1 /* 2 * Copyright (c) 2001, 2016, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "classfile/classLoaderData.hpp" 27 #include "classfile/stringTable.hpp" 28 #include "classfile/symbolTable.hpp" 29 #include "classfile/systemDictionary.hpp" 30 #include "code/codeCache.hpp" 31 #include "gc/cms/cmsCollectorPolicy.hpp" 32 #include "gc/cms/cmsOopClosures.inline.hpp" 33 #include "gc/cms/compactibleFreeListSpace.hpp" 34 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 35 #include "gc/cms/concurrentMarkSweepThread.hpp" 36 #include "gc/cms/parNewGeneration.hpp" 37 #include "gc/cms/vmCMSOperations.hpp" 38 #include "gc/serial/genMarkSweep.hpp" 39 #include "gc/serial/tenuredGeneration.hpp" 40 #include "gc/shared/adaptiveSizePolicy.hpp" 41 #include "gc/shared/cardGeneration.inline.hpp" 42 #include "gc/shared/cardTableRS.hpp" 43 #include "gc/shared/collectedHeap.inline.hpp" 44 #include "gc/shared/collectorCounters.hpp" 45 #include "gc/shared/collectorPolicy.hpp" 46 #include "gc/shared/gcLocker.inline.hpp" 47 #include "gc/shared/gcPolicyCounters.hpp" 48 #include "gc/shared/gcTimer.hpp" 49 #include "gc/shared/gcTrace.hpp" 50 #include "gc/shared/gcTraceTime.inline.hpp" 51 #include "gc/shared/genCollectedHeap.hpp" 52 #include "gc/shared/genOopClosures.inline.hpp" 53 #include "gc/shared/isGCActiveMark.hpp" 54 #include "gc/shared/referencePolicy.hpp" 55 #include "gc/shared/strongRootsScope.hpp" 56 #include "gc/shared/taskqueue.inline.hpp" 57 #include "logging/log.hpp" 58 #include "memory/allocation.hpp" 59 #include "memory/iterator.inline.hpp" 60 #include "memory/padded.hpp" 61 #include "memory/resourceArea.hpp" 62 #include "oops/oop.inline.hpp" 63 #include "prims/jvmtiExport.hpp" 64 #include "runtime/atomic.hpp" 65 #include "runtime/globals_extension.hpp" 66 #include "runtime/handles.inline.hpp" 67 #include "runtime/java.hpp" 68 #include "runtime/orderAccess.inline.hpp" 69 #include "runtime/timer.hpp" 70 #include "runtime/vmThread.hpp" 71 #include "services/memoryService.hpp" 72 #include "services/runtimeService.hpp" 73 #include "utilities/stack.inline.hpp" 74 75 // statics 76 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 77 bool CMSCollector::_full_gc_requested = false; 78 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 79 80 ////////////////////////////////////////////////////////////////// 81 // In support of CMS/VM thread synchronization 82 ////////////////////////////////////////////////////////////////// 83 // We split use of the CGC_lock into 2 "levels". 84 // The low-level locking is of the usual CGC_lock monitor. We introduce 85 // a higher level "token" (hereafter "CMS token") built on top of the 86 // low level monitor (hereafter "CGC lock"). 87 // The token-passing protocol gives priority to the VM thread. The 88 // CMS-lock doesn't provide any fairness guarantees, but clients 89 // should ensure that it is only held for very short, bounded 90 // durations. 91 // 92 // When either of the CMS thread or the VM thread is involved in 93 // collection operations during which it does not want the other 94 // thread to interfere, it obtains the CMS token. 95 // 96 // If either thread tries to get the token while the other has 97 // it, that thread waits. However, if the VM thread and CMS thread 98 // both want the token, then the VM thread gets priority while the 99 // CMS thread waits. This ensures, for instance, that the "concurrent" 100 // phases of the CMS thread's work do not block out the VM thread 101 // for long periods of time as the CMS thread continues to hog 102 // the token. (See bug 4616232). 103 // 104 // The baton-passing functions are, however, controlled by the 105 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 106 // and here the low-level CMS lock, not the high level token, 107 // ensures mutual exclusion. 108 // 109 // Two important conditions that we have to satisfy: 110 // 1. if a thread does a low-level wait on the CMS lock, then it 111 // relinquishes the CMS token if it were holding that token 112 // when it acquired the low-level CMS lock. 113 // 2. any low-level notifications on the low-level lock 114 // should only be sent when a thread has relinquished the token. 115 // 116 // In the absence of either property, we'd have potential deadlock. 117 // 118 // We protect each of the CMS (concurrent and sequential) phases 119 // with the CMS _token_, not the CMS _lock_. 120 // 121 // The only code protected by CMS lock is the token acquisition code 122 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 123 // baton-passing code. 124 // 125 // Unfortunately, i couldn't come up with a good abstraction to factor and 126 // hide the naked CGC_lock manipulation in the baton-passing code 127 // further below. That's something we should try to do. Also, the proof 128 // of correctness of this 2-level locking scheme is far from obvious, 129 // and potentially quite slippery. We have an uneasy suspicion, for instance, 130 // that there may be a theoretical possibility of delay/starvation in the 131 // low-level lock/wait/notify scheme used for the baton-passing because of 132 // potential interference with the priority scheme embodied in the 133 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 134 // invocation further below and marked with "XXX 20011219YSR". 135 // Indeed, as we note elsewhere, this may become yet more slippery 136 // in the presence of multiple CMS and/or multiple VM threads. XXX 137 138 class CMSTokenSync: public StackObj { 139 private: 140 bool _is_cms_thread; 141 public: 142 CMSTokenSync(bool is_cms_thread): 143 _is_cms_thread(is_cms_thread) { 144 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 145 "Incorrect argument to constructor"); 146 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 147 } 148 149 ~CMSTokenSync() { 150 assert(_is_cms_thread ? 151 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 152 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 153 "Incorrect state"); 154 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 155 } 156 }; 157 158 // Convenience class that does a CMSTokenSync, and then acquires 159 // upto three locks. 160 class CMSTokenSyncWithLocks: public CMSTokenSync { 161 private: 162 // Note: locks are acquired in textual declaration order 163 // and released in the opposite order 164 MutexLockerEx _locker1, _locker2, _locker3; 165 public: 166 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 167 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 168 CMSTokenSync(is_cms_thread), 169 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 170 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 171 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 172 { } 173 }; 174 175 176 ////////////////////////////////////////////////////////////////// 177 // Concurrent Mark-Sweep Generation ///////////////////////////// 178 ////////////////////////////////////////////////////////////////// 179 180 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 181 182 // This struct contains per-thread things necessary to support parallel 183 // young-gen collection. 184 class CMSParGCThreadState: public CHeapObj<mtGC> { 185 public: 186 CompactibleFreeListSpaceLAB lab; 187 PromotionInfo promo; 188 189 // Constructor. 190 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 191 promo.setSpace(cfls); 192 } 193 }; 194 195 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 196 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 197 CardGeneration(rs, initial_byte_size, ct), 198 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 199 _did_compact(false) 200 { 201 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 202 HeapWord* end = (HeapWord*) _virtual_space.high(); 203 204 _direct_allocated_words = 0; 205 NOT_PRODUCT( 206 _numObjectsPromoted = 0; 207 _numWordsPromoted = 0; 208 _numObjectsAllocated = 0; 209 _numWordsAllocated = 0; 210 ) 211 212 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 213 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 214 _cmsSpace->_old_gen = this; 215 216 _gc_stats = new CMSGCStats(); 217 218 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 219 // offsets match. The ability to tell free chunks from objects 220 // depends on this property. 221 debug_only( 222 FreeChunk* junk = NULL; 223 assert(UseCompressedClassPointers || 224 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 225 "Offset of FreeChunk::_prev within FreeChunk must match" 226 " that of OopDesc::_klass within OopDesc"); 227 ) 228 229 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 230 for (uint i = 0; i < ParallelGCThreads; i++) { 231 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 232 } 233 234 _incremental_collection_failed = false; 235 // The "dilatation_factor" is the expansion that can occur on 236 // account of the fact that the minimum object size in the CMS 237 // generation may be larger than that in, say, a contiguous young 238 // generation. 239 // Ideally, in the calculation below, we'd compute the dilatation 240 // factor as: MinChunkSize/(promoting_gen's min object size) 241 // Since we do not have such a general query interface for the 242 // promoting generation, we'll instead just use the minimum 243 // object size (which today is a header's worth of space); 244 // note that all arithmetic is in units of HeapWords. 245 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 246 assert(_dilatation_factor >= 1.0, "from previous assert"); 247 } 248 249 250 // The field "_initiating_occupancy" represents the occupancy percentage 251 // at which we trigger a new collection cycle. Unless explicitly specified 252 // via CMSInitiatingOccupancyFraction (argument "io" below), it 253 // is calculated by: 254 // 255 // Let "f" be MinHeapFreeRatio in 256 // 257 // _initiating_occupancy = 100-f + 258 // f * (CMSTriggerRatio/100) 259 // where CMSTriggerRatio is the argument "tr" below. 260 // 261 // That is, if we assume the heap is at its desired maximum occupancy at the 262 // end of a collection, we let CMSTriggerRatio of the (purported) free 263 // space be allocated before initiating a new collection cycle. 264 // 265 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 266 assert(io <= 100 && tr <= 100, "Check the arguments"); 267 if (io >= 0) { 268 _initiating_occupancy = (double)io / 100.0; 269 } else { 270 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 271 (double)(tr * MinHeapFreeRatio) / 100.0) 272 / 100.0; 273 } 274 } 275 276 void ConcurrentMarkSweepGeneration::ref_processor_init() { 277 assert(collector() != NULL, "no collector"); 278 collector()->ref_processor_init(); 279 } 280 281 void CMSCollector::ref_processor_init() { 282 if (_ref_processor == NULL) { 283 // Allocate and initialize a reference processor 284 _ref_processor = 285 new ReferenceProcessor(_span, // span 286 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 287 ParallelGCThreads, // mt processing degree 288 _cmsGen->refs_discovery_is_mt(), // mt discovery 289 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 290 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 291 &_is_alive_closure); // closure for liveness info 292 // Initialize the _ref_processor field of CMSGen 293 _cmsGen->set_ref_processor(_ref_processor); 294 295 } 296 } 297 298 AdaptiveSizePolicy* CMSCollector::size_policy() { 299 GenCollectedHeap* gch = GenCollectedHeap::heap(); 300 return gch->gen_policy()->size_policy(); 301 } 302 303 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 304 305 const char* gen_name = "old"; 306 GenCollectorPolicy* gcp = GenCollectedHeap::heap()->gen_policy(); 307 // Generation Counters - generation 1, 1 subspace 308 _gen_counters = new GenerationCounters(gen_name, 1, 1, 309 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 310 311 _space_counters = new GSpaceCounters(gen_name, 0, 312 _virtual_space.reserved_size(), 313 this, _gen_counters); 314 } 315 316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 317 _cms_gen(cms_gen) 318 { 319 assert(alpha <= 100, "bad value"); 320 _saved_alpha = alpha; 321 322 // Initialize the alphas to the bootstrap value of 100. 323 _gc0_alpha = _cms_alpha = 100; 324 325 _cms_begin_time.update(); 326 _cms_end_time.update(); 327 328 _gc0_duration = 0.0; 329 _gc0_period = 0.0; 330 _gc0_promoted = 0; 331 332 _cms_duration = 0.0; 333 _cms_period = 0.0; 334 _cms_allocated = 0; 335 336 _cms_used_at_gc0_begin = 0; 337 _cms_used_at_gc0_end = 0; 338 _allow_duty_cycle_reduction = false; 339 _valid_bits = 0; 340 } 341 342 double CMSStats::cms_free_adjustment_factor(size_t free) const { 343 // TBD: CR 6909490 344 return 1.0; 345 } 346 347 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 348 } 349 350 // If promotion failure handling is on use 351 // the padded average size of the promotion for each 352 // young generation collection. 353 double CMSStats::time_until_cms_gen_full() const { 354 size_t cms_free = _cms_gen->cmsSpace()->free(); 355 GenCollectedHeap* gch = GenCollectedHeap::heap(); 356 size_t expected_promotion = MIN2(gch->young_gen()->capacity(), 357 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 358 if (cms_free > expected_promotion) { 359 // Start a cms collection if there isn't enough space to promote 360 // for the next young collection. Use the padded average as 361 // a safety factor. 362 cms_free -= expected_promotion; 363 364 // Adjust by the safety factor. 365 double cms_free_dbl = (double)cms_free; 366 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 367 // Apply a further correction factor which tries to adjust 368 // for recent occurance of concurrent mode failures. 369 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 370 cms_free_dbl = cms_free_dbl * cms_adjustment; 371 372 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 373 cms_free, expected_promotion); 374 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 375 // Add 1 in case the consumption rate goes to zero. 376 return cms_free_dbl / (cms_consumption_rate() + 1.0); 377 } 378 return 0.0; 379 } 380 381 // Compare the duration of the cms collection to the 382 // time remaining before the cms generation is empty. 383 // Note that the time from the start of the cms collection 384 // to the start of the cms sweep (less than the total 385 // duration of the cms collection) can be used. This 386 // has been tried and some applications experienced 387 // promotion failures early in execution. This was 388 // possibly because the averages were not accurate 389 // enough at the beginning. 390 double CMSStats::time_until_cms_start() const { 391 // We add "gc0_period" to the "work" calculation 392 // below because this query is done (mostly) at the 393 // end of a scavenge, so we need to conservatively 394 // account for that much possible delay 395 // in the query so as to avoid concurrent mode failures 396 // due to starting the collection just a wee bit too 397 // late. 398 double work = cms_duration() + gc0_period(); 399 double deadline = time_until_cms_gen_full(); 400 // If a concurrent mode failure occurred recently, we want to be 401 // more conservative and halve our expected time_until_cms_gen_full() 402 if (work > deadline) { 403 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 404 cms_duration(), gc0_period(), time_until_cms_gen_full()); 405 return 0.0; 406 } 407 return work - deadline; 408 } 409 410 #ifndef PRODUCT 411 void CMSStats::print_on(outputStream *st) const { 412 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 413 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 414 gc0_duration(), gc0_period(), gc0_promoted()); 415 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 416 cms_duration(), cms_period(), cms_allocated()); 417 st->print(",cms_since_beg=%g,cms_since_end=%g", 418 cms_time_since_begin(), cms_time_since_end()); 419 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 420 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 421 422 if (valid()) { 423 st->print(",promo_rate=%g,cms_alloc_rate=%g", 424 promotion_rate(), cms_allocation_rate()); 425 st->print(",cms_consumption_rate=%g,time_until_full=%g", 426 cms_consumption_rate(), time_until_cms_gen_full()); 427 } 428 st->cr(); 429 } 430 #endif // #ifndef PRODUCT 431 432 CMSCollector::CollectorState CMSCollector::_collectorState = 433 CMSCollector::Idling; 434 bool CMSCollector::_foregroundGCIsActive = false; 435 bool CMSCollector::_foregroundGCShouldWait = false; 436 437 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 438 CardTableRS* ct, 439 ConcurrentMarkSweepPolicy* cp): 440 _cmsGen(cmsGen), 441 _ct(ct), 442 _ref_processor(NULL), // will be set later 443 _conc_workers(NULL), // may be set later 444 _abort_preclean(false), 445 _start_sampling(false), 446 _between_prologue_and_epilogue(false), 447 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 448 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), 449 -1 /* lock-free */, "No_lock" /* dummy */), 450 _modUnionClosurePar(&_modUnionTable), 451 // Adjust my span to cover old (cms) gen 452 _span(cmsGen->reserved()), 453 // Construct the is_alive_closure with _span & markBitMap 454 _is_alive_closure(_span, &_markBitMap), 455 _restart_addr(NULL), 456 _overflow_list(NULL), 457 _stats(cmsGen), 458 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 459 //verify that this lock should be acquired with safepoint check. 460 Monitor::_safepoint_check_sometimes)), 461 _eden_chunk_array(NULL), // may be set in ctor body 462 _eden_chunk_capacity(0), // -- ditto -- 463 _eden_chunk_index(0), // -- ditto -- 464 _survivor_plab_array(NULL), // -- ditto -- 465 _survivor_chunk_array(NULL), // -- ditto -- 466 _survivor_chunk_capacity(0), // -- ditto -- 467 _survivor_chunk_index(0), // -- ditto -- 468 _ser_pmc_preclean_ovflw(0), 469 _ser_kac_preclean_ovflw(0), 470 _ser_pmc_remark_ovflw(0), 471 _par_pmc_remark_ovflw(0), 472 _ser_kac_ovflw(0), 473 _par_kac_ovflw(0), 474 #ifndef PRODUCT 475 _num_par_pushes(0), 476 #endif 477 _collection_count_start(0), 478 _verifying(false), 479 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 480 _completed_initialization(false), 481 _collector_policy(cp), 482 _should_unload_classes(CMSClassUnloadingEnabled), 483 _concurrent_cycles_since_last_unload(0), 484 _roots_scanning_options(GenCollectedHeap::SO_None), 485 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 486 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 487 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 488 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 489 _cms_start_registered(false) 490 { 491 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { 492 ExplicitGCInvokesConcurrent = true; 493 } 494 // Now expand the span and allocate the collection support structures 495 // (MUT, marking bit map etc.) to cover both generations subject to 496 // collection. 497 498 // For use by dirty card to oop closures. 499 _cmsGen->cmsSpace()->set_collector(this); 500 501 // Allocate MUT and marking bit map 502 { 503 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 504 if (!_markBitMap.allocate(_span)) { 505 log_warning(gc)("Failed to allocate CMS Bit Map"); 506 return; 507 } 508 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 509 } 510 { 511 _modUnionTable.allocate(_span); 512 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 513 } 514 515 if (!_markStack.allocate(MarkStackSize)) { 516 log_warning(gc)("Failed to allocate CMS Marking Stack"); 517 return; 518 } 519 520 // Support for multi-threaded concurrent phases 521 if (CMSConcurrentMTEnabled) { 522 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 523 // just for now 524 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 525 } 526 if (ConcGCThreads > 1) { 527 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 528 ConcGCThreads, true); 529 if (_conc_workers == NULL) { 530 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 531 CMSConcurrentMTEnabled = false; 532 } else { 533 _conc_workers->initialize_workers(); 534 } 535 } else { 536 CMSConcurrentMTEnabled = false; 537 } 538 } 539 if (!CMSConcurrentMTEnabled) { 540 ConcGCThreads = 0; 541 } else { 542 // Turn off CMSCleanOnEnter optimization temporarily for 543 // the MT case where it's not fixed yet; see 6178663. 544 CMSCleanOnEnter = false; 545 } 546 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 547 "Inconsistency"); 548 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 549 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 550 551 // Parallel task queues; these are shared for the 552 // concurrent and stop-world phases of CMS, but 553 // are not shared with parallel scavenge (ParNew). 554 { 555 uint i; 556 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 557 558 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 559 || ParallelRefProcEnabled) 560 && num_queues > 0) { 561 _task_queues = new OopTaskQueueSet(num_queues); 562 if (_task_queues == NULL) { 563 log_warning(gc)("task_queues allocation failure."); 564 return; 565 } 566 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 567 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 568 for (i = 0; i < num_queues; i++) { 569 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 570 if (q == NULL) { 571 log_warning(gc)("work_queue allocation failure."); 572 return; 573 } 574 _task_queues->register_queue(i, q); 575 } 576 for (i = 0; i < num_queues; i++) { 577 _task_queues->queue(i)->initialize(); 578 _hash_seed[i] = 17; // copied from ParNew 579 } 580 } 581 } 582 583 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 584 585 // Clip CMSBootstrapOccupancy between 0 and 100. 586 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 587 588 // Now tell CMS generations the identity of their collector 589 ConcurrentMarkSweepGeneration::set_collector(this); 590 591 // Create & start a CMS thread for this CMS collector 592 _cmsThread = ConcurrentMarkSweepThread::start(this); 593 assert(cmsThread() != NULL, "CMS Thread should have been created"); 594 assert(cmsThread()->collector() == this, 595 "CMS Thread should refer to this gen"); 596 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 597 598 // Support for parallelizing young gen rescan 599 GenCollectedHeap* gch = GenCollectedHeap::heap(); 600 assert(gch->young_gen()->kind() == Generation::ParNew, "CMS can only be used with ParNew"); 601 _young_gen = (ParNewGeneration*)gch->young_gen(); 602 if (gch->supports_inline_contig_alloc()) { 603 _top_addr = gch->top_addr(); 604 _end_addr = gch->end_addr(); 605 assert(_young_gen != NULL, "no _young_gen"); 606 _eden_chunk_index = 0; 607 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 608 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 609 } 610 611 // Support for parallelizing survivor space rescan 612 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 613 const size_t max_plab_samples = 614 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 615 616 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 617 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 618 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 619 _survivor_chunk_capacity = max_plab_samples; 620 for (uint i = 0; i < ParallelGCThreads; i++) { 621 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 622 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 623 assert(cur->end() == 0, "Should be 0"); 624 assert(cur->array() == vec, "Should be vec"); 625 assert(cur->capacity() == max_plab_samples, "Error"); 626 } 627 } 628 629 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 630 _gc_counters = new CollectorCounters("CMS", 1); 631 _completed_initialization = true; 632 _inter_sweep_timer.start(); // start of time 633 } 634 635 const char* ConcurrentMarkSweepGeneration::name() const { 636 return "concurrent mark-sweep generation"; 637 } 638 void ConcurrentMarkSweepGeneration::update_counters() { 639 if (UsePerfData) { 640 _space_counters->update_all(); 641 _gen_counters->update_all(); 642 } 643 } 644 645 // this is an optimized version of update_counters(). it takes the 646 // used value as a parameter rather than computing it. 647 // 648 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 649 if (UsePerfData) { 650 _space_counters->update_used(used); 651 _space_counters->update_capacity(); 652 _gen_counters->update_all(); 653 } 654 } 655 656 void ConcurrentMarkSweepGeneration::print() const { 657 Generation::print(); 658 cmsSpace()->print(); 659 } 660 661 #ifndef PRODUCT 662 void ConcurrentMarkSweepGeneration::print_statistics() { 663 cmsSpace()->printFLCensus(0); 664 } 665 #endif 666 667 size_t 668 ConcurrentMarkSweepGeneration::contiguous_available() const { 669 // dld proposes an improvement in precision here. If the committed 670 // part of the space ends in a free block we should add that to 671 // uncommitted size in the calculation below. Will make this 672 // change later, staying with the approximation below for the 673 // time being. -- ysr. 674 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 675 } 676 677 size_t 678 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 679 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 680 } 681 682 size_t ConcurrentMarkSweepGeneration::max_available() const { 683 return free() + _virtual_space.uncommitted_size(); 684 } 685 686 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 687 size_t available = max_available(); 688 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 689 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 690 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 691 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 692 return res; 693 } 694 695 // At a promotion failure dump information on block layout in heap 696 // (cms old generation). 697 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 698 Log(gc, promotion) log; 699 if (log.is_trace()) { 700 ResourceMark rm; 701 cmsSpace()->dump_at_safepoint_with_locks(collector(), log.trace_stream()); 702 } 703 } 704 705 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 706 // Clear the promotion information. These pointers can be adjusted 707 // along with all the other pointers into the heap but 708 // compaction is expected to be a rare event with 709 // a heap using cms so don't do it without seeing the need. 710 for (uint i = 0; i < ParallelGCThreads; i++) { 711 _par_gc_thread_states[i]->promo.reset(); 712 } 713 } 714 715 void ConcurrentMarkSweepGeneration::compute_new_size() { 716 assert_locked_or_safepoint(Heap_lock); 717 718 // If incremental collection failed, we just want to expand 719 // to the limit. 720 if (incremental_collection_failed()) { 721 clear_incremental_collection_failed(); 722 grow_to_reserved(); 723 return; 724 } 725 726 // The heap has been compacted but not reset yet. 727 // Any metric such as free() or used() will be incorrect. 728 729 CardGeneration::compute_new_size(); 730 731 // Reset again after a possible resizing 732 if (did_compact()) { 733 cmsSpace()->reset_after_compaction(); 734 } 735 } 736 737 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 738 assert_locked_or_safepoint(Heap_lock); 739 740 // If incremental collection failed, we just want to expand 741 // to the limit. 742 if (incremental_collection_failed()) { 743 clear_incremental_collection_failed(); 744 grow_to_reserved(); 745 return; 746 } 747 748 double free_percentage = ((double) free()) / capacity(); 749 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 750 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 751 752 // compute expansion delta needed for reaching desired free percentage 753 if (free_percentage < desired_free_percentage) { 754 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 755 assert(desired_capacity >= capacity(), "invalid expansion size"); 756 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 757 Log(gc) log; 758 if (log.is_trace()) { 759 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 760 log.trace("From compute_new_size: "); 761 log.trace(" Free fraction %f", free_percentage); 762 log.trace(" Desired free fraction %f", desired_free_percentage); 763 log.trace(" Maximum free fraction %f", maximum_free_percentage); 764 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 765 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 766 GenCollectedHeap* gch = GenCollectedHeap::heap(); 767 assert(gch->is_old_gen(this), "The CMS generation should always be the old generation"); 768 size_t young_size = gch->young_gen()->capacity(); 769 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 770 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 771 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 772 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 773 } 774 // safe if expansion fails 775 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 776 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 777 } else { 778 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 779 assert(desired_capacity <= capacity(), "invalid expansion size"); 780 size_t shrink_bytes = capacity() - desired_capacity; 781 // Don't shrink unless the delta is greater than the minimum shrink we want 782 if (shrink_bytes >= MinHeapDeltaBytes) { 783 shrink_free_list_by(shrink_bytes); 784 } 785 } 786 } 787 788 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 789 return cmsSpace()->freelistLock(); 790 } 791 792 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 793 CMSSynchronousYieldRequest yr; 794 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 795 return have_lock_and_allocate(size, tlab); 796 } 797 798 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 799 bool tlab /* ignored */) { 800 assert_lock_strong(freelistLock()); 801 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 802 HeapWord* res = cmsSpace()->allocate(adjustedSize); 803 // Allocate the object live (grey) if the background collector has 804 // started marking. This is necessary because the marker may 805 // have passed this address and consequently this object will 806 // not otherwise be greyed and would be incorrectly swept up. 807 // Note that if this object contains references, the writing 808 // of those references will dirty the card containing this object 809 // allowing the object to be blackened (and its references scanned) 810 // either during a preclean phase or at the final checkpoint. 811 if (res != NULL) { 812 // We may block here with an uninitialized object with 813 // its mark-bit or P-bits not yet set. Such objects need 814 // to be safely navigable by block_start(). 815 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 816 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 817 collector()->direct_allocated(res, adjustedSize); 818 _direct_allocated_words += adjustedSize; 819 // allocation counters 820 NOT_PRODUCT( 821 _numObjectsAllocated++; 822 _numWordsAllocated += (int)adjustedSize; 823 ) 824 } 825 return res; 826 } 827 828 // In the case of direct allocation by mutators in a generation that 829 // is being concurrently collected, the object must be allocated 830 // live (grey) if the background collector has started marking. 831 // This is necessary because the marker may 832 // have passed this address and consequently this object will 833 // not otherwise be greyed and would be incorrectly swept up. 834 // Note that if this object contains references, the writing 835 // of those references will dirty the card containing this object 836 // allowing the object to be blackened (and its references scanned) 837 // either during a preclean phase or at the final checkpoint. 838 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 839 assert(_markBitMap.covers(start, size), "Out of bounds"); 840 if (_collectorState >= Marking) { 841 MutexLockerEx y(_markBitMap.lock(), 842 Mutex::_no_safepoint_check_flag); 843 // [see comments preceding SweepClosure::do_blk() below for details] 844 // 845 // Can the P-bits be deleted now? JJJ 846 // 847 // 1. need to mark the object as live so it isn't collected 848 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 849 // 3. need to mark the end of the object so marking, precleaning or sweeping 850 // can skip over uninitialized or unparsable objects. An allocated 851 // object is considered uninitialized for our purposes as long as 852 // its klass word is NULL. All old gen objects are parsable 853 // as soon as they are initialized.) 854 _markBitMap.mark(start); // object is live 855 _markBitMap.mark(start + 1); // object is potentially uninitialized? 856 _markBitMap.mark(start + size - 1); 857 // mark end of object 858 } 859 // check that oop looks uninitialized 860 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 861 } 862 863 void CMSCollector::promoted(bool par, HeapWord* start, 864 bool is_obj_array, size_t obj_size) { 865 assert(_markBitMap.covers(start), "Out of bounds"); 866 // See comment in direct_allocated() about when objects should 867 // be allocated live. 868 if (_collectorState >= Marking) { 869 // we already hold the marking bit map lock, taken in 870 // the prologue 871 if (par) { 872 _markBitMap.par_mark(start); 873 } else { 874 _markBitMap.mark(start); 875 } 876 // We don't need to mark the object as uninitialized (as 877 // in direct_allocated above) because this is being done with the 878 // world stopped and the object will be initialized by the 879 // time the marking, precleaning or sweeping get to look at it. 880 // But see the code for copying objects into the CMS generation, 881 // where we need to ensure that concurrent readers of the 882 // block offset table are able to safely navigate a block that 883 // is in flux from being free to being allocated (and in 884 // transition while being copied into) and subsequently 885 // becoming a bona-fide object when the copy/promotion is complete. 886 assert(SafepointSynchronize::is_at_safepoint(), 887 "expect promotion only at safepoints"); 888 889 if (_collectorState < Sweeping) { 890 // Mark the appropriate cards in the modUnionTable, so that 891 // this object gets scanned before the sweep. If this is 892 // not done, CMS generation references in the object might 893 // not get marked. 894 // For the case of arrays, which are otherwise precisely 895 // marked, we need to dirty the entire array, not just its head. 896 if (is_obj_array) { 897 // The [par_]mark_range() method expects mr.end() below to 898 // be aligned to the granularity of a bit's representation 899 // in the heap. In the case of the MUT below, that's a 900 // card size. 901 MemRegion mr(start, 902 (HeapWord*)round_to((intptr_t)(start + obj_size), 903 CardTableModRefBS::card_size /* bytes */)); 904 if (par) { 905 _modUnionTable.par_mark_range(mr); 906 } else { 907 _modUnionTable.mark_range(mr); 908 } 909 } else { // not an obj array; we can just mark the head 910 if (par) { 911 _modUnionTable.par_mark(start); 912 } else { 913 _modUnionTable.mark(start); 914 } 915 } 916 } 917 } 918 } 919 920 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 921 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 922 // allocate, copy and if necessary update promoinfo -- 923 // delegate to underlying space. 924 assert_lock_strong(freelistLock()); 925 926 #ifndef PRODUCT 927 if (GenCollectedHeap::heap()->promotion_should_fail()) { 928 return NULL; 929 } 930 #endif // #ifndef PRODUCT 931 932 oop res = _cmsSpace->promote(obj, obj_size); 933 if (res == NULL) { 934 // expand and retry 935 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 936 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 937 // Since this is the old generation, we don't try to promote 938 // into a more senior generation. 939 res = _cmsSpace->promote(obj, obj_size); 940 } 941 if (res != NULL) { 942 // See comment in allocate() about when objects should 943 // be allocated live. 944 assert(obj->is_oop(), "Will dereference klass pointer below"); 945 collector()->promoted(false, // Not parallel 946 (HeapWord*)res, obj->is_objArray(), obj_size); 947 // promotion counters 948 NOT_PRODUCT( 949 _numObjectsPromoted++; 950 _numWordsPromoted += 951 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 952 ) 953 } 954 return res; 955 } 956 957 958 // IMPORTANT: Notes on object size recognition in CMS. 959 // --------------------------------------------------- 960 // A block of storage in the CMS generation is always in 961 // one of three states. A free block (FREE), an allocated 962 // object (OBJECT) whose size() method reports the correct size, 963 // and an intermediate state (TRANSIENT) in which its size cannot 964 // be accurately determined. 965 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 966 // ----------------------------------------------------- 967 // FREE: klass_word & 1 == 1; mark_word holds block size 968 // 969 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 970 // obj->size() computes correct size 971 // 972 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 973 // 974 // STATE IDENTIFICATION: (64 bit+COOPS) 975 // ------------------------------------ 976 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 977 // 978 // OBJECT: klass_word installed; klass_word != 0; 979 // obj->size() computes correct size 980 // 981 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 982 // 983 // 984 // STATE TRANSITION DIAGRAM 985 // 986 // mut / parnew mut / parnew 987 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 988 // ^ | 989 // |------------------------ DEAD <------------------------------------| 990 // sweep mut 991 // 992 // While a block is in TRANSIENT state its size cannot be determined 993 // so readers will either need to come back later or stall until 994 // the size can be determined. Note that for the case of direct 995 // allocation, P-bits, when available, may be used to determine the 996 // size of an object that may not yet have been initialized. 997 998 // Things to support parallel young-gen collection. 999 oop 1000 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1001 oop old, markOop m, 1002 size_t word_sz) { 1003 #ifndef PRODUCT 1004 if (GenCollectedHeap::heap()->promotion_should_fail()) { 1005 return NULL; 1006 } 1007 #endif // #ifndef PRODUCT 1008 1009 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1010 PromotionInfo* promoInfo = &ps->promo; 1011 // if we are tracking promotions, then first ensure space for 1012 // promotion (including spooling space for saving header if necessary). 1013 // then allocate and copy, then track promoted info if needed. 1014 // When tracking (see PromotionInfo::track()), the mark word may 1015 // be displaced and in this case restoration of the mark word 1016 // occurs in the (oop_since_save_marks_)iterate phase. 1017 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1018 // Out of space for allocating spooling buffers; 1019 // try expanding and allocating spooling buffers. 1020 if (!expand_and_ensure_spooling_space(promoInfo)) { 1021 return NULL; 1022 } 1023 } 1024 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1025 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1026 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1027 if (obj_ptr == NULL) { 1028 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1029 if (obj_ptr == NULL) { 1030 return NULL; 1031 } 1032 } 1033 oop obj = oop(obj_ptr); 1034 OrderAccess::storestore(); 1035 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1036 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1037 // IMPORTANT: See note on object initialization for CMS above. 1038 // Otherwise, copy the object. Here we must be careful to insert the 1039 // klass pointer last, since this marks the block as an allocated object. 