decrease_limits(); 3853 } 3854 3855 void CMTask::print_stats() { 3856 gclog_or_tty->print_cr("Marking Stats, task = %u, calls = %d", 3857 _worker_id, _calls); 3858 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", 3859 _elapsed_time_ms, _termination_time_ms); 3860 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 3861 _step_times_ms.num(), _step_times_ms.avg(), 3862 _step_times_ms.sd()); 3863 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 3864 _step_times_ms.maximum(), _step_times_ms.sum()); 3865 3866 #if _MARKING_STATS_ 3867 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", 3868 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(), 3869 _all_clock_intervals_ms.sd()); 3870 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", 3871 _all_clock_intervals_ms.maximum(), 3872 _all_clock_intervals_ms.sum()); 3873 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d", 3874 _clock_due_to_scanning, _clock_due_to_marking); 3875 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d", 3876 _objs_scanned, _objs_found_on_bitmap); 3877 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d", 3878 _local_pushes, _local_pops, _local_max_size); 3879 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d", 3880 _global_pushes, _global_pops, _global_max_size); 3881 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d", 3882 _global_transfers_to,_global_transfers_from); 3883 gclog_or_tty->print_cr(" Regions: claimed = %d", _regions_claimed); 3884 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed); 3885 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d", 3886 _steal_attempts, _steals); 3887 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted); 3888 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d", 3889 _aborted_overflow, _aborted_cm_aborted, _aborted_yield); 3890 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d", 3891 _aborted_timed_out, _aborted_satb, _aborted_termination); 3892 #endif // _MARKING_STATS_ 3893 } 3894 3895 /***************************************************************************** 3896 3897 The do_marking_step(time_target_ms) method is the building block 3898 of the parallel marking framework. It can be called in parallel 3899 with other invocations of do_marking_step() on different tasks 3900 (but only one per task, obviously) and concurrently with the 3901 mutator threads, or during remark, hence it eliminates the need 3902 for two versions of the code. When called during remark, it will 3903 pick up from where the task left off during the concurrent marking 3904 phase. Interestingly, tasks are also claimable during evacuation 3905 pauses too, since do_marking_step() ensures that it aborts before 3906 it needs to yield. 3907 3908 The data structures that is uses to do marking work are the 3909 following: 3910 3911 (1) Marking Bitmap. If there are gray objects that appear only 3912 on the bitmap (this happens either when dealing with an overflow 3913 or when the initial marking phase has simply marked the roots 3914 and didn't push them on the stack), then tasks claim heap 3915 regions whose bitmap they then scan to find gray objects. A 3916 global finger indicates where the end of the last claimed region 3917 is. A local finger indicates how far into the region a task has 3918 scanned. The two fingers are used to determine how to gray an 3919 object (i.e. whether simply marking it is OK, as it will be 3920 visited by a task in the future, or whether it needs to be also 3921 pushed on a stack). 3922 3923 (2) Local Queue. The local queue of the task which is accessed 3924 reasonably efficiently by the task. Other tasks can steal from 3925 it when they run out of work. Throughout the marking phase, a 3926 task attempts to keep its local queue short but not totally 3927 empty, so that entries are available for stealing by other 3928 tasks. Only when there is no more work, a task will totally 3929 drain its local queue. 