1040 // Except with compressed oops it's the mark word. 1041 HeapWord* old_ptr = (HeapWord*)old; 1042 // Restore the mark word copied above. 1043 obj->set_mark(m); 1044 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1045 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1046 OrderAccess::storestore(); 1047 1048 if (UseCompressedClassPointers) { 1049 // Copy gap missed by (aligned) header size calculation below 1050 obj->set_klass_gap(old->klass_gap()); 1051 } 1052 if (word_sz > (size_t)oopDesc::header_size()) { 1053 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1054 obj_ptr + oopDesc::header_size(), 1055 word_sz - oopDesc::header_size()); 1056 } 1057 1058 // Now we can track the promoted object, if necessary. We take care 1059 // to delay the transition from uninitialized to full object 1060 // (i.e., insertion of klass pointer) until after, so that it 1061 // atomically becomes a promoted object. 1062 if (promoInfo->tracking()) { 1063 promoInfo->track((PromotedObject*)obj, old->klass()); 1064 } 1065 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1066 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1067 assert(old->is_oop(), "Will use and dereference old klass ptr below"); 1068 1069 // Finally, install the klass pointer (this should be volatile). 1070 OrderAccess::storestore(); 1071 obj->set_klass(old->klass()); 1072 // We should now be able to calculate the right size for this object 1073 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1074 1075 collector()->promoted(true, // parallel 1076 obj_ptr, old->is_objArray(), word_sz); 1077 1078 NOT_PRODUCT( 1079 Atomic::inc_ptr(&_numObjectsPromoted); 1080 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); 1081 ) 1082 1083 return obj; 1084 } 1085 1086 void 1087 ConcurrentMarkSweepGeneration:: 1088 par_promote_alloc_done(int thread_num) { 1089 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1090 ps->lab.retire(thread_num); 1091 } 1092 1093 void 1094 ConcurrentMarkSweepGeneration:: 1095 par_oop_since_save_marks_iterate_done(int thread_num) { 1096 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1097 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1098 ps->promo.promoted_oops_iterate_nv(dummy_cl); 1099 1100 // Because card-scanning has been completed, subsequent phases 1101 // (e.g., reference processing) will not need to recognize which 1102 // objects have been promoted during this GC. So, we can now disable 1103 // promotion tracking. 1104 ps->promo.stopTrackingPromotions(); 1105 } 1106 1107 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1108 size_t size, 1109 bool tlab) 1110 { 1111 // We allow a STW collection only if a full 1112 // collection was requested. 1113 return full || should_allocate(size, tlab); // FIX ME !!! 1114 // This and promotion failure handling are connected at the 1115 // hip and should be fixed by untying them. 1116 } 1117 1118 bool CMSCollector::shouldConcurrentCollect() { 1119 LogTarget(Trace, gc) log; 1120 1121 if (_full_gc_requested) { 1122 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1123 return true; 1124 } 1125 1126 FreelistLocker x(this); 1127 // ------------------------------------------------------------------ 1128 // Print out lots of information which affects the initiation of 1129 // a collection. 1130 if (log.is_enabled() && stats().valid()) { 1131 log.print("CMSCollector shouldConcurrentCollect: "); 1132 1133 LogStream out(log); 1134 stats().print_on(&out); 1135 1136 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1137 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1138 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1139 log.print("promotion_rate=%g", stats().promotion_rate()); 1140 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1141 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1142 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1143 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1144 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1145 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1146 } 1147 // ------------------------------------------------------------------ 1148 1149 // If the estimated time to complete a cms collection (cms_duration()) 1150 // is less than the estimated time remaining until the cms generation 1151 // is full, start a collection. 1152 if (!UseCMSInitiatingOccupancyOnly) { 1153 if (stats().valid()) { 1154 if (stats().time_until_cms_start() == 0.0) { 1155 return true; 1156 } 1157 } else { 1158 // We want to conservatively collect somewhat early in order 1159 // to try and "bootstrap" our CMS/promotion statistics; 1160 // this branch will not fire after the first successful CMS 1161 // collection because the stats should then be valid. 1162 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1163 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1164 _cmsGen->occupancy(), _bootstrap_occupancy); 1165 return true; 1166 } 1167 } 1168 } 1169 1170 // Otherwise, we start a collection cycle if 1171 // old gen want a collection cycle started. Each may use 1172 // an appropriate criterion for making this decision. 1173 // XXX We need to make sure that the gen expansion 1174 // criterion dovetails well with this. XXX NEED TO FIX THIS 1175 if (_cmsGen->should_concurrent_collect()) { 1176 log.print("CMS old gen initiated"); 1177 return true; 1178 } 1179 1180 // We start a collection if we believe an incremental collection may fail; 1181 // this is not likely to be productive in practice because it's probably too 1182 // late anyway. 1183 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1184 assert(gch->collector_policy()->is_generation_policy(), 1185 "You may want to check the correctness of the following"); 1186 if (gch->incremental_collection_will_fail(true /* consult_young */)) { 1187 log.print("CMSCollector: collect because incremental collection will fail "); 1188 return true; 1189 } 1190 1191 if (MetaspaceGC::should_concurrent_collect()) { 1192 log.print("CMSCollector: collect for metadata allocation "); 1193 return true; 1194 } 1195 1196 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1197 if (CMSTriggerInterval >= 0) { 1198 if (CMSTriggerInterval == 0) { 1199 // Trigger always 1200 return true; 1201 } 1202 1203 // Check the CMS time since begin (we do not check the stats validity 1204 // as we want to be able to trigger the first CMS cycle as well) 1205 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1206 if (stats().valid()) { 1207 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1208 stats().cms_time_since_begin()); 1209 } else { 1210 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1211 } 1212 return true; 1213 } 1214 } 1215 1216 return false; 1217 } 1218 1219 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1220 1221 // Clear _expansion_cause fields of constituent generations 1222 void CMSCollector::clear_expansion_cause() { 1223 _cmsGen->clear_expansion_cause(); 1224 } 1225 1226 // We should be conservative in starting a collection cycle. To 1227 // start too eagerly runs the risk of collecting too often in the 1228 // extreme. To collect too rarely falls back on full collections, 1229 // which works, even if not optimum in terms of concurrent work. 1230 // As a work around for too eagerly collecting, use the flag 1231 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1232 // giving the user an easily understandable way of controlling the 1233 // collections. 1234 // We want to start a new collection cycle if any of the following 1235 // conditions hold: 1236 // . our current occupancy exceeds the configured initiating occupancy 1237 // for this generation, or 1238 // . we recently needed to expand this space and have not, since that 1239 // expansion, done a collection of this generation, or 1240 // . the underlying space believes that it may be a good idea to initiate 1241 // a concurrent collection (this may be based on criteria such as the 1242 // following: the space uses linear allocation and linear allocation is 1243 // going to fail, or there is believed to be excessive fragmentation in 1244 // the generation, etc... or ... 1245 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1246 // the case of the old generation; see CR 6543076): 1247 // we may be approaching a point at which allocation requests may fail because 1248 // we will be out of sufficient free space given allocation rate estimates.] 1249 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1250 1251 assert_lock_strong(freelistLock()); 1252 if (occupancy() > initiating_occupancy()) { 1253 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1254 short_name(), occupancy(), initiating_occupancy()); 1255 return true; 1256 } 1257 if (UseCMSInitiatingOccupancyOnly) { 1258 return false; 1259 } 1260 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1261 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1262 return true; 1263 } 1264 return false; 1265 } 1266 1267 void ConcurrentMarkSweepGeneration::collect(bool full, 1268 bool clear_all_soft_refs, 1269 size_t size, 1270 bool tlab) 1271 { 1272 collector()->collect(full, clear_all_soft_refs, size, tlab); 1273 } 1274 1275 void CMSCollector::collect(bool full, 1276 bool clear_all_soft_refs, 1277 size_t size, 1278 bool tlab) 1279 { 1280 // The following "if" branch is present for defensive reasons. 1281 // In the current uses of this interface, it can be replaced with: 1282 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1283 // But I am not placing that assert here to allow future 1284 // generality in invoking this interface. 1285 if (GCLocker::is_active()) { 1286 // A consistency test for GCLocker 1287 assert(GCLocker::needs_gc(), "Should have been set already"); 1288 // Skip this foreground collection, instead 1289 // expanding the heap if necessary. 1290 // Need the free list locks for the call to free() in compute_new_size() 1291 compute_new_size(); 1292 return; 1293 } 1294 acquire_control_and_collect(full, clear_all_soft_refs); 1295 } 1296 1297 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1298 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1299 unsigned int gc_count = gch->total_full_collections(); 1300 if (gc_count == full_gc_count) { 1301 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1302 _full_gc_requested = true; 1303 _full_gc_cause = cause; 1304 CGC_lock->notify(); // nudge CMS thread 1305 } else { 1306 assert(gc_count > full_gc_count, "Error: causal loop"); 1307 } 1308 } 1309 1310 bool CMSCollector::is_external_interruption() { 1311 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); 1312 return GCCause::is_user_requested_gc(cause) || 1313 GCCause::is_serviceability_requested_gc(cause); 1314 } 1315 1316 void CMSCollector::report_concurrent_mode_interruption() { 1317 if (is_external_interruption()) { 1318 log_debug(gc)("Concurrent mode interrupted"); 1319 } else { 1320 log_debug(gc)("Concurrent mode failure"); 1321 _gc_tracer_cm->report_concurrent_mode_failure(); 1322 } 1323 } 1324 1325 1326 // The foreground and background collectors need to coordinate in order 1327 // to make sure that they do not mutually interfere with CMS collections. 1328 // When a background collection is active, 1329 // the foreground collector may need to take over (preempt) and 1330 // synchronously complete an ongoing collection. Depending on the 1331 // frequency of the background collections and the heap usage 1332 // of the application, this preemption can be seldom or frequent. 1333 // There are only certain 1334 // points in the background collection that the "collection-baton" 1335 // can be passed to the foreground collector. 1336 // 1337 // The foreground collector will wait for the baton before 1338 // starting any part of the collection. The foreground collector 1339 // will only wait at one location. 1340 // 1341 // The background collector will yield the baton before starting a new 1342 // phase of the collection (e.g., before initial marking, marking from roots, 1343 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1344 // of the loop which switches the phases. The background collector does some 1345 // of the phases (initial mark, final re-mark) with the world stopped. 1346 // Because of locking involved in stopping the world, 1347 // the foreground collector should not block waiting for the background 1348 // collector when it is doing a stop-the-world phase. The background 1349 // collector will yield the baton at an additional point just before 1350 // it enters a stop-the-world phase. Once the world is stopped, the 1351 // background collector checks the phase of the collection. If the 1352 // phase has not changed, it proceeds with the collection. If the 1353 // phase has changed, it skips that phase of the collection. See 1354 // the comments on the use of the Heap_lock in collect_in_background(). 1355 // 1356 // Variable used in baton passing. 1357 // _foregroundGCIsActive - Set to true by the foreground collector when 1358 // it wants the baton. The foreground clears it when it has finished 1359 // the collection. 1360 // _foregroundGCShouldWait - Set to true by the background collector 1361 // when it is running. The foreground collector waits while 1362 // _foregroundGCShouldWait is true. 1363 // CGC_lock - monitor used to protect access to the above variables 1364 // and to notify the foreground and background collectors. 1365 // _collectorState - current state of the CMS collection. 1366 // 1367 // The foreground collector 1368 // acquires the CGC_lock 1369 // sets _foregroundGCIsActive 1370 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1371 // various locks acquired in preparation for the collection 1372 // are released so as not to block the background collector 1373 // that is in the midst of a collection 1374 // proceeds with the collection 1375 // clears _foregroundGCIsActive 1376 // returns 1377 // 1378 // The background collector in a loop iterating on the phases of the 1379 // collection 1380 // acquires the CGC_lock 1381 // sets _foregroundGCShouldWait 1382 // if _foregroundGCIsActive is set 1383 // clears _foregroundGCShouldWait, notifies _CGC_lock 1384 // waits on _CGC_lock for _foregroundGCIsActive to become false 1385 // and exits the loop. 1386 // otherwise 1387 // proceed with that phase of the collection 1388 // if the phase is a stop-the-world phase, 1389 // yield the baton once more just before enqueueing 1390 // the stop-world CMS operation (executed by the VM thread). 1391 // returns after all phases of the collection are done 1392 // 1393 1394 void CMSCollector::acquire_control_and_collect(bool full, 1395 bool clear_all_soft_refs) { 1396 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1397 assert(!Thread::current()->is_ConcurrentGC_thread(), 1398 "shouldn't try to acquire control from self!"); 1399 1400 // Start the protocol for acquiring control of the 1401 // collection from the background collector (aka CMS thread). 1402 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1403 "VM thread should have CMS token"); 1404 // Remember the possibly interrupted state of an ongoing 1405 // concurrent collection 1406 CollectorState first_state = _collectorState; 1407 1408 // Signal to a possibly ongoing concurrent collection that 1409 // we want to do a foreground collection. 1410 _foregroundGCIsActive = true; 1411 1412 // release locks and wait for a notify from the background collector 1413 // releasing the locks in only necessary for phases which 1414 // do yields to improve the granularity of the collection. 1415 assert_lock_strong(bitMapLock()); 1416 // We need to lock the Free list lock for the space that we are 1417 // currently collecting. 1418 assert(haveFreelistLocks(), "Must be holding free list locks"); 1419 bitMapLock()->unlock(); 1420 releaseFreelistLocks(); 1421 { 1422 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1423 if (_foregroundGCShouldWait) { 1424 // We are going to be waiting for action for the CMS thread; 1425 // it had better not be gone (for instance at shutdown)! 1426 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1427 "CMS thread must be running"); 1428 // Wait here until the background collector gives us the go-ahead 1429 ConcurrentMarkSweepThread::clear_CMS_flag( 1430 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1431 // Get a possibly blocked CMS thread going: 1432 // Note that we set _foregroundGCIsActive true above, 1433 // without protection of the CGC_lock. 1434 CGC_lock->notify(); 1435 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1436 "Possible deadlock"); 1437 while (_foregroundGCShouldWait) { 1438 // wait for notification 1439 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1440 // Possibility of delay/starvation here, since CMS token does 1441 // not know to give priority to VM thread? Actually, i think 1442 // there wouldn't be any delay/starvation, but the proof of 1443 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1444 } 1445 ConcurrentMarkSweepThread::set_CMS_flag( 1446 ConcurrentMarkSweepThread::CMS_vm_has_token); 1447 } 1448 } 1449 // The CMS_token is already held. Get back the other locks. 1450 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1451 "VM thread should have CMS token"); 1452 getFreelistLocks(); 1453 bitMapLock()->lock_without_safepoint_check(); 1454 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1455 p2i(Thread::current()), first_state); 1456 log_debug(gc, state)(" gets control with state %d", _collectorState); 1457 1458 // Inform cms gen if this was due to partial collection failing. 1459 // The CMS gen may use this fact to determine its expansion policy. 1460 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1461 if (gch->incremental_collection_will_fail(false /* don't consult_young */)) { 1462 assert(!_cmsGen->incremental_collection_failed(), 1463 "Should have been noticed, reacted to and cleared"); 1464 _cmsGen->set_incremental_collection_failed(); 1465 } 1466 1467 if (first_state > Idling) { 1468 report_concurrent_mode_interruption(); 1469 } 1470 1471 set_did_compact(true); 1472 1473 // If the collection is being acquired from the background 1474 // collector, there may be references on the discovered 1475 // references lists. Abandon those references, since some 1476 // of them may have become unreachable after concurrent 1477 // discovery; the STW compacting collector will redo discovery 1478 // more precisely, without being subject to floating garbage. 1479 // Leaving otherwise unreachable references in the discovered 1480 // lists would require special handling. 1481 ref_processor()->disable_discovery(); 1482 ref_processor()->abandon_partial_discovery(); 1483 ref_processor()->verify_no_references_recorded(); 1484 1485 if (first_state > Idling) { 1486 save_heap_summary(); 1487 } 1488 1489 do_compaction_work(clear_all_soft_refs); 1490 1491 // Has the GC time limit been exceeded? 1492 size_t max_eden_size = _young_gen->max_eden_size(); 1493 GCCause::Cause gc_cause = gch->gc_cause(); 1494 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1495 _young_gen->eden()->used(), 1496 _cmsGen->max_capacity(), 1497 max_eden_size, 1498 full, 1499 gc_cause, 1500 gch->collector_policy()); 1501 1502 // Reset the expansion cause, now that we just completed 1503 // a collection cycle. 1504 clear_expansion_cause(); 1505 _foregroundGCIsActive = false; 1506 return; 1507 } 1508 1509 // Resize the tenured generation 1510 // after obtaining the free list locks for the 1511 // two generations. 1512 void CMSCollector::compute_new_size() { 1513 assert_locked_or_safepoint(Heap_lock); 1514 FreelistLocker z(this); 1515 MetaspaceGC::compute_new_size(); 1516 _cmsGen->compute_new_size_free_list(); 1517 } 1518 1519 // A work method used by the foreground collector to do 1520 // a mark-sweep-compact. 1521 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1522 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1523 1524 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1525 gc_timer->register_gc_start(); 1526 1527 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1528 gc_tracer->report_gc_start(gch->gc_cause(), gc_timer->gc_start()); 1529 1530 gch->pre_full_gc_dump(gc_timer); 1531 1532 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1533 1534 // Temporarily widen the span of the weak reference processing to 1535 // the entire heap. 1536 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); 1537 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1538 // Temporarily, clear the "is_alive_non_header" field of the 1539 // reference processor. 1540 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1541 // Temporarily make reference _processing_ single threaded (non-MT). 1542 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1543 // Temporarily make refs discovery atomic 1544 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1545 // Temporarily make reference _discovery_ single threaded (non-MT) 1546 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1547 1548 ref_processor()->set_enqueuing_is_done(false); 1549 ref_processor()->enable_discovery(); 1550 ref_processor()->setup_policy(clear_all_soft_refs); 1551 // If an asynchronous collection finishes, the _modUnionTable is 1552 // all clear. If we are assuming the collection from an asynchronous 1553 // collection, clear the _modUnionTable. 1554 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1555 "_modUnionTable should be clear if the baton was not passed"); 1556 _modUnionTable.clear_all(); 1557 assert(_collectorState != Idling || _ct->klass_rem_set()->mod_union_is_clear(), 1558 "mod union for klasses should be clear if the baton was passed"); 1559 _ct->klass_rem_set()->clear_mod_union(); 1560 1561 // We must adjust the allocation statistics being maintained 1562 // in the free list space. We do so by reading and clearing 1563 // the sweep timer and updating the block flux rate estimates below. 1564 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1565 if (_inter_sweep_timer.is_active()) { 1566 _inter_sweep_timer.stop(); 1567 // Note that we do not use this sample to update the _inter_sweep_estimate. 1568 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1569 _inter_sweep_estimate.padded_average(), 1570 _intra_sweep_estimate.padded_average()); 1571 } 1572 1573 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1574 #ifdef ASSERT 1575 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1576 size_t free_size = cms_space->free(); 1577 assert(free_size == 1578 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1579 * HeapWordSize, 1580 "All the free space should be compacted into one chunk at top"); 1581 assert(cms_space->dictionary()->total_chunk_size( 1582 debug_only(cms_space->freelistLock())) == 0 || 1583 cms_space->totalSizeInIndexedFreeLists() == 0, 1584 "All the free space should be in a single chunk"); 1585 size_t num = cms_space->totalCount(); 1586 assert((free_size == 0 && num == 0) || 1587 (free_size > 0 && (num == 1 || num == 2)), 1588 "There should be at most 2 free chunks after compaction"); 1589 #endif // ASSERT 1590 _collectorState = Resetting; 1591 assert(_restart_addr == NULL, 1592 "Should have been NULL'd before baton was passed"); 1593 reset_stw(); 1594 _cmsGen->reset_after_compaction(); 1595 _concurrent_cycles_since_last_unload = 0; 1596 1597 // Clear any data recorded in the PLAB chunk arrays. 1598 if (_survivor_plab_array != NULL) { 1599 reset_survivor_plab_arrays(); 1600 } 1601 1602 // Adjust the per-size allocation stats for the next epoch. 1603 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1604 // Restart the "inter sweep timer" for the next epoch. 1605 _inter_sweep_timer.reset(); 1606 _inter_sweep_timer.start(); 1607 1608 // No longer a need to do a concurrent collection for Metaspace. 1609 MetaspaceGC::set_should_concurrent_collect(false); 1610 1611 gch->post_full_gc_dump(gc_timer); 1612 1613 gc_timer->register_gc_end(); 1614 1615 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1616 1617 // For a mark-sweep-compact, compute_new_size() will be called 1618 // in the heap's do_collection() method. 1619 } 1620 1621 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1622 Log(gc, heap) log; 1623 if (!log.is_trace()) { 1624 return; 1625 } 1626 1627 ContiguousSpace* eden_space = _young_gen->eden(); 1628 ContiguousSpace* from_space = _young_gen->from(); 1629 ContiguousSpace* to_space = _young_gen->to(); 1630 // Eden 1631 if (_eden_chunk_array != NULL) { 1632 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1633 p2i(eden_space->bottom()), p2i(eden_space->top()), 1634 p2i(eden_space->end()), eden_space->capacity()); 1635 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1636 _eden_chunk_index, _eden_chunk_capacity); 1637 for (size_t i = 0; i < _eden_chunk_index; i++) { 1638 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1639 } 1640 } 1641 // Survivor 1642 if (_survivor_chunk_array != NULL) { 1643 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1644 p2i(from_space->bottom()), p2i(from_space->top()), 1645 p2i(from_space->end()), from_space->capacity()); 1646 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1647 _survivor_chunk_index, _survivor_chunk_capacity); 1648 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1649 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1650 } 1651 } 1652 } 1653 1654 void CMSCollector::getFreelistLocks() const { 1655 // Get locks for all free lists in all generations that this 1656 // collector is responsible for 1657 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1658 } 1659 1660 void CMSCollector::releaseFreelistLocks() const { 1661 // Release locks for all free lists in all generations that this 1662 // collector is responsible for 1663 _cmsGen->freelistLock()->unlock(); 1664 } 1665 1666 bool CMSCollector::haveFreelistLocks() const { 1667 // Check locks for all free lists in all generations that this 1668 // collector is responsible for 1669 assert_lock_strong(_cmsGen->freelistLock()); 1670 PRODUCT_ONLY(ShouldNotReachHere()); 1671 return true; 1672 } 1673 1674 // A utility class that is used by the CMS collector to 1675 // temporarily "release" the foreground collector from its 1676 // usual obligation to wait for the background collector to 1677 // complete an ongoing phase before proceeding. 1678 class ReleaseForegroundGC: public StackObj { 1679 private: 1680 CMSCollector* _c; 1681 public: 1682 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1683 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1684 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1685 // allow a potentially blocked foreground collector to proceed 1686 _c->_foregroundGCShouldWait = false; 1687 if (_c->_foregroundGCIsActive) { 1688 CGC_lock->notify(); 1689 } 1690 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1691 "Possible deadlock"); 1692 } 1693 1694 ~ReleaseForegroundGC() { 1695 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1696 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1697 _c->_foregroundGCShouldWait = true; 1698 } 1699 }; 1700 1701 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1702 assert(Thread::current()->is_ConcurrentGC_thread(), 1703 "A CMS asynchronous collection is only allowed on a CMS thread."); 1704 1705 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1706 { 1707 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1708 MutexLockerEx hl(Heap_lock, safepoint_check); 1709 FreelistLocker fll(this); 1710 MutexLockerEx x(CGC_lock, safepoint_check); 1711 if (_foregroundGCIsActive) { 1712 // The foreground collector is. Skip this 1713 // background collection. 1714 assert(!_foregroundGCShouldWait, "Should be clear"); 1715 return; 1716 } else { 1717 assert(_collectorState == Idling, "Should be idling before start."); 1718 _collectorState = InitialMarking; 1719 register_gc_start(cause); 1720 // Reset the expansion cause, now that we are about to begin 1721 // a new cycle. 1722 clear_expansion_cause(); 1723 1724 // Clear the MetaspaceGC flag since a concurrent collection 1725 // is starting but also clear it after the collection. 1726 MetaspaceGC::set_should_concurrent_collect(false); 1727 } 1728 // Decide if we want to enable class unloading as part of the 1729 // ensuing concurrent GC cycle. 1730 update_should_unload_classes(); 1731 _full_gc_requested = false; // acks all outstanding full gc requests 1732 _full_gc_cause = GCCause::_no_gc; 1733 // Signal that we are about to start a collection 1734 gch->increment_total_full_collections(); // ... starting a collection cycle 1735 _collection_count_start = gch->total_full_collections(); 1736 } 1737 1738 size_t prev_used = _cmsGen->used(); 1739 1740 // The change of the collection state is normally done at this level; 1741 // the exceptions are phases that are executed while the world is 1742 // stopped. For those phases the change of state is done while the 1743 // world is stopped. For baton passing purposes this allows the 1744 // background collector to finish the phase and change state atomically. 1745 // The foreground collector cannot wait on a phase that is done 1746 // while the world is stopped because the foreground collector already 1747 // has the world stopped and would deadlock. 1748 while (_collectorState != Idling) { 1749 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1750 p2i(Thread::current()), _collectorState); 1751 // The foreground collector 1752 // holds the Heap_lock throughout its collection. 1753 // holds the CMS token (but not the lock) 1754 // except while it is waiting for the background collector to yield. 1755 // 1756 // The foreground collector should be blocked (not for long) 1757 // if the background collector is about to start a phase 1758 // executed with world stopped. If the background 1759 // collector has already started such a phase, the 1760 // foreground collector is blocked waiting for the 1761 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1762 // are executed in the VM thread. 1763 // 1764 // The locking order is 1765 // PendingListLock (PLL) -- if applicable (FinalMarking) 1766 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1767 // CMS token (claimed in 1768 // stop_world_and_do() --> 1769 // safepoint_synchronize() --> 1770 // CMSThread::synchronize()) 1771 1772 { 1773 // Check if the FG collector wants us to yield. 1774 CMSTokenSync x(true); // is cms thread 1775 if (waitForForegroundGC()) { 1776 // We yielded to a foreground GC, nothing more to be 1777 // done this round. 1778 assert(_foregroundGCShouldWait == false, "We set it to false in " 1779 "waitForForegroundGC()"); 1780 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1781 p2i(Thread::current()), _collectorState); 1782 return; 1783 } else { 1784 // The background collector can run but check to see if the 1785 // foreground collector has done a collection while the 1786 // background collector was waiting to get the CGC_lock 1787 // above. If yes, break so that _foregroundGCShouldWait 1788 // is cleared before returning. 1789 if (_collectorState == Idling) { 1790 break; 1791 } 1792 } 1793 } 1794 1795 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1796 "should be waiting"); 1797 1798 switch (_collectorState) { 1799 case InitialMarking: 1800 { 1801 ReleaseForegroundGC x(this); 1802 stats().record_cms_begin(); 1803 VM_CMS_Initial_Mark initial_mark_op(this); 1804 VMThread::execute(&initial_mark_op); 1805 } 1806 // The collector state may be any legal state at this point 1807 // since the background collector may have yielded to the 1808 // foreground collector. 1809 break; 1810 case Marking: 1811 // initial marking in checkpointRootsInitialWork has been completed 1812 if (markFromRoots()) { // we were successful 1813 assert(_collectorState == Precleaning, "Collector state should " 1814 "have changed"); 1815 } else { 1816 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1817 } 1818 break; 1819 case Precleaning: 1820 // marking from roots in markFromRoots has been completed 1821 preclean(); 1822 assert(_collectorState == AbortablePreclean || 1823 _collectorState == FinalMarking, 1824 "Collector state should have changed"); 1825 break; 1826 case AbortablePreclean: 1827 abortable_preclean(); 1828 assert(_collectorState == FinalMarking, "Collector state should " 1829 "have changed"); 1830 break; 1831 case FinalMarking: 1832 { 1833 ReleaseForegroundGC x(this); 1834 1835 VM_CMS_Final_Remark final_remark_op(this); 1836 VMThread::execute(&final_remark_op); 1837 } 1838 assert(_foregroundGCShouldWait, "block post-condition"); 1839 break; 1840 case Sweeping: 1841 // final marking in checkpointRootsFinal has been completed 1842 sweep(); 1843 assert(_collectorState == Resizing, "Collector state change " 1844 "to Resizing must be done under the free_list_lock"); 1845 1846 case Resizing: { 1847 // Sweeping has been completed... 1848 // At this point the background collection has completed. 1849 // Don't move the call to compute_new_size() down 1850 // into code that might be executed if the background 1851 // collection was preempted. 1852 { 1853 ReleaseForegroundGC x(this); // unblock FG collection 1854 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1855 CMSTokenSync z(true); // not strictly needed. 1856 if (_collectorState == Resizing) { 1857 compute_new_size(); 1858 save_heap_summary(); 1859 _collectorState = Resetting; 1860 } else { 1861 assert(_collectorState == Idling, "The state should only change" 1862 " because the foreground collector has finished the collection"); 1863 } 1864 } 1865 break; 1866 } 1867 case Resetting: 1868 // CMS heap resizing has been completed 1869 reset_concurrent(); 1870 assert(_collectorState == Idling, "Collector state should " 1871 "have changed"); 1872 1873 MetaspaceGC::set_should_concurrent_collect(false); 1874 1875 stats().record_cms_end(); 1876 // Don't move the concurrent_phases_end() and compute_new_size() 1877 // calls to here because a preempted background collection 1878 // has it's state set to "Resetting". 1879 break; 1880 case Idling: 1881 default: 1882 ShouldNotReachHere(); 1883 break; 1884 } 1885 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1886 p2i(Thread::current()), _collectorState); 1887 assert(_foregroundGCShouldWait, "block post-condition"); 1888 } 1889 1890 // Should this be in gc_epilogue? 1891 collector_policy()->counters()->update_counters(); 1892 1893 { 1894 // Clear _foregroundGCShouldWait and, in the event that the 1895 // foreground collector is waiting, notify it, before 1896 // returning. 1897 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1898 _foregroundGCShouldWait = false; 1899 if (_foregroundGCIsActive) { 1900 CGC_lock->notify(); 1901 } 1902 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1903 "Possible deadlock"); 1904 } 1905 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1906 p2i(Thread::current()), _collectorState); 1907 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1908 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1909 } 1910 1911 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1912 _cms_start_registered = true; 1913 _gc_timer_cm->register_gc_start(); 1914 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1915 } 1916 1917 void CMSCollector::register_gc_end() { 1918 if (_cms_start_registered) { 1919 report_heap_summary(GCWhen::AfterGC); 1920 1921 _gc_timer_cm->register_gc_end(); 1922 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1923 _cms_start_registered = false; 1924 } 1925 } 1926 1927 void CMSCollector::save_heap_summary() { 1928 GenCollectedHeap* gch = GenCollectedHeap::heap(); 1929 _last_heap_summary = gch->create_heap_summary(); 1930 _last_metaspace_summary = gch->create_metaspace_summary(); 1931 } 1932 1933 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1934 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1935 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1936 } 1937 1938 bool CMSCollector::waitForForegroundGC() { 1939 bool res = false; 1940 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1941 "CMS thread should have CMS token"); 1942 // Block the foreground collector until the 1943 // background collectors decides whether to 1944 // yield. 1945 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1946 _foregroundGCShouldWait = true; 1947 if (_foregroundGCIsActive) { 1948 // The background collector yields to the 1949 // foreground collector and returns a value 1950 // indicating that it has yielded. The foreground 1951 // collector can proceed. 1952 res = true; 1953 _foregroundGCShouldWait = false; 1954 ConcurrentMarkSweepThread::clear_CMS_flag( 1955 ConcurrentMarkSweepThread::CMS_cms_has_token); 1956 ConcurrentMarkSweepThread::set_CMS_flag( 1957 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1958 // Get a possibly blocked foreground thread going 1959 CGC_lock->notify(); 1960 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1961 p2i(Thread::current()), _collectorState); 1962 while (_foregroundGCIsActive) { 1963 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1964 } 1965 ConcurrentMarkSweepThread::set_CMS_flag( 1966 ConcurrentMarkSweepThread::CMS_cms_has_token); 1967 ConcurrentMarkSweepThread::clear_CMS_flag( 1968 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1969 } 1970 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1971 p2i(Thread::current()), _collectorState); 1972 return res; 1973 } 1974 1975 // Because of the need to lock the free lists and other structures in 1976 // the collector, common to all the generations that the collector is 1977 // collecting, we need the gc_prologues of individual CMS generations 1978 // delegate to their collector. It may have been simpler had the 1979 // current infrastructure allowed one to call a prologue on a 1980 // collector. In the absence of that we have the generation's 1981 // prologue delegate to the collector, which delegates back 1982 // some "local" work to a worker method in the individual generations 1983 // that it's responsible for collecting, while itself doing any 1984 // work common to all generations it's responsible for. A similar 1985 // comment applies to the gc_epilogue()'s. 1986 // The role of the variable _between_prologue_and_epilogue is to 1987 // enforce the invocation protocol. 1988 void CMSCollector::gc_prologue(bool full) { 1989 // Call gc_prologue_work() for the CMSGen 1990 // we are responsible for. 1991 1992 // The following locking discipline assumes that we are only called 1993 // when the world is stopped. 1994 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 1995 1996 // The CMSCollector prologue must call the gc_prologues for the 1997 // "generations" that it's responsible 1998 // for. 1999 2000 assert( Thread::current()->is_VM_thread() 2001 || ( CMSScavengeBeforeRemark 2002 && Thread::current()->is_ConcurrentGC_thread()), 2003 "Incorrect thread type for prologue execution"); 2004 2005 if (_between_prologue_and_epilogue) { 2006 // We have already been invoked; this is a gc_prologue delegation 2007 // from yet another CMS generation that we are responsible for, just 2008 // ignore it since all relevant work has already been done. 2009 return; 2010 } 2011 2012 // set a bit saying prologue has been called; cleared in epilogue 2013 _between_prologue_and_epilogue = true; 2014 // Claim locks for common data structures, then call gc_prologue_work() 2015 // for each CMSGen. 2016 2017 getFreelistLocks(); // gets free list locks on constituent spaces 2018 bitMapLock()->lock_without_safepoint_check(); 2019 2020 // Should call gc_prologue_work() for all cms gens we are responsible for 2021 bool duringMarking = _collectorState >= Marking 2022 && _collectorState < Sweeping; 2023 2024 // The young collections clear the modified oops state, which tells if 2025 // there are any modified oops in the class. The remark phase also needs 2026 // that information. Tell the young collection to save the union of all 2027 // modified klasses. 