3930 3931 (3) Global Mark Stack. This handles local queue overflow. During 3932 marking only sets of entries are moved between it and the local 3933 queues, as access to it requires a mutex and more fine-grain 3934 interaction with it which might cause contention. If it 3935 overflows, then the marking phase should restart and iterate 3936 over the bitmap to identify gray objects. Throughout the marking 3937 phase, tasks attempt to keep the global mark stack at a small 3938 length but not totally empty, so that entries are available for 3939 popping by other tasks. Only when there is no more work, tasks 3940 will totally drain the global mark stack. 3941 3942 (4) SATB Buffer Queue. This is where completed SATB buffers are 3943 made available. Buffers are regularly removed from this queue 3944 and scanned for roots, so that the queue doesn't get too 3945 long. During remark, all completed buffers are processed, as 3946 well as the filled in parts of any uncompleted buffers. 3947 3948 The do_marking_step() method tries to abort when the time target 3949 has been reached. There are a few other cases when the 3950 do_marking_step() method also aborts: 3951 3952 (1) When the marking phase has been aborted (after a Full GC). 3953 3954 (2) When a global overflow (on the global stack) has been 3955 triggered. Before the task aborts, it will actually sync up with 3956 the other tasks to ensure that all the marking data structures 3957 (local queues, stacks, fingers etc.) are re-initialised so that 3958 when do_marking_step() completes, the marking phase can 3959 immediately restart. 3960 3961 (3) When enough completed SATB buffers are available. The 3962 do_marking_step() method only tries to drain SATB buffers right 3963 at the beginning. So, if enough buffers are available, the 3964 marking step aborts and the SATB buffers are processed at 3965 the beginning of the next invocation. 3966 3967 (4) To yield. when we have to yield then we abort and yield 3968 right at the end of do_marking_step(). This saves us from a lot 3969 of hassle as, by yielding we might allow a Full GC. If this 3970 happens then objects will be compacted underneath our feet, the 3971 heap might shrink, etc. We save checking for this by just 3972 aborting and doing the yield right at the end. 3973 3974 From the above it follows that the do_marking_step() method should 3975 be called in a loop (or, otherwise, regularly) until it completes. 3976 3977 If a marking step completes without its has_aborted() flag being 3978 true, it means it has completed the current marking phase (and 3979 also all other marking tasks have done so and have all synced up). 3980 3981 A method called regular_clock_call() is invoked "regularly" (in 3982 sub ms intervals) throughout marking. It is this clock method that 3983 checks all the abort conditions which were mentioned above and 3984 decides when the task should abort. A work-based scheme is used to 3985 trigger this clock method: when the number of object words the 3986 marking phase has scanned or the number of references the marking 3987 phase has visited reach a given limit. Additional invocations to 3988 the method clock have been planted in a few other strategic places 3989 too. The initial reason for the clock method was to avoid calling 3990 vtime too regularly, as it is quite expensive. So, once it was in 3991 place, it was natural to piggy-back all the other conditions on it 3992 too and not constantly check them throughout the code. 3993 3994 *****************************************************************************/ 3995 3996 void CMTask::do_marking_step(double time_target_ms, 3997 bool do_stealing, 3998 bool do_termination) { 3999 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); 4000 assert(concurrent() == _cm->concurrent(), "they should be the same"); 4001 4002 G1CollectorPolicy* g1_policy = _g1h->g1_policy(); 4003 assert(_task_queues != NULL, "invariant"); 4004 assert(_task_queue != NULL, "invariant"); 4005 assert(_task_queues->queue(_worker_id) == _task_queue, "invariant"); 4006 4007 assert(!_claimed, 4008 "only one thread should claim this task at any one time"); 4009 4010 // OK, this doesn't safeguard again all possible scenarios, as it is 4011 // possible for two threads to set the _claimed flag at the same 4012 // time. But it is only for debugging purposes anyway and it will 4013 // catch most problems. 4014 _claimed = true; 4015 4016 _start_time_ms = os::elapsedVTime() * 1000.0; 4017 statsOnly( _interval_start_time_ms = _start_time_ms ); 4018 4019 double diff_prediction_ms = 4020 g1_policy->get_new_prediction(&_marking_step_diffs_ms); 4021 _time_target_ms = time_target_ms - diff_prediction_ms; 4022 4023 // set up the variables that are used in the work-based scheme to 4024 // call the regular clock method 4025 _words_scanned = 0; 4026 _refs_reached = 0; 4027 recalculate_limits(); 4028 4029 // clear all flags 4030 clear_has_aborted(); 4031 _has_timed_out = false; 4032 _draining_satb_buffers = false; 4033 4034 ++_calls; 4035 4036 if (_cm->verbose_low()) { 4037 gclog_or_tty->print_cr("[%u] >>>>>>>>>> START, call = %d, " 4038 "target = %1.2lfms >>>>>>>>>>", 4039 _worker_id, _calls, _time_target_ms); 4040 } 4041 4042 // Set up the bitmap and oop closures. Anything that uses them is 4043 // eventually called from this method, so it is OK to allocate these 4044 // statically. 4045 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); 4046 G1CMOopClosure cm_oop_closure(_g1h, _cm, this); 4047 set_cm_oop_closure(&cm_oop_closure); 4048 4049 if (_cm->has_overflown()) { 4050 // This can happen if the mark stack overflows during a GC pause 4051 // and this task, after a yield point, restarts. We have to abort 4052 // as we need to get into the overflow protocol which happens 4053 // right at the end of this task. 4054 set_has_aborted(); 4055 } 4056 4057 // First drain any available SATB buffers. After this, we will not 4058 // look at SATB buffers before the next invocation of this method. 4059 // If enough completed SATB buffers are queued up, the regular clock 4060 // will abort this task so that it restarts. 4061 drain_satb_buffers(); 4062 // ...then partially drain the local queue and the global stack 4063 drain_local_queue(true); 4064 drain_global_stack(true); 4065 4066 do { 4067 if (!has_aborted() && _curr_region != NULL) { 4068 // This means that we're already holding on to a region. 4069 assert(_finger != NULL, "if region is not NULL, then the finger " 4070 "should not be NULL either"); 4071 4072 // We might have restarted this task after an evacuation pause 4073 // which might have evacuated the region we're holding on to 4074 // underneath our feet. Let's read its limit again to make sure 4075 // that we do not iterate over a region of the heap that 4076 // contains garbage (update_region_limit() will also move 4077 // _finger to the start of the region if it is found empty). 4078 update_region_limit(); 4079 // We will start from _finger not from the start of the region, 4080 // as we might be restarting this task after aborting half-way 4081 // through scanning this region. In this case, _finger points to 4082 // the address where we last found a marked object. If this is a 4083 // fresh region, _finger points to start(). 4084 MemRegion mr = MemRegion(_finger, _region_limit); 4085 4086 if (_cm->verbose_low()) { 4087 gclog_or_tty->print_cr("[%u] we're scanning part " 4088 "["PTR_FORMAT", "PTR_FORMAT") " 4089 "of region "HR_FORMAT, 4090 _worker_id, _finger, _region_limit, 4091 HR_FORMAT_PARAMS(_curr_region)); 4092 } 4093 4094 assert(!_curr_region->isHumongous() || mr.start() == _curr_region->bottom(), 4095 "humongous regions should go around loop once only"); 4096 4097 // Some special cases: 4098 // If the memory region is empty, we can just give up the region. 