2028 if (duringMarking) { 2029 _ct->klass_rem_set()->set_accumulate_modified_oops(true); 2030 } 2031 2032 bool registerClosure = duringMarking; 2033 2034 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2035 2036 if (!full) { 2037 stats().record_gc0_begin(); 2038 } 2039 } 2040 2041 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2042 2043 _capacity_at_prologue = capacity(); 2044 _used_at_prologue = used(); 2045 2046 // We enable promotion tracking so that card-scanning can recognize 2047 // which objects have been promoted during this GC and skip them. 2048 for (uint i = 0; i < ParallelGCThreads; i++) { 2049 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2050 } 2051 2052 // Delegate to CMScollector which knows how to coordinate between 2053 // this and any other CMS generations that it is responsible for 2054 // collecting. 2055 collector()->gc_prologue(full); 2056 } 2057 2058 // This is a "private" interface for use by this generation's CMSCollector. 2059 // Not to be called directly by any other entity (for instance, 2060 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2061 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2062 bool registerClosure, ModUnionClosure* modUnionClosure) { 2063 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2064 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2065 "Should be NULL"); 2066 if (registerClosure) { 2067 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2068 } 2069 cmsSpace()->gc_prologue(); 2070 // Clear stat counters 2071 NOT_PRODUCT( 2072 assert(_numObjectsPromoted == 0, "check"); 2073 assert(_numWordsPromoted == 0, "check"); 2074 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2075 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2076 _numObjectsAllocated = 0; 2077 _numWordsAllocated = 0; 2078 ) 2079 } 2080 2081 void CMSCollector::gc_epilogue(bool full) { 2082 // The following locking discipline assumes that we are only called 2083 // when the world is stopped. 2084 assert(SafepointSynchronize::is_at_safepoint(), 2085 "world is stopped assumption"); 2086 2087 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2088 // if linear allocation blocks need to be appropriately marked to allow the 2089 // the blocks to be parsable. We also check here whether we need to nudge the 2090 // CMS collector thread to start a new cycle (if it's not already active). 2091 assert( Thread::current()->is_VM_thread() 2092 || ( CMSScavengeBeforeRemark 2093 && Thread::current()->is_ConcurrentGC_thread()), 2094 "Incorrect thread type for epilogue execution"); 2095 2096 if (!_between_prologue_and_epilogue) { 2097 // We have already been invoked; this is a gc_epilogue delegation 2098 // from yet another CMS generation that we are responsible for, just 2099 // ignore it since all relevant work has already been done. 2100 return; 2101 } 2102 assert(haveFreelistLocks(), "must have freelist locks"); 2103 assert_lock_strong(bitMapLock()); 2104 2105 _ct->klass_rem_set()->set_accumulate_modified_oops(false); 2106 2107 _cmsGen->gc_epilogue_work(full); 2108 2109 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2110 // in case sampling was not already enabled, enable it 2111 _start_sampling = true; 2112 } 2113 // reset _eden_chunk_array so sampling starts afresh 2114 _eden_chunk_index = 0; 2115 2116 size_t cms_used = _cmsGen->cmsSpace()->used(); 2117 2118 // update performance counters - this uses a special version of 2119 // update_counters() that allows the utilization to be passed as a 2120 // parameter, avoiding multiple calls to used(). 2121 // 2122 _cmsGen->update_counters(cms_used); 2123 2124 bitMapLock()->unlock(); 2125 releaseFreelistLocks(); 2126 2127 if (!CleanChunkPoolAsync) { 2128 Chunk::clean_chunk_pool(); 2129 } 2130 2131 set_did_compact(false); 2132 _between_prologue_and_epilogue = false; // ready for next cycle 2133 } 2134 2135 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2136 collector()->gc_epilogue(full); 2137 2138 // When using ParNew, promotion tracking should have already been 2139 // disabled. However, the prologue (which enables promotion 2140 // tracking) and epilogue are called irrespective of the type of 2141 // GC. So they will also be called before and after Full GCs, during 2142 // which promotion tracking will not be explicitly disabled. So, 2143 // it's safer to also disable it here too (to be symmetric with 2144 // enabling it in the prologue). 2145 for (uint i = 0; i < ParallelGCThreads; i++) { 2146 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2147 } 2148 } 2149 2150 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2151 assert(!incremental_collection_failed(), "Should have been cleared"); 2152 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2153 cmsSpace()->gc_epilogue(); 2154 // Print stat counters 2155 NOT_PRODUCT( 2156 assert(_numObjectsAllocated == 0, "check"); 2157 assert(_numWordsAllocated == 0, "check"); 2158 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2159 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2160 _numObjectsPromoted = 0; 2161 _numWordsPromoted = 0; 2162 ) 2163 2164 // Call down the chain in contiguous_available needs the freelistLock 2165 // so print this out before releasing the freeListLock. 2166 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2167 } 2168 2169 #ifndef PRODUCT 2170 bool CMSCollector::have_cms_token() { 2171 Thread* thr = Thread::current(); 2172 if (thr->is_VM_thread()) { 2173 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2174 } else if (thr->is_ConcurrentGC_thread()) { 2175 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2176 } else if (thr->is_GC_task_thread()) { 2177 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2178 ParGCRareEvent_lock->owned_by_self(); 2179 } 2180 return false; 2181 } 2182 2183 // Check reachability of the given heap address in CMS generation, 2184 // treating all other generations as roots. 2185 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2186 // We could "guarantee" below, rather than assert, but I'll 2187 // leave these as "asserts" so that an adventurous debugger 2188 // could try this in the product build provided some subset of 2189 // the conditions were met, provided they were interested in the 2190 // results and knew that the computation below wouldn't interfere 2191 // with other concurrent computations mutating the structures 2192 // being read or written. 2193 assert(SafepointSynchronize::is_at_safepoint(), 2194 "Else mutations in object graph will make answer suspect"); 2195 assert(have_cms_token(), "Should hold cms token"); 2196 assert(haveFreelistLocks(), "must hold free list locks"); 2197 assert_lock_strong(bitMapLock()); 2198 2199 // Clear the marking bit map array before starting, but, just 2200 // for kicks, first report if the given address is already marked 2201 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2202 _markBitMap.isMarked(addr) ? "" : " not"); 2203 2204 if (verify_after_remark()) { 2205 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2206 bool result = verification_mark_bm()->isMarked(addr); 2207 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2208 result ? "IS" : "is NOT"); 2209 return result; 2210 } else { 2211 tty->print_cr("Could not compute result"); 2212 return false; 2213 } 2214 } 2215 #endif 2216 2217 void 2218 CMSCollector::print_on_error(outputStream* st) { 2219 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2220 if (collector != NULL) { 2221 CMSBitMap* bitmap = &collector->_markBitMap; 2222 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2223 bitmap->print_on_error(st, " Bits: "); 2224 2225 st->cr(); 2226 2227 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2228 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2229 mut_bitmap->print_on_error(st, " Bits: "); 2230 } 2231 } 2232 2233 //////////////////////////////////////////////////////// 2234 // CMS Verification Support 2235 //////////////////////////////////////////////////////// 2236 // Following the remark phase, the following invariant 2237 // should hold -- each object in the CMS heap which is 2238 // marked in markBitMap() should be marked in the verification_mark_bm(). 2239 2240 class VerifyMarkedClosure: public BitMapClosure { 2241 CMSBitMap* _marks; 2242 bool _failed; 2243 2244 public: 2245 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2246 2247 bool do_bit(size_t offset) { 2248 HeapWord* addr = _marks->offsetToHeapWord(offset); 2249 if (!_marks->isMarked(addr)) { 2250 Log(gc, verify) log; 2251 ResourceMark rm; 2252 oop(addr)->print_on(log.error_stream()); 2253 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2254 _failed = true; 2255 } 2256 return true; 2257 } 2258 2259 bool failed() { return _failed; } 2260 }; 2261 2262 bool CMSCollector::verify_after_remark() { 2263 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2264 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2265 static bool init = false; 2266 2267 assert(SafepointSynchronize::is_at_safepoint(), 2268 "Else mutations in object graph will make answer suspect"); 2269 assert(have_cms_token(), 2270 "Else there may be mutual interference in use of " 2271 " verification data structures"); 2272 assert(_collectorState > Marking && _collectorState <= Sweeping, 2273 "Else marking info checked here may be obsolete"); 2274 assert(haveFreelistLocks(), "must hold free list locks"); 2275 assert_lock_strong(bitMapLock()); 2276 2277 2278 // Allocate marking bit map if not already allocated 2279 if (!init) { // first time 2280 if (!verification_mark_bm()->allocate(_span)) { 2281 return false; 2282 } 2283 init = true; 2284 } 2285 2286 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2287 2288 // Turn off refs discovery -- so we will be tracing through refs. 2289 // This is as intended, because by this time 2290 // GC must already have cleared any refs that need to be cleared, 2291 // and traced those that need to be marked; moreover, 2292 // the marking done here is not going to interfere in any 2293 // way with the marking information used by GC. 2294 NoRefDiscovery no_discovery(ref_processor()); 2295 2296 #if defined(COMPILER2) || INCLUDE_JVMCI 2297 DerivedPointerTableDeactivate dpt_deact; 2298 #endif 2299 2300 // Clear any marks from a previous round 2301 verification_mark_bm()->clear_all(); 2302 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2303 verify_work_stacks_empty(); 2304 2305 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2306 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2307 // Update the saved marks which may affect the root scans. 2308 gch->save_marks(); 2309 2310 if (CMSRemarkVerifyVariant == 1) { 2311 // In this first variant of verification, we complete 2312 // all marking, then check if the new marks-vector is 2313 // a subset of the CMS marks-vector. 2314 verify_after_remark_work_1(); 2315 } else { 2316 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2317 // In this second variant of verification, we flag an error 2318 // (i.e. an object reachable in the new marks-vector not reachable 2319 // in the CMS marks-vector) immediately, also indicating the 2320 // identify of an object (A) that references the unmarked object (B) -- 2321 // presumably, a mutation to A failed to be picked up by preclean/remark? 2322 verify_after_remark_work_2(); 2323 } 2324 2325 return true; 2326 } 2327 2328 void CMSCollector::verify_after_remark_work_1() { 2329 ResourceMark rm; 2330 HandleMark hm; 2331 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2332 2333 // Get a clear set of claim bits for the roots processing to work with. 2334 ClassLoaderDataGraph::clear_claimed_marks(); 2335 2336 // Mark from roots one level into CMS 2337 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2338 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2339 2340 { 2341 StrongRootsScope srs(1); 2342 2343 gch->old_process_roots(&srs, 2344 true, // young gen as roots 2345 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2346 should_unload_classes(), 2347 ¬Older, 2348 NULL); 2349 } 2350 2351 // Now mark from the roots 2352 MarkFromRootsClosure markFromRootsClosure(this, _span, 2353 verification_mark_bm(), verification_mark_stack(), 2354 false /* don't yield */, true /* verifying */); 2355 assert(_restart_addr == NULL, "Expected pre-condition"); 2356 verification_mark_bm()->iterate(&markFromRootsClosure); 2357 while (_restart_addr != NULL) { 2358 // Deal with stack overflow: by restarting at the indicated 2359 // address. 2360 HeapWord* ra = _restart_addr; 2361 markFromRootsClosure.reset(ra); 2362 _restart_addr = NULL; 2363 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2364 } 2365 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2366 verify_work_stacks_empty(); 2367 2368 // Marking completed -- now verify that each bit marked in 2369 // verification_mark_bm() is also marked in markBitMap(); flag all 2370 // errors by printing corresponding objects. 2371 VerifyMarkedClosure vcl(markBitMap()); 2372 verification_mark_bm()->iterate(&vcl); 2373 if (vcl.failed()) { 2374 Log(gc, verify) log; 2375 log.error("Failed marking verification after remark"); 2376 ResourceMark rm; 2377 gch->print_on(log.error_stream()); 2378 fatal("CMS: failed marking verification after remark"); 2379 } 2380 } 2381 2382 class VerifyKlassOopsKlassClosure : public KlassClosure { 2383 class VerifyKlassOopsClosure : public OopClosure { 2384 CMSBitMap* _bitmap; 2385 public: 2386 VerifyKlassOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2387 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2388 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2389 } _oop_closure; 2390 public: 2391 VerifyKlassOopsKlassClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2392 void do_klass(Klass* k) { 2393 k->oops_do(&_oop_closure); 2394 } 2395 }; 2396 2397 void CMSCollector::verify_after_remark_work_2() { 2398 ResourceMark rm; 2399 HandleMark hm; 2400 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2401 2402 // Get a clear set of claim bits for the roots processing to work with. 2403 ClassLoaderDataGraph::clear_claimed_marks(); 2404 2405 // Mark from roots one level into CMS 2406 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2407 markBitMap()); 2408 CLDToOopClosure cld_closure(¬Older, true); 2409 2410 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2411 2412 { 2413 StrongRootsScope srs(1); 2414 2415 gch->old_process_roots(&srs, 2416 true, // young gen as roots 2417 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2418 should_unload_classes(), 2419 ¬Older, 2420 &cld_closure); 2421 } 2422 2423 // Now mark from the roots 2424 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2425 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2426 assert(_restart_addr == NULL, "Expected pre-condition"); 2427 verification_mark_bm()->iterate(&markFromRootsClosure); 2428 while (_restart_addr != NULL) { 2429 // Deal with stack overflow: by restarting at the indicated 2430 // address. 2431 HeapWord* ra = _restart_addr; 2432 markFromRootsClosure.reset(ra); 2433 _restart_addr = NULL; 2434 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2435 } 2436 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2437 verify_work_stacks_empty(); 2438 2439 VerifyKlassOopsKlassClosure verify_klass_oops(verification_mark_bm()); 2440 ClassLoaderDataGraph::classes_do(&verify_klass_oops); 2441 2442 // Marking completed -- now verify that each bit marked in 2443 // verification_mark_bm() is also marked in markBitMap(); flag all 2444 // errors by printing corresponding objects. 2445 VerifyMarkedClosure vcl(markBitMap()); 2446 verification_mark_bm()->iterate(&vcl); 2447 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2448 } 2449 2450 void ConcurrentMarkSweepGeneration::save_marks() { 2451 // delegate to CMS space 2452 cmsSpace()->save_marks(); 2453 } 2454 2455 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2456 return cmsSpace()->no_allocs_since_save_marks(); 2457 } 2458 2459 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ 2460 \ 2461 void ConcurrentMarkSweepGeneration:: \ 2462 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ 2463 cl->set_generation(this); \ 2464 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ 2465 cl->reset_generation(); \ 2466 save_marks(); \ 2467 } 2468 2469 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) 2470 2471 void 2472 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2473 if (freelistLock()->owned_by_self()) { 2474 Generation::oop_iterate(cl); 2475 } else { 2476 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2477 Generation::oop_iterate(cl); 2478 } 2479 } 2480 2481 void 2482 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2483 if (freelistLock()->owned_by_self()) { 2484 Generation::object_iterate(cl); 2485 } else { 2486 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2487 Generation::object_iterate(cl); 2488 } 2489 } 2490 2491 void 2492 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2493 if (freelistLock()->owned_by_self()) { 2494 Generation::safe_object_iterate(cl); 2495 } else { 2496 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2497 Generation::safe_object_iterate(cl); 2498 } 2499 } 2500 2501 void 2502 ConcurrentMarkSweepGeneration::post_compact() { 2503 } 2504 2505 void 2506 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2507 // Fix the linear allocation blocks to look like free blocks. 2508 2509 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2510 // are not called when the heap is verified during universe initialization and 2511 // at vm shutdown. 2512 if (freelistLock()->owned_by_self()) { 2513 cmsSpace()->prepare_for_verify(); 2514 } else { 2515 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2516 cmsSpace()->prepare_for_verify(); 2517 } 2518 } 2519 2520 void 2521 ConcurrentMarkSweepGeneration::verify() { 2522 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2523 // are not called when the heap is verified during universe initialization and 2524 // at vm shutdown. 2525 if (freelistLock()->owned_by_self()) { 2526 cmsSpace()->verify(); 2527 } else { 2528 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2529 cmsSpace()->verify(); 2530 } 2531 } 2532 2533 void CMSCollector::verify() { 2534 _cmsGen->verify(); 2535 } 2536 2537 #ifndef PRODUCT 2538 bool CMSCollector::overflow_list_is_empty() const { 2539 assert(_num_par_pushes >= 0, "Inconsistency"); 2540 if (_overflow_list == NULL) { 2541 assert(_num_par_pushes == 0, "Inconsistency"); 2542 } 2543 return _overflow_list == NULL; 2544 } 2545 2546 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2547 // merely consolidate assertion checks that appear to occur together frequently. 2548 void CMSCollector::verify_work_stacks_empty() const { 2549 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2550 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2551 } 2552 2553 void CMSCollector::verify_overflow_empty() const { 2554 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2555 assert(no_preserved_marks(), "No preserved marks"); 2556 } 2557 #endif // PRODUCT 2558 2559 // Decide if we want to enable class unloading as part of the 2560 // ensuing concurrent GC cycle. We will collect and 2561 // unload classes if it's the case that: 2562 // (1) an explicit gc request has been made and the flag 2563 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR 2564 // (2) (a) class unloading is enabled at the command line, and 2565 // (b) old gen is getting really full 2566 // NOTE: Provided there is no change in the state of the heap between 2567 // calls to this method, it should have idempotent results. Moreover, 2568 // its results should be monotonically increasing (i.e. going from 0 to 1, 2569 // but not 1 to 0) between successive calls between which the heap was 2570 // not collected. For the implementation below, it must thus rely on 2571 // the property that concurrent_cycles_since_last_unload() 2572 // will not decrease unless a collection cycle happened and that 2573 // _cmsGen->is_too_full() are 2574 // themselves also monotonic in that sense. See check_monotonicity() 2575 // below. 2576 void CMSCollector::update_should_unload_classes() { 2577 _should_unload_classes = false; 2578 // Condition 1 above 2579 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { 2580 _should_unload_classes = true; 2581 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above 2582 // Disjuncts 2.b.(i,ii,iii) above 2583 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2584 CMSClassUnloadingMaxInterval) 2585 || _cmsGen->is_too_full(); 2586 } 2587 } 2588 2589 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2590 bool res = should_concurrent_collect(); 2591 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2592 return res; 2593 } 2594 2595 void CMSCollector::setup_cms_unloading_and_verification_state() { 2596 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2597 || VerifyBeforeExit; 2598 const int rso = GenCollectedHeap::SO_AllCodeCache; 2599 2600 // We set the proper root for this CMS cycle here. 2601 if (should_unload_classes()) { // Should unload classes this cycle 2602 remove_root_scanning_option(rso); // Shrink the root set appropriately 2603 set_verifying(should_verify); // Set verification state for this cycle 2604 return; // Nothing else needs to be done at this time 2605 } 2606 2607 // Not unloading classes this cycle 2608 assert(!should_unload_classes(), "Inconsistency!"); 2609 2610 // If we are not unloading classes then add SO_AllCodeCache to root 2611 // scanning options. 2612 add_root_scanning_option(rso); 2613 2614 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2615 set_verifying(true); 2616 } else if (verifying() && !should_verify) { 2617 // We were verifying, but some verification flags got disabled. 2618 set_verifying(false); 2619 // Exclude symbols, strings and code cache elements from root scanning to 2620 // reduce IM and RM pauses. 2621 remove_root_scanning_option(rso); 2622 } 2623 } 2624 2625 2626 #ifndef PRODUCT 2627 HeapWord* CMSCollector::block_start(const void* p) const { 2628 const HeapWord* addr = (HeapWord*)p; 2629 if (_span.contains(p)) { 2630 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2631 return _cmsGen->cmsSpace()->block_start(p); 2632 } 2633 } 2634 return NULL; 2635 } 2636 #endif 2637 2638 HeapWord* 2639 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2640 bool tlab, 2641 bool parallel) { 2642 CMSSynchronousYieldRequest yr; 2643 assert(!tlab, "Can't deal with TLAB allocation"); 2644 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2645 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2646 if (GCExpandToAllocateDelayMillis > 0) { 2647 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2648 } 2649 return have_lock_and_allocate(word_size, tlab); 2650 } 2651 2652 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2653 size_t bytes, 2654 size_t expand_bytes, 2655 CMSExpansionCause::Cause cause) 2656 { 2657 2658 bool success = expand(bytes, expand_bytes); 2659 2660 // remember why we expanded; this information is used 2661 // by shouldConcurrentCollect() when making decisions on whether to start 2662 // a new CMS cycle. 2663 if (success) { 2664 set_expansion_cause(cause); 2665 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2666 } 2667 } 2668 2669 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2670 HeapWord* res = NULL; 2671 MutexLocker x(ParGCRareEvent_lock); 2672 while (true) { 2673 // Expansion by some other thread might make alloc OK now: 2674 res = ps->lab.alloc(word_sz); 2675 if (res != NULL) return res; 2676 // If there's not enough expansion space available, give up. 2677 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2678 return NULL; 2679 } 2680 // Otherwise, we try expansion. 2681 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2682 // Now go around the loop and try alloc again; 2683 // A competing par_promote might beat us to the expansion space, 2684 // so we may go around the loop again if promotion fails again. 2685 if (GCExpandToAllocateDelayMillis > 0) { 2686 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2687 } 2688 } 2689 } 2690 2691 2692 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2693 PromotionInfo* promo) { 2694 MutexLocker x(ParGCRareEvent_lock); 2695 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2696 while (true) { 2697 // Expansion by some other thread might make alloc OK now: 2698 if (promo->ensure_spooling_space()) { 2699 assert(promo->has_spooling_space(), 2700 "Post-condition of successful ensure_spooling_space()"); 2701 return true; 2702 } 2703 // If there's not enough expansion space available, give up. 2704 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2705 return false; 2706 } 2707 // Otherwise, we try expansion. 2708 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2709 // Now go around the loop and try alloc again; 2710 // A competing allocation might beat us to the expansion space, 2711 // so we may go around the loop again if allocation fails again. 2712 if (GCExpandToAllocateDelayMillis > 0) { 2713 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2714 } 2715 } 2716 } 2717 2718 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2719 // Only shrink if a compaction was done so that all the free space 2720 // in the generation is in a contiguous block at the end. 2721 if (did_compact()) { 2722 CardGeneration::shrink(bytes); 2723 } 2724 } 2725 2726 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2727 assert_locked_or_safepoint(Heap_lock); 2728 } 2729 2730 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2731 assert_locked_or_safepoint(Heap_lock); 2732 assert_lock_strong(freelistLock()); 2733 log_trace(gc)("Shrinking of CMS not yet implemented"); 2734 return; 2735 } 2736 2737 2738 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2739 // phases. 2740 class CMSPhaseAccounting: public StackObj { 2741 public: 2742 CMSPhaseAccounting(CMSCollector *collector, 2743 const char *title); 2744 ~CMSPhaseAccounting(); 2745 2746 private: 2747 CMSCollector *_collector; 2748 const char *_title; 2749 GCTraceConcTime(Info, gc) _trace_time; 2750 2751 public: 2752 // Not MT-safe; so do not pass around these StackObj's 2753 // where they may be accessed by other threads. 2754 double wallclock_millis() { 2755 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2756 } 2757 }; 2758 2759 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2760 const char *title) : 2761 _collector(collector), _title(title), _trace_time(title) { 2762 2763 _collector->resetYields(); 2764 _collector->resetTimer(); 2765 _collector->startTimer(); 2766 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2767 } 2768 2769 CMSPhaseAccounting::~CMSPhaseAccounting() { 2770 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2771 _collector->stopTimer(); 2772 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2773 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2774 } 2775 2776 // CMS work 2777 2778 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2779 class CMSParMarkTask : public AbstractGangTask { 2780 protected: 2781 CMSCollector* _collector; 2782 uint _n_workers; 2783 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2784 AbstractGangTask(name), 2785 _collector(collector), 2786 _n_workers(n_workers) {} 2787 // Work method in support of parallel rescan ... of young gen spaces 2788 void do_young_space_rescan(OopsInGenClosure* cl, 2789 ContiguousSpace* space, 2790 HeapWord** chunk_array, size_t chunk_top); 2791 void work_on_young_gen_roots(OopsInGenClosure* cl); 2792 }; 2793 2794 // Parallel initial mark task 2795 class CMSParInitialMarkTask: public CMSParMarkTask { 2796 StrongRootsScope* _strong_roots_scope; 2797 public: 2798 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2799 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2800 _strong_roots_scope(strong_roots_scope) {} 2801 void work(uint worker_id); 2802 }; 2803 2804 // Checkpoint the roots into this generation from outside 2805 // this generation. [Note this initial checkpoint need only 2806 // be approximate -- we'll do a catch up phase subsequently.] 2807 void CMSCollector::checkpointRootsInitial() { 2808 assert(_collectorState == InitialMarking, "Wrong collector state"); 2809 check_correct_thread_executing(); 2810 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 2811 2812 save_heap_summary(); 2813 report_heap_summary(GCWhen::BeforeGC); 2814 2815 ReferenceProcessor* rp = ref_processor(); 2816 assert(_restart_addr == NULL, "Control point invariant"); 2817 { 2818 // acquire locks for subsequent manipulations 2819 MutexLockerEx x(bitMapLock(), 2820 Mutex::_no_safepoint_check_flag); 2821 checkpointRootsInitialWork(); 2822 // enable ("weak") refs discovery 2823 rp->enable_discovery(); 2824 _collectorState = Marking; 2825 } 2826 } 2827 2828 void CMSCollector::checkpointRootsInitialWork() { 2829 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2830 assert(_collectorState == InitialMarking, "just checking"); 2831 2832 // Already have locks. 2833 assert_lock_strong(bitMapLock()); 2834 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2835 2836 // Setup the verification and class unloading state for this 2837 // CMS collection cycle. 2838 setup_cms_unloading_and_verification_state(); 2839 2840 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2841 2842 // Reset all the PLAB chunk arrays if necessary. 2843 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2844 reset_survivor_plab_arrays(); 2845 } 2846 2847 ResourceMark rm; 2848 HandleMark hm; 2849 2850 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2851 GenCollectedHeap* gch = GenCollectedHeap::heap(); 2852 2853 verify_work_stacks_empty(); 2854 verify_overflow_empty(); 2855 2856 gch->ensure_parsability(false); // fill TLABs, but no need to retire them 2857 // Update the saved marks which may affect the root scans. 2858 gch->save_marks(); 2859 2860 // weak reference processing has not started yet. 2861 ref_processor()->set_enqueuing_is_done(false); 2862 2863 // Need to remember all newly created CLDs, 2864 // so that we can guarantee that the remark finds them. 2865 ClassLoaderDataGraph::remember_new_clds(true); 2866 2867 // Whenever a CLD is found, it will be claimed before proceeding to mark 2868 // the klasses. The claimed marks need to be cleared before marking starts. 2869 ClassLoaderDataGraph::clear_claimed_marks(); 2870 2871 print_eden_and_survivor_chunk_arrays(); 2872 2873 { 2874 #if defined(COMPILER2) || INCLUDE_JVMCI 2875 DerivedPointerTableDeactivate dpt_deact; 2876 #endif 2877 if (CMSParallelInitialMarkEnabled) { 2878 // The parallel version. 2879 WorkGang* workers = gch->workers(); 2880 assert(workers != NULL, "Need parallel worker threads."); 2881 uint n_workers = workers->active_workers(); 2882 2883 StrongRootsScope srs(n_workers); 2884 2885 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2886 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2887 // If the total workers is greater than 1, then multiple workers 2888 // may be used at some time and the initialization has been set 2889 // such that the single threaded path cannot be used. 2890 if (workers->total_workers() > 1) { 2891 workers->run_task(&tsk); 2892 } else { 2893 tsk.work(0); 2894 } 2895 } else { 2896 // The serial version. 2897 CLDToOopClosure cld_closure(¬Older, true); 2898 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2899 2900 StrongRootsScope srs(1); 2901 2902 gch->old_process_roots(&srs, 2903 true, // young gen as roots 2904 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2905 should_unload_classes(), 2906 ¬Older, 2907 &cld_closure); 2908 } 2909 } 2910 2911 // Clear mod-union table; it will be dirtied in the prologue of 2912 // CMS generation per each young generation collection. 2913 2914 assert(_modUnionTable.isAllClear(), 2915 "Was cleared in most recent final checkpoint phase" 2916 " or no bits are set in the gc_prologue before the start of the next " 2917 "subsequent marking phase."); 2918 2919 assert(_ct->klass_rem_set()->mod_union_is_clear(), "Must be"); 2920 2921 // Save the end of the used_region of the constituent generations 2922 // to be used to limit the extent of sweep in each generation. 2923 save_sweep_limits(); 2924 verify_overflow_empty(); 2925 } 2926 2927 bool CMSCollector::markFromRoots() { 2928 // we might be tempted to assert that: 2929 // assert(!SafepointSynchronize::is_at_safepoint(), 2930 // "inconsistent argument?"); 2931 // However that wouldn't be right, because it's possible that 2932 // a safepoint is indeed in progress as a young generation 2933 // stop-the-world GC happens even as we mark in this generation. 2934 assert(_collectorState == Marking, "inconsistent state?"); 2935 check_correct_thread_executing(); 2936 verify_overflow_empty(); 2937 2938 // Weak ref discovery note: We may be discovering weak 2939 // refs in this generation concurrent (but interleaved) with 2940 // weak ref discovery by the young generation collector. 2941 2942 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2943 GCTraceCPUTime tcpu; 2944 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2945 bool res = markFromRootsWork(); 2946 if (res) { 2947 _collectorState = Precleaning; 2948 } else { // We failed and a foreground collection wants to take over 2949 assert(_foregroundGCIsActive, "internal state inconsistency"); 2950 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2951 log_debug(gc)("bailing out to foreground collection"); 2952 } 2953 verify_overflow_empty(); 2954 return res; 2955 } 2956 2957 bool CMSCollector::markFromRootsWork() { 2958 // iterate over marked bits in bit map, doing a full scan and mark 2959 // from these roots using the following algorithm: 2960 // . if oop is to the right of the current scan pointer, 2961 // mark corresponding bit (we'll process it later) 2962 // . else (oop is to left of current scan pointer) 2963 // push oop on marking stack 2964 // . drain the marking stack 2965 2966 // Note that when we do a marking step we need to hold the 2967 // bit map lock -- recall that direct allocation (by mutators) 2968 // and promotion (by the young generation collector) is also 2969 // marking the bit map. [the so-called allocate live policy.] 2970 // Because the implementation of bit map marking is not 2971 // robust wrt simultaneous marking of bits in the same word, 2972 // we need to make sure that there is no such interference 2973 // between concurrent such updates. 2974 2975 // already have locks 2976 assert_lock_strong(bitMapLock()); 2977 2978 verify_work_stacks_empty(); 2979 verify_overflow_empty(); 2980 bool result = false; 2981 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2982 result = do_marking_mt(); 2983 } else { 2984 result = do_marking_st(); 2985 } 2986 return result; 2987 } 2988 2989 // Forward decl 2990 class CMSConcMarkingTask; 2991 2992 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2993 CMSCollector* _collector; 2994 CMSConcMarkingTask* _task; 2995 public: 2996 virtual void yield(); 2997 2998 // "n_threads" is the number of threads to be terminated. 2999 // "queue_set" is a set of work queues of other threads. 3000 // "collector" is the CMS collector associated with this task terminator. 3001 // "yield" indicates whether we need the gang as a whole to yield. 3002 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 3003 ParallelTaskTerminator(n_threads, queue_set), 3004 _collector(collector) { } 3005 3006 void set_task(CMSConcMarkingTask* task) { 3007 _task = task; 3008 } 3009 }; 3010 3011 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3012 CMSConcMarkingTask* _task; 3013 public: 3014 bool should_exit_termination(); 3015 void set_task(CMSConcMarkingTask* task) { 3016 _task = task; 3017 } 3018 }; 3019 3020 // MT Concurrent Marking Task 3021 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3022 CMSCollector* _collector; 3023 uint _n_workers; // requested/desired # workers 3024 bool _result; 3025 CompactibleFreeListSpace* _cms_space; 3026 char _pad_front[64]; // padding to ... 3027 HeapWord* _global_finger; // ... avoid sharing cache line 3028 char _pad_back[64]; 3029 HeapWord* _restart_addr; 3030 3031 // Exposed here for yielding support 3032 Mutex* const _bit_map_lock; 3033 3034 // The per thread work queues, available here for stealing 3035 OopTaskQueueSet* _task_queues; 3036 3037 // Termination (and yielding) support 3038 CMSConcMarkingTerminator _term; 3039 CMSConcMarkingTerminatorTerminator _term_term; 3040 3041 public: 3042 CMSConcMarkingTask(CMSCollector* collector, 3043 CompactibleFreeListSpace* cms_space, 3044 YieldingFlexibleWorkGang* workers, 3045 OopTaskQueueSet* task_queues): 3046 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3047 _collector(collector), 3048 _cms_space(cms_space), 3049 _n_workers(0), _result(true), 3050 _task_queues(task_queues), 3051 _term(_n_workers, task_queues, _collector), 3052 _bit_map_lock(collector->bitMapLock()) 3053 { 3054 _requested_size = _n_workers; 3055 _term.set_task(this); 3056 _term_term.set_task(this); 3057 _restart_addr = _global_finger = _cms_space->bottom(); 3058 } 3059 3060 3061 OopTaskQueueSet* task_queues() { return _task_queues; } 3062 3063 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3064 3065 HeapWord** global_finger_addr() { return &_global_finger; } 3066 3067 CMSConcMarkingTerminator* terminator() { return &_term; } 3068 3069 virtual void set_for_termination(uint active_workers) { 3070 terminator()->reset_for_reuse(active_workers); 3071 } 3072 3073 void work(uint worker_id); 3074 bool should_yield() { 3075 return ConcurrentMarkSweepThread::should_yield() 3076 && !_collector->foregroundGCIsActive(); 3077 } 3078 3079 virtual void coordinator_yield(); // stuff done by coordinator 3080 bool result() { return _result; } 3081 3082 void reset(HeapWord* ra) { 3083 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3084 _restart_addr = _global_finger = ra; 3085 _term.reset_for_reuse(); 3086 } 3087 3088 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3089 OopTaskQueue* work_q); 3090 3091 private: 3092 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3093 void do_work_steal(int i); 3094 void bump_global_finger(HeapWord* f); 3095 }; 3096 3097 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3098 assert(_task != NULL, "Error"); 3099 return _task->yielding(); 3100 // Note that we do not need the disjunct || _task->should_yield() above 3101 // because we want terminating threads to yield only if the task 3102 // is already in the midst of yielding, which happens only after at least one 3103 // thread has yielded. 3104 } 3105 3106 void CMSConcMarkingTerminator::yield() { 3107 if (_task->should_yield()) { 3108 _task->yield(); 3109 } else { 3110 ParallelTaskTerminator::yield(); 3111 } 3112 } 3113 3114 //////////////////////////////////////////////////////////////// 3115 // Concurrent Marking Algorithm Sketch 3116 //////////////////////////////////////////////////////////////// 3117 // Until all tasks exhausted (both spaces): 3118 // -- claim next available chunk 3119 // -- bump global finger via CAS 3120 // -- find first object that starts in this chunk 3121 // and start scanning bitmap from that position 3122 // -- scan marked objects for oops 3123 // -- CAS-mark target, and if successful: 3124 // . if target oop is above global finger (volatile read) 3125 // nothing to do 3126 // . if target oop is in chunk and above local finger 3127 // then nothing to do 3128 // . else push on work-queue 3129 // -- Deal with possible overflow issues: 3130 // . local work-queue overflow causes stuff to be pushed on 3131 // global (common) overflow queue 3132 // . always first empty local work queue 3133 // . then get a batch of oops from global work queue if any 3134 // . then do work stealing 3135 // -- When all tasks claimed (both spaces) 3136 // and local work queue empty, 3137 // then in a loop do: 3138 // . check global overflow stack; steal a batch of oops and trace 3139 // . try to steal from other threads oif GOS is empty 3140 // . if neither is available, offer termination 3141 // -- Terminate and return result 3142 // 3143 void CMSConcMarkingTask::work(uint worker_id) { 3144 elapsedTimer _timer; 3145 ResourceMark rm; 3146 HandleMark hm; 3147 3148 DEBUG_ONLY(_collector->verify_overflow_empty();) 3149 3150 // Before we begin work, our work queue should be empty 3151 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3152 // Scan the bitmap covering _cms_space, tracing through grey objects. 3153 _timer.start(); 3154 do_scan_and_mark(worker_id, _cms_space); 3155 _timer.stop(); 3156 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3157 3158 // ... do work stealing 3159 _timer.reset(); 3160 _timer.start(); 3161 do_work_steal(worker_id); 3162 _timer.stop(); 3163 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3164 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3165 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3166 // Note that under the current task protocol, the 3167 // following assertion is true even of the spaces 3168 // expanded since the completion of the concurrent 3169 // marking. XXX This will likely change under a strict 3170 // ABORT semantics. 3171 // After perm removal the comparison was changed to 3172 // greater than or equal to from strictly greater than. 3173 // Before perm removal the highest address sweep would 3174 // have been at the end of perm gen but now is at the 3175 // end of the tenured gen. 3176 assert(_global_finger >= _cms_space->end(), 3177 "All tasks have been completed"); 3178 DEBUG_ONLY(_collector->verify_overflow_empty();) 3179 } 3180 3181 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3182 HeapWord* read = _global_finger; 3183 HeapWord* cur = read; 3184 while (f > read) { 3185 cur = read; 3186 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); 3187 if (cur == read) { 3188 // our cas succeeded 3189 assert(_global_finger >= f, "protocol consistency"); 3190 break; 3191 } 3192 } 3193 } 3194 3195 // This is really inefficient, and should be redone by 3196 // using (not yet available) block-read and -write interfaces to the 3197 // stack and the work_queue. XXX FIX ME !!! 3198 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3199 OopTaskQueue* work_q) { 3200 // Fast lock-free check 3201 if (ovflw_stk->length() == 0) { 3202 return false; 3203 } 3204 assert(work_q->size() == 0, "Shouldn't steal"); 3205 MutexLockerEx ml(ovflw_stk->par_lock(), 3206 Mutex::_no_safepoint_check_flag); 3207 // Grab up to 1/4 the size of the work queue 3208 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3209 (size_t)ParGCDesiredObjsFromOverflowList); 3210 num = MIN2(num, ovflw_stk->length()); 3211 for (int i = (int) num; i > 0; i--) { 3212 oop cur = ovflw_stk->pop(); 3213 assert(cur != NULL, "Counted wrong?"); 3214 work_q->push(cur); 3215 } 3216 return num > 0; 3217 } 3218 3219 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3220 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3221 int n_tasks = pst->n_tasks(); 3222 // We allow that there may be no tasks to do here because 3223 // we are restarting after a stack overflow. 3224 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3225 uint nth_task = 0; 3226 3227 HeapWord* aligned_start = sp->bottom(); 3228 if (sp->used_region().contains(_restart_addr)) { 3229 // Align down to a card boundary for the start of 0th task 3230 // for this space. 3231 aligned_start = 3232 (HeapWord*)align_size_down((uintptr_t)_restart_addr, 3233 CardTableModRefBS::card_size); 3234 } 3235 3236 size_t chunk_size = sp->marking_task_size(); 3237 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3238 // Having claimed the nth task in this space, 3239 // compute the chunk that it corresponds to: 3240 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3241 aligned_start + (nth_task+1)*chunk_size); 3242 // Try and bump the global finger via a CAS; 3243 // note that we need to do the global finger bump 3244 // _before_ taking the intersection below, because 3245 // the task corresponding to that region will be 3246 // deemed done even if the used_region() expands 3247 // because of allocation -- as it almost certainly will 3248 // during start-up while the threads yield in the 3249 // closure below. 3250 HeapWord* finger = span.end(); 3251 bump_global_finger(finger); // atomically 3252 // There are null tasks here corresponding to chunks 3253 // beyond the "top" address of the space. 3254 span = span.intersection(sp->used_region()); 3255 if (!span.is_empty()) { // Non-null task 3256 HeapWord* prev_obj; 3257 assert(!span.contains(_restart_addr) || nth_task == 0, 3258 "Inconsistency"); 3259 if (nth_task == 0) { 3260 // For the 0th task, we'll not need to compute a block_start. 3261 if (span.contains(_restart_addr)) { 3262 // In the case of a restart because of stack overflow, 3263 // we might additionally skip a chunk prefix. 3264 prev_obj = _restart_addr; 3265 } else { 3266 prev_obj = span.start(); 3267 } 3268 } else { 3269 // We want to skip the first object because 3270 // the protocol is to scan any object in its entirety 3271 // that _starts_ in this span; a fortiori, any 3272 // object starting in an earlier span is scanned 3273 // as part of an earlier claimed task. 3274 // Below we use the "careful" version of block_start 3275 // so we do not try to navigate uninitialized objects. 3276 prev_obj = sp->block_start_careful(span.start()); 3277 // Below we use a variant of block_size that uses the 3278 // Printezis bits to avoid waiting for allocated 3279 // objects to become initialized/parsable. 3280 while (prev_obj < span.start()) { 3281 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3282 if (sz > 0) { 3283 prev_obj += sz; 3284 } else { 3285 // In this case we may end up doing a bit of redundant 3286 // scanning, but that appears unavoidable, short of 3287 // locking the free list locks; see bug 6324141. 3288 break; 3289 } 3290 } 3291 } 3292 if (prev_obj < span.end()) { 3293 MemRegion my_span = MemRegion(prev_obj, span.end()); 3294 // Do the marking work within a non-empty span -- 3295 // the last argument to the constructor indicates whether the 3296 // iteration should be incremental with periodic yields. 3297 ParMarkFromRootsClosure cl(this, _collector, my_span, 3298 &_collector->_markBitMap, 3299 work_queue(i), 3300 &_collector->_markStack); 3301 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3302 } // else nothing to do for this task 3303 } // else nothing to do for this task 3304 } 3305 // We'd be tempted to assert here that since there are no 3306 // more tasks left to claim in this space, the global_finger 3307 // must exceed space->top() and a fortiori space->end(). However, 3308 // that would not quite be correct because the bumping of 3309 // global_finger occurs strictly after the claiming of a task, 3310 // so by the time we reach here the global finger may not yet 3311 // have been bumped up by the thread that claimed the last 3312 // task. 3313 pst->all_tasks_completed(); 3314 } 3315 3316 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3317 private: 3318 CMSCollector* _collector; 3319 CMSConcMarkingTask* _task; 3320 MemRegion _span; 3321 CMSBitMap* _bit_map; 3322 CMSMarkStack* _overflow_stack; 3323 OopTaskQueue* _work_queue; 3324 protected: 3325 DO_OOP_WORK_DEFN 3326 public: 3327 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3328 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3329 MetadataAwareOopClosure(collector->ref_processor()), 3330 _collector(collector), 3331 _task(task), 3332 _span(collector->_span), 3333 _work_queue(work_queue), 3334 _bit_map(bit_map), 3335 _overflow_stack(overflow_stack) 3336 { } 3337 virtual void do_oop(oop* p); 3338 virtual void do_oop(narrowOop* p); 3339 3340 void trim_queue(size_t max); 3341 void handle_stack_overflow(HeapWord* lost); 3342 void do_yield_check() { 3343 if (_task->should_yield()) { 3344 _task->yield(); 3345 } 3346 } 3347 }; 3348 3349 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3350 3351 // Grey object scanning during work stealing phase -- 3352 // the salient assumption here is that any references 3353 // that are in these stolen objects being scanned must 3354 // already have been initialized (else they would not have 3355 // been published), so we do not need to check for 3356 // uninitialized objects before pushing here. 3357 void ParConcMarkingClosure::do_oop(oop obj) { 3358 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3359 HeapWord* addr = (HeapWord*)obj; 3360 // Check if oop points into the CMS generation 3361 // and is not marked 3362 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3363 // a white object ... 3364 // If we manage to "claim" the object, by being the 3365 // first thread to mark it, then we push it on our 3366 // marking stack 3367 if (_bit_map->par_mark(addr)) { // ... now grey 3368 // push on work queue (grey set) 3369 bool simulate_overflow = false; 3370 NOT_PRODUCT( 3371 if (CMSMarkStackOverflowALot && 3372 _collector->simulate_overflow()) { 3373 // simulate a stack overflow 3374 simulate_overflow = true; 3375 } 3376 ) 3377 if (simulate_overflow || 3378 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3379 // stack overflow 3380 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3381 // We cannot assert that the overflow stack is full because 3382 // it may have been emptied since. 3383 assert(simulate_overflow || 3384 _work_queue->size() == _work_queue->max_elems(), 3385 "Else push should have succeeded"); 3386 handle_stack_overflow(addr); 3387 } 3388 } // Else, some other thread got there first 3389 do_yield_check(); 3390 } 3391 } 3392 3393 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3394 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3395 3396 void ParConcMarkingClosure::trim_queue(size_t max) { 3397 while (_work_queue->size() > max) { 3398 oop new_oop; 3399 if (_work_queue->pop_local(new_oop)) { 3400 assert(new_oop->is_oop(), "Should be an oop"); 3401 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3402 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3403 new_oop->oop_iterate(this); // do_oop() above 3404 do_yield_check(); 3405 } 3406 } 3407 } 3408 3409 // Upon stack overflow, we discard (part of) the stack, 3410 // remembering the least address amongst those discarded 3411 // in CMSCollector's _restart_address. 3412 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3413 // We need to do this under a mutex to prevent other 3414 // workers from interfering with the work done below. 3415 MutexLockerEx ml(_overflow_stack->par_lock(), 3416 Mutex::_no_safepoint_check_flag); 3417 // Remember the least grey address discarded 3418 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3419 _collector->lower_restart_addr(ra); 3420 _overflow_stack->reset(); // discard stack contents 3421 _overflow_stack->expand(); // expand the stack if possible 3422 } 3423 3424 3425 void CMSConcMarkingTask::do_work_steal(int i) { 3426 OopTaskQueue* work_q = work_queue(i); 3427 oop obj_to_scan; 3428 CMSBitMap* bm = &(_collector->_markBitMap); 3429 CMSMarkStack* ovflw = &(_collector->_markStack); 3430 int* seed = _collector->hash_seed(i); 3431 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3432 while (true) { 3433 cl.trim_queue(0); 3434 assert(work_q->size() == 0, "Should have been emptied above"); 3435 if (get_work_from_overflow_stack(ovflw, work_q)) { 3436 // Can't assert below because the work obtained from the 3437 // overflow stack may already have been stolen from us. 3438 // assert(work_q->size() > 0, "Work from overflow stack"); 3439 continue; 3440 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3441 assert(obj_to_scan->is_oop(), "Should be an oop"); 3442 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3443 obj_to_scan->oop_iterate(&cl); 3444 } else if (terminator()->offer_termination(&_term_term)) { 3445 assert(work_q->size() == 0, "Impossible!"); 3446 break; 3447 } else if (yielding() || should_yield()) { 3448 yield(); 3449 } 3450 } 3451 } 3452 3453 // This is run by the CMS (coordinator) thread. 3454 void CMSConcMarkingTask::coordinator_yield() { 3455 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3456 "CMS thread should hold CMS token"); 3457 // First give up the locks, then yield, then re-lock 3458 // We should probably use a constructor/destructor idiom to 3459 // do this unlock/lock or modify the MutexUnlocker class to 3460 // serve our purpose. XXX 3461 assert_lock_strong(_bit_map_lock); 3462 _bit_map_lock->unlock(); 3463 ConcurrentMarkSweepThread::desynchronize(true); 3464 _collector->stopTimer(); 3465 _collector->incrementYields(); 3466 3467 // It is possible for whichever thread initiated the yield request 3468 // not to get a chance to wake up and take the bitmap lock between 3469 // this thread releasing it and reacquiring it. So, while the 3470 // should_yield() flag is on, let's sleep for a bit to give the 3471 // other thread a chance to wake up. The limit imposed on the number 3472 // of iterations is defensive, to avoid any unforseen circumstances 3473 // putting us into an infinite loop. Since it's always been this 3474 // (coordinator_yield()) method that was observed to cause the 3475 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3476 // which is by default non-zero. For the other seven methods that 3477 // also perform the yield operation, as are using a different 3478 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3479 // can enable the sleeping for those methods too, if necessary. 3480 // See 6442774. 3481 // 3482 // We really need to reconsider the synchronization between the GC 3483 // thread and the yield-requesting threads in the future and we 3484 // should really use wait/notify, which is the recommended 3485 // way of doing this type of interaction. Additionally, we should 3486 // consolidate the eight methods that do the yield operation and they 3487 // are almost identical into one for better maintainability and 3488 // readability. See 6445193. 3489 // 3490 // Tony 2006.06.29 3491 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3492 ConcurrentMarkSweepThread::should_yield() && 3493 !CMSCollector::foregroundGCIsActive(); ++i) { 3494 os::sleep(Thread::current(), 1, false); 3495 } 3496 3497 ConcurrentMarkSweepThread::synchronize(true); 3498 _bit_map_lock->lock_without_safepoint_check(); 3499 _collector->startTimer(); 3500 } 3501 3502 bool CMSCollector::do_marking_mt() { 3503 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3504 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3505 conc_workers()->active_workers(), 3506 Threads::number_of_non_daemon_threads()); 3507 num_workers = conc_workers()->update_active_workers(num_workers); 3508 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3509 3510 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3511 3512 CMSConcMarkingTask tsk(this, 3513 cms_space, 3514 conc_workers(), 3515 task_queues()); 3516 3517 // Since the actual number of workers we get may be different 3518 // from the number we requested above, do we need to do anything different 3519 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3520 // class?? XXX 3521 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3522 3523 // Refs discovery is already non-atomic. 3524 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3525 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3526 conc_workers()->start_task(&tsk); 3527 while (tsk.yielded()) { 3528 tsk.coordinator_yield(); 3529 conc_workers()->continue_task(&tsk); 3530 } 3531 // If the task was aborted, _restart_addr will be non-NULL 3532 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3533 while (_restart_addr != NULL) { 3534 // XXX For now we do not make use of ABORTED state and have not 3535 // yet implemented the right abort semantics (even in the original 3536 // single-threaded CMS case). That needs some more investigation 3537 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3538 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3539 // If _restart_addr is non-NULL, a marking stack overflow 3540 // occurred; we need to do a fresh marking iteration from the 3541 // indicated restart address. 3542 if (_foregroundGCIsActive) { 3543 // We may be running into repeated stack overflows, having 3544 // reached the limit of the stack size, while making very 3545 // slow forward progress. It may be best to bail out and 3546 // let the foreground collector do its job. 3547 // Clear _restart_addr, so that foreground GC 3548 // works from scratch. This avoids the headache of 3549 // a "rescan" which would otherwise be needed because 3550 // of the dirty mod union table & card table. 3551 _restart_addr = NULL; 3552 return false; 3553 } 3554 // Adjust the task to restart from _restart_addr 3555 tsk.reset(_restart_addr); 3556 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3557 _restart_addr); 3558 _restart_addr = NULL; 3559 // Get the workers going again 3560 conc_workers()->start_task(&tsk); 3561 while (tsk.yielded()) { 3562 tsk.coordinator_yield(); 3563 conc_workers()->continue_task(&tsk); 3564 } 3565 } 3566 assert(tsk.completed(), "Inconsistency"); 3567 assert(tsk.result() == true, "Inconsistency"); 3568 return true; 3569 } 3570 3571 bool CMSCollector::do_marking_st() { 3572 ResourceMark rm; 3573 HandleMark hm; 3574 3575 // Temporarily make refs discovery single threaded (non-MT) 3576 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3577 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3578 &_markStack, CMSYield); 3579 // the last argument to iterate indicates whether the iteration 3580 // should be incremental with periodic yields. 3581 _markBitMap.iterate(&markFromRootsClosure); 3582 // If _restart_addr is non-NULL, a marking stack overflow 3583 // occurred; we need to do a fresh iteration from the 3584 // indicated restart address. 3585 while (_restart_addr != NULL) { 3586 if (_foregroundGCIsActive) { 3587 // We may be running into repeated stack overflows, having 3588 // reached the limit of the stack size, while making very 3589 // slow forward progress. It may be best to bail out and 3590 // let the foreground collector do its job. 3591 // Clear _restart_addr, so that foreground GC 3592 // works from scratch. This avoids the headache of 3593 // a "rescan" which would otherwise be needed because 3594 // of the dirty mod union table & card table. 3595 _restart_addr = NULL; 3596 return false; // indicating failure to complete marking 3597 } 3598 // Deal with stack overflow: 3599 // we restart marking from _restart_addr 3600 HeapWord* ra = _restart_addr; 3601 markFromRootsClosure.reset(ra); 3602 _restart_addr = NULL; 3603 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3604 } 3605 return true; 3606 } 3607 3608 void CMSCollector::preclean() { 3609 check_correct_thread_executing(); 3610 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3611 verify_work_stacks_empty(); 3612 verify_overflow_empty(); 3613 _abort_preclean = false; 3614 if (CMSPrecleaningEnabled) { 3615 if (!CMSEdenChunksRecordAlways) { 3616 _eden_chunk_index = 0; 3617 } 3618 size_t used = get_eden_used(); 3619 size_t capacity = get_eden_capacity(); 3620 // Don't start sampling unless we will get sufficiently 3621 // many samples. 3622 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3623 * CMSScheduleRemarkEdenPenetration)) { 3624 _start_sampling = true; 3625 } else { 3626 _start_sampling = false; 3627 } 3628 GCTraceCPUTime tcpu; 3629 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3630 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3631 } 3632 CMSTokenSync x(true); // is cms thread 3633 if (CMSPrecleaningEnabled) { 3634 sample_eden(); 3635 _collectorState = AbortablePreclean; 3636 } else { 3637 _collectorState = FinalMarking; 3638 } 3639 verify_work_stacks_empty(); 3640 verify_overflow_empty(); 3641 } 3642 3643 // Try and schedule the remark such that young gen 3644 // occupancy is CMSScheduleRemarkEdenPenetration %. 3645 void CMSCollector::abortable_preclean() { 3646 check_correct_thread_executing(); 3647 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3648 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3649 3650 // If Eden's current occupancy is below this threshold, 3651 // immediately schedule the remark; else preclean 3652 // past the next scavenge in an effort to 3653 // schedule the pause as described above. By choosing 3654 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3655 // we will never do an actual abortable preclean cycle. 3656 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3657 GCTraceCPUTime tcpu; 3658 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3659 // We need more smarts in the abortable preclean 3660 // loop below to deal with cases where allocation 3661 // in young gen is very very slow, and our precleaning 3662 // is running a losing race against a horde of 3663 // mutators intent on flooding us with CMS updates 3664 // (dirty cards). 3665 // One, admittedly dumb, strategy is to give up 3666 // after a certain number of abortable precleaning loops 3667 // or after a certain maximum time. We want to make 3668 // this smarter in the next iteration. 3669 // XXX FIX ME!!! YSR 3670 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3671 while (!(should_abort_preclean() || 3672 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3673 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3674 cumworkdone += workdone; 3675 loops++; 3676 // Voluntarily terminate abortable preclean phase if we have 3677 // been at it for too long. 3678 if ((CMSMaxAbortablePrecleanLoops != 0) && 3679 loops >= CMSMaxAbortablePrecleanLoops) { 3680 log_debug(gc)(" CMS: abort preclean due to loops "); 3681 break; 3682 } 3683 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3684 log_debug(gc)(" CMS: abort preclean due to time "); 3685 break; 3686 } 3687 // If we are doing little work each iteration, we should 3688 // take a short break. 3689 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3690 // Sleep for some time, waiting for work to accumulate 3691 stopTimer(); 3692 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3693 startTimer(); 3694 waited++; 3695 } 3696 } 3697 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3698 loops, waited, cumworkdone); 3699 } 3700 CMSTokenSync x(true); // is cms thread 3701 if (_collectorState != Idling) { 3702 assert(_collectorState == AbortablePreclean, 3703 "Spontaneous state transition?"); 3704 _collectorState = FinalMarking; 3705 } // Else, a foreground collection completed this CMS cycle. 3706 return; 3707 } 3708 3709 // Respond to an Eden sampling opportunity 3710 void CMSCollector::sample_eden() { 3711 // Make sure a young gc cannot sneak in between our 3712 // reading and recording of a sample. 3713 assert(Thread::current()->is_ConcurrentGC_thread(), 3714 "Only the cms thread may collect Eden samples"); 3715 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3716 "Should collect samples while holding CMS token"); 3717 if (!_start_sampling) { 3718 return; 3719 } 3720 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3721 // is populated by the young generation. 3722 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3723 if (_eden_chunk_index < _eden_chunk_capacity) { 3724 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3725 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3726 "Unexpected state of Eden"); 3727 // We'd like to check that what we just sampled is an oop-start address; 3728 // however, we cannot do that here since the object may not yet have been 3729 // initialized. So we'll instead do the check when we _use_ this sample 3730 // later. 3731 if (_eden_chunk_index == 0 || 3732 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3733 _eden_chunk_array[_eden_chunk_index-1]) 3734 >= CMSSamplingGrain)) { 3735 _eden_chunk_index++; // commit sample 3736 } 3737 } 3738 } 3739 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3740 size_t used = get_eden_used(); 3741 size_t capacity = get_eden_capacity(); 3742 assert(used <= capacity, "Unexpected state of Eden"); 3743 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3744 _abort_preclean = true; 3745 } 3746 } 3747 } 3748 3749 3750 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3751 assert(_collectorState == Precleaning || 3752 _collectorState == AbortablePreclean, "incorrect state"); 3753 ResourceMark rm; 3754 HandleMark hm; 3755 3756 // Precleaning is currently not MT but the reference processor 3757 // may be set for MT. Disable it temporarily here. 3758 ReferenceProcessor* rp = ref_processor(); 3759 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3760 3761 // Do one pass of scrubbing the discovered reference lists 3762 // to remove any reference objects with strongly-reachable 3763 // referents. 3764 if (clean_refs) { 3765 CMSPrecleanRefsYieldClosure yield_cl(this); 3766 assert(rp->span().equals(_span), "Spans should be equal"); 3767 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3768 &_markStack, true /* preclean */); 3769 CMSDrainMarkingStackClosure complete_trace(this, 3770 _span, &_markBitMap, &_markStack, 3771 &keep_alive, true /* preclean */); 3772 3773 // We don't want this step to interfere with a young 3774 // collection because we don't want to take CPU 3775 // or memory bandwidth away from the young GC threads 3776 // (which may be as many as there are CPUs). 3777 // Note that we don't need to protect ourselves from 3778 // interference with mutators because they can't 3779 // manipulate the discovered reference lists nor affect 3780 // the computed reachability of the referents, the 3781 // only properties manipulated by the precleaning 3782 // of these reference lists. 3783 stopTimer(); 3784 CMSTokenSyncWithLocks x(true /* is cms thread */, 3785 bitMapLock()); 3786 startTimer(); 3787 sample_eden(); 3788 3789 // The following will yield to allow foreground 3790 // collection to proceed promptly. XXX YSR: 3791 // The code in this method may need further 3792 // tweaking for better performance and some restructuring 3793 // for cleaner interfaces. 3794 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3795 rp->preclean_discovered_references( 3796 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3797 gc_timer); 3798 } 3799 3800 if (clean_survivor) { // preclean the active survivor space(s) 3801 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3802 &_markBitMap, &_modUnionTable, 3803 &_markStack, true /* precleaning phase */); 3804 stopTimer(); 3805 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3806 bitMapLock()); 3807 startTimer(); 3808 unsigned int before_count = 3809 GenCollectedHeap::heap()->total_collections(); 3810 SurvivorSpacePrecleanClosure 3811 sss_cl(this, _span, &_markBitMap, &_markStack, 3812 &pam_cl, before_count, CMSYield); 3813 _young_gen->from()->object_iterate_careful(&sss_cl); 3814 _young_gen->to()->object_iterate_careful(&sss_cl); 3815 } 3816 MarkRefsIntoAndScanClosure 3817 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3818 &_markStack, this, CMSYield, 3819 true /* precleaning phase */); 3820 // CAUTION: The following closure has persistent state that may need to 3821 // be reset upon a decrease in the sequence of addresses it 3822 // processes. 3823 ScanMarkedObjectsAgainCarefullyClosure 3824 smoac_cl(this, _span, 3825 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3826 3827 // Preclean dirty cards in ModUnionTable and CardTable using 3828 // appropriate convergence criterion; 3829 // repeat CMSPrecleanIter times unless we find that 3830 // we are losing. 3831 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3832 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3833 "Bad convergence multiplier"); 3834 assert(CMSPrecleanThreshold >= 100, 3835 "Unreasonably low CMSPrecleanThreshold"); 3836 3837 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3838 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3839 numIter < CMSPrecleanIter; 3840 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3841 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3842 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3843 // Either there are very few dirty cards, so re-mark 3844 // pause will be small anyway, or our pre-cleaning isn't 3845 // that much faster than the rate at which cards are being 3846 // dirtied, so we might as well stop and re-mark since 3847 // precleaning won't improve our re-mark time by much. 3848 if (curNumCards <= CMSPrecleanThreshold || 3849 (numIter > 0 && 3850 (curNumCards * CMSPrecleanDenominator > 3851 lastNumCards * CMSPrecleanNumerator))) { 3852 numIter++; 3853 cumNumCards += curNumCards; 3854 break; 3855 } 3856 } 3857 3858 preclean_klasses(&mrias_cl, _cmsGen->freelistLock()); 3859 3860 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3861 cumNumCards += curNumCards; 3862 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3863 curNumCards, cumNumCards, numIter); 3864 return cumNumCards; // as a measure of useful work done 3865 } 3866 3867 // PRECLEANING NOTES: 3868 // Precleaning involves: 3869 // . reading the bits of the modUnionTable and clearing the set bits. 3870 // . For the cards corresponding to the set bits, we scan the 3871 // objects on those cards. This means we need the free_list_lock 3872 // so that we can safely iterate over the CMS space when scanning 3873 // for oops. 3874 // . When we scan the objects, we'll be both reading and setting 3875 // marks in the marking bit map, so we'll need the marking bit map. 3876 // . For protecting _collector_state transitions, we take the CGC_lock. 3877 // Note that any races in the reading of of card table entries by the 3878 // CMS thread on the one hand and the clearing of those entries by the 3879 // VM thread or the setting of those entries by the mutator threads on the 3880 // other are quite benign. However, for efficiency it makes sense to keep 3881 // the VM thread from racing with the CMS thread while the latter is 3882 // dirty card info to the modUnionTable. We therefore also use the 3883 // CGC_lock to protect the reading of the card table and the mod union 3884 // table by the CM thread. 3885 // . We run concurrently with mutator updates, so scanning 3886 // needs to be done carefully -- we should not try to scan 3887 // potentially uninitialized objects. 3888 // 3889 // Locking strategy: While holding the CGC_lock, we scan over and 3890 // reset a maximal dirty range of the mod union / card tables, then lock 3891 // the free_list_lock and bitmap lock to do a full marking, then 3892 // release these locks; and repeat the cycle. This allows for a 3893 // certain amount of fairness in the sharing of these locks between 3894 // the CMS collector on the one hand, and the VM thread and the 3895 // mutators on the other. 3896 3897 // NOTE: preclean_mod_union_table() and preclean_card_table() 3898 // further below are largely identical; if you need to modify 3899 // one of these methods, please check the other method too. 3900 3901 size_t CMSCollector::preclean_mod_union_table( 3902 ConcurrentMarkSweepGeneration* old_gen, 3903 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3904 verify_work_stacks_empty(); 3905 verify_overflow_empty(); 3906 3907 // strategy: starting with the first card, accumulate contiguous 3908 // ranges of dirty cards; clear these cards, then scan the region 3909 // covered by these cards. 3910 3911 // Since all of the MUT is committed ahead, we can just use 3912 // that, in case the generations expand while we are precleaning. 3913 // It might also be fine to just use the committed part of the 3914 // generation, but we might potentially miss cards when the 3915 // generation is rapidly expanding while we are in the midst 3916 // of precleaning. 3917 HeapWord* startAddr = old_gen->reserved().start(); 3918 HeapWord* endAddr = old_gen->reserved().end(); 3919 3920 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3921 3922 size_t numDirtyCards, cumNumDirtyCards; 3923 HeapWord *nextAddr, *lastAddr; 3924 for (cumNumDirtyCards = numDirtyCards = 0, 3925 nextAddr = lastAddr = startAddr; 3926 nextAddr < endAddr; 3927 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3928 3929 ResourceMark rm; 3930 HandleMark hm; 3931 3932 MemRegion dirtyRegion; 3933 { 3934 stopTimer(); 3935 // Potential yield point 3936 CMSTokenSync ts(true); 3937 startTimer(); 3938 sample_eden(); 3939 // Get dirty region starting at nextOffset (inclusive), 3940 // simultaneously clearing it. 3941 dirtyRegion = 3942 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3943 assert(dirtyRegion.start() >= nextAddr, 3944 "returned region inconsistent?"); 3945 } 3946 // Remember where the next search should begin. 3947 // The returned region (if non-empty) is a right open interval, 3948 // so lastOffset is obtained from the right end of that 3949 // interval. 3950 lastAddr = dirtyRegion.end(); 3951 // Should do something more transparent and less hacky XXX 3952 numDirtyCards = 3953 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3954 3955 // We'll scan the cards in the dirty region (with periodic 3956 // yields for foreground GC as needed). 3957 if (!dirtyRegion.is_empty()) { 3958 assert(numDirtyCards > 0, "consistency check"); 3959 HeapWord* stop_point = NULL; 3960 stopTimer(); 3961 // Potential yield point 3962 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3963 bitMapLock()); 3964 startTimer(); 3965 { 3966 verify_work_stacks_empty(); 3967 verify_overflow_empty(); 3968 sample_eden(); 3969 stop_point = 3970 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3971 } 3972 if (stop_point != NULL) { 3973 // The careful iteration stopped early either because it found an 3974 // uninitialized object, or because we were in the midst of an 3975 // "abortable preclean", which should now be aborted. Redirty 3976 // the bits corresponding to the partially-scanned or unscanned 3977 // cards. We'll either restart at the next block boundary or 3978 // abort the preclean. 3979 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3980 "Should only be AbortablePreclean."); 3981 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3982 if (should_abort_preclean()) { 3983 break; // out of preclean loop 3984 } else { 3985 // Compute the next address at which preclean should pick up; 3986 // might need bitMapLock in order to read P-bits. 3987 lastAddr = next_card_start_after_block(stop_point); 3988 } 3989 } 3990 } else { 3991 assert(lastAddr == endAddr, "consistency check"); 3992 assert(numDirtyCards == 0, "consistency check"); 3993 break; 3994 } 3995 } 3996 verify_work_stacks_empty(); 3997 verify_overflow_empty(); 3998 return cumNumDirtyCards; 3999 } 4000 4001 // NOTE: preclean_mod_union_table() above and preclean_card_table() 4002 // below are largely identical; if you need to modify 4003 // one of these methods, please check the other method too. 4004 4005 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 4006 ScanMarkedObjectsAgainCarefullyClosure* cl) { 4007 // strategy: it's similar to precleamModUnionTable above, in that 4008 // we accumulate contiguous ranges of dirty cards, mark these cards 4009 // precleaned, then scan the region covered by these cards. 4010 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4011 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4012 4013 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4014 4015 size_t numDirtyCards, cumNumDirtyCards; 4016 HeapWord *lastAddr, *nextAddr; 4017 4018 for (cumNumDirtyCards = numDirtyCards = 0, 4019 nextAddr = lastAddr = startAddr; 4020 nextAddr < endAddr; 4021 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4022 4023 ResourceMark rm; 4024 HandleMark hm; 4025 4026 MemRegion dirtyRegion; 4027 { 4028 // See comments in "Precleaning notes" above on why we 4029 // do this locking. XXX Could the locking overheads be 4030 // too high when dirty cards are sparse? [I don't think so.] 4031 stopTimer(); 4032 CMSTokenSync x(true); // is cms thread 4033 startTimer(); 4034 sample_eden(); 4035 // Get and clear dirty region from card table 4036 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( 4037 MemRegion(nextAddr, endAddr), 4038 true, 4039 CardTableModRefBS::precleaned_card_val()); 4040 4041 assert(dirtyRegion.start() >= nextAddr, 4042 "returned region inconsistent?"); 4043 } 4044 lastAddr = dirtyRegion.end(); 4045 numDirtyCards = 4046 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; 4047 4048 if (!dirtyRegion.is_empty()) { 4049 stopTimer(); 4050 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4051 startTimer(); 4052 sample_eden(); 4053 verify_work_stacks_empty(); 4054 verify_overflow_empty(); 4055 HeapWord* stop_point = 4056 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4057 if (stop_point != NULL) { 4058 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4059 "Should only be AbortablePreclean."); 4060 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4061 if (should_abort_preclean()) { 4062 break; // out of preclean loop 4063 } else { 4064 // Compute the next address at which preclean should pick up. 4065 lastAddr = next_card_start_after_block(stop_point); 4066 } 4067 } 4068 } else { 4069 break; 4070 } 4071 } 4072 verify_work_stacks_empty(); 4073 verify_overflow_empty(); 4074 return cumNumDirtyCards; 4075 } 4076 4077 class PrecleanKlassClosure : public KlassClosure { 4078 KlassToOopClosure _cm_klass_closure; 4079 public: 4080 PrecleanKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4081 void do_klass(Klass* k) { 4082 if (k->has_accumulated_modified_oops()) { 4083 k->clear_accumulated_modified_oops(); 4084 4085 _cm_klass_closure.do_klass(k); 4086 } 4087 } 4088 }; 4089 4090 // The freelist lock is needed to prevent asserts, is it really needed? 