4099 // If the current region is humongous then we only need to check 4100 // the bitmap for the bit associated with the start of the object, 4101 // scan the object if it's live, and give up the region. 4102 // Otherwise, let's iterate over the bitmap of the part of the region 4103 // that is left. 4104 // If the iteration is successful, give up the region. 4105 if (mr.is_empty()) { 4106 giveup_current_region(); 4107 regular_clock_call(); 4108 } else if (_curr_region->isHumongous() && mr.start() == _curr_region->bottom()) { 4109 if (_nextMarkBitMap->isMarked(mr.start())) { 4110 // The object is marked - apply the closure 4111 BitMap::idx_t offset = _nextMarkBitMap->heapWordToOffset(mr.start()); 4112 bitmap_closure.do_bit(offset); 4113 } 4114 // Even if this task aborted while scanning the humongous object 4115 // we can (and should) give up the current region. 4116 giveup_current_region(); 4117 regular_clock_call(); 4118 } else if (_nextMarkBitMap->iterate(&bitmap_closure, mr)) { 4119 giveup_current_region(); 4120 regular_clock_call(); 4121 } else { 4122 assert(has_aborted(), "currently the only way to do so"); 4123 // The only way to abort the bitmap iteration is to return 4124 // false from the do_bit() method. However, inside the 4125 // do_bit() method we move the _finger to point to the 4126 // object currently being looked at. So, if we bail out, we 4127 // have definitely set _finger to something non-null. 4128 assert(_finger != NULL, "invariant"); 4129 4130 // Region iteration was actually aborted. So now _finger 4131 // points to the address of the object we last scanned. If we 4132 // leave it there, when we restart this task, we will rescan 4133 // the object. It is easy to avoid this. We move the finger by 4134 // enough to point to the next possible object header (the 4135 // bitmap knows by how much we need to move it as it knows its 4136 // granularity). 4137 assert(_finger < _region_limit, "invariant"); 4138 HeapWord* new_finger = _nextMarkBitMap->nextObject(_finger); 4139 // Check if bitmap iteration was aborted while scanning the last object 4140 if (new_finger >= _region_limit) { 4141 giveup_current_region(); 4142 } else { 4143 move_finger_to(new_finger); 4144 } 4145 } 4146 } 4147 // At this point we have either completed iterating over the 4148 // region we were holding on to, or we have aborted. 4149 4150 // We then partially drain the local queue and the global stack. 4151 // (Do we really need this?) 4152 drain_local_queue(true); 4153 drain_global_stack(true); 4154 4155 // Read the note on the claim_region() method on why it might 4156 // return NULL with potentially more regions available for 4157 // claiming and why we have to check out_of_regions() to determine 4158 // whether we're done or not. 4159 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { 4160 // We are going to try to claim a new region. We should have 4161 // given up on the previous one. 4162 // Separated the asserts so that we know which one fires. 4163 assert(_curr_region == NULL, "invariant"); 4164 assert(_finger == NULL, "invariant"); 4165 assert(_region_limit == NULL, "invariant"); 4166 if (_cm->verbose_low()) { 4167 gclog_or_tty->print_cr("[%u] trying to claim a new region", _worker_id); 4168 } 4169 HeapRegion* claimed_region = _cm->claim_region(_worker_id); 4170 if (claimed_region != NULL) { 4171 // Yes, we managed to claim one 4172 statsOnly( ++_regions_claimed ); 4173 4174 if (_cm->verbose_low()) { 4175 gclog_or_tty->print_cr("[%u] we successfully claimed " 4176 "region "PTR_FORMAT, 4177 _worker_id, claimed_region); 4178 } 4179 4180 setup_for_region(claimed_region); 4181 assert(_curr_region == claimed_region, "invariant"); 4182 } 4183 // It is important to call the regular clock here. It might take 4184 // a while to claim a region if, for example, we hit a large 4185 // block of empty regions. So we need to call the regular clock 4186 // method once round the loop to make sure it's called 4187 // frequently enough. 