4091 void CMSCollector::preclean_klasses(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4092 4093 cl->set_freelistLock(freelistLock); 4094 4095 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4096 4097 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4098 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4099 PrecleanKlassClosure preclean_klass_closure(cl); 4100 ClassLoaderDataGraph::classes_do(&preclean_klass_closure); 4101 4102 verify_work_stacks_empty(); 4103 verify_overflow_empty(); 4104 } 4105 4106 void CMSCollector::checkpointRootsFinal() { 4107 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4108 check_correct_thread_executing(); 4109 // world is stopped at this checkpoint 4110 assert(SafepointSynchronize::is_at_safepoint(), 4111 "world should be stopped"); 4112 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 4113 4114 verify_work_stacks_empty(); 4115 verify_overflow_empty(); 4116 4117 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4118 _young_gen->used() / K, _young_gen->capacity() / K); 4119 { 4120 if (CMSScavengeBeforeRemark) { 4121 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4122 // Temporarily set flag to false, GCH->do_collection will 4123 // expect it to be false and set to true 4124 FlagSetting fl(gch->_is_gc_active, false); 4125 4126 gch->do_collection(true, // full (i.e. force, see below) 4127 false, // !clear_all_soft_refs 4128 0, // size 4129 false, // is_tlab 4130 GenCollectedHeap::YoungGen // type 4131 ); 4132 } 4133 FreelistLocker x(this); 4134 MutexLockerEx y(bitMapLock(), 4135 Mutex::_no_safepoint_check_flag); 4136 checkpointRootsFinalWork(); 4137 } 4138 verify_work_stacks_empty(); 4139 verify_overflow_empty(); 4140 } 4141 4142 void CMSCollector::checkpointRootsFinalWork() { 4143 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4144 4145 assert(haveFreelistLocks(), "must have free list locks"); 4146 assert_lock_strong(bitMapLock()); 4147 4148 ResourceMark rm; 4149 HandleMark hm; 4150 4151 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4152 4153 if (should_unload_classes()) { 4154 CodeCache::gc_prologue(); 4155 } 4156 assert(haveFreelistLocks(), "must have free list locks"); 4157 assert_lock_strong(bitMapLock()); 4158 4159 // We might assume that we need not fill TLAB's when 4160 // CMSScavengeBeforeRemark is set, because we may have just done 4161 // a scavenge which would have filled all TLAB's -- and besides 4162 // Eden would be empty. This however may not always be the case -- 4163 // for instance although we asked for a scavenge, it may not have 4164 // happened because of a JNI critical section. We probably need 4165 // a policy for deciding whether we can in that case wait until 4166 // the critical section releases and then do the remark following 4167 // the scavenge, and skip it here. In the absence of that policy, 4168 // or of an indication of whether the scavenge did indeed occur, 4169 // we cannot rely on TLAB's having been filled and must do 4170 // so here just in case a scavenge did not happen. 4171 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them 4172 // Update the saved marks which may affect the root scans. 4173 gch->save_marks(); 4174 4175 print_eden_and_survivor_chunk_arrays(); 4176 4177 { 4178 #if defined(COMPILER2) || INCLUDE_JVMCI 4179 DerivedPointerTableDeactivate dpt_deact; 4180 #endif 4181 4182 // Note on the role of the mod union table: 4183 // Since the marker in "markFromRoots" marks concurrently with 4184 // mutators, it is possible for some reachable objects not to have been 4185 // scanned. For instance, an only reference to an object A was 4186 // placed in object B after the marker scanned B. Unless B is rescanned, 4187 // A would be collected. Such updates to references in marked objects 4188 // are detected via the mod union table which is the set of all cards 4189 // dirtied since the first checkpoint in this GC cycle and prior to 4190 // the most recent young generation GC, minus those cleaned up by the 4191 // concurrent precleaning. 4192 if (CMSParallelRemarkEnabled) { 4193 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4194 do_remark_parallel(); 4195 } else { 4196 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4197 do_remark_non_parallel(); 4198 } 4199 } 4200 verify_work_stacks_empty(); 4201 verify_overflow_empty(); 4202 4203 { 4204 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4205 refProcessingWork(); 4206 } 4207 verify_work_stacks_empty(); 4208 verify_overflow_empty(); 4209 4210 if (should_unload_classes()) { 4211 CodeCache::gc_epilogue(); 4212 } 4213 JvmtiExport::gc_epilogue(); 4214 4215 // If we encountered any (marking stack / work queue) overflow 4216 // events during the current CMS cycle, take appropriate 4217 // remedial measures, where possible, so as to try and avoid 4218 // recurrence of that condition. 4219 assert(_markStack.isEmpty(), "No grey objects"); 4220 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4221 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4222 if (ser_ovflw > 0) { 4223 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4224 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4225 _markStack.expand(); 4226 _ser_pmc_remark_ovflw = 0; 4227 _ser_pmc_preclean_ovflw = 0; 4228 _ser_kac_preclean_ovflw = 0; 4229 _ser_kac_ovflw = 0; 4230 } 4231 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4232 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4233 _par_pmc_remark_ovflw, _par_kac_ovflw); 4234 _par_pmc_remark_ovflw = 0; 4235 _par_kac_ovflw = 0; 4236 } 4237 if (_markStack._hit_limit > 0) { 4238 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4239 _markStack._hit_limit); 4240 } 4241 if (_markStack._failed_double > 0) { 4242 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4243 _markStack._failed_double, _markStack.capacity()); 4244 } 4245 _markStack._hit_limit = 0; 4246 _markStack._failed_double = 0; 4247 4248 if ((VerifyAfterGC || VerifyDuringGC) && 4249 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4250 verify_after_remark(); 4251 } 4252 4253 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4254 4255 // Change under the freelistLocks. 4256 _collectorState = Sweeping; 4257 // Call isAllClear() under bitMapLock 4258 assert(_modUnionTable.isAllClear(), 4259 "Should be clear by end of the final marking"); 4260 assert(_ct->klass_rem_set()->mod_union_is_clear(), 4261 "Should be clear by end of the final marking"); 4262 } 4263 4264 void CMSParInitialMarkTask::work(uint worker_id) { 4265 elapsedTimer _timer; 4266 ResourceMark rm; 4267 HandleMark hm; 4268 4269 // ---------- scan from roots -------------- 4270 _timer.start(); 4271 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4272 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4273 4274 // ---------- young gen roots -------------- 4275 { 4276 work_on_young_gen_roots(&par_mri_cl); 4277 _timer.stop(); 4278 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4279 } 4280 4281 // ---------- remaining roots -------------- 4282 _timer.reset(); 4283 _timer.start(); 4284 4285 CLDToOopClosure cld_closure(&par_mri_cl, true); 4286 4287 gch->old_process_roots(_strong_roots_scope, 4288 false, // yg was scanned above 4289 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4290 _collector->should_unload_classes(), 4291 &par_mri_cl, 4292 &cld_closure); 4293 assert(_collector->should_unload_classes() 4294 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4295 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4296 _timer.stop(); 4297 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4298 } 4299 4300 // Parallel remark task 4301 class CMSParRemarkTask: public CMSParMarkTask { 4302 CompactibleFreeListSpace* _cms_space; 4303 4304 // The per-thread work queues, available here for stealing. 4305 OopTaskQueueSet* _task_queues; 4306 ParallelTaskTerminator _term; 4307 StrongRootsScope* _strong_roots_scope; 4308 4309 public: 4310 // A value of 0 passed to n_workers will cause the number of 4311 // workers to be taken from the active workers in the work gang. 4312 CMSParRemarkTask(CMSCollector* collector, 4313 CompactibleFreeListSpace* cms_space, 4314 uint n_workers, WorkGang* workers, 4315 OopTaskQueueSet* task_queues, 4316 StrongRootsScope* strong_roots_scope): 4317 CMSParMarkTask("Rescan roots and grey objects in parallel", 4318 collector, n_workers), 4319 _cms_space(cms_space), 4320 _task_queues(task_queues), 4321 _term(n_workers, task_queues), 4322 _strong_roots_scope(strong_roots_scope) { } 4323 4324 OopTaskQueueSet* task_queues() { return _task_queues; } 4325 4326 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4327 4328 ParallelTaskTerminator* terminator() { return &_term; } 4329 uint n_workers() { return _n_workers; } 4330 4331 void work(uint worker_id); 4332 4333 private: 4334 // ... of dirty cards in old space 4335 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4336 ParMarkRefsIntoAndScanClosure* cl); 4337 4338 // ... work stealing for the above 4339 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4340 }; 4341 4342 class RemarkKlassClosure : public KlassClosure { 4343 KlassToOopClosure _cm_klass_closure; 4344 public: 4345 RemarkKlassClosure(OopClosure* oop_closure) : _cm_klass_closure(oop_closure) {} 4346 void do_klass(Klass* k) { 4347 // Check if we have modified any oops in the Klass during the concurrent marking. 4348 if (k->has_accumulated_modified_oops()) { 4349 k->clear_accumulated_modified_oops(); 4350 4351 // We could have transfered the current modified marks to the accumulated marks, 4352 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4353 } else if (k->has_modified_oops()) { 4354 // Don't clear anything, this info is needed by the next young collection. 4355 } else { 4356 // No modified oops in the Klass. 4357 return; 4358 } 4359 4360 // The klass has modified fields, need to scan the klass. 4361 _cm_klass_closure.do_klass(k); 4362 } 4363 }; 4364 4365 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4366 ParNewGeneration* young_gen = _collector->_young_gen; 4367 ContiguousSpace* eden_space = young_gen->eden(); 4368 ContiguousSpace* from_space = young_gen->from(); 4369 ContiguousSpace* to_space = young_gen->to(); 4370 4371 HeapWord** eca = _collector->_eden_chunk_array; 4372 size_t ect = _collector->_eden_chunk_index; 4373 HeapWord** sca = _collector->_survivor_chunk_array; 4374 size_t sct = _collector->_survivor_chunk_index; 4375 4376 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4377 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4378 4379 do_young_space_rescan(cl, to_space, NULL, 0); 4380 do_young_space_rescan(cl, from_space, sca, sct); 4381 do_young_space_rescan(cl, eden_space, eca, ect); 4382 } 4383 4384 // work_queue(i) is passed to the closure 4385 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4386 // also is passed to do_dirty_card_rescan_tasks() and to 4387 // do_work_steal() to select the i-th task_queue. 4388 4389 void CMSParRemarkTask::work(uint worker_id) { 4390 elapsedTimer _timer; 4391 ResourceMark rm; 4392 HandleMark hm; 4393 4394 // ---------- rescan from roots -------------- 4395 _timer.start(); 4396 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4397 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4398 _collector->_span, _collector->ref_processor(), 4399 &(_collector->_markBitMap), 4400 work_queue(worker_id)); 4401 4402 // Rescan young gen roots first since these are likely 4403 // coarsely partitioned and may, on that account, constitute 4404 // the critical path; thus, it's best to start off that 4405 // work first. 4406 // ---------- young gen roots -------------- 4407 { 4408 work_on_young_gen_roots(&par_mrias_cl); 4409 _timer.stop(); 4410 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4411 } 4412 4413 // ---------- remaining roots -------------- 4414 _timer.reset(); 4415 _timer.start(); 4416 gch->old_process_roots(_strong_roots_scope, 4417 false, // yg was scanned above 4418 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4419 _collector->should_unload_classes(), 4420 &par_mrias_cl, 4421 NULL); // The dirty klasses will be handled below 4422 4423 assert(_collector->should_unload_classes() 4424 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4425 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4426 _timer.stop(); 4427 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4428 4429 // ---------- unhandled CLD scanning ---------- 4430 if (worker_id == 0) { // Single threaded at the moment. 4431 _timer.reset(); 4432 _timer.start(); 4433 4434 // Scan all new class loader data objects and new dependencies that were 4435 // introduced during concurrent marking. 4436 ResourceMark rm; 4437 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4438 for (int i = 0; i < array->length(); i++) { 4439 par_mrias_cl.do_cld_nv(array->at(i)); 4440 } 4441 4442 // We don't need to keep track of new CLDs anymore. 4443 ClassLoaderDataGraph::remember_new_clds(false); 4444 4445 _timer.stop(); 4446 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4447 } 4448 4449 // ---------- dirty klass scanning ---------- 4450 if (worker_id == 0) { // Single threaded at the moment. 4451 _timer.reset(); 4452 _timer.start(); 4453 4454 // Scan all classes that was dirtied during the concurrent marking phase. 4455 RemarkKlassClosure remark_klass_closure(&par_mrias_cl); 4456 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 4457 4458 _timer.stop(); 4459 log_trace(gc, task)("Finished dirty klass scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4460 } 4461 4462 // We might have added oops to ClassLoaderData::_handles during the 4463 // concurrent marking phase. These oops point to newly allocated objects 4464 // that are guaranteed to be kept alive. Either by the direct allocation 4465 // code, or when the young collector processes the roots. Hence, 4466 // we don't have to revisit the _handles block during the remark phase. 4467 4468 // ---------- rescan dirty cards ------------ 4469 _timer.reset(); 4470 _timer.start(); 4471 4472 // Do the rescan tasks for each of the two spaces 4473 // (cms_space) in turn. 4474 // "worker_id" is passed to select the task_queue for "worker_id" 4475 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4476 _timer.stop(); 4477 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4478 4479 // ---------- steal work from other threads ... 4480 // ---------- ... and drain overflow list. 4481 _timer.reset(); 4482 _timer.start(); 4483 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4484 _timer.stop(); 4485 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4486 } 4487 4488 void 4489 CMSParMarkTask::do_young_space_rescan( 4490 OopsInGenClosure* cl, ContiguousSpace* space, 4491 HeapWord** chunk_array, size_t chunk_top) { 4492 // Until all tasks completed: 4493 // . claim an unclaimed task 4494 // . compute region boundaries corresponding to task claimed 4495 // using chunk_array 4496 // . par_oop_iterate(cl) over that region 4497 4498 ResourceMark rm; 4499 HandleMark hm; 4500 4501 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4502 4503 uint nth_task = 0; 4504 uint n_tasks = pst->n_tasks(); 4505 4506 if (n_tasks > 0) { 4507 assert(pst->valid(), "Uninitialized use?"); 4508 HeapWord *start, *end; 4509 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4510 // We claimed task # nth_task; compute its boundaries. 4511 if (chunk_top == 0) { // no samples were taken 4512 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4513 start = space->bottom(); 4514 end = space->top(); 4515 } else if (nth_task == 0) { 4516 start = space->bottom(); 4517 end = chunk_array[nth_task]; 4518 } else if (nth_task < (uint)chunk_top) { 4519 assert(nth_task >= 1, "Control point invariant"); 4520 start = chunk_array[nth_task - 1]; 4521 end = chunk_array[nth_task]; 4522 } else { 4523 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4524 start = chunk_array[chunk_top - 1]; 4525 end = space->top(); 4526 } 4527 MemRegion mr(start, end); 4528 // Verify that mr is in space 4529 assert(mr.is_empty() || space->used_region().contains(mr), 4530 "Should be in space"); 4531 // Verify that "start" is an object boundary 4532 assert(mr.is_empty() || oop(mr.start())->is_oop(), 4533 "Should be an oop"); 4534 space->par_oop_iterate(mr, cl); 4535 } 4536 pst->all_tasks_completed(); 4537 } 4538 } 4539 4540 void 4541 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4542 CompactibleFreeListSpace* sp, int i, 4543 ParMarkRefsIntoAndScanClosure* cl) { 4544 // Until all tasks completed: 4545 // . claim an unclaimed task 4546 // . compute region boundaries corresponding to task claimed 4547 // . transfer dirty bits ct->mut for that region 4548 // . apply rescanclosure to dirty mut bits for that region 4549 4550 ResourceMark rm; 4551 HandleMark hm; 4552 4553 OopTaskQueue* work_q = work_queue(i); 4554 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4555 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4556 // CAUTION: This closure has state that persists across calls to 4557 // the work method dirty_range_iterate_clear() in that it has 4558 // embedded in it a (subtype of) UpwardsObjectClosure. The 4559 // use of that state in the embedded UpwardsObjectClosure instance 4560 // assumes that the cards are always iterated (even if in parallel 4561 // by several threads) in monotonically increasing order per each 4562 // thread. This is true of the implementation below which picks 4563 // card ranges (chunks) in monotonically increasing order globally 4564 // and, a-fortiori, in monotonically increasing order per thread 4565 // (the latter order being a subsequence of the former). 4566 // If the work code below is ever reorganized into a more chaotic 4567 // work-partitioning form than the current "sequential tasks" 4568 // paradigm, the use of that persistent state will have to be 4569 // revisited and modified appropriately. See also related 4570 // bug 4756801 work on which should examine this code to make 4571 // sure that the changes there do not run counter to the 4572 // assumptions made here and necessary for correctness and 4573 // efficiency. Note also that this code might yield inefficient 4574 // behavior in the case of very large objects that span one or 4575 // more work chunks. Such objects would potentially be scanned 4576 // several times redundantly. Work on 4756801 should try and 4577 // address that performance anomaly if at all possible. XXX 4578 MemRegion full_span = _collector->_span; 4579 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4580 MarkFromDirtyCardsClosure 4581 greyRescanClosure(_collector, full_span, // entire span of interest 4582 sp, bm, work_q, cl); 4583 4584 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4585 assert(pst->valid(), "Uninitialized use?"); 4586 uint nth_task = 0; 4587 const int alignment = CardTableModRefBS::card_size * BitsPerWord; 4588 MemRegion span = sp->used_region(); 4589 HeapWord* start_addr = span.start(); 4590 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), 4591 alignment); 4592 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4593 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == 4594 start_addr, "Check alignment"); 4595 assert((size_t)round_to((intptr_t)chunk_size, alignment) == 4596 chunk_size, "Check alignment"); 4597 4598 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4599 // Having claimed the nth_task, compute corresponding mem-region, 4600 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4601 // The alignment restriction ensures that we do not need any 4602 // synchronization with other gang-workers while setting or 4603 // clearing bits in thus chunk of the MUT. 4604 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4605 start_addr + (nth_task+1)*chunk_size); 4606 // The last chunk's end might be way beyond end of the 4607 // used region. In that case pull back appropriately. 4608 if (this_span.end() > end_addr) { 4609 this_span.set_end(end_addr); 4610 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4611 } 4612 // Iterate over the dirty cards covering this chunk, marking them 4613 // precleaned, and setting the corresponding bits in the mod union 4614 // table. Since we have been careful to partition at Card and MUT-word 4615 // boundaries no synchronization is needed between parallel threads. 4616 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, 4617 &modUnionClosure); 4618 4619 // Having transferred these marks into the modUnionTable, 4620 // rescan the marked objects on the dirty cards in the modUnionTable. 4621 // Even if this is at a synchronous collection, the initial marking 4622 // may have been done during an asynchronous collection so there 4623 // may be dirty bits in the mod-union table. 4624 _collector->_modUnionTable.dirty_range_iterate_clear( 4625 this_span, &greyRescanClosure); 4626 _collector->_modUnionTable.verifyNoOneBitsInRange( 4627 this_span.start(), 4628 this_span.end()); 4629 } 4630 pst->all_tasks_completed(); // declare that i am done 4631 } 4632 4633 // . see if we can share work_queues with ParNew? XXX 4634 void 4635 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4636 int* seed) { 4637 OopTaskQueue* work_q = work_queue(i); 4638 NOT_PRODUCT(int num_steals = 0;) 4639 oop obj_to_scan; 4640 CMSBitMap* bm = &(_collector->_markBitMap); 4641 4642 while (true) { 4643 // Completely finish any left over work from (an) earlier round(s) 4644 cl->trim_queue(0); 4645 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4646 (size_t)ParGCDesiredObjsFromOverflowList); 4647 // Now check if there's any work in the overflow list 4648 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4649 // only affects the number of attempts made to get work from the 4650 // overflow list and does not affect the number of workers. Just 4651 // pass ParallelGCThreads so this behavior is unchanged. 4652 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4653 work_q, 4654 ParallelGCThreads)) { 4655 // found something in global overflow list; 4656 // not yet ready to go stealing work from others. 4657 // We'd like to assert(work_q->size() != 0, ...) 4658 // because we just took work from the overflow list, 4659 // but of course we can't since all of that could have 4660 // been already stolen from us. 4661 // "He giveth and He taketh away." 4662 continue; 4663 } 4664 // Verify that we have no work before we resort to stealing 4665 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4666 // Try to steal from other queues that have work 4667 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4668 NOT_PRODUCT(num_steals++;) 4669 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 4670 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4671 // Do scanning work 4672 obj_to_scan->oop_iterate(cl); 4673 // Loop around, finish this work, and try to steal some more 4674 } else if (terminator()->offer_termination()) { 4675 break; // nirvana from the infinite cycle 4676 } 4677 } 4678 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4679 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4680 "Else our work is not yet done"); 4681 } 4682 4683 // Record object boundaries in _eden_chunk_array by sampling the eden 4684 // top in the slow-path eden object allocation code path and record 4685 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4686 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4687 // sampling in sample_eden() that activates during the part of the 4688 // preclean phase. 4689 void CMSCollector::sample_eden_chunk() { 4690 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4691 if (_eden_chunk_lock->try_lock()) { 4692 // Record a sample. This is the critical section. The contents 4693 // of the _eden_chunk_array have to be non-decreasing in the 4694 // address order. 4695 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4696 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4697 "Unexpected state of Eden"); 4698 if (_eden_chunk_index == 0 || 4699 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4700 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4701 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4702 _eden_chunk_index++; // commit sample 4703 } 4704 _eden_chunk_lock->unlock(); 4705 } 4706 } 4707 } 4708 4709 // Return a thread-local PLAB recording array, as appropriate. 4710 void* CMSCollector::get_data_recorder(int thr_num) { 4711 if (_survivor_plab_array != NULL && 4712 (CMSPLABRecordAlways || 4713 (_collectorState > Marking && _collectorState < FinalMarking))) { 4714 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4715 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4716 ca->reset(); // clear it so that fresh data is recorded 4717 return (void*) ca; 4718 } else { 4719 return NULL; 4720 } 4721 } 4722 4723 // Reset all the thread-local PLAB recording arrays 4724 void CMSCollector::reset_survivor_plab_arrays() { 4725 for (uint i = 0; i < ParallelGCThreads; i++) { 4726 _survivor_plab_array[i].reset(); 4727 } 4728 } 4729 4730 // Merge the per-thread plab arrays into the global survivor chunk 4731 // array which will provide the partitioning of the survivor space 4732 // for CMS initial scan and rescan. 4733 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4734 int no_of_gc_threads) { 4735 assert(_survivor_plab_array != NULL, "Error"); 4736 assert(_survivor_chunk_array != NULL, "Error"); 4737 assert(_collectorState == FinalMarking || 4738 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4739 for (int j = 0; j < no_of_gc_threads; j++) { 4740 _cursor[j] = 0; 4741 } 4742 HeapWord* top = surv->top(); 4743 size_t i; 4744 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4745 HeapWord* min_val = top; // Higher than any PLAB address 4746 uint min_tid = 0; // position of min_val this round 4747 for (int j = 0; j < no_of_gc_threads; j++) { 4748 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4749 if (_cursor[j] == cur_sca->end()) { 4750 continue; 4751 } 4752 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4753 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4754 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4755 if (cur_val < min_val) { 4756 min_tid = j; 4757 min_val = cur_val; 4758 } else { 4759 assert(cur_val < top, "All recorded addresses should be less"); 4760 } 4761 } 4762 // At this point min_val and min_tid are respectively 4763 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4764 // and the thread (j) that witnesses that address. 4765 // We record this address in the _survivor_chunk_array[i] 4766 // and increment _cursor[min_tid] prior to the next round i. 4767 if (min_val == top) { 4768 break; 4769 } 4770 _survivor_chunk_array[i] = min_val; 4771 _cursor[min_tid]++; 4772 } 4773 // We are all done; record the size of the _survivor_chunk_array 4774 _survivor_chunk_index = i; // exclusive: [0, i) 4775 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4776 // Verify that we used up all the recorded entries 4777 #ifdef ASSERT 4778 size_t total = 0; 4779 for (int j = 0; j < no_of_gc_threads; j++) { 4780 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4781 total += _cursor[j]; 4782 } 4783 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4784 // Check that the merged array is in sorted order 4785 if (total > 0) { 4786 for (size_t i = 0; i < total - 1; i++) { 4787 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4788 i, p2i(_survivor_chunk_array[i])); 4789 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4790 "Not sorted"); 4791 } 4792 } 4793 #endif // ASSERT 4794 } 4795 4796 // Set up the space's par_seq_tasks structure for work claiming 4797 // for parallel initial scan and rescan of young gen. 4798 // See ParRescanTask where this is currently used. 4799 void 4800 CMSCollector:: 4801 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4802 assert(n_threads > 0, "Unexpected n_threads argument"); 4803 4804 // Eden space 4805 if (!_young_gen->eden()->is_empty()) { 4806 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4807 assert(!pst->valid(), "Clobbering existing data?"); 4808 // Each valid entry in [0, _eden_chunk_index) represents a task. 4809 size_t n_tasks = _eden_chunk_index + 1; 4810 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4811 // Sets the condition for completion of the subtask (how many threads 4812 // need to finish in order to be done). 4813 pst->set_n_threads(n_threads); 4814 pst->set_n_tasks((int)n_tasks); 4815 } 4816 4817 // Merge the survivor plab arrays into _survivor_chunk_array 4818 if (_survivor_plab_array != NULL) { 4819 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4820 } else { 4821 assert(_survivor_chunk_index == 0, "Error"); 4822 } 4823 4824 // To space 4825 { 4826 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4827 assert(!pst->valid(), "Clobbering existing data?"); 4828 // Sets the condition for completion of the subtask (how many threads 4829 // need to finish in order to be done). 4830 pst->set_n_threads(n_threads); 4831 pst->set_n_tasks(1); 4832 assert(pst->valid(), "Error"); 4833 } 4834 4835 // From space 4836 { 4837 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4838 assert(!pst->valid(), "Clobbering existing data?"); 4839 size_t n_tasks = _survivor_chunk_index + 1; 4840 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4841 // Sets the condition for completion of the subtask (how many threads 4842 // need to finish in order to be done). 4843 pst->set_n_threads(n_threads); 4844 pst->set_n_tasks((int)n_tasks); 4845 assert(pst->valid(), "Error"); 4846 } 4847 } 4848 4849 // Parallel version of remark 4850 void CMSCollector::do_remark_parallel() { 4851 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4852 WorkGang* workers = gch->workers(); 4853 assert(workers != NULL, "Need parallel worker threads."); 4854 // Choose to use the number of GC workers most recently set 4855 // into "active_workers". 4856 uint n_workers = workers->active_workers(); 4857 4858 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4859 4860 StrongRootsScope srs(n_workers); 4861 4862 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4863 4864 // We won't be iterating over the cards in the card table updating 4865 // the younger_gen cards, so we shouldn't call the following else 4866 // the verification code as well as subsequent younger_refs_iterate 4867 // code would get confused. XXX 4868 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4869 4870 // The young gen rescan work will not be done as part of 4871 // process_roots (which currently doesn't know how to 4872 // parallelize such a scan), but rather will be broken up into 4873 // a set of parallel tasks (via the sampling that the [abortable] 4874 // preclean phase did of eden, plus the [two] tasks of 4875 // scanning the [two] survivor spaces. Further fine-grain 4876 // parallelization of the scanning of the survivor spaces 4877 // themselves, and of precleaning of the young gen itself 4878 // is deferred to the future. 4879 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4880 4881 // The dirty card rescan work is broken up into a "sequence" 4882 // of parallel tasks (per constituent space) that are dynamically 4883 // claimed by the parallel threads. 4884 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4885 4886 // It turns out that even when we're using 1 thread, doing the work in a 4887 // separate thread causes wide variance in run times. We can't help this 4888 // in the multi-threaded case, but we special-case n=1 here to get 4889 // repeatable measurements of the 1-thread overhead of the parallel code. 4890 if (n_workers > 1) { 4891 // Make refs discovery MT-safe, if it isn't already: it may not 4892 // necessarily be so, since it's possible that we are doing 4893 // ST marking. 4894 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4895 workers->run_task(&tsk); 4896 } else { 4897 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4898 tsk.work(0); 4899 } 4900 4901 // restore, single-threaded for now, any preserved marks 4902 // as a result of work_q overflow 4903 restore_preserved_marks_if_any(); 4904 } 4905 4906 // Non-parallel version of remark 4907 void CMSCollector::do_remark_non_parallel() { 4908 ResourceMark rm; 4909 HandleMark hm; 4910 GenCollectedHeap* gch = GenCollectedHeap::heap(); 4911 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4912 4913 MarkRefsIntoAndScanClosure 4914 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4915 &_markStack, this, 4916 false /* should_yield */, false /* not precleaning */); 4917 MarkFromDirtyCardsClosure 4918 markFromDirtyCardsClosure(this, _span, 4919 NULL, // space is set further below 4920 &_markBitMap, &_markStack, &mrias_cl); 4921 { 4922 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4923 // Iterate over the dirty cards, setting the corresponding bits in the 4924 // mod union table. 4925 { 4926 ModUnionClosure modUnionClosure(&_modUnionTable); 4927 _ct->ct_bs()->dirty_card_iterate( 4928 _cmsGen->used_region(), 4929 &modUnionClosure); 4930 } 4931 // Having transferred these marks into the modUnionTable, we just need 4932 // to rescan the marked objects on the dirty cards in the modUnionTable. 4933 // The initial marking may have been done during an asynchronous 4934 // collection so there may be dirty bits in the mod-union table. 4935 const int alignment = 4936 CardTableModRefBS::card_size * BitsPerWord; 4937 { 4938 // ... First handle dirty cards in CMS gen 4939 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4940 MemRegion ur = _cmsGen->used_region(); 4941 HeapWord* lb = ur.start(); 4942 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); 4943 MemRegion cms_span(lb, ub); 4944 _modUnionTable.dirty_range_iterate_clear(cms_span, 4945 &markFromDirtyCardsClosure); 4946 verify_work_stacks_empty(); 4947 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4948 } 4949 } 4950 if (VerifyDuringGC && 4951 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { 4952 HandleMark hm; // Discard invalid handles created during verification 4953 Universe::verify(); 4954 } 4955 { 4956 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4957 4958 verify_work_stacks_empty(); 4959 4960 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4961 StrongRootsScope srs(1); 4962 4963 gch->old_process_roots(&srs, 4964 true, // young gen as roots 4965 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4966 should_unload_classes(), 4967 &mrias_cl, 4968 NULL); // The dirty klasses will be handled below 4969 4970 assert(should_unload_classes() 4971 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4972 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4973 } 4974 4975 { 4976 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4977 4978 verify_work_stacks_empty(); 4979 4980 // Scan all class loader data objects that might have been introduced 4981 // during concurrent marking. 4982 ResourceMark rm; 4983 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4984 for (int i = 0; i < array->length(); i++) { 4985 mrias_cl.do_cld_nv(array->at(i)); 4986 } 4987 4988 // We don't need to keep track of new CLDs anymore. 4989 ClassLoaderDataGraph::remember_new_clds(false); 4990 4991 verify_work_stacks_empty(); 4992 } 4993 4994 { 4995 GCTraceTime(Trace, gc, phases) t("Dirty Klass Scan", _gc_timer_cm); 4996 4997 verify_work_stacks_empty(); 4998 4999 RemarkKlassClosure remark_klass_closure(&mrias_cl); 5000 ClassLoaderDataGraph::classes_do(&remark_klass_closure); 5001 5002 verify_work_stacks_empty(); 5003 } 5004 5005 // We might have added oops to ClassLoaderData::_handles during the 5006 // concurrent marking phase. These oops point to newly allocated objects 5007 // that are guaranteed to be kept alive. Either by the direct allocation 5008 // code, or when the young collector processes the roots. Hence, 5009 // we don't have to revisit the _handles block during the remark phase. 5010 5011 verify_work_stacks_empty(); 5012 // Restore evacuated mark words, if any, used for overflow list links 5013 restore_preserved_marks_if_any(); 5014 5015 verify_overflow_empty(); 5016 } 5017 5018 //////////////////////////////////////////////////////// 5019 // Parallel Reference Processing Task Proxy Class 5020 //////////////////////////////////////////////////////// 5021 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5022 OopTaskQueueSet* _queues; 5023 ParallelTaskTerminator _terminator; 5024 public: 5025 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5026 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5027 ParallelTaskTerminator* terminator() { return &_terminator; } 5028 OopTaskQueueSet* queues() { return _queues; } 5029 }; 5030 5031 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5032 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5033 CMSCollector* _collector; 5034 CMSBitMap* _mark_bit_map; 5035 const MemRegion _span; 5036 ProcessTask& _task; 5037 5038 public: 5039 CMSRefProcTaskProxy(ProcessTask& task, 5040 CMSCollector* collector, 5041 const MemRegion& span, 5042 CMSBitMap* mark_bit_map, 5043 AbstractWorkGang* workers, 5044 OopTaskQueueSet* task_queues): 5045 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5046 task_queues, 5047 workers->active_workers()), 5048 _task(task), 5049 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5050 { 5051 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5052 "Inconsistency in _span"); 5053 } 5054 5055 OopTaskQueueSet* task_queues() { return queues(); } 5056 5057 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5058 5059 void do_work_steal(int i, 5060 CMSParDrainMarkingStackClosure* drain, 5061 CMSParKeepAliveClosure* keep_alive, 5062 int* seed); 5063 5064 virtual void work(uint worker_id); 5065 }; 5066 5067 void CMSRefProcTaskProxy::work(uint worker_id) { 5068 ResourceMark rm; 5069 HandleMark hm; 5070 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5071 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5072 _mark_bit_map, 5073 work_queue(worker_id)); 5074 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5075 _mark_bit_map, 5076 work_queue(worker_id)); 5077 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5078 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5079 if (_task.marks_oops_alive()) { 5080 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5081 _collector->hash_seed(worker_id)); 5082 } 5083 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5084 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5085 } 5086 5087 class CMSRefEnqueueTaskProxy: public AbstractGangTask { 5088 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5089 EnqueueTask& _task; 5090 5091 public: 5092 CMSRefEnqueueTaskProxy(EnqueueTask& task) 5093 : AbstractGangTask("Enqueue reference objects in parallel"), 5094 _task(task) 5095 { } 5096 5097 virtual void work(uint worker_id) 5098 { 5099 _task.work(worker_id); 5100 } 5101 }; 5102 5103 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5104 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5105 _span(span), 5106 _bit_map(bit_map), 5107 _work_queue(work_queue), 5108 _mark_and_push(collector, span, bit_map, work_queue), 5109 _low_water_mark(MIN2((work_queue->max_elems()/4), 5110 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5111 { } 5112 5113 // . see if we can share work_queues with ParNew? XXX 5114 void CMSRefProcTaskProxy::do_work_steal(int i, 5115 CMSParDrainMarkingStackClosure* drain, 5116 CMSParKeepAliveClosure* keep_alive, 5117 int* seed) { 5118 OopTaskQueue* work_q = work_queue(i); 5119 NOT_PRODUCT(int num_steals = 0;) 5120 oop obj_to_scan; 5121 5122 while (true) { 5123 // Completely finish any left over work from (an) earlier round(s) 5124 drain->trim_queue(0); 5125 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5126 (size_t)ParGCDesiredObjsFromOverflowList); 5127 // Now check if there's any work in the overflow list 5128 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5129 // only affects the number of attempts made to get work from the 5130 // overflow list and does not affect the number of workers. Just 5131 // pass ParallelGCThreads so this behavior is unchanged. 5132 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5133 work_q, 5134 ParallelGCThreads)) { 5135 // Found something in global overflow list; 5136 // not yet ready to go stealing work from others. 5137 // We'd like to assert(work_q->size() != 0, ...) 5138 // because we just took work from the overflow list, 5139 // but of course we can't, since all of that might have 5140 // been already stolen from us. 5141 continue; 5142 } 5143 // Verify that we have no work before we resort to stealing 5144 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5145 // Try to steal from other queues that have work 5146 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5147 NOT_PRODUCT(num_steals++;) 5148 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); 5149 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5150 // Do scanning work 5151 obj_to_scan->oop_iterate(keep_alive); 5152 // Loop around, finish this work, and try to steal some more 5153 } else if (terminator()->offer_termination()) { 5154 break; // nirvana from the infinite cycle 5155 } 5156 } 5157 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5158 } 5159 5160 void CMSRefProcTaskExecutor::execute(ProcessTask& task) 5161 { 5162 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5163 WorkGang* workers = gch->workers(); 5164 assert(workers != NULL, "Need parallel worker threads."); 5165 CMSRefProcTaskProxy rp_task(task, &_collector, 5166 _collector.ref_processor()->span(), 5167 _collector.markBitMap(), 5168 workers, _collector.task_queues()); 5169 workers->run_task(&rp_task); 5170 } 5171 5172 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) 5173 { 5174 5175 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5176 WorkGang* workers = gch->workers(); 5177 assert(workers != NULL, "Need parallel worker threads."); 5178 CMSRefEnqueueTaskProxy enq_task(task); 5179 workers->run_task(&enq_task); 5180 } 5181 5182 void CMSCollector::refProcessingWork() { 5183 ResourceMark rm; 5184 HandleMark hm; 5185 5186 ReferenceProcessor* rp = ref_processor(); 5187 assert(rp->span().equals(_span), "Spans should be equal"); 5188 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5189 // Process weak references. 5190 rp->setup_policy(false); 5191 verify_work_stacks_empty(); 5192 5193 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5194 &_markStack, false /* !preclean */); 5195 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5196 _span, &_markBitMap, &_markStack, 5197 &cmsKeepAliveClosure, false /* !preclean */); 5198 { 5199 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5200 5201 ReferenceProcessorStats stats; 5202 if (rp->processing_is_mt()) { 5203 // Set the degree of MT here. If the discovery is done MT, there 5204 // may have been a different number of threads doing the discovery 5205 // and a different number of discovered lists may have Ref objects. 5206 // That is OK as long as the Reference lists are balanced (see 5207 // balance_all_queues() and balance_queues()). 5208 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5209 uint active_workers = ParallelGCThreads; 5210 WorkGang* workers = gch->workers(); 5211 if (workers != NULL) { 5212 active_workers = workers->active_workers(); 5213 // The expectation is that active_workers will have already 5214 // been set to a reasonable value. If it has not been set, 5215 // investigate. 5216 assert(active_workers > 0, "Should have been set during scavenge"); 5217 } 5218 rp->set_active_mt_degree(active_workers); 5219 CMSRefProcTaskExecutor task_executor(*this); 5220 stats = rp->process_discovered_references(&_is_alive_closure, 5221 &cmsKeepAliveClosure, 5222 &cmsDrainMarkingStackClosure, 5223 &task_executor, 5224 _gc_timer_cm); 5225 } else { 5226 stats = rp->process_discovered_references(&_is_alive_closure, 5227 &cmsKeepAliveClosure, 5228 &cmsDrainMarkingStackClosure, 5229 NULL, 5230 _gc_timer_cm); 5231 } 5232 _gc_tracer_cm->report_gc_reference_stats(stats); 5233 5234 } 5235 5236 // This is the point where the entire marking should have completed. 5237 verify_work_stacks_empty(); 5238 5239 if (should_unload_classes()) { 5240 { 5241 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5242 5243 // Unload classes and purge the SystemDictionary. 5244 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); 5245 5246 // Unload nmethods. 5247 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5248 5249 // Prune dead klasses from subklass/sibling/implementor lists. 5250 Klass::clean_weak_klass_links(&_is_alive_closure); 5251 } 5252 5253 { 5254 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5255 // Clean up unreferenced symbols in symbol table. 5256 SymbolTable::unlink(); 5257 } 5258 5259 { 5260 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5261 // Delete entries for dead interned strings. 5262 StringTable::unlink(&_is_alive_closure); 5263 } 5264 } 5265 5266 5267 // Restore any preserved marks as a result of mark stack or 5268 // work queue overflow 5269 restore_preserved_marks_if_any(); // done single-threaded for now 5270 5271 rp->set_enqueuing_is_done(true); 5272 if (rp->processing_is_mt()) { 5273 rp->balance_all_queues(); 5274 CMSRefProcTaskExecutor task_executor(*this); 5275 rp->enqueue_discovered_references(&task_executor); 5276 } else { 5277 rp->enqueue_discovered_references(NULL); 5278 } 5279 rp->verify_no_references_recorded(); 5280 assert(!rp->discovery_enabled(), "should have been disabled"); 5281 } 5282 5283 #ifndef PRODUCT 5284 void CMSCollector::check_correct_thread_executing() { 5285 Thread* t = Thread::current(); 5286 // Only the VM thread or the CMS thread should be here. 5287 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5288 "Unexpected thread type"); 5289 // If this is the vm thread, the foreground process 5290 // should not be waiting. Note that _foregroundGCIsActive is 5291 // true while the foreground collector is waiting. 5292 if (_foregroundGCShouldWait) { 5293 // We cannot be the VM thread 5294 assert(t->is_ConcurrentGC_thread(), 5295 "Should be CMS thread"); 5296 } else { 5297 // We can be the CMS thread only if we are in a stop-world 5298 // phase of CMS collection. 5299 if (t->is_ConcurrentGC_thread()) { 5300 assert(_collectorState == InitialMarking || 5301 _collectorState == FinalMarking, 5302 "Should be a stop-world phase"); 5303 // The CMS thread should be holding the CMS_token. 5304 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5305 "Potential interference with concurrently " 5306 "executing VM thread"); 5307 } 5308 } 5309 } 5310 #endif 5311 5312 void CMSCollector::sweep() { 5313 assert(_collectorState == Sweeping, "just checking"); 5314 check_correct_thread_executing(); 5315 verify_work_stacks_empty(); 5316 verify_overflow_empty(); 5317 increment_sweep_count(); 5318 TraceCMSMemoryManagerStats tms(_collectorState,GenCollectedHeap::heap()->gc_cause()); 5319 5320 _inter_sweep_timer.stop(); 5321 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5322 5323 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5324 _intra_sweep_timer.reset(); 5325 _intra_sweep_timer.start(); 5326 { 5327 GCTraceCPUTime tcpu; 5328 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5329 // First sweep the old gen 5330 { 5331 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5332 bitMapLock()); 5333 sweepWork(_cmsGen); 5334 } 5335 5336 // Update Universe::_heap_*_at_gc figures. 5337 // We need all the free list locks to make the abstract state 5338 // transition from Sweeping to Resetting. See detailed note 5339 // further below. 5340 { 5341 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5342 // Update heap occupancy information which is used as 5343 // input to soft ref clearing policy at the next gc. 5344 Universe::update_heap_info_at_gc(); 5345 _collectorState = Resizing; 5346 } 5347 } 5348 verify_work_stacks_empty(); 5349 verify_overflow_empty(); 5350 5351 if (should_unload_classes()) { 5352 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5353 // requires that the virtual spaces are stable and not deleted. 5354 ClassLoaderDataGraph::set_should_purge(true); 5355 } 5356 5357 _intra_sweep_timer.stop(); 5358 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5359 5360 _inter_sweep_timer.reset(); 5361 _inter_sweep_timer.start(); 5362 5363 // We need to use a monotonically non-decreasing time in ms 5364 // or we will see time-warp warnings and os::javaTimeMillis() 5365 // does not guarantee monotonicity. 5366 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5367 update_time_of_last_gc(now); 5368 5369 // NOTE on abstract state transitions: 5370 // Mutators allocate-live and/or mark the mod-union table dirty 5371 // based on the state of the collection. The former is done in 5372 // the interval [Marking, Sweeping] and the latter in the interval 5373 // [Marking, Sweeping). Thus the transitions into the Marking state 5374 // and out of the Sweeping state must be synchronously visible 5375 // globally to the mutators. 5376 // The transition into the Marking state happens with the world 5377 // stopped so the mutators will globally see it. Sweeping is 5378 // done asynchronously by the background collector so the transition 5379 // from the Sweeping state to the Resizing state must be done 5380 // under the freelistLock (as is the check for whether to 5381 // allocate-live and whether to dirty the mod-union table). 5382 assert(_collectorState == Resizing, "Change of collector state to" 5383 " Resizing must be done under the freelistLocks (plural)"); 5384 5385 // Now that sweeping has been completed, we clear 5386 // the incremental_collection_failed flag, 5387 // thus inviting a younger gen collection to promote into 5388 // this generation. If such a promotion may still fail, 5389 // the flag will be set again when a young collection is 5390 // attempted. 5391 GenCollectedHeap* gch = GenCollectedHeap::heap(); 5392 gch->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5393 gch->update_full_collections_completed(_collection_count_start); 5394 } 5395 5396 // FIX ME!!! Looks like this belongs in CFLSpace, with 5397 // CMSGen merely delegating to it. 5398 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5399 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5400 HeapWord* minAddr = _cmsSpace->bottom(); 5401 HeapWord* largestAddr = 5402 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5403 if (largestAddr == NULL) { 5404 // The dictionary appears to be empty. In this case 5405 // try to coalesce at the end of the heap. 5406 largestAddr = _cmsSpace->end(); 5407 } 5408 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5409 size_t nearLargestOffset = 5410 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5411 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5412 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5413 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5414 } 5415 5416 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5417 return addr >= _cmsSpace->nearLargestChunk(); 5418 } 5419 5420 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5421 return _cmsSpace->find_chunk_at_end(); 5422 } 5423 5424 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5425 bool full) { 5426 // If the young generation has been collected, gather any statistics 5427 // that are of interest at this point. 5428 bool current_is_young = GenCollectedHeap::heap()->is_young_gen(current_generation); 5429 if (!full && current_is_young) { 5430 // Gather statistics on the young generation collection. 5431 collector()->stats().record_gc0_end(used()); 5432 } 5433 } 5434 5435 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5436 // We iterate over the space(s) underlying this generation, 5437 // checking the mark bit map to see if the bits corresponding 5438 // to specific blocks are marked or not. Blocks that are 5439 // marked are live and are not swept up. All remaining blocks 5440 // are swept up, with coalescing on-the-fly as we sweep up 5441 // contiguous free and/or garbage blocks: 5442 // We need to ensure that the sweeper synchronizes with allocators 5443 // and stop-the-world collectors. In particular, the following 5444 // locks are used: 5445 // . CMS token: if this is held, a stop the world collection cannot occur 5446 // . freelistLock: if this is held no allocation can occur from this 5447 // generation by another thread 5448 // . bitMapLock: if this is held, no other thread can access or update 5449 // 5450 5451 // Note that we need to hold the freelistLock if we use 5452 // block iterate below; else the iterator might go awry if 5453 // a mutator (or promotion) causes block contents to change 5454 // (for instance if the allocator divvies up a block). 5455 // If we hold the free list lock, for all practical purposes 5456 // young generation GC's can't occur (they'll usually need to 5457 // promote), so we might as well prevent all young generation 5458 // GC's while we do a sweeping step. For the same reason, we might 5459 // as well take the bit map lock for the entire duration 5460 5461 // check that we hold the requisite locks 5462 assert(have_cms_token(), "Should hold cms token"); 5463 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5464 assert_lock_strong(old_gen->freelistLock()); 5465 assert_lock_strong(bitMapLock()); 5466 5467 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5468 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5469 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5470 _inter_sweep_estimate.padded_average(), 5471 _intra_sweep_estimate.padded_average()); 5472 old_gen->setNearLargestChunk(); 5473 5474 { 5475 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5476 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5477 // We need to free-up/coalesce garbage/blocks from a 5478 // co-terminal free run. This is done in the SweepClosure 5479 // destructor; so, do not remove this scope, else the 5480 // end-of-sweep-census below will be off by a little bit. 5481 } 5482 old_gen->cmsSpace()->sweep_completed(); 5483 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5484 if (should_unload_classes()) { // unloaded classes this cycle, 5485 _concurrent_cycles_since_last_unload = 0; // ... reset count 5486 } else { // did not unload classes, 5487 _concurrent_cycles_since_last_unload++; // ... increment count 5488 } 5489 } 5490 5491 // Reset CMS data structures (for now just the marking bit map) 5492 // preparatory for the next cycle. 5493 void CMSCollector::reset_concurrent() { 5494 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5495 5496 // If the state is not "Resetting", the foreground thread 5497 // has done a collection and the resetting. 5498 if (_collectorState != Resetting) { 5499 assert(_collectorState == Idling, "The state should only change" 5500 " because the foreground collector has finished the collection"); 5501 return; 5502 } 5503 5504 { 5505 // Clear the mark bitmap (no grey objects to start with) 5506 // for the next cycle. 5507 GCTraceCPUTime tcpu; 5508 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5509 5510 HeapWord* curAddr = _markBitMap.startWord(); 5511 while (curAddr < _markBitMap.endWord()) { 5512 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5513 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5514 _markBitMap.clear_large_range(chunk); 5515 if (ConcurrentMarkSweepThread::should_yield() && 5516 !foregroundGCIsActive() && 5517 CMSYield) { 5518 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5519 "CMS thread should hold CMS token"); 5520 assert_lock_strong(bitMapLock()); 5521 bitMapLock()->unlock(); 5522 ConcurrentMarkSweepThread::desynchronize(true); 5523 stopTimer(); 5524 incrementYields(); 5525 5526 // See the comment in coordinator_yield() 5527 for (unsigned i = 0; i < CMSYieldSleepCount && 5528 ConcurrentMarkSweepThread::should_yield() && 5529 !CMSCollector::foregroundGCIsActive(); ++i) { 5530 os::sleep(Thread::current(), 1, false); 5531 } 5532 5533 ConcurrentMarkSweepThread::synchronize(true); 5534 bitMapLock()->lock_without_safepoint_check(); 5535 startTimer(); 5536 } 5537 curAddr = chunk.end(); 5538 } 5539 // A successful mostly concurrent collection has been done. 5540 // Because only the full (i.e., concurrent mode failure) collections 5541 // are being measured for gc overhead limits, clean the "near" flag 5542 // and count. 5543 size_policy()->reset_gc_overhead_limit_count(); 5544 _collectorState = Idling; 5545 } 5546 5547 register_gc_end(); 5548 } 5549 5550 // Same as above but for STW paths 5551 void CMSCollector::reset_stw() { 5552 // already have the lock 5553 assert(_collectorState == Resetting, "just checking"); 5554 assert_lock_strong(bitMapLock()); 5555 GCIdMarkAndRestore gc_id_mark(_cmsThread->gc_id()); 5556 _markBitMap.clear_all(); 5557 _collectorState = Idling; 5558 register_gc_end(); 5559 } 5560 5561 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5562 GCTraceCPUTime tcpu; 5563 TraceCollectorStats tcs(counters()); 5564 5565 switch (op) { 5566 case CMS_op_checkpointRootsInitial: { 5567 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5568 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5569 checkpointRootsInitial(); 5570 break; 5571 } 5572 case CMS_op_checkpointRootsFinal: { 5573 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5574 SvcGCMarker sgcm(SvcGCMarker::OTHER); 5575 checkpointRootsFinal(); 5576 break; 5577 } 5578 default: 5579 fatal("No such CMS_op"); 5580 } 5581 } 5582 5583 #ifndef PRODUCT 5584 size_t const CMSCollector::skip_header_HeapWords() { 5585 return FreeChunk::header_size(); 5586 } 5587 5588 // Try and collect here conditions that should hold when 5589 // CMS thread is exiting. The idea is that the foreground GC 5590 // thread should not be blocked if it wants to terminate 5591 // the CMS thread and yet continue to run the VM for a while 5592 // after that. 5593 void CMSCollector::verify_ok_to_terminate() const { 5594 assert(Thread::current()->is_ConcurrentGC_thread(), 5595 "should be called by CMS thread"); 5596 assert(!_foregroundGCShouldWait, "should be false"); 5597 // We could check here that all the various low-level locks 5598 // are not held by the CMS thread, but that is overkill; see 5599 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5600 // is checked. 5601 } 5602 #endif 5603 5604 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5605 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5606 "missing Printezis mark?"); 5607 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5608 size_t size = pointer_delta(nextOneAddr + 1, addr); 5609 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5610 "alignment problem"); 5611 assert(size >= 3, "Necessary for Printezis marks to work"); 5612 return size; 5613 } 5614 5615 // A variant of the above (block_size_using_printezis_bits()) except 5616 // that we return 0 if the P-bits are not yet set. 5617 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5618 if (_markBitMap.isMarked(addr + 1)) { 5619 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5620 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5621 size_t size = pointer_delta(nextOneAddr + 1, addr); 5622 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5623 "alignment problem"); 5624 assert(size >= 3, "Necessary for Printezis marks to work"); 5625 return size; 5626 } 5627 return 0; 5628 } 5629 5630 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5631 size_t sz = 0; 5632 oop p = (oop)addr; 5633 if (p->klass_or_null() != NULL) { 5634 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5635 } else { 5636 sz = block_size_using_printezis_bits(addr); 5637 } 5638 assert(sz > 0, "size must be nonzero"); 5639 HeapWord* next_block = addr + sz; 5640 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, 5641 CardTableModRefBS::card_size); 5642 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < 5643 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), 5644 "must be different cards"); 5645 return next_card; 5646 } 5647 5648 5649 // CMS Bit Map Wrapper ///////////////////////////////////////// 5650 5651 // Construct a CMS bit map infrastructure, but don't create the 5652 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5653 // further below. 5654 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5655 _bm(), 5656 _shifter(shifter), 5657 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5658 Monitor::_safepoint_check_sometimes) : NULL) 5659 { 5660 _bmStartWord = 0; 5661 _bmWordSize = 0; 5662 } 5663 5664 bool CMSBitMap::allocate(MemRegion mr) { 5665 _bmStartWord = mr.start(); 5666 _bmWordSize = mr.word_size(); 5667 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5668 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5669 if (!brs.is_reserved()) { 5670 log_warning(gc)("CMS bit map allocation failure"); 5671 return false; 5672 } 5673 // For now we'll just commit all of the bit map up front. 5674 // Later on we'll try to be more parsimonious with swap. 5675 if (!_virtual_space.initialize(brs, brs.size())) { 5676 log_warning(gc)("CMS bit map backing store failure"); 5677 return false; 5678 } 5679 assert(_virtual_space.committed_size() == brs.size(), 5680 "didn't reserve backing store for all of CMS bit map?"); 5681 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5682 _bmWordSize, "inconsistency in bit map sizing"); 5683 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5684 5685 // bm.clear(); // can we rely on getting zero'd memory? verify below 5686 assert(isAllClear(), 5687 "Expected zero'd memory from ReservedSpace constructor"); 5688 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5689 "consistency check"); 5690 return true; 5691 } 5692 5693 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5694 HeapWord *next_addr, *end_addr, *last_addr; 5695 assert_locked(); 5696 assert(covers(mr), "out-of-range error"); 5697 // XXX assert that start and end are appropriately aligned 5698 for (next_addr = mr.start(), end_addr = mr.end(); 5699 next_addr < end_addr; next_addr = last_addr) { 5700 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5701 last_addr = dirty_region.end(); 5702 if (!dirty_region.is_empty()) { 5703 cl->do_MemRegion(dirty_region); 5704 } else { 5705 assert(last_addr == end_addr, "program logic"); 5706 return; 5707 } 5708 } 5709 } 5710 5711 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5712 _bm.print_on_error(st, prefix); 5713 } 5714 5715 #ifndef PRODUCT 5716 void CMSBitMap::assert_locked() const { 5717 CMSLockVerifier::assert_locked(lock()); 5718 } 5719 5720 bool CMSBitMap::covers(MemRegion mr) const { 5721 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5722 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5723 "size inconsistency"); 5724 return (mr.start() >= _bmStartWord) && 5725 (mr.end() <= endWord()); 5726 } 5727 5728 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5729 return (start >= _bmStartWord && (start + size) <= endWord()); 5730 } 5731 5732 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5733 // verify that there are no 1 bits in the interval [left, right) 5734 FalseBitMapClosure falseBitMapClosure; 5735 iterate(&falseBitMapClosure, left, right); 5736 } 5737 5738 void CMSBitMap::region_invariant(MemRegion mr) 5739 { 5740 assert_locked(); 5741 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5742 assert(!mr.is_empty(), "unexpected empty region"); 5743 assert(covers(mr), "mr should be covered by bit map"); 5744 // convert address range into offset range 5745 size_t start_ofs = heapWordToOffset(mr.start()); 5746 // Make sure that end() is appropriately aligned 5747 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), 5748 (1 << (_shifter+LogHeapWordSize))), 5749 "Misaligned mr.end()"); 5750 size_t end_ofs = heapWordToOffset(mr.end()); 5751 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5752 } 5753 5754 #endif 5755 5756 bool CMSMarkStack::allocate(size_t size) { 5757 // allocate a stack of the requisite depth 5758 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5759 size * sizeof(oop))); 5760 if (!rs.is_reserved()) { 5761 log_warning(gc)("CMSMarkStack allocation failure"); 5762 return false; 5763 } 5764 if (!_virtual_space.initialize(rs, rs.size())) { 5765 log_warning(gc)("CMSMarkStack backing store failure"); 5766 return false; 5767 } 5768 assert(_virtual_space.committed_size() == rs.size(), 5769 "didn't reserve backing store for all of CMS stack?"); 5770 _base = (oop*)(_virtual_space.low()); 5771 _index = 0; 5772 _capacity = size; 5773 NOT_PRODUCT(_max_depth = 0); 5774 return true; 5775 } 5776 5777 // XXX FIX ME !!! In the MT case we come in here holding a 5778 // leaf lock. For printing we need to take a further lock 5779 // which has lower rank. We need to recalibrate the two 5780 // lock-ranks involved in order to be able to print the 5781 // messages below. (Or defer the printing to the caller. 5782 // For now we take the expedient path of just disabling the 5783 // messages for the problematic case.) 5784 void CMSMarkStack::expand() { 5785 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5786 if (_capacity == MarkStackSizeMax) { 5787 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5788 // We print a warning message only once per CMS cycle. 5789 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5790 } 5791 return; 5792 } 5793 // Double capacity if possible 5794 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5795 // Do not give up existing stack until we have managed to 5796 // get the double capacity that we desired. 5797 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5798 new_capacity * sizeof(oop))); 5799 if (rs.is_reserved()) { 5800 // Release the backing store associated with old stack 5801 _virtual_space.release(); 5802 // Reinitialize virtual space for new stack 5803 if (!_virtual_space.initialize(rs, rs.size())) { 5804 fatal("Not enough swap for expanded marking stack"); 5805 } 5806 _base = (oop*)(_virtual_space.low()); 5807 _index = 0; 5808 _capacity = new_capacity; 5809 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5810 // Failed to double capacity, continue; 5811 // we print a detail message only once per CMS cycle. 5812 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5813 _capacity / K, new_capacity / K); 5814 } 5815 } 5816 5817 5818 // Closures 5819 // XXX: there seems to be a lot of code duplication here; 5820 // should refactor and consolidate common code. 5821 5822 // This closure is used to mark refs into the CMS generation in 5823 // the CMS bit map. Called at the first checkpoint. This closure 5824 // assumes that we do not need to re-mark dirty cards; if the CMS 5825 // generation on which this is used is not an oldest 5826 // generation then this will lose younger_gen cards! 5827 5828 MarkRefsIntoClosure::MarkRefsIntoClosure( 5829 MemRegion span, CMSBitMap* bitMap): 5830 _span(span), 5831 _bitMap(bitMap) 5832 { 5833 assert(ref_processor() == NULL, "deliberately left NULL"); 5834 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5835 } 5836 5837 void MarkRefsIntoClosure::do_oop(oop obj) { 5838 // if p points into _span, then mark corresponding bit in _markBitMap 5839 assert(obj->is_oop(), "expected an oop"); 5840 HeapWord* addr = (HeapWord*)obj; 5841 if (_span.contains(addr)) { 5842 // this should be made more efficient 5843 _bitMap->mark(addr); 5844 } 5845 } 5846 5847 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5848 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5849 5850 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5851 MemRegion span, CMSBitMap* bitMap): 5852 _span(span), 5853 _bitMap(bitMap) 5854 { 5855 assert(ref_processor() == NULL, "deliberately left NULL"); 5856 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5857 } 5858 5859 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5860 // if p points into _span, then mark corresponding bit in _markBitMap 5861 assert(obj->is_oop(), "expected an oop"); 5862 HeapWord* addr = (HeapWord*)obj; 5863 if (_span.contains(addr)) { 5864 // this should be made more efficient 5865 _bitMap->par_mark(addr); 5866 } 5867 } 5868 5869 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5870 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5871 5872 // A variant of the above, used for CMS marking verification. 5873 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5874 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5875 _span(span), 5876 _verification_bm(verification_bm), 5877 _cms_bm(cms_bm) 5878 { 5879 assert(ref_processor() == NULL, "deliberately left NULL"); 5880 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5881 } 5882 5883 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5884 // if p points into _span, then mark corresponding bit in _markBitMap 5885 assert(obj->is_oop(), "expected an oop"); 5886 HeapWord* addr = (HeapWord*)obj; 5887 if (_span.contains(addr)) { 5888 _verification_bm->mark(addr); 5889 if (!_cms_bm->isMarked(addr)) { 5890 Log(gc, verify) log; 5891 ResourceMark rm; 5892 oop(addr)->print_on(log.error_stream()); 5893 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5894 fatal("... aborting"); 5895 } 5896 } 5897 } 5898 5899 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5900 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5901 5902 ////////////////////////////////////////////////// 5903 // MarkRefsIntoAndScanClosure 5904 ////////////////////////////////////////////////// 5905 5906 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5907 ReferenceProcessor* rp, 5908 CMSBitMap* bit_map, 5909 CMSBitMap* mod_union_table, 5910 CMSMarkStack* mark_stack, 5911 CMSCollector* collector, 5912 bool should_yield, 5913 bool concurrent_precleaning): 5914 _collector(collector), 5915 _span(span), 5916 _bit_map(bit_map), 5917 _mark_stack(mark_stack), 5918 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, 5919 mark_stack, concurrent_precleaning), 5920 _yield(should_yield), 5921 _concurrent_precleaning(concurrent_precleaning), 5922 _freelistLock(NULL) 5923 { 5924 // FIXME: Should initialize in base class constructor. 5925 assert(rp != NULL, "ref_processor shouldn't be NULL"); 5926 set_ref_processor_internal(rp); 5927 } 5928 5929 // This closure is used to mark refs into the CMS generation at the 5930 // second (final) checkpoint, and to scan and transitively follow 5931 // the unmarked oops. It is also used during the concurrent precleaning 5932 // phase while scanning objects on dirty cards in the CMS generation. 5933 // The marks are made in the marking bit map and the marking stack is 5934 // used for keeping the (newly) grey objects during the scan. 5935 // The parallel version (Par_...) appears further below. 5936 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5937 if (obj != NULL) { 5938 assert(obj->is_oop(), "expected an oop"); 5939 HeapWord* addr = (HeapWord*)obj; 5940 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5941 assert(_collector->overflow_list_is_empty(), 5942 "overflow list should be empty"); 5943 if (_span.contains(addr) && 5944 !_bit_map->isMarked(addr)) { 5945 // mark bit map (object is now grey) 5946 _bit_map->mark(addr); 5947 // push on marking stack (stack should be empty), and drain the 5948 // stack by applying this closure to the oops in the oops popped 5949 // from the stack (i.e. blacken the grey objects) 5950 bool res = _mark_stack->push(obj); 5951 assert(res, "Should have space to push on empty stack"); 5952 do { 5953 oop new_oop = _mark_stack->pop(); 5954 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 5955 assert(_bit_map->isMarked((HeapWord*)new_oop), 5956 "only grey objects on this stack"); 5957 // iterate over the oops in this oop, marking and pushing 5958 // the ones in CMS heap (i.e. in _span). 5959 new_oop->oop_iterate(&_pushAndMarkClosure); 5960 // check if it's time to yield 5961 do_yield_check(); 5962 } while (!_mark_stack->isEmpty() || 5963 (!_concurrent_precleaning && take_from_overflow_list())); 5964 // if marking stack is empty, and we are not doing this 5965 // during precleaning, then check the overflow list 5966 } 5967 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5968 assert(_collector->overflow_list_is_empty(), 5969 "overflow list was drained above"); 5970 5971 assert(_collector->no_preserved_marks(), 5972 "All preserved marks should have been restored above"); 5973 } 5974 } 5975 5976 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5977 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5978 5979 void MarkRefsIntoAndScanClosure::do_yield_work() { 5980 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5981 "CMS thread should hold CMS token"); 5982 assert_lock_strong(_freelistLock); 5983 assert_lock_strong(_bit_map->lock()); 5984 // relinquish the free_list_lock and bitMaplock() 5985 _bit_map->lock()->unlock(); 5986 _freelistLock->unlock(); 5987 ConcurrentMarkSweepThread::desynchronize(true); 5988 _collector->stopTimer(); 5989 _collector->incrementYields(); 5990 5991 // See the comment in coordinator_yield() 5992 for (unsigned i = 0; 5993 i < CMSYieldSleepCount && 5994 ConcurrentMarkSweepThread::should_yield() && 5995 !CMSCollector::foregroundGCIsActive(); 5996 ++i) { 5997 os::sleep(Thread::current(), 1, false); 5998 } 5999 6000 ConcurrentMarkSweepThread::synchronize(true); 6001 _freelistLock->lock_without_safepoint_check(); 6002 _bit_map->lock()->lock_without_safepoint_check(); 6003 _collector->startTimer(); 6004 } 6005 6006 /////////////////////////////////////////////////////////// 6007 // ParMarkRefsIntoAndScanClosure: a parallel version of 6008 // MarkRefsIntoAndScanClosure 6009 /////////////////////////////////////////////////////////// 6010 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 6011 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, 6012 CMSBitMap* bit_map, OopTaskQueue* work_queue): 6013 _span(span), 6014 _bit_map(bit_map), 6015 _work_queue(work_queue), 6016 _low_water_mark(MIN2((work_queue->max_elems()/4), 6017 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 6018 _parPushAndMarkClosure(collector, span, rp, bit_map, work_queue) 6019 { 6020 // FIXME: Should initialize in base class constructor. 6021 assert(rp != NULL, "ref_processor shouldn't be NULL"); 6022 set_ref_processor_internal(rp); 6023 } 6024 6025 // This closure is used to mark refs into the CMS generation at the 6026 // second (final) checkpoint, and to scan and transitively follow 6027 // the unmarked oops. The marks are made in the marking bit map and 6028 // the work_queue is used for keeping the (newly) grey objects during 6029 // the scan phase whence they are also available for stealing by parallel 6030 // threads. Since the marking bit map is shared, updates are 6031 // synchronized (via CAS). 6032 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 6033 if (obj != NULL) { 6034 // Ignore mark word because this could be an already marked oop 6035 // that may be chained at the end of the overflow list. 6036 assert(obj->is_oop(true), "expected an oop"); 6037 HeapWord* addr = (HeapWord*)obj; 6038 if (_span.contains(addr) && 6039 !_bit_map->isMarked(addr)) { 6040 // mark bit map (object will become grey): 6041 // It is possible for several threads to be 6042 // trying to "claim" this object concurrently; 6043 // the unique thread that succeeds in marking the 6044 // object first will do the subsequent push on 6045 // to the work queue (or overflow list). 6046 if (_bit_map->par_mark(addr)) { 6047 // push on work_queue (which may not be empty), and trim the 6048 // queue to an appropriate length by applying this closure to 6049 // the oops in the oops popped from the stack (i.e. blacken the 6050 // grey objects) 6051 bool res = _work_queue->push(obj); 6052 assert(res, "Low water mark should be less than capacity?"); 6053 trim_queue(_low_water_mark); 6054 } // Else, another thread claimed the object 6055 } 6056 } 6057 } 6058 6059 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6060 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6061 6062 // This closure is used to rescan the marked objects on the dirty cards 6063 // in the mod union table and the card table proper. 6064 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6065 oop p, MemRegion mr) { 6066 6067 size_t size = 0; 6068 HeapWord* addr = (HeapWord*)p; 6069 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6070 assert(_span.contains(addr), "we are scanning the CMS generation"); 6071 // check if it's time to yield 6072 if (do_yield_check()) { 6073 // We yielded for some foreground stop-world work, 6074 // and we have been asked to abort this ongoing preclean cycle. 6075 return 0; 6076 } 6077 if (_bitMap->isMarked(addr)) { 6078 // it's marked; is it potentially uninitialized? 6079 if (p->klass_or_null() != NULL) { 6080 // an initialized object; ignore mark word in verification below 6081 // since we are running concurrent with mutators 6082 assert(p->is_oop(true), "should be an oop"); 6083 if (p->is_objArray()) { 6084 // objArrays are precisely marked; restrict scanning 6085 // to dirty cards only. 6086 size = CompactibleFreeListSpace::adjustObjectSize( 6087 p->oop_iterate_size(_scanningClosure, mr)); 6088 } else { 6089 // A non-array may have been imprecisely marked; we need 6090 // to scan object in its entirety. 6091 size = CompactibleFreeListSpace::adjustObjectSize( 6092 p->oop_iterate_size(_scanningClosure)); 6093 } 6094 #ifdef ASSERT 6095 size_t direct_size = 6096 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6097 assert(size == direct_size, "Inconsistency in size"); 6098 assert(size >= 3, "Necessary for Printezis marks to work"); 6099 HeapWord* start_pbit = addr + 1; 6100 HeapWord* end_pbit = addr + size - 1; 6101 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6102 "inconsistent Printezis mark"); 6103 // Verify inner mark bits (between Printezis bits) are clear, 6104 // but don't repeat if there are multiple dirty regions for 6105 // the same object, to avoid potential O(N^2) performance. 6106 if (addr != _last_scanned_object) { 6107 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6108 _last_scanned_object = addr; 6109 } 6110 #endif // ASSERT 6111 } else { 6112 // An uninitialized object. 6113 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6114 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6115 size = pointer_delta(nextOneAddr + 1, addr); 6116 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6117 "alignment problem"); 6118 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6119 // will dirty the card when the klass pointer is installed in the 6120 // object (signaling the completion of initialization). 6121 } 6122 } else { 6123 // Either a not yet marked object or an uninitialized object 6124 if (p->klass_or_null() == NULL) { 6125 // An uninitialized object, skip to the next card, since 6126 // we may not be able to read its P-bits yet. 6127 assert(size == 0, "Initial value"); 6128 } else { 6129 // An object not (yet) reached by marking: we merely need to 6130 // compute its size so as to go look at the next block. 6131 assert(p->is_oop(true), "should be an oop"); 6132 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6133 } 6134 } 6135 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6136 return size; 6137 } 6138 6139 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6140 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6141 "CMS thread should hold CMS token"); 6142 assert_lock_strong(_freelistLock); 6143 assert_lock_strong(_bitMap->lock()); 6144 // relinquish the free_list_lock and bitMaplock() 6145 _bitMap->lock()->unlock(); 6146 _freelistLock->unlock(); 6147 ConcurrentMarkSweepThread::desynchronize(true); 6148 _collector->stopTimer(); 6149 _collector->incrementYields(); 6150 6151 // See the comment in coordinator_yield() 6152 for (unsigned i = 0; i < CMSYieldSleepCount && 6153 ConcurrentMarkSweepThread::should_yield() && 6154 !CMSCollector::foregroundGCIsActive(); ++i) { 6155 os::sleep(Thread::current(), 1, false); 6156 } 6157 6158 ConcurrentMarkSweepThread::synchronize(true); 6159 _freelistLock->lock_without_safepoint_check(); 6160 _bitMap->lock()->lock_without_safepoint_check(); 6161 _collector->startTimer(); 6162 } 6163 6164 6165 ////////////////////////////////////////////////////////////////// 6166 // SurvivorSpacePrecleanClosure 6167 ////////////////////////////////////////////////////////////////// 6168 // This (single-threaded) closure is used to preclean the oops in 6169 // the survivor spaces. 