4188 regular_clock_call(); 4189 } 4190 4191 if (!has_aborted() && _curr_region == NULL) { 4192 assert(_cm->out_of_regions(), 4193 "at this point we should be out of regions"); 4194 } 4195 } while ( _curr_region != NULL && !has_aborted()); 4196 4197 if (!has_aborted()) { 4198 // We cannot check whether the global stack is empty, since other 4199 // tasks might be pushing objects to it concurrently. 4200 assert(_cm->out_of_regions(), 4201 "at this point we should be out of regions"); 4202 4203 if (_cm->verbose_low()) { 4204 gclog_or_tty->print_cr("[%u] all regions claimed", _worker_id); 4205 } 4206 4207 // Try to reduce the number of available SATB buffers so that 4208 // remark has less work to do. 4209 drain_satb_buffers(); 4210 } 4211 4212 // Since we've done everything else, we can now totally drain the 4213 // local queue and global stack. 4214 drain_local_queue(false); 4215 drain_global_stack(false); 4216 4217 // Attempt at work stealing from other task's queues. 4218 if (do_stealing && !has_aborted()) { 4219 // We have not aborted. This means that we have finished all that 4220 // we could. Let's try to do some stealing... 4221 4222 // We cannot check whether the global stack is empty, since other 4223 // tasks might be pushing objects to it concurrently. 4224 assert(_cm->out_of_regions() && _task_queue->size() == 0, 4225 "only way to reach here"); 4226 4227 if (_cm->verbose_low()) { 4228 gclog_or_tty->print_cr("[%u] starting to steal", _worker_id); 4229 } 4230 4231 while (!has_aborted()) { 4232 oop obj; 4233 statsOnly( ++_steal_attempts ); 4234 4235 if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) { 4236 if (_cm->verbose_medium()) { 4237 gclog_or_tty->print_cr("[%u] stolen "PTR_FORMAT" successfully", 4238 _worker_id, (void*) obj); 4239 } 4240 4241 statsOnly( ++_steals ); 4242 4243 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), 4244 "any stolen object should be marked"); 4245 scan_object(obj); 4246 4247 // And since we're towards the end, let's totally drain the 4248 // local queue and global stack. 4249 drain_local_queue(false); 4250 drain_global_stack(false); 4251 } else { 4252 break; 4253 } 4254 } 4255 } 4256 4257 // If we are about to wrap up and go into termination, check if we 4258 // should raise the overflow flag. 4259 if (do_termination && !has_aborted()) { 4260 if (_cm->force_overflow()->should_force()) { 4261 _cm->set_has_overflown(); 4262 regular_clock_call(); 4263 } 4264 } 4265 4266 // We still haven't aborted. Now, let's try to get into the 4267 // termination protocol. 4268 if (do_termination && !has_aborted()) { 4269 // We cannot check whether the global stack is empty, since other 4270 // tasks might be concurrently pushing objects on it. 4271 // Separated the asserts so that we know which one fires. 4272 assert(_cm->out_of_regions(), "only way to reach here"); 4273 assert(_task_queue->size() == 0, "only way to reach here"); 4274 4275 if (_cm->verbose_low()) { 4276 gclog_or_tty->print_cr("[%u] starting termination protocol", _worker_id); 4277 } 4278 4279 _termination_start_time_ms = os::elapsedVTime() * 1000.0; 4280 // The CMTask class also extends the TerminatorTerminator class, 4281 // hence its should_exit_termination() method will also decide 4282 // whether to exit the termination protocol or not. 4283 bool finished = _cm->terminator()->offer_termination(this); 4284 double termination_end_time_ms = os::elapsedVTime() * 1000.0; 4285 _termination_time_ms += 4286 termination_end_time_ms - _termination_start_time_ms; 4287 4288 if (finished) { 4289 // We're all done. 4290 4291 if (_worker_id == 0) { 4292 // let's allow task 0 to do this 4293 if (concurrent()) { 4294 assert(_cm->concurrent_marking_in_progress(), "invariant"); 4295 // we need to set this to false before the next 4296 // safepoint. This way we ensure that the marking phase 4297 // doesn't observe any more heap expansions. 