6170 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6171 6172 HeapWord* addr = (HeapWord*)p; 6173 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6174 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6175 assert(p->klass_or_null() != NULL, "object should be initialized"); 6176 // an initialized object; ignore mark word in verification below 6177 // since we are running concurrent with mutators 6178 assert(p->is_oop(true), "should be an oop"); 6179 // Note that we do not yield while we iterate over 6180 // the interior oops of p, pushing the relevant ones 6181 // on our marking stack. 6182 size_t size = p->oop_iterate_size(_scanning_closure); 6183 do_yield_check(); 6184 // Observe that below, we do not abandon the preclean 6185 // phase as soon as we should; rather we empty the 6186 // marking stack before returning. This is to satisfy 6187 // some existing assertions. In general, it may be a 6188 // good idea to abort immediately and complete the marking 6189 // from the grey objects at a later time. 6190 while (!_mark_stack->isEmpty()) { 6191 oop new_oop = _mark_stack->pop(); 6192 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 6193 assert(_bit_map->isMarked((HeapWord*)new_oop), 6194 "only grey objects on this stack"); 6195 // iterate over the oops in this oop, marking and pushing 6196 // the ones in CMS heap (i.e. in _span). 6197 new_oop->oop_iterate(_scanning_closure); 6198 // check if it's time to yield 6199 do_yield_check(); 6200 } 6201 unsigned int after_count = 6202 GenCollectedHeap::heap()->total_collections(); 6203 bool abort = (_before_count != after_count) || 6204 _collector->should_abort_preclean(); 6205 return abort ? 0 : size; 6206 } 6207 6208 void SurvivorSpacePrecleanClosure::do_yield_work() { 6209 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6210 "CMS thread should hold CMS token"); 6211 assert_lock_strong(_bit_map->lock()); 6212 // Relinquish the bit map lock 6213 _bit_map->lock()->unlock(); 6214 ConcurrentMarkSweepThread::desynchronize(true); 6215 _collector->stopTimer(); 6216 _collector->incrementYields(); 6217 6218 // See the comment in coordinator_yield() 6219 for (unsigned i = 0; i < CMSYieldSleepCount && 6220 ConcurrentMarkSweepThread::should_yield() && 6221 !CMSCollector::foregroundGCIsActive(); ++i) { 6222 os::sleep(Thread::current(), 1, false); 6223 } 6224 6225 ConcurrentMarkSweepThread::synchronize(true); 6226 _bit_map->lock()->lock_without_safepoint_check(); 6227 _collector->startTimer(); 6228 } 6229 6230 // This closure is used to rescan the marked objects on the dirty cards 6231 // in the mod union table and the card table proper. In the parallel 6232 // case, although the bitMap is shared, we do a single read so the 6233 // isMarked() query is "safe". 6234 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6235 // Ignore mark word because we are running concurrent with mutators 6236 assert(p->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6237 HeapWord* addr = (HeapWord*)p; 6238 assert(_span.contains(addr), "we are scanning the CMS generation"); 6239 bool is_obj_array = false; 6240 #ifdef ASSERT 6241 if (!_parallel) { 6242 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6243 assert(_collector->overflow_list_is_empty(), 6244 "overflow list should be empty"); 6245 6246 } 6247 #endif // ASSERT 6248 if (_bit_map->isMarked(addr)) { 6249 // Obj arrays are precisely marked, non-arrays are not; 6250 // so we scan objArrays precisely and non-arrays in their 6251 // entirety. 6252 if (p->is_objArray()) { 6253 is_obj_array = true; 6254 if (_parallel) { 6255 p->oop_iterate(_par_scan_closure, mr); 6256 } else { 6257 p->oop_iterate(_scan_closure, mr); 6258 } 6259 } else { 6260 if (_parallel) { 6261 p->oop_iterate(_par_scan_closure); 6262 } else { 6263 p->oop_iterate(_scan_closure); 6264 } 6265 } 6266 } 6267 #ifdef ASSERT 6268 if (!_parallel) { 6269 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6270 assert(_collector->overflow_list_is_empty(), 6271 "overflow list should be empty"); 6272 6273 } 6274 #endif // ASSERT 6275 return is_obj_array; 6276 } 6277 6278 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6279 MemRegion span, 6280 CMSBitMap* bitMap, CMSMarkStack* markStack, 6281 bool should_yield, bool verifying): 6282 _collector(collector), 6283 _span(span), 6284 _bitMap(bitMap), 6285 _mut(&collector->_modUnionTable), 6286 _markStack(markStack), 6287 _yield(should_yield), 6288 _skipBits(0) 6289 { 6290 assert(_markStack->isEmpty(), "stack should be empty"); 6291 _finger = _bitMap->startWord(); 6292 _threshold = _finger; 6293 assert(_collector->_restart_addr == NULL, "Sanity check"); 6294 assert(_span.contains(_finger), "Out of bounds _finger?"); 6295 DEBUG_ONLY(_verifying = verifying;) 6296 } 6297 6298 void MarkFromRootsClosure::reset(HeapWord* addr) { 6299 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6300 assert(_span.contains(addr), "Out of bounds _finger?"); 6301 _finger = addr; 6302 _threshold = (HeapWord*)round_to( 6303 (intptr_t)_finger, CardTableModRefBS::card_size); 6304 } 6305 6306 // Should revisit to see if this should be restructured for 6307 // greater efficiency. 6308 bool MarkFromRootsClosure::do_bit(size_t offset) { 6309 if (_skipBits > 0) { 6310 _skipBits--; 6311 return true; 6312 } 6313 // convert offset into a HeapWord* 6314 HeapWord* addr = _bitMap->startWord() + offset; 6315 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6316 "address out of range"); 6317 assert(_bitMap->isMarked(addr), "tautology"); 6318 if (_bitMap->isMarked(addr+1)) { 6319 // this is an allocated but not yet initialized object 6320 assert(_skipBits == 0, "tautology"); 6321 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6322 oop p = oop(addr); 6323 if (p->klass_or_null() == NULL) { 6324 DEBUG_ONLY(if (!_verifying) {) 6325 // We re-dirty the cards on which this object lies and increase 6326 // the _threshold so that we'll come back to scan this object 6327 // during the preclean or remark phase. (CMSCleanOnEnter) 6328 if (CMSCleanOnEnter) { 6329 size_t sz = _collector->block_size_using_printezis_bits(addr); 6330 HeapWord* end_card_addr = (HeapWord*)round_to( 6331 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6332 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6333 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6334 // Bump _threshold to end_card_addr; note that 6335 // _threshold cannot possibly exceed end_card_addr, anyhow. 6336 // This prevents future clearing of the card as the scan proceeds 6337 // to the right. 6338 assert(_threshold <= end_card_addr, 6339 "Because we are just scanning into this object"); 6340 if (_threshold < end_card_addr) { 6341 _threshold = end_card_addr; 6342 } 6343 if (p->klass_or_null() != NULL) { 6344 // Redirty the range of cards... 6345 _mut->mark_range(redirty_range); 6346 } // ...else the setting of klass will dirty the card anyway. 6347 } 6348 DEBUG_ONLY(}) 6349 return true; 6350 } 6351 } 6352 scanOopsInOop(addr); 6353 return true; 6354 } 6355 6356 // We take a break if we've been at this for a while, 6357 // so as to avoid monopolizing the locks involved. 6358 void MarkFromRootsClosure::do_yield_work() { 6359 // First give up the locks, then yield, then re-lock 6360 // We should probably use a constructor/destructor idiom to 6361 // do this unlock/lock or modify the MutexUnlocker class to 6362 // serve our purpose. XXX 6363 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6364 "CMS thread should hold CMS token"); 6365 assert_lock_strong(_bitMap->lock()); 6366 _bitMap->lock()->unlock(); 6367 ConcurrentMarkSweepThread::desynchronize(true); 6368 _collector->stopTimer(); 6369 _collector->incrementYields(); 6370 6371 // See the comment in coordinator_yield() 6372 for (unsigned i = 0; i < CMSYieldSleepCount && 6373 ConcurrentMarkSweepThread::should_yield() && 6374 !CMSCollector::foregroundGCIsActive(); ++i) { 6375 os::sleep(Thread::current(), 1, false); 6376 } 6377 6378 ConcurrentMarkSweepThread::synchronize(true); 6379 _bitMap->lock()->lock_without_safepoint_check(); 6380 _collector->startTimer(); 6381 } 6382 6383 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6384 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6385 assert(_markStack->isEmpty(), 6386 "should drain stack to limit stack usage"); 6387 // convert ptr to an oop preparatory to scanning 6388 oop obj = oop(ptr); 6389 // Ignore mark word in verification below, since we 6390 // may be running concurrent with mutators. 6391 assert(obj->is_oop(true), "should be an oop"); 6392 assert(_finger <= ptr, "_finger runneth ahead"); 6393 // advance the finger to right end of this object 6394 _finger = ptr + obj->size(); 6395 assert(_finger > ptr, "we just incremented it above"); 6396 // On large heaps, it may take us some time to get through 6397 // the marking phase. During 6398 // this time it's possible that a lot of mutations have 6399 // accumulated in the card table and the mod union table -- 6400 // these mutation records are redundant until we have 6401 // actually traced into the corresponding card. 6402 // Here, we check whether advancing the finger would make 6403 // us cross into a new card, and if so clear corresponding 6404 // cards in the MUT (preclean them in the card-table in the 6405 // future). 6406 6407 DEBUG_ONLY(if (!_verifying) {) 6408 // The clean-on-enter optimization is disabled by default, 6409 // until we fix 6178663. 6410 if (CMSCleanOnEnter && (_finger > _threshold)) { 6411 // [_threshold, _finger) represents the interval 6412 // of cards to be cleared in MUT (or precleaned in card table). 6413 // The set of cards to be cleared is all those that overlap 6414 // with the interval [_threshold, _finger); note that 6415 // _threshold is always kept card-aligned but _finger isn't 6416 // always card-aligned. 6417 HeapWord* old_threshold = _threshold; 6418 assert(old_threshold == (HeapWord*)round_to( 6419 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6420 "_threshold should always be card-aligned"); 6421 _threshold = (HeapWord*)round_to( 6422 (intptr_t)_finger, CardTableModRefBS::card_size); 6423 MemRegion mr(old_threshold, _threshold); 6424 assert(!mr.is_empty(), "Control point invariant"); 6425 assert(_span.contains(mr), "Should clear within span"); 6426 _mut->clear_range(mr); 6427 } 6428 DEBUG_ONLY(}) 6429 // Note: the finger doesn't advance while we drain 6430 // the stack below. 6431 PushOrMarkClosure pushOrMarkClosure(_collector, 6432 _span, _bitMap, _markStack, 6433 _finger, this); 6434 bool res = _markStack->push(obj); 6435 assert(res, "Empty non-zero size stack should have space for single push"); 6436 while (!_markStack->isEmpty()) { 6437 oop new_oop = _markStack->pop(); 6438 // Skip verifying header mark word below because we are 6439 // running concurrent with mutators. 6440 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6441 // now scan this oop's oops 6442 new_oop->oop_iterate(&pushOrMarkClosure); 6443 do_yield_check(); 6444 } 6445 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6446 } 6447 6448 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6449 CMSCollector* collector, MemRegion span, 6450 CMSBitMap* bit_map, 6451 OopTaskQueue* work_queue, 6452 CMSMarkStack* overflow_stack): 6453 _collector(collector), 6454 _whole_span(collector->_span), 6455 _span(span), 6456 _bit_map(bit_map), 6457 _mut(&collector->_modUnionTable), 6458 _work_queue(work_queue), 6459 _overflow_stack(overflow_stack), 6460 _skip_bits(0), 6461 _task(task) 6462 { 6463 assert(_work_queue->size() == 0, "work_queue should be empty"); 6464 _finger = span.start(); 6465 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6466 assert(_span.contains(_finger), "Out of bounds _finger?"); 6467 } 6468 6469 // Should revisit to see if this should be restructured for 6470 // greater efficiency. 6471 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6472 if (_skip_bits > 0) { 6473 _skip_bits--; 6474 return true; 6475 } 6476 // convert offset into a HeapWord* 6477 HeapWord* addr = _bit_map->startWord() + offset; 6478 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6479 "address out of range"); 6480 assert(_bit_map->isMarked(addr), "tautology"); 6481 if (_bit_map->isMarked(addr+1)) { 6482 // this is an allocated object that might not yet be initialized 6483 assert(_skip_bits == 0, "tautology"); 6484 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6485 oop p = oop(addr); 6486 if (p->klass_or_null() == NULL) { 6487 // in the case of Clean-on-Enter optimization, redirty card 6488 // and avoid clearing card by increasing the threshold. 6489 return true; 6490 } 6491 } 6492 scan_oops_in_oop(addr); 6493 return true; 6494 } 6495 6496 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6497 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6498 // Should we assert that our work queue is empty or 6499 // below some drain limit? 6500 assert(_work_queue->size() == 0, 6501 "should drain stack to limit stack usage"); 6502 // convert ptr to an oop preparatory to scanning 6503 oop obj = oop(ptr); 6504 // Ignore mark word in verification below, since we 6505 // may be running concurrent with mutators. 6506 assert(obj->is_oop(true), "should be an oop"); 6507 assert(_finger <= ptr, "_finger runneth ahead"); 6508 // advance the finger to right end of this object 6509 _finger = ptr + obj->size(); 6510 assert(_finger > ptr, "we just incremented it above"); 6511 // On large heaps, it may take us some time to get through 6512 // the marking phase. During 6513 // this time it's possible that a lot of mutations have 6514 // accumulated in the card table and the mod union table -- 6515 // these mutation records are redundant until we have 6516 // actually traced into the corresponding card. 6517 // Here, we check whether advancing the finger would make 6518 // us cross into a new card, and if so clear corresponding 6519 // cards in the MUT (preclean them in the card-table in the 6520 // future). 6521 6522 // The clean-on-enter optimization is disabled by default, 6523 // until we fix 6178663. 6524 if (CMSCleanOnEnter && (_finger > _threshold)) { 6525 // [_threshold, _finger) represents the interval 6526 // of cards to be cleared in MUT (or precleaned in card table). 6527 // The set of cards to be cleared is all those that overlap 6528 // with the interval [_threshold, _finger); note that 6529 // _threshold is always kept card-aligned but _finger isn't 6530 // always card-aligned. 6531 HeapWord* old_threshold = _threshold; 6532 assert(old_threshold == (HeapWord*)round_to( 6533 (intptr_t)old_threshold, CardTableModRefBS::card_size), 6534 "_threshold should always be card-aligned"); 6535 _threshold = (HeapWord*)round_to( 6536 (intptr_t)_finger, CardTableModRefBS::card_size); 6537 MemRegion mr(old_threshold, _threshold); 6538 assert(!mr.is_empty(), "Control point invariant"); 6539 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6540 _mut->clear_range(mr); 6541 } 6542 6543 // Note: the local finger doesn't advance while we drain 6544 // the stack below, but the global finger sure can and will. 6545 HeapWord** gfa = _task->global_finger_addr(); 6546 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6547 _span, _bit_map, 6548 _work_queue, 6549 _overflow_stack, 6550 _finger, 6551 gfa, this); 6552 bool res = _work_queue->push(obj); // overflow could occur here 6553 assert(res, "Will hold once we use workqueues"); 6554 while (true) { 6555 oop new_oop; 6556 if (!_work_queue->pop_local(new_oop)) { 6557 // We emptied our work_queue; check if there's stuff that can 6558 // be gotten from the overflow stack. 6559 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6560 _overflow_stack, _work_queue)) { 6561 do_yield_check(); 6562 continue; 6563 } else { // done 6564 break; 6565 } 6566 } 6567 // Skip verifying header mark word below because we are 6568 // running concurrent with mutators. 6569 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); 6570 // now scan this oop's oops 6571 new_oop->oop_iterate(&pushOrMarkClosure); 6572 do_yield_check(); 6573 } 6574 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6575 } 6576 6577 // Yield in response to a request from VM Thread or 6578 // from mutators. 6579 void ParMarkFromRootsClosure::do_yield_work() { 6580 assert(_task != NULL, "sanity"); 6581 _task->yield(); 6582 } 6583 6584 // A variant of the above used for verifying CMS marking work. 6585 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6586 MemRegion span, 6587 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6588 CMSMarkStack* mark_stack): 6589 _collector(collector), 6590 _span(span), 6591 _verification_bm(verification_bm), 6592 _cms_bm(cms_bm), 6593 _mark_stack(mark_stack), 6594 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6595 mark_stack) 6596 { 6597 assert(_mark_stack->isEmpty(), "stack should be empty"); 6598 _finger = _verification_bm->startWord(); 6599 assert(_collector->_restart_addr == NULL, "Sanity check"); 6600 assert(_span.contains(_finger), "Out of bounds _finger?"); 6601 } 6602 6603 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6604 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6605 assert(_span.contains(addr), "Out of bounds _finger?"); 6606 _finger = addr; 6607 } 6608 6609 // Should revisit to see if this should be restructured for 6610 // greater efficiency. 6611 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6612 // convert offset into a HeapWord* 6613 HeapWord* addr = _verification_bm->startWord() + offset; 6614 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6615 "address out of range"); 6616 assert(_verification_bm->isMarked(addr), "tautology"); 6617 assert(_cms_bm->isMarked(addr), "tautology"); 6618 6619 assert(_mark_stack->isEmpty(), 6620 "should drain stack to limit stack usage"); 6621 // convert addr to an oop preparatory to scanning 6622 oop obj = oop(addr); 6623 assert(obj->is_oop(), "should be an oop"); 6624 assert(_finger <= addr, "_finger runneth ahead"); 6625 // advance the finger to right end of this object 6626 _finger = addr + obj->size(); 6627 assert(_finger > addr, "we just incremented it above"); 6628 // Note: the finger doesn't advance while we drain 6629 // the stack below. 6630 bool res = _mark_stack->push(obj); 6631 assert(res, "Empty non-zero size stack should have space for single push"); 6632 while (!_mark_stack->isEmpty()) { 6633 oop new_oop = _mark_stack->pop(); 6634 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); 6635 // now scan this oop's oops 6636 new_oop->oop_iterate(&_pam_verify_closure); 6637 } 6638 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6639 return true; 6640 } 6641 6642 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6643 CMSCollector* collector, MemRegion span, 6644 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6645 CMSMarkStack* mark_stack): 6646 MetadataAwareOopClosure(collector->ref_processor()), 6647 _collector(collector), 6648 _span(span), 6649 _verification_bm(verification_bm), 6650 _cms_bm(cms_bm), 6651 _mark_stack(mark_stack) 6652 { } 6653 6654 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6655 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6656 6657 // Upon stack overflow, we discard (part of) the stack, 6658 // remembering the least address amongst those discarded 6659 // in CMSCollector's _restart_address. 6660 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6661 // Remember the least grey address discarded 6662 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6663 _collector->lower_restart_addr(ra); 6664 _mark_stack->reset(); // discard stack contents 6665 _mark_stack->expand(); // expand the stack if possible 6666 } 6667 6668 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6669 assert(obj->is_oop_or_null(), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6670 HeapWord* addr = (HeapWord*)obj; 6671 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6672 // Oop lies in _span and isn't yet grey or black 6673 _verification_bm->mark(addr); // now grey 6674 if (!_cms_bm->isMarked(addr)) { 6675 Log(gc, verify) log; 6676 ResourceMark rm; 6677 oop(addr)->print_on(log.error_stream()); 6678 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6679 fatal("... aborting"); 6680 } 6681 6682 if (!_mark_stack->push(obj)) { // stack overflow 6683 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6684 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6685 handle_stack_overflow(addr); 6686 } 6687 // anything including and to the right of _finger 6688 // will be scanned as we iterate over the remainder of the 6689 // bit map 6690 } 6691 } 6692 6693 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6694 MemRegion span, 6695 CMSBitMap* bitMap, CMSMarkStack* markStack, 6696 HeapWord* finger, MarkFromRootsClosure* parent) : 6697 MetadataAwareOopClosure(collector->ref_processor()), 6698 _collector(collector), 6699 _span(span), 6700 _bitMap(bitMap), 6701 _markStack(markStack), 6702 _finger(finger), 6703 _parent(parent) 6704 { } 6705 6706 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6707 MemRegion span, 6708 CMSBitMap* bit_map, 6709 OopTaskQueue* work_queue, 6710 CMSMarkStack* overflow_stack, 6711 HeapWord* finger, 6712 HeapWord** global_finger_addr, 6713 ParMarkFromRootsClosure* parent) : 6714 MetadataAwareOopClosure(collector->ref_processor()), 6715 _collector(collector), 6716 _whole_span(collector->_span), 6717 _span(span), 6718 _bit_map(bit_map), 6719 _work_queue(work_queue), 6720 _overflow_stack(overflow_stack), 6721 _finger(finger), 6722 _global_finger_addr(global_finger_addr), 6723 _parent(parent) 6724 { } 6725 6726 // Assumes thread-safe access by callers, who are 6727 // responsible for mutual exclusion. 6728 void CMSCollector::lower_restart_addr(HeapWord* low) { 6729 assert(_span.contains(low), "Out of bounds addr"); 6730 if (_restart_addr == NULL) { 6731 _restart_addr = low; 6732 } else { 6733 _restart_addr = MIN2(_restart_addr, low); 6734 } 6735 } 6736 6737 // Upon stack overflow, we discard (part of) the stack, 6738 // remembering the least address amongst those discarded 6739 // in CMSCollector's _restart_address. 6740 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6741 // Remember the least grey address discarded 6742 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6743 _collector->lower_restart_addr(ra); 6744 _markStack->reset(); // discard stack contents 6745 _markStack->expand(); // expand the stack if possible 6746 } 6747 6748 // Upon stack overflow, we discard (part of) the stack, 6749 // remembering the least address amongst those discarded 6750 // in CMSCollector's _restart_address. 6751 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6752 // We need to do this under a mutex to prevent other 6753 // workers from interfering with the work done below. 6754 MutexLockerEx ml(_overflow_stack->par_lock(), 6755 Mutex::_no_safepoint_check_flag); 6756 // Remember the least grey address discarded 6757 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6758 _collector->lower_restart_addr(ra); 6759 _overflow_stack->reset(); // discard stack contents 6760 _overflow_stack->expand(); // expand the stack if possible 6761 } 6762 6763 void PushOrMarkClosure::do_oop(oop obj) { 6764 // Ignore mark word because we are running concurrent with mutators. 6765 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6766 HeapWord* addr = (HeapWord*)obj; 6767 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6768 // Oop lies in _span and isn't yet grey or black 6769 _bitMap->mark(addr); // now grey 6770 if (addr < _finger) { 6771 // the bit map iteration has already either passed, or 6772 // sampled, this bit in the bit map; we'll need to 6773 // use the marking stack to scan this oop's oops. 6774 bool simulate_overflow = false; 6775 NOT_PRODUCT( 6776 if (CMSMarkStackOverflowALot && 6777 _collector->simulate_overflow()) { 6778 // simulate a stack overflow 6779 simulate_overflow = true; 6780 } 6781 ) 6782 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6783 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6784 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6785 handle_stack_overflow(addr); 6786 } 6787 } 6788 // anything including and to the right of _finger 6789 // will be scanned as we iterate over the remainder of the 6790 // bit map 6791 do_yield_check(); 6792 } 6793 } 6794 6795 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6796 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6797 6798 void ParPushOrMarkClosure::do_oop(oop obj) { 6799 // Ignore mark word because we are running concurrent with mutators. 6800 assert(obj->is_oop_or_null(true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6801 HeapWord* addr = (HeapWord*)obj; 6802 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6803 // Oop lies in _span and isn't yet grey or black 6804 // We read the global_finger (volatile read) strictly after marking oop 6805 bool res = _bit_map->par_mark(addr); // now grey 6806 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6807 // Should we push this marked oop on our stack? 6808 // -- if someone else marked it, nothing to do 6809 // -- if target oop is above global finger nothing to do 6810 // -- if target oop is in chunk and above local finger 6811 // then nothing to do 6812 // -- else push on work queue 6813 if ( !res // someone else marked it, they will deal with it 6814 || (addr >= *gfa) // will be scanned in a later task 6815 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6816 return; 6817 } 6818 // the bit map iteration has already either passed, or 6819 // sampled, this bit in the bit map; we'll need to 6820 // use the marking stack to scan this oop's oops. 6821 bool simulate_overflow = false; 6822 NOT_PRODUCT( 6823 if (CMSMarkStackOverflowALot && 6824 _collector->simulate_overflow()) { 6825 // simulate a stack overflow 6826 simulate_overflow = true; 6827 } 6828 ) 6829 if (simulate_overflow || 6830 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6831 // stack overflow 6832 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6833 // We cannot assert that the overflow stack is full because 6834 // it may have been emptied since. 6835 assert(simulate_overflow || 6836 _work_queue->size() == _work_queue->max_elems(), 6837 "Else push should have succeeded"); 6838 handle_stack_overflow(addr); 6839 } 6840 do_yield_check(); 6841 } 6842 } 6843 6844 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6845 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6846 6847 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6848 MemRegion span, 6849 ReferenceProcessor* rp, 6850 CMSBitMap* bit_map, 6851 CMSBitMap* mod_union_table, 6852 CMSMarkStack* mark_stack, 6853 bool concurrent_precleaning): 6854 MetadataAwareOopClosure(rp), 6855 _collector(collector), 6856 _span(span), 6857 _bit_map(bit_map), 6858 _mod_union_table(mod_union_table), 6859 _mark_stack(mark_stack), 6860 _concurrent_precleaning(concurrent_precleaning) 6861 { 6862 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6863 } 6864 6865 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6866 // the non-parallel version (the parallel version appears further below.) 6867 void PushAndMarkClosure::do_oop(oop obj) { 6868 // Ignore mark word verification. If during concurrent precleaning, 6869 // the object monitor may be locked. If during the checkpoint 6870 // phases, the object may already have been reached by a different 6871 // path and may be at the end of the global overflow list (so 6872 // the mark word may be NULL). 6873 assert(obj->is_oop_or_null(true /* ignore mark word */), 6874 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6875 HeapWord* addr = (HeapWord*)obj; 6876 // Check if oop points into the CMS generation 6877 // and is not marked 6878 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6879 // a white object ... 6880 _bit_map->mark(addr); // ... now grey 6881 // push on the marking stack (grey set) 6882 bool simulate_overflow = false; 6883 NOT_PRODUCT( 6884 if (CMSMarkStackOverflowALot && 6885 _collector->simulate_overflow()) { 6886 // simulate a stack overflow 6887 simulate_overflow = true; 6888 } 6889 ) 6890 if (simulate_overflow || !_mark_stack->push(obj)) { 6891 if (_concurrent_precleaning) { 6892 // During precleaning we can just dirty the appropriate card(s) 6893 // in the mod union table, thus ensuring that the object remains 6894 // in the grey set and continue. In the case of object arrays 6895 // we need to dirty all of the cards that the object spans, 6896 // since the rescan of object arrays will be limited to the 6897 // dirty cards. 6898 // Note that no one can be interfering with us in this action 6899 // of dirtying the mod union table, so no locking or atomics 6900 // are required. 6901 if (obj->is_objArray()) { 6902 size_t sz = obj->size(); 6903 HeapWord* end_card_addr = (HeapWord*)round_to( 6904 (intptr_t)(addr+sz), CardTableModRefBS::card_size); 6905 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6906 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6907 _mod_union_table->mark_range(redirty_range); 6908 } else { 6909 _mod_union_table->mark(addr); 6910 } 6911 _collector->_ser_pmc_preclean_ovflw++; 6912 } else { 6913 // During the remark phase, we need to remember this oop 6914 // in the overflow list. 6915 _collector->push_on_overflow_list(obj); 6916 _collector->_ser_pmc_remark_ovflw++; 6917 } 6918 } 6919 } 6920 } 6921 6922 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6923 MemRegion span, 6924 ReferenceProcessor* rp, 6925 CMSBitMap* bit_map, 6926 OopTaskQueue* work_queue): 6927 MetadataAwareOopClosure(rp), 6928 _collector(collector), 6929 _span(span), 6930 _bit_map(bit_map), 6931 _work_queue(work_queue) 6932 { 6933 assert(ref_processor() != NULL, "ref_processor shouldn't be NULL"); 6934 } 6935 6936 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6937 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6938 6939 // Grey object rescan during second checkpoint phase -- 6940 // the parallel version. 6941 void ParPushAndMarkClosure::do_oop(oop obj) { 6942 // In the assert below, we ignore the mark word because 6943 // this oop may point to an already visited object that is 6944 // on the overflow stack (in which case the mark word has 6945 // been hijacked for chaining into the overflow stack -- 6946 // if this is the last object in the overflow stack then 6947 // its mark word will be NULL). Because this object may 6948 // have been subsequently popped off the global overflow 6949 // stack, and the mark word possibly restored to the prototypical 6950 // value, by the time we get to examined this failing assert in 6951 // the debugger, is_oop_or_null(false) may subsequently start 6952 // to hold. 6953 assert(obj->is_oop_or_null(true), 6954 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6955 HeapWord* addr = (HeapWord*)obj; 6956 // Check if oop points into the CMS generation 6957 // and is not marked 6958 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6959 // a white object ... 6960 // If we manage to "claim" the object, by being the 6961 // first thread to mark it, then we push it on our 6962 // marking stack 6963 if (_bit_map->par_mark(addr)) { // ... now grey 6964 // push on work queue (grey set) 6965 bool simulate_overflow = false; 6966 NOT_PRODUCT( 6967 if (CMSMarkStackOverflowALot && 6968 _collector->par_simulate_overflow()) { 6969 // simulate a stack overflow 6970 simulate_overflow = true; 6971 } 6972 ) 6973 if (simulate_overflow || !_work_queue->push(obj)) { 6974 _collector->par_push_on_overflow_list(obj); 6975 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6976 } 6977 } // Else, some other thread got there first 6978 } 6979 } 6980 6981 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6982 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6983 6984 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6985 Mutex* bml = _collector->bitMapLock(); 6986 assert_lock_strong(bml); 6987 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6988 "CMS thread should hold CMS token"); 6989 6990 bml->unlock(); 6991 ConcurrentMarkSweepThread::desynchronize(true); 6992 6993 _collector->stopTimer(); 6994 _collector->incrementYields(); 6995 6996 // See the comment in coordinator_yield() 6997 for (unsigned i = 0; i < CMSYieldSleepCount && 6998 ConcurrentMarkSweepThread::should_yield() && 6999 !CMSCollector::foregroundGCIsActive(); ++i) { 7000 os::sleep(Thread::current(), 1, false); 7001 } 7002 7003 ConcurrentMarkSweepThread::synchronize(true); 7004 bml->lock(); 7005 7006 _collector->startTimer(); 7007 } 7008 7009 bool CMSPrecleanRefsYieldClosure::should_return() { 7010 if (ConcurrentMarkSweepThread::should_yield()) { 7011 do_yield_work(); 7012 } 7013 return _collector->foregroundGCIsActive(); 7014 } 7015 7016 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 7017 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, 7018 "mr should be aligned to start at a card boundary"); 7019 // We'd like to assert: 7020 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, 7021 // "mr should be a range of cards"); 7022 // However, that would be too strong in one case -- the last 7023 // partition ends at _unallocated_block which, in general, can be 7024 // an arbitrary boundary, not necessarily card aligned. 7025 _num_dirty_cards += mr.word_size()/CardTableModRefBS::card_size_in_words; 7026 _space->object_iterate_mem(mr, &_scan_cl); 7027 } 7028 7029 SweepClosure::SweepClosure(CMSCollector* collector, 7030 ConcurrentMarkSweepGeneration* g, 7031 CMSBitMap* bitMap, bool should_yield) : 7032 _collector(collector), 7033 _g(g), 7034 _sp(g->cmsSpace()), 7035 _limit(_sp->sweep_limit()), 7036 _freelistLock(_sp->freelistLock()), 7037 _bitMap(bitMap), 7038 _yield(should_yield), 7039 _inFreeRange(false), // No free range at beginning of sweep 7040 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7041 _lastFreeRangeCoalesced(false), 7042 _freeFinger(g->used_region().start()) 7043 { 7044 NOT_PRODUCT( 7045 _numObjectsFreed = 0; 7046 _numWordsFreed = 0; 7047 _numObjectsLive = 0; 7048 _numWordsLive = 0; 7049 _numObjectsAlreadyFree = 0; 7050 _numWordsAlreadyFree = 0; 7051 _last_fc = NULL; 7052 7053 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7054 _sp->dictionary()->initialize_dict_returned_bytes(); 7055 ) 7056 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7057 "sweep _limit out of bounds"); 7058 log_develop_trace(gc, sweep)("===================="); 7059 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7060 } 7061 7062 void SweepClosure::print_on(outputStream* st) const { 7063 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7064 p2i(_sp->bottom()), p2i(_sp->end())); 7065 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7066 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7067 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7068 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7069 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7070 } 7071 7072 #ifndef PRODUCT 7073 // Assertion checking only: no useful work in product mode -- 7074 // however, if any of the flags below become product flags, 7075 // you may need to review this code to see if it needs to be 7076 // enabled in product mode. 7077 SweepClosure::~SweepClosure() { 7078 assert_lock_strong(_freelistLock); 7079 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7080 "sweep _limit out of bounds"); 7081 if (inFreeRange()) { 7082 Log(gc, sweep) log; 7083 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7084 ResourceMark rm; 7085 print_on(log.error_stream()); 7086 ShouldNotReachHere(); 7087 } 7088 7089 if (log_is_enabled(Debug, gc, sweep)) { 7090 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7091 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7092 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7093 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7094 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7095 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7096 } 7097 7098 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7099 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7100 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7101 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7102 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7103 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7104 } 7105 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7106 log_develop_trace(gc, sweep)("================"); 7107 } 7108 #endif // PRODUCT 7109 7110 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7111 bool freeRangeInFreeLists) { 7112 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7113 p2i(freeFinger), freeRangeInFreeLists); 7114 assert(!inFreeRange(), "Trampling existing free range"); 7115 set_inFreeRange(true); 7116 set_lastFreeRangeCoalesced(false); 7117 7118 set_freeFinger(freeFinger); 7119 set_freeRangeInFreeLists(freeRangeInFreeLists); 7120 if (CMSTestInFreeList) { 7121 if (freeRangeInFreeLists) { 7122 FreeChunk* fc = (FreeChunk*) freeFinger; 7123 assert(fc->is_free(), "A chunk on the free list should be free."); 7124 assert(fc->size() > 0, "Free range should have a size"); 7125 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7126 } 7127 } 7128 } 7129 7130 // Note that the sweeper runs concurrently with mutators. Thus, 7131 // it is possible for direct allocation in this generation to happen 7132 // in the middle of the sweep. Note that the sweeper also coalesces 7133 // contiguous free blocks. Thus, unless the sweeper and the allocator 7134 // synchronize appropriately freshly allocated blocks may get swept up. 7135 // This is accomplished by the sweeper locking the free lists while 7136 // it is sweeping. Thus blocks that are determined to be free are 7137 // indeed free. There is however one additional complication: 7138 // blocks that have been allocated since the final checkpoint and 7139 // mark, will not have been marked and so would be treated as 7140 // unreachable and swept up. To prevent this, the allocator marks 7141 // the bit map when allocating during the sweep phase. This leads, 7142 // however, to a further complication -- objects may have been allocated 7143 // but not yet initialized -- in the sense that the header isn't yet 7144 // installed. The sweeper can not then determine the size of the block 7145 // in order to skip over it. To deal with this case, we use a technique 7146 // (due to Printezis) to encode such uninitialized block sizes in the 7147 // bit map. Since the bit map uses a bit per every HeapWord, but the 7148 // CMS generation has a minimum object size of 3 HeapWords, it follows 7149 // that "normal marks" won't be adjacent in the bit map (there will 7150 // always be at least two 0 bits between successive 1 bits). We make use 7151 // of these "unused" bits to represent uninitialized blocks -- the bit 7152 // corresponding to the start of the uninitialized object and the next 7153 // bit are both set. Finally, a 1 bit marks the end of the object that 7154 // started with the two consecutive 1 bits to indicate its potentially 7155 // uninitialized state. 