4298 _cm->clear_concurrent_marking_in_progress(); 4299 } 4300 } 4301 4302 // We can now guarantee that the global stack is empty, since 4303 // all other tasks have finished. We separated the guarantees so 4304 // that, if a condition is false, we can immediately find out 4305 // which one. 4306 guarantee(_cm->out_of_regions(), "only way to reach here"); 4307 guarantee(_cm->mark_stack_empty(), "only way to reach here"); 4308 guarantee(_task_queue->size() == 0, "only way to reach here"); 4309 guarantee(!_cm->has_overflown(), "only way to reach here"); 4310 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); 4311 4312 if (_cm->verbose_low()) { 4313 gclog_or_tty->print_cr("[%u] all tasks terminated", _worker_id); 4314 } 4315 } else { 4316 // Apparently there's more work to do. Let's abort this task. It 4317 // will restart it and we can hopefully find more things to do. 4318 4319 if (_cm->verbose_low()) { 4320 gclog_or_tty->print_cr("[%u] apparently there is more work to do", 4321 _worker_id); 4322 } 4323 4324 set_has_aborted(); 4325 statsOnly( ++_aborted_termination ); 4326 } 4327 } 4328 4329 // Mainly for debugging purposes to make sure that a pointer to the 4330 // closure which was statically allocated in this frame doesn't 4331 // escape it by accident. 4332 set_cm_oop_closure(NULL); 4333 double end_time_ms = os::elapsedVTime() * 1000.0; 4334 double elapsed_time_ms = end_time_ms - _start_time_ms; 4335 // Update the step history. 4336 _step_times_ms.add(elapsed_time_ms); 4337 4338 if (has_aborted()) { 4339 // The task was aborted for some reason. 4340 4341 statsOnly( ++_aborted ); 4342 4343 if (_has_timed_out) { 4344 double diff_ms = elapsed_time_ms - _time_target_ms; 4345 // Keep statistics of how well we did with respect to hitting 4346 // our target only if we actually timed out (if we aborted for 4347 // other reasons, then the results might get skewed). 4348 _marking_step_diffs_ms.add(diff_ms); 4349 } 4350 4351 if (_cm->has_overflown()) { 4352 // This is the interesting one. We aborted because a global 4353 // overflow was raised. This means we have to restart the 4354 // marking phase and start iterating over regions. However, in 4355 // order to do this we have to make sure that all tasks stop 4356 // what they are doing and re-initialise in a safe manner. We 4357 // will achieve this with the use of two barrier sync points. 4358 4359 if (_cm->verbose_low()) { 4360 gclog_or_tty->print_cr("[%u] detected overflow", _worker_id); 4361 } 4362 4363 _cm->enter_first_sync_barrier(_worker_id); 4364 // When we exit this sync barrier we know that all tasks have 4365 // stopped doing marking work. So, it's now safe to 4366 // re-initialise our data structures. At the end of this method, 4367 // task 0 will clear the global data structures. 4368 4369 statsOnly( ++_aborted_overflow ); 4370 4371 // We clear the local state of this task... 4372 clear_region_fields(); 4373 4374 // ...and enter the second barrier. 4375 _cm->enter_second_sync_barrier(_worker_id); 4376 // At this point everything has bee re-initialised and we're 4377 // ready to restart. 4378 } 4379 4380 if (_cm->verbose_low()) { 4381 gclog_or_tty->print_cr("[%u] <<<<<<<<<< ABORTING, target = %1.2lfms, " 4382 "elapsed = %1.2lfms <<<<<<<<<<", 4383 _worker_id, _time_target_ms, elapsed_time_ms); 4384 if (_cm->has_aborted()) { 4385 gclog_or_tty->print_cr("[%u] ========== MARKING ABORTED ==========", 4386 _worker_id); 4387 } 4388 } 4389 } else { 4390 if (_cm->verbose_low()) { 4391 gclog_or_tty->print_cr("[%u] <<<<<<<<<< FINISHED, target = %1.2lfms, " 4392 "elapsed = %1.2lfms <<<<<<<<<<", 4393 _worker_id, _time_target_ms, elapsed_time_ms); 4394 } 4395 } 4396 4397 _claimed = false; 4398 } 4399 4400 CMTask::CMTask(uint worker_id, 4401 ConcurrentMark* cm, 4402 size_t* marked_bytes, 4403 BitMap* card_bm, 4404 CMTaskQueue* task_queue, 4405 CMTaskQueueSet* task_queues) 4406 : _g1h(G1CollectedHeap::heap()), 4407 _worker_id(worker_id), _cm(cm), 4408 _claimed(false), 4409 _nextMarkBitMap(NULL), _hash_seed(17), 4410 _task_queue(task_queue), 4411 _task_queues(task_queues), 4412 _cm_oop_closure(NULL), 4413 _marked_bytes_array(marked_bytes), 4414 _card_bm(card_bm) { 4415 guarantee(task_queue != NULL, "invariant"); 4416 guarantee(task_queues != NULL, "invariant"); 4417 4418 statsOnly( _clock_due_to_scanning = 0; 4419 _clock_due_to_marking = 0 ); 4420 4421 _marking_step_diffs_ms.add(0.5); 4422 } 4423 4424 // These are formatting macros that are used below to ensure 4425 // consistent formatting. The *_H_* versions are used to format the 4426 // header for a particular value and they should be kept consistent 4427 // with the corresponding macro. Also note that most of the macros add 4428 // the necessary white space (as a prefix) which makes them a bit 4429 // easier to compose. 4430 4431 // All the output lines are prefixed with this string to be able to 4432 // identify them easily in a large log file. 4433 #define G1PPRL_LINE_PREFIX "###" 4434 4435 #define G1PPRL_ADDR_BASE_FORMAT " "PTR_FORMAT"-"PTR_FORMAT 4436 #ifdef _LP64 4437 #define G1PPRL_ADDR_BASE_H_FORMAT " %37s" 4438 #else // _LP64 4439 #define G1PPRL_ADDR_BASE_H_FORMAT " %21s" 4440 #endif // _LP64 4441 4442 // For per-region info 4443 #define G1PPRL_TYPE_FORMAT " %-4s" 4444 #define G1PPRL_TYPE_H_FORMAT " %4s" 4445 #define G1PPRL_BYTE_FORMAT " "SIZE_FORMAT_W(9) 4446 #define G1PPRL_BYTE_H_FORMAT " %9s" 4447 #define G1PPRL_DOUBLE_FORMAT " %14.1f" 4448 #define G1PPRL_DOUBLE_H_FORMAT " %14s" 4449 4450 // For summary info 4451 #define G1PPRL_SUM_ADDR_FORMAT(tag) " "tag":"G1PPRL_ADDR_BASE_FORMAT 4452 #define G1PPRL_SUM_BYTE_FORMAT(tag) " "tag": "SIZE_FORMAT 4453 #define G1PPRL_SUM_MB_FORMAT(tag) " "tag": %1.2f MB" 4454 #define G1PPRL_SUM_MB_PERC_FORMAT(tag) G1PPRL_SUM_MB_FORMAT(tag)" / %1.2f %%" 4455 4456 G1PrintRegionLivenessInfoClosure:: 4457 G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name) 4458 : _out(out), 4459 _total_used_bytes(0), _total_capacity_bytes(0), 4460 _total_prev_live_bytes(0), _total_next_live_bytes(0), 4461 _hum_used_bytes(0), _hum_capacity_bytes(0), 4462 _hum_prev_live_bytes(0), _hum_next_live_bytes(0) { 4463 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 4464 MemRegion g1_committed = g1h->g1_committed(); 4465 MemRegion g1_reserved = g1h->g1_reserved(); 4466 double now = os::elapsedTime(); 4467 4468 // Print the header of the output. 4469 _out->cr(); 4470 _out->print_cr(G1PPRL_LINE_PREFIX" PHASE %s @ %1.3f", phase_name, now); 4471 _out->print_cr(G1PPRL_LINE_PREFIX" HEAP" 4472 G1PPRL_SUM_ADDR_FORMAT("committed") 4473 G1PPRL_SUM_ADDR_FORMAT("reserved") 4474 G1PPRL_SUM_BYTE_FORMAT("region-size"), 4475 g1_committed.start(), g1_committed.end(), 4476 g1_reserved.start(), g1_reserved.end(), 4477 HeapRegion::GrainBytes); 4478 _out->print_cr(G1PPRL_LINE_PREFIX); 4479 _out->print_cr(G1PPRL_LINE_PREFIX 4480 G1PPRL_TYPE_H_FORMAT 4481 G1PPRL_ADDR_BASE_H_FORMAT 4482 G1PPRL_BYTE_H_FORMAT 4483 G1PPRL_BYTE_H_FORMAT 4484 G1PPRL_BYTE_H_FORMAT 4485 G1PPRL_DOUBLE_H_FORMAT, 4486 "type", "address-range", 4487 "used", "prev-live", "next-live", "gc-eff"); 4488 _out->print_cr(G1PPRL_LINE_PREFIX 4489 G1PPRL_TYPE_H_FORMAT 4490 G1PPRL_ADDR_BASE_H_FORMAT 4491 G1PPRL_BYTE_H_FORMAT 4492 G1PPRL_BYTE_H_FORMAT 4493 G1PPRL_BYTE_H_FORMAT 4494 G1PPRL_DOUBLE_H_FORMAT, 4495 "", "", 4496 "(bytes)", "(bytes)", "(bytes)", "(bytes/ms)"); 4497 } 4498 4499 // It takes as a parameter a reference to one of the _hum_* fields, it 4500 // deduces the corresponding value for a region in a humongous region 4501 // series (either the region size, or what's left if the _hum_* field 4502 // is < the region size), and updates the _hum_* field accordingly. 