7156 7157 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7158 FreeChunk* fc = (FreeChunk*)addr; 7159 size_t res; 7160 7161 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7162 // than "addr == _limit" because although _limit was a block boundary when 7163 // we started the sweep, it may no longer be one because heap expansion 7164 // may have caused us to coalesce the block ending at the address _limit 7165 // with a newly expanded chunk (this happens when _limit was set to the 7166 // previous _end of the space), so we may have stepped past _limit: 7167 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7168 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7169 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7170 "sweep _limit out of bounds"); 7171 assert(addr < _sp->end(), "addr out of bounds"); 7172 // Flush any free range we might be holding as a single 7173 // coalesced chunk to the appropriate free list. 7174 if (inFreeRange()) { 7175 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7176 "freeFinger() " PTR_FORMAT " is out-of-bounds", p2i(freeFinger())); 7177 flush_cur_free_chunk(freeFinger(), 7178 pointer_delta(addr, freeFinger())); 7179 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7180 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7181 lastFreeRangeCoalesced() ? 1 : 0); 7182 } 7183 7184 // help the iterator loop finish 7185 return pointer_delta(_sp->end(), addr); 7186 } 7187 7188 assert(addr < _limit, "sweep invariant"); 7189 // check if we should yield 7190 do_yield_check(addr); 7191 if (fc->is_free()) { 7192 // Chunk that is already free 7193 res = fc->size(); 7194 do_already_free_chunk(fc); 7195 debug_only(_sp->verifyFreeLists()); 7196 // If we flush the chunk at hand in lookahead_and_flush() 7197 // and it's coalesced with a preceding chunk, then the 7198 // process of "mangling" the payload of the coalesced block 7199 // will cause erasure of the size information from the 7200 // (erstwhile) header of all the coalesced blocks but the 7201 // first, so the first disjunct in the assert will not hold 7202 // in that specific case (in which case the second disjunct 7203 // will hold). 7204 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7205 "Otherwise the size info doesn't change at this step"); 7206 NOT_PRODUCT( 7207 _numObjectsAlreadyFree++; 7208 _numWordsAlreadyFree += res; 7209 ) 7210 NOT_PRODUCT(_last_fc = fc;) 7211 } else if (!_bitMap->isMarked(addr)) { 7212 // Chunk is fresh garbage 7213 res = do_garbage_chunk(fc); 7214 debug_only(_sp->verifyFreeLists()); 7215 NOT_PRODUCT( 7216 _numObjectsFreed++; 7217 _numWordsFreed += res; 7218 ) 7219 } else { 7220 // Chunk that is alive. 7221 res = do_live_chunk(fc); 7222 debug_only(_sp->verifyFreeLists()); 7223 NOT_PRODUCT( 7224 _numObjectsLive++; 7225 _numWordsLive += res; 7226 ) 7227 } 7228 return res; 7229 } 7230 7231 // For the smart allocation, record following 7232 // split deaths - a free chunk is removed from its free list because 7233 // it is being split into two or more chunks. 7234 // split birth - a free chunk is being added to its free list because 7235 // a larger free chunk has been split and resulted in this free chunk. 7236 // coal death - a free chunk is being removed from its free list because 7237 // it is being coalesced into a large free chunk. 7238 // coal birth - a free chunk is being added to its free list because 7239 // it was created when two or more free chunks where coalesced into 7240 // this free chunk. 7241 // 7242 // These statistics are used to determine the desired number of free 7243 // chunks of a given size. The desired number is chosen to be relative 7244 // to the end of a CMS sweep. The desired number at the end of a sweep 7245 // is the 7246 // count-at-end-of-previous-sweep (an amount that was enough) 7247 // - count-at-beginning-of-current-sweep (the excess) 7248 // + split-births (gains in this size during interval) 7249 // - split-deaths (demands on this size during interval) 7250 // where the interval is from the end of one sweep to the end of the 7251 // next. 7252 // 7253 // When sweeping the sweeper maintains an accumulated chunk which is 7254 // the chunk that is made up of chunks that have been coalesced. That 7255 // will be termed the left-hand chunk. A new chunk of garbage that 7256 // is being considered for coalescing will be referred to as the 7257 // right-hand chunk. 7258 // 7259 // When making a decision on whether to coalesce a right-hand chunk with 7260 // the current left-hand chunk, the current count vs. the desired count 7261 // of the left-hand chunk is considered. Also if the right-hand chunk 7262 // is near the large chunk at the end of the heap (see 7263 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7264 // left-hand chunk is coalesced. 7265 // 7266 // When making a decision about whether to split a chunk, the desired count 7267 // vs. the current count of the candidate to be split is also considered. 7268 // If the candidate is underpopulated (currently fewer chunks than desired) 7269 // a chunk of an overpopulated (currently more chunks than desired) size may 7270 // be chosen. The "hint" associated with a free list, if non-null, points 7271 // to a free list which may be overpopulated. 7272 // 7273 7274 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7275 const size_t size = fc->size(); 7276 // Chunks that cannot be coalesced are not in the 7277 // free lists. 7278 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7279 assert(_sp->verify_chunk_in_free_list(fc), 7280 "free chunk should be in free lists"); 7281 } 7282 // a chunk that is already free, should not have been 7283 // marked in the bit map 7284 HeapWord* const addr = (HeapWord*) fc; 7285 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7286 // Verify that the bit map has no bits marked between 7287 // addr and purported end of this block. 7288 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7289 7290 // Some chunks cannot be coalesced under any circumstances. 7291 // See the definition of cantCoalesce(). 7292 if (!fc->cantCoalesce()) { 7293 // This chunk can potentially be coalesced. 7294 // All the work is done in 7295 do_post_free_or_garbage_chunk(fc, size); 7296 // Note that if the chunk is not coalescable (the else arm 7297 // below), we unconditionally flush, without needing to do 7298 // a "lookahead," as we do below. 7299 if (inFreeRange()) lookahead_and_flush(fc, size); 7300 } else { 7301 // Code path common to both original and adaptive free lists. 7302 7303 // cant coalesce with previous block; this should be treated 7304 // as the end of a free run if any 7305 if (inFreeRange()) { 7306 // we kicked some butt; time to pick up the garbage 7307 assert(freeFinger() < addr, "freeFinger points too high"); 7308 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7309 } 7310 // else, nothing to do, just continue 7311 } 7312 } 7313 7314 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7315 // This is a chunk of garbage. It is not in any free list. 7316 // Add it to a free list or let it possibly be coalesced into 7317 // a larger chunk. 7318 HeapWord* const addr = (HeapWord*) fc; 7319 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7320 7321 // Verify that the bit map has no bits marked between 7322 // addr and purported end of just dead object. 7323 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7324 do_post_free_or_garbage_chunk(fc, size); 7325 7326 assert(_limit >= addr + size, 7327 "A freshly garbage chunk can't possibly straddle over _limit"); 7328 if (inFreeRange()) lookahead_and_flush(fc, size); 7329 return size; 7330 } 7331 7332 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7333 HeapWord* addr = (HeapWord*) fc; 7334 // The sweeper has just found a live object. Return any accumulated 7335 // left hand chunk to the free lists. 7336 if (inFreeRange()) { 7337 assert(freeFinger() < addr, "freeFinger points too high"); 7338 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7339 } 7340 7341 // This object is live: we'd normally expect this to be 7342 // an oop, and like to assert the following: 7343 // assert(oop(addr)->is_oop(), "live block should be an oop"); 7344 // However, as we commented above, this may be an object whose 7345 // header hasn't yet been initialized. 7346 size_t size; 7347 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7348 if (_bitMap->isMarked(addr + 1)) { 7349 // Determine the size from the bit map, rather than trying to 7350 // compute it from the object header. 7351 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7352 size = pointer_delta(nextOneAddr + 1, addr); 7353 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7354 "alignment problem"); 7355 7356 #ifdef ASSERT 7357 if (oop(addr)->klass_or_null() != NULL) { 7358 // Ignore mark word because we are running concurrent with mutators 7359 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7360 assert(size == 7361 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7362 "P-mark and computed size do not agree"); 7363 } 7364 #endif 7365 7366 } else { 7367 // This should be an initialized object that's alive. 7368 assert(oop(addr)->klass_or_null() != NULL, 7369 "Should be an initialized object"); 7370 // Ignore mark word because we are running concurrent with mutators 7371 assert(oop(addr)->is_oop(true), "live block should be an oop"); 7372 // Verify that the bit map has no bits marked between 7373 // addr and purported end of this block. 7374 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7375 assert(size >= 3, "Necessary for Printezis marks to work"); 7376 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7377 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7378 } 7379 return size; 7380 } 7381 7382 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7383 size_t chunkSize) { 7384 // do_post_free_or_garbage_chunk() should only be called in the case 7385 // of the adaptive free list allocator. 7386 const bool fcInFreeLists = fc->is_free(); 7387 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7388 if (CMSTestInFreeList && fcInFreeLists) { 7389 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7390 } 7391 7392 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7393 7394 HeapWord* const fc_addr = (HeapWord*) fc; 7395 7396 bool coalesce = false; 7397 const size_t left = pointer_delta(fc_addr, freeFinger()); 7398 const size_t right = chunkSize; 7399 switch (FLSCoalescePolicy) { 7400 // numeric value forms a coalition aggressiveness metric 7401 case 0: { // never coalesce 7402 coalesce = false; 7403 break; 7404 } 7405 case 1: { // coalesce if left & right chunks on overpopulated lists 7406 coalesce = _sp->coalOverPopulated(left) && 7407 _sp->coalOverPopulated(right); 7408 break; 7409 } 7410 case 2: { // coalesce if left chunk on overpopulated list (default) 7411 coalesce = _sp->coalOverPopulated(left); 7412 break; 7413 } 7414 case 3: { // coalesce if left OR right chunk on overpopulated list 7415 coalesce = _sp->coalOverPopulated(left) || 7416 _sp->coalOverPopulated(right); 7417 break; 7418 } 7419 case 4: { // always coalesce 7420 coalesce = true; 7421 break; 7422 } 7423 default: 7424 ShouldNotReachHere(); 7425 } 7426 7427 // Should the current free range be coalesced? 7428 // If the chunk is in a free range and either we decided to coalesce above 7429 // or the chunk is near the large block at the end of the heap 7430 // (isNearLargestChunk() returns true), then coalesce this chunk. 7431 const bool doCoalesce = inFreeRange() 7432 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7433 if (doCoalesce) { 7434 // Coalesce the current free range on the left with the new 7435 // chunk on the right. If either is on a free list, 7436 // it must be removed from the list and stashed in the closure. 7437 if (freeRangeInFreeLists()) { 7438 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7439 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7440 "Size of free range is inconsistent with chunk size."); 7441 if (CMSTestInFreeList) { 7442 assert(_sp->verify_chunk_in_free_list(ffc), 7443 "Chunk is not in free lists"); 7444 } 7445 _sp->coalDeath(ffc->size()); 7446 _sp->removeFreeChunkFromFreeLists(ffc); 7447 set_freeRangeInFreeLists(false); 7448 } 7449 if (fcInFreeLists) { 7450 _sp->coalDeath(chunkSize); 7451 assert(fc->size() == chunkSize, 7452 "The chunk has the wrong size or is not in the free lists"); 7453 _sp->removeFreeChunkFromFreeLists(fc); 7454 } 7455 set_lastFreeRangeCoalesced(true); 7456 print_free_block_coalesced(fc); 7457 } else { // not in a free range and/or should not coalesce 7458 // Return the current free range and start a new one. 7459 if (inFreeRange()) { 7460 // In a free range but cannot coalesce with the right hand chunk. 7461 // Put the current free range into the free lists. 7462 flush_cur_free_chunk(freeFinger(), 7463 pointer_delta(fc_addr, freeFinger())); 7464 } 7465 // Set up for new free range. Pass along whether the right hand 7466 // chunk is in the free lists. 7467 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7468 } 7469 } 7470 7471 // Lookahead flush: 7472 // If we are tracking a free range, and this is the last chunk that 7473 // we'll look at because its end crosses past _limit, we'll preemptively 7474 // flush it along with any free range we may be holding on to. Note that 7475 // this can be the case only for an already free or freshly garbage 7476 // chunk. If this block is an object, it can never straddle 7477 // over _limit. The "straddling" occurs when _limit is set at 7478 // the previous end of the space when this cycle started, and 7479 // a subsequent heap expansion caused the previously co-terminal 7480 // free block to be coalesced with the newly expanded portion, 7481 // thus rendering _limit a non-block-boundary making it dangerous 7482 // for the sweeper to step over and examine. 7483 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7484 assert(inFreeRange(), "Should only be called if currently in a free range."); 7485 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7486 assert(_sp->used_region().contains(eob - 1), 7487 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7488 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7489 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7490 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7491 if (eob >= _limit) { 7492 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7493 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7494 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7495 "[" PTR_FORMAT "," PTR_FORMAT ")", 7496 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7497 // Return the storage we are tracking back into the free lists. 7498 log_develop_trace(gc, sweep)("Flushing ... "); 7499 assert(freeFinger() < eob, "Error"); 7500 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7501 } 7502 } 7503 7504 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7505 assert(inFreeRange(), "Should only be called if currently in a free range."); 7506 assert(size > 0, 7507 "A zero sized chunk cannot be added to the free lists."); 7508 if (!freeRangeInFreeLists()) { 7509 if (CMSTestInFreeList) { 7510 FreeChunk* fc = (FreeChunk*) chunk; 7511 fc->set_size(size); 7512 assert(!_sp->verify_chunk_in_free_list(fc), 7513 "chunk should not be in free lists yet"); 7514 } 7515 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7516 // A new free range is going to be starting. The current 7517 // free range has not been added to the free lists yet or 7518 // was removed so add it back. 7519 // If the current free range was coalesced, then the death 7520 // of the free range was recorded. Record a birth now. 7521 if (lastFreeRangeCoalesced()) { 7522 _sp->coalBirth(size); 7523 } 7524 _sp->addChunkAndRepairOffsetTable(chunk, size, 7525 lastFreeRangeCoalesced()); 7526 } else { 7527 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7528 } 7529 set_inFreeRange(false); 7530 set_freeRangeInFreeLists(false); 7531 } 7532 7533 // We take a break if we've been at this for a while, 7534 // so as to avoid monopolizing the locks involved. 7535 void SweepClosure::do_yield_work(HeapWord* addr) { 7536 // Return current free chunk being used for coalescing (if any) 7537 // to the appropriate freelist. After yielding, the next 7538 // free block encountered will start a coalescing range of 7539 // free blocks. If the next free block is adjacent to the 7540 // chunk just flushed, they will need to wait for the next 7541 // sweep to be coalesced. 7542 if (inFreeRange()) { 7543 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7544 } 7545 7546 // First give up the locks, then yield, then re-lock. 7547 // We should probably use a constructor/destructor idiom to 7548 // do this unlock/lock or modify the MutexUnlocker class to 7549 // serve our purpose. XXX 7550 assert_lock_strong(_bitMap->lock()); 7551 assert_lock_strong(_freelistLock); 7552 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7553 "CMS thread should hold CMS token"); 7554 _bitMap->lock()->unlock(); 7555 _freelistLock->unlock(); 7556 ConcurrentMarkSweepThread::desynchronize(true); 7557 _collector->stopTimer(); 7558 _collector->incrementYields(); 7559 7560 // See the comment in coordinator_yield() 7561 for (unsigned i = 0; i < CMSYieldSleepCount && 7562 ConcurrentMarkSweepThread::should_yield() && 7563 !CMSCollector::foregroundGCIsActive(); ++i) { 7564 os::sleep(Thread::current(), 1, false); 7565 } 7566 7567 ConcurrentMarkSweepThread::synchronize(true); 7568 _freelistLock->lock(); 7569 _bitMap->lock()->lock_without_safepoint_check(); 7570 _collector->startTimer(); 7571 } 7572 7573 #ifndef PRODUCT 7574 // This is actually very useful in a product build if it can 7575 // be called from the debugger. Compile it into the product 7576 // as needed. 7577 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7578 return debug_cms_space->verify_chunk_in_free_list(fc); 7579 } 7580 #endif 7581 7582 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7583 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7584 p2i(fc), fc->size()); 7585 } 7586 7587 // CMSIsAliveClosure 7588 bool CMSIsAliveClosure::do_object_b(oop obj) { 7589 HeapWord* addr = (HeapWord*)obj; 7590 return addr != NULL && 7591 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7592 } 7593 7594 7595 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7596 MemRegion span, 7597 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7598 bool cpc): 7599 _collector(collector), 7600 _span(span), 7601 _bit_map(bit_map), 7602 _mark_stack(mark_stack), 7603 _concurrent_precleaning(cpc) { 7604 assert(!_span.is_empty(), "Empty span could spell trouble"); 7605 } 7606 7607 7608 // CMSKeepAliveClosure: the serial version 7609 void CMSKeepAliveClosure::do_oop(oop obj) { 7610 HeapWord* addr = (HeapWord*)obj; 7611 if (_span.contains(addr) && 7612 !_bit_map->isMarked(addr)) { 7613 _bit_map->mark(addr); 7614 bool simulate_overflow = false; 7615 NOT_PRODUCT( 7616 if (CMSMarkStackOverflowALot && 7617 _collector->simulate_overflow()) { 7618 // simulate a stack overflow 7619 simulate_overflow = true; 7620 } 7621 ) 7622 if (simulate_overflow || !_mark_stack->push(obj)) { 7623 if (_concurrent_precleaning) { 7624 // We dirty the overflown object and let the remark 7625 // phase deal with it. 7626 assert(_collector->overflow_list_is_empty(), "Error"); 7627 // In the case of object arrays, we need to dirty all of 7628 // the cards that the object spans. No locking or atomics 7629 // are needed since no one else can be mutating the mod union 7630 // table. 7631 if (obj->is_objArray()) { 7632 size_t sz = obj->size(); 7633 HeapWord* end_card_addr = 7634 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); 7635 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7636 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7637 _collector->_modUnionTable.mark_range(redirty_range); 7638 } else { 7639 _collector->_modUnionTable.mark(addr); 7640 } 7641 _collector->_ser_kac_preclean_ovflw++; 7642 } else { 7643 _collector->push_on_overflow_list(obj); 7644 _collector->_ser_kac_ovflw++; 7645 } 7646 } 7647 } 7648 } 7649 7650 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7651 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7652 7653 // CMSParKeepAliveClosure: a parallel version of the above. 7654 // The work queues are private to each closure (thread), 7655 // but (may be) available for stealing by other threads. 7656 void CMSParKeepAliveClosure::do_oop(oop obj) { 7657 HeapWord* addr = (HeapWord*)obj; 7658 if (_span.contains(addr) && 7659 !_bit_map->isMarked(addr)) { 7660 // In general, during recursive tracing, several threads 7661 // may be concurrently getting here; the first one to 7662 // "tag" it, claims it. 7663 if (_bit_map->par_mark(addr)) { 7664 bool res = _work_queue->push(obj); 7665 assert(res, "Low water mark should be much less than capacity"); 7666 // Do a recursive trim in the hope that this will keep 7667 // stack usage lower, but leave some oops for potential stealers 7668 trim_queue(_low_water_mark); 7669 } // Else, another thread got there first 7670 } 7671 } 7672 7673 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7674 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7675 7676 void CMSParKeepAliveClosure::trim_queue(uint max) { 7677 while (_work_queue->size() > max) { 7678 oop new_oop; 7679 if (_work_queue->pop_local(new_oop)) { 7680 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); 7681 assert(_bit_map->isMarked((HeapWord*)new_oop), 7682 "no white objects on this stack!"); 7683 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7684 // iterate over the oops in this oop, marking and pushing 7685 // the ones in CMS heap (i.e. in _span). 7686 new_oop->oop_iterate(&_mark_and_push); 7687 } 7688 } 7689 } 7690 7691 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7692 CMSCollector* collector, 7693 MemRegion span, CMSBitMap* bit_map, 7694 OopTaskQueue* work_queue): 7695 _collector(collector), 7696 _span(span), 7697 _bit_map(bit_map), 7698 _work_queue(work_queue) { } 7699 7700 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7701 HeapWord* addr = (HeapWord*)obj; 7702 if (_span.contains(addr) && 7703 !_bit_map->isMarked(addr)) { 7704 if (_bit_map->par_mark(addr)) { 7705 bool simulate_overflow = false; 7706 NOT_PRODUCT( 7707 if (CMSMarkStackOverflowALot && 7708 _collector->par_simulate_overflow()) { 7709 // simulate a stack overflow 7710 simulate_overflow = true; 7711 } 7712 ) 7713 if (simulate_overflow || !_work_queue->push(obj)) { 7714 _collector->par_push_on_overflow_list(obj); 7715 _collector->_par_kac_ovflw++; 7716 } 7717 } // Else another thread got there already 7718 } 7719 } 7720 7721 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7722 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7723 7724 ////////////////////////////////////////////////////////////////// 7725 // CMSExpansionCause ///////////////////////////// 7726 ////////////////////////////////////////////////////////////////// 7727 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7728 switch (cause) { 7729 case _no_expansion: 7730 return "No expansion"; 7731 case _satisfy_free_ratio: 7732 return "Free ratio"; 7733 case _satisfy_promotion: 7734 return "Satisfy promotion"; 7735 case _satisfy_allocation: 7736 return "allocation"; 7737 case _allocate_par_lab: 7738 return "Par LAB"; 7739 case _allocate_par_spooling_space: 7740 return "Par Spooling Space"; 7741 case _adaptive_size_policy: 7742 return "Ergonomics"; 7743 default: 7744 return "unknown"; 7745 } 7746 } 7747 7748 void CMSDrainMarkingStackClosure::do_void() { 7749 // the max number to take from overflow list at a time 7750 const size_t num = _mark_stack->capacity()/4; 7751 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7752 "Overflow list should be NULL during concurrent phases"); 7753 while (!_mark_stack->isEmpty() || 7754 // if stack is empty, check the overflow list 7755 _collector->take_from_overflow_list(num, _mark_stack)) { 7756 oop obj = _mark_stack->pop(); 7757 HeapWord* addr = (HeapWord*)obj; 7758 assert(_span.contains(addr), "Should be within span"); 7759 assert(_bit_map->isMarked(addr), "Should be marked"); 7760 assert(obj->is_oop(), "Should be an oop"); 7761 obj->oop_iterate(_keep_alive); 7762 } 7763 } 7764 7765 void CMSParDrainMarkingStackClosure::do_void() { 7766 // drain queue 7767 trim_queue(0); 7768 } 7769 7770 // Trim our work_queue so its length is below max at return 7771 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7772 while (_work_queue->size() > max) { 7773 oop new_oop; 7774 if (_work_queue->pop_local(new_oop)) { 7775 assert(new_oop->is_oop(), "Expected an oop"); 7776 assert(_bit_map->isMarked((HeapWord*)new_oop), 7777 "no white objects on this stack!"); 7778 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7779 // iterate over the oops in this oop, marking and pushing 7780 // the ones in CMS heap (i.e. in _span). 7781 new_oop->oop_iterate(&_mark_and_push); 7782 } 7783 } 7784 } 7785 7786 //////////////////////////////////////////////////////////////////// 7787 // Support for Marking Stack Overflow list handling and related code 7788 //////////////////////////////////////////////////////////////////// 7789 // Much of the following code is similar in shape and spirit to the 7790 // code used in ParNewGC. We should try and share that code 7791 // as much as possible in the future. 7792 7793 #ifndef PRODUCT 7794 // Debugging support for CMSStackOverflowALot 7795 7796 // It's OK to call this multi-threaded; the worst thing 7797 // that can happen is that we'll get a bunch of closely 7798 // spaced simulated overflows, but that's OK, in fact 7799 // probably good as it would exercise the overflow code 7800 // under contention. 7801 bool CMSCollector::simulate_overflow() { 7802 if (_overflow_counter-- <= 0) { // just being defensive 7803 _overflow_counter = CMSMarkStackOverflowInterval; 7804 return true; 7805 } else { 7806 return false; 7807 } 7808 } 7809 7810 bool CMSCollector::par_simulate_overflow() { 7811 return simulate_overflow(); 7812 } 7813 #endif 7814 7815 // Single-threaded 7816 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7817 assert(stack->isEmpty(), "Expected precondition"); 7818 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7819 size_t i = num; 7820 oop cur = _overflow_list; 7821 const markOop proto = markOopDesc::prototype(); 7822 NOT_PRODUCT(ssize_t n = 0;) 7823 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7824 next = oop(cur->mark()); 7825 cur->set_mark(proto); // until proven otherwise 7826 assert(cur->is_oop(), "Should be an oop"); 7827 bool res = stack->push(cur); 7828 assert(res, "Bit off more than can chew?"); 7829 NOT_PRODUCT(n++;) 7830 } 7831 _overflow_list = cur; 7832 #ifndef PRODUCT 7833 assert(_num_par_pushes >= n, "Too many pops?"); 7834 _num_par_pushes -=n; 7835 #endif 7836 return !stack->isEmpty(); 7837 } 7838 7839 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7840 // (MT-safe) Get a prefix of at most "num" from the list. 7841 // The overflow list is chained through the mark word of 7842 // each object in the list. We fetch the entire list, 7843 // break off a prefix of the right size and return the 7844 // remainder. If other threads try to take objects from 7845 // the overflow list at that time, they will wait for 7846 // some time to see if data becomes available. If (and 7847 // only if) another thread places one or more object(s) 7848 // on the global list before we have returned the suffix 7849 // to the global list, we will walk down our local list 7850 // to find its end and append the global list to 7851 // our suffix before returning it. This suffix walk can 7852 // prove to be expensive (quadratic in the amount of traffic) 7853 // when there are many objects in the overflow list and 7854 // there is much producer-consumer contention on the list. 7855 // *NOTE*: The overflow list manipulation code here and 7856 // in ParNewGeneration:: are very similar in shape, 7857 // except that in the ParNew case we use the old (from/eden) 7858 // copy of the object to thread the list via its klass word. 7859 // Because of the common code, if you make any changes in 7860 // the code below, please check the ParNew version to see if 7861 // similar changes might be needed. 7862 // CR 6797058 has been filed to consolidate the common code. 7863 bool CMSCollector::par_take_from_overflow_list(size_t num, 7864 OopTaskQueue* work_q, 7865 int no_of_gc_threads) { 7866 assert(work_q->size() == 0, "First empty local work queue"); 7867 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7868 if (_overflow_list == NULL) { 7869 return false; 7870 } 7871 // Grab the entire list; we'll put back a suffix 7872 oop prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7873 Thread* tid = Thread::current(); 7874 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7875 // set to ParallelGCThreads. 7876 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7877 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7878 // If the list is busy, we spin for a short while, 7879 // sleeping between attempts to get the list. 7880 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7881 os::sleep(tid, sleep_time_millis, false); 7882 if (_overflow_list == NULL) { 7883 // Nothing left to take 7884 return false; 7885 } else if (_overflow_list != BUSY) { 7886 // Try and grab the prefix 7887 prefix = cast_to_oop(Atomic::xchg_ptr(BUSY, &_overflow_list)); 7888 } 7889 } 7890 // If the list was found to be empty, or we spun long 7891 // enough, we give up and return empty-handed. If we leave 7892 // the list in the BUSY state below, it must be the case that 7893 // some other thread holds the overflow list and will set it 7894 // to a non-BUSY state in the future. 7895 if (prefix == NULL || prefix == BUSY) { 7896 // Nothing to take or waited long enough 7897 if (prefix == NULL) { 7898 // Write back the NULL in case we overwrote it with BUSY above 7899 // and it is still the same value. 7900 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7901 } 7902 return false; 7903 } 7904 assert(prefix != NULL && prefix != BUSY, "Error"); 7905 size_t i = num; 7906 oop cur = prefix; 7907 // Walk down the first "num" objects, unless we reach the end. 7908 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); 7909 if (cur->mark() == NULL) { 7910 // We have "num" or fewer elements in the list, so there 7911 // is nothing to return to the global list. 7912 // Write back the NULL in lieu of the BUSY we wrote 7913 // above, if it is still the same value. 7914 if (_overflow_list == BUSY) { 7915 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); 7916 } 7917 } else { 7918 // Chop off the suffix and return it to the global list. 7919 assert(cur->mark() != BUSY, "Error"); 7920 oop suffix_head = cur->mark(); // suffix will be put back on global list 7921 cur->set_mark(NULL); // break off suffix 7922 // It's possible that the list is still in the empty(busy) state 7923 // we left it in a short while ago; in that case we may be 7924 // able to place back the suffix without incurring the cost 7925 // of a walk down the list. 7926 oop observed_overflow_list = _overflow_list; 7927 oop cur_overflow_list = observed_overflow_list; 7928 bool attached = false; 7929 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7930 observed_overflow_list = 7931 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7932 if (cur_overflow_list == observed_overflow_list) { 7933 attached = true; 7934 break; 7935 } else cur_overflow_list = observed_overflow_list; 7936 } 7937 if (!attached) { 7938 // Too bad, someone else sneaked in (at least) an element; we'll need 7939 // to do a splice. Find tail of suffix so we can prepend suffix to global 7940 // list. 7941 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); 7942 oop suffix_tail = cur; 7943 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, 7944 "Tautology"); 7945 observed_overflow_list = _overflow_list; 7946 do { 7947 cur_overflow_list = observed_overflow_list; 7948 if (cur_overflow_list != BUSY) { 7949 // Do the splice ... 7950 suffix_tail->set_mark(markOop(cur_overflow_list)); 7951 } else { // cur_overflow_list == BUSY 7952 suffix_tail->set_mark(NULL); 7953 } 7954 // ... and try to place spliced list back on overflow_list ... 7955 observed_overflow_list = 7956 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); 7957 } while (cur_overflow_list != observed_overflow_list); 7958 // ... until we have succeeded in doing so. 7959 } 7960 } 7961 7962 // Push the prefix elements on work_q 7963 assert(prefix != NULL, "control point invariant"); 7964 const markOop proto = markOopDesc::prototype(); 7965 oop next; 7966 NOT_PRODUCT(ssize_t n = 0;) 7967 for (cur = prefix; cur != NULL; cur = next) { 7968 next = oop(cur->mark()); 7969 cur->set_mark(proto); // until proven otherwise 7970 assert(cur->is_oop(), "Should be an oop"); 7971 bool res = work_q->push(cur); 7972 assert(res, "Bit off more than we can chew?"); 7973 NOT_PRODUCT(n++;) 7974 } 7975 #ifndef PRODUCT 7976 assert(_num_par_pushes >= n, "Too many pops?"); 7977 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); 7978 #endif 7979 return true; 7980 } 7981 7982 // Single-threaded 7983 void CMSCollector::push_on_overflow_list(oop p) { 7984 NOT_PRODUCT(_num_par_pushes++;) 7985 assert(p->is_oop(), "Not an oop"); 7986 preserve_mark_if_necessary(p); 7987 p->set_mark((markOop)_overflow_list); 7988 _overflow_list = p; 7989 } 7990 7991 // Multi-threaded; use CAS to prepend to overflow list 7992 void CMSCollector::par_push_on_overflow_list(oop p) { 7993 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) 7994 assert(p->is_oop(), "Not an oop"); 7995 par_preserve_mark_if_necessary(p); 7996 oop observed_overflow_list = _overflow_list; 7997 oop cur_overflow_list; 7998 do { 7999 cur_overflow_list = observed_overflow_list; 8000 if (cur_overflow_list != BUSY) { 8001 p->set_mark(markOop(cur_overflow_list)); 8002 } else { 8003 p->set_mark(NULL); 8004 } 8005 observed_overflow_list = 8006 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); 8007 } while (cur_overflow_list != observed_overflow_list); 8008 } 8009 #undef BUSY 8010 8011 // Single threaded 8012 // General Note on GrowableArray: pushes may silently fail 8013 // because we are (temporarily) out of C-heap for expanding 8014 // the stack. The problem is quite ubiquitous and affects 8015 // a lot of code in the JVM. The prudent thing for GrowableArray 8016 // to do (for now) is to exit with an error. However, that may 8017 // be too draconian in some cases because the caller may be 8018 // able to recover without much harm. For such cases, we 8019 // should probably introduce a "soft_push" method which returns 8020 // an indication of success or failure with the assumption that 8021 // the caller may be able to recover from a failure; code in 8022 // the VM can then be changed, incrementally, to deal with such 8023 // failures where possible, thus, incrementally hardening the VM 8024 // in such low resource situations. 8025 void CMSCollector::preserve_mark_work(oop p, markOop m) { 8026 _preserved_oop_stack.push(p); 8027 _preserved_mark_stack.push(m); 8028 assert(m == p->mark(), "Mark word changed"); 8029 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8030 "bijection"); 8031 } 8032 8033 // Single threaded 8034 void CMSCollector::preserve_mark_if_necessary(oop p) { 8035 markOop m = p->mark(); 8036 if (m->must_be_preserved(p)) { 8037 preserve_mark_work(p, m); 8038 } 8039 } 8040 8041 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8042 markOop m = p->mark(); 8043 if (m->must_be_preserved(p)) { 8044 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8045 // Even though we read the mark word without holding 8046 // the lock, we are assured that it will not change 8047 // because we "own" this oop, so no other thread can 8048 // be trying to push it on the overflow list; see 8049 // the assertion in preserve_mark_work() that checks 8050 // that m == p->mark(). 8051 preserve_mark_work(p, m); 8052 } 8053 } 8054 8055 // We should be able to do this multi-threaded, 8056 // a chunk of stack being a task (this is 8057 // correct because each oop only ever appears 8058 // once in the overflow list. However, it's 8059 // not very easy to completely overlap this with 8060 // other operations, so will generally not be done 8061 // until all work's been completed. Because we 8062 // expect the preserved oop stack (set) to be small, 8063 // it's probably fine to do this single-threaded. 8064 // We can explore cleverer concurrent/overlapped/parallel 8065 // processing of preserved marks if we feel the 8066 // need for this in the future. Stack overflow should 8067 // be so rare in practice and, when it happens, its 8068 // effect on performance so great that this will 8069 // likely just be in the noise anyway. 8070 void CMSCollector::restore_preserved_marks_if_any() { 8071 assert(SafepointSynchronize::is_at_safepoint(), 8072 "world should be stopped"); 8073 assert(Thread::current()->is_ConcurrentGC_thread() || 8074 Thread::current()->is_VM_thread(), 8075 "should be single-threaded"); 8076 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8077 "bijection"); 8078 8079 while (!_preserved_oop_stack.is_empty()) { 8080 oop p = _preserved_oop_stack.pop(); 8081 assert(p->is_oop(), "Should be an oop"); 8082 assert(_span.contains(p), "oop should be in _span"); 8083 assert(p->mark() == markOopDesc::prototype(), 8084 "Set when taken from overflow list"); 8085 markOop m = _preserved_mark_stack.pop(); 8086 p->set_mark(m); 8087 } 8088 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8089 "stacks were cleared above"); 8090 } 8091 8092 #ifndef PRODUCT 8093 bool CMSCollector::no_preserved_marks() const { 8094 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8095 } 8096 #endif 8097 8098 // Transfer some number of overflown objects to usual marking 8099 // stack. Return true if some objects were transferred. 8100 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8101 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8102 (size_t)ParGCDesiredObjsFromOverflowList); 8103 8104 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8105 assert(_collector->overflow_list_is_empty() || res, 8106 "If list is not empty, we should have taken something"); 8107 assert(!res || !_mark_stack->isEmpty(), 8108 "If we took something, it should now be on our stack"); 8109 return res; 8110 } 8111 8112 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8113 size_t res = _sp->block_size_no_stall(addr, _collector); 8114 if (_sp->block_is_obj(addr)) { 8115 if (_live_bit_map->isMarked(addr)) { 8116 // It can't have been dead in a previous cycle 8117 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8118 } else { 8119 _dead_bit_map->mark(addr); // mark the dead object 8120 } 8121 } 8122 // Could be 0, if the block size could not be computed without stalling. 8123 return res; 8124 } 8125 8126 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8127 8128 switch (phase) { 8129 case CMSCollector::InitialMarking: 8130 initialize(true /* fullGC */ , 8131 cause /* cause of the GC */, 8132 true /* recordGCBeginTime */, 8133 true /* recordPreGCUsage */, 8134 false /* recordPeakUsage */, 8135 false /* recordPostGCusage */, 8136 true /* recordAccumulatedGCTime */, 8137 false /* recordGCEndTime */, 8138 false /* countCollection */ ); 8139 break; 8140 8141 case CMSCollector::FinalMarking: 8142 initialize(true /* fullGC */ , 8143 cause /* cause of the GC */, 8144 false /* recordGCBeginTime */, 8145 false /* recordPreGCUsage */, 8146 false /* recordPeakUsage */, 8147 false /* recordPostGCusage */, 8148 true /* recordAccumulatedGCTime */, 8149 false /* recordGCEndTime */, 8150 false /* countCollection */ ); 8151 break; 8152 8153 case CMSCollector::Sweeping: 8154 initialize(true /* fullGC */ , 8155 cause /* cause of the GC */, 8156 false /* recordGCBeginTime */, 8157 false /* recordPreGCUsage */, 8158 true /* recordPeakUsage */, 8159 true /* recordPostGCusage */, 8160 false /* recordAccumulatedGCTime */, 8161 true /* recordGCEndTime */, 8162 true /* countCollection */ ); 8163 break; 8164 8165 default: 8166 ShouldNotReachHere(); 8167 } 8168 }