4503 size_t G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* hum_bytes) { 4504 size_t bytes = 0; 4505 // The > 0 check is to deal with the prev and next live bytes which 4506 // could be 0. 4507 if (*hum_bytes > 0) { 4508 bytes = MIN2(HeapRegion::GrainBytes, *hum_bytes); 4509 *hum_bytes -= bytes; 4510 } 4511 return bytes; 4512 } 4513 4514 // It deduces the values for a region in a humongous region series 4515 // from the _hum_* fields and updates those accordingly. It assumes 4516 // that that _hum_* fields have already been set up from the "starts 4517 // humongous" region and we visit the regions in address order. 4518 void G1PrintRegionLivenessInfoClosure::get_hum_bytes(size_t* used_bytes, 4519 size_t* capacity_bytes, 4520 size_t* prev_live_bytes, 4521 size_t* next_live_bytes) { 4522 assert(_hum_used_bytes > 0 && _hum_capacity_bytes > 0, "pre-condition"); 4523 *used_bytes = get_hum_bytes(&_hum_used_bytes); 4524 *capacity_bytes = get_hum_bytes(&_hum_capacity_bytes); 4525 *prev_live_bytes = get_hum_bytes(&_hum_prev_live_bytes); 4526 *next_live_bytes = get_hum_bytes(&_hum_next_live_bytes); 4527 } 4528 4529 bool G1PrintRegionLivenessInfoClosure::doHeapRegion(HeapRegion* r) { 4530 const char* type = ""; 4531 HeapWord* bottom = r->bottom(); 4532 HeapWord* end = r->end(); 4533 size_t capacity_bytes = r->capacity(); 4534 size_t used_bytes = r->used(); 4535 size_t prev_live_bytes = r->live_bytes(); 4536 size_t next_live_bytes = r->next_live_bytes(); 4537 double gc_eff = r->gc_efficiency(); 4538 if (r->used() == 0) { 4539 type = "FREE"; 4540 } else if (r->is_survivor()) { 4541 type = "SURV"; 4542 } else if (r->is_young()) { 4543 type = "EDEN"; 4544 } else if (r->startsHumongous()) { 4545 type = "HUMS"; 4546 4547 assert(_hum_used_bytes == 0 && _hum_capacity_bytes == 0 && 4548 _hum_prev_live_bytes == 0 && _hum_next_live_bytes == 0, 4549 "they should have been zeroed after the last time we used them"); 4550 // Set up the _hum_* fields. 4551 _hum_capacity_bytes = capacity_bytes; 4552 _hum_used_bytes = used_bytes; 4553 _hum_prev_live_bytes = prev_live_bytes; 4554 _hum_next_live_bytes = next_live_bytes; 4555 get_hum_bytes(&used_bytes, &capacity_bytes, 4556 &prev_live_bytes, &next_live_bytes); 4557 end = bottom + HeapRegion::GrainWords; 4558 } else if (r->continuesHumongous()) { 4559 type = "HUMC"; 4560 get_hum_bytes(&used_bytes, &capacity_bytes, 4561 &prev_live_bytes, &next_live_bytes); 4562 assert(end == bottom + HeapRegion::GrainWords, "invariant"); 4563 } else { 4564 type = "OLD"; 4565 } 4566 4567 _total_used_bytes += used_bytes; 4568 _total_capacity_bytes += capacity_bytes; 4569 _total_prev_live_bytes += prev_live_bytes; 4570 _total_next_live_bytes += next_live_bytes; 4571 4572 // Print a line for this particular region. 4573 _out->print_cr(G1PPRL_LINE_PREFIX 4574 G1PPRL_TYPE_FORMAT 4575 G1PPRL_ADDR_BASE_FORMAT 4576 G1PPRL_BYTE_FORMAT 4577 G1PPRL_BYTE_FORMAT 4578 G1PPRL_BYTE_FORMAT 4579 G1PPRL_DOUBLE_FORMAT, 4580 type, bottom, end, 4581 used_bytes, prev_live_bytes, next_live_bytes, gc_eff); 4582 4583 return false; 4584 } 4585 4586 G1PrintRegionLivenessInfoClosure::~G1PrintRegionLivenessInfoClosure() { 4587 // Print the footer of the output. 4588 _out->print_cr(G1PPRL_LINE_PREFIX); 4589 _out->print_cr(G1PPRL_LINE_PREFIX 4590 " SUMMARY" 4591 G1PPRL_SUM_MB_FORMAT("capacity") 4592 G1PPRL_SUM_MB_PERC_FORMAT("used") 4593 G1PPRL_SUM_MB_PERC_FORMAT("prev-live") 4594 G1PPRL_SUM_MB_PERC_FORMAT("next-live"), 4595 bytes_to_mb(_total_capacity_bytes), 4596 bytes_to_mb(_total_used_bytes), 4597 perc(_total_used_bytes, _total_capacity_bytes), 4598 bytes_to_mb(_total_prev_live_bytes), 4599 perc(_total_prev_live_bytes, _total_capacity_bytes), 4600 bytes_to_mb(_total_next_live_bytes), 4601 perc(_total_next_live_bytes, _total_capacity_bytes)); 4602 _out->cr(); 4603 }