1 /* 2 * Copyright (c) 2001, 2018, 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/cmsGCStats.hpp" 33 #include "gc/cms/cmsHeap.hpp" 34 #include "gc/cms/cmsOopClosures.inline.hpp" 35 #include "gc/cms/compactibleFreeListSpace.hpp" 36 #include "gc/cms/concurrentMarkSweepGeneration.inline.hpp" 37 #include "gc/cms/concurrentMarkSweepThread.hpp" 38 #include "gc/cms/parNewGeneration.hpp" 39 #include "gc/cms/promotionInfo.inline.hpp" 40 #include "gc/cms/vmCMSOperations.hpp" 41 #include "gc/serial/genMarkSweep.hpp" 42 #include "gc/serial/tenuredGeneration.hpp" 43 #include "gc/shared/adaptiveSizePolicy.hpp" 44 #include "gc/shared/cardGeneration.inline.hpp" 45 #include "gc/shared/cardTableRS.hpp" 46 #include "gc/shared/collectedHeap.inline.hpp" 47 #include "gc/shared/collectorCounters.hpp" 48 #include "gc/shared/collectorPolicy.hpp" 49 #include "gc/shared/gcLocker.hpp" 50 #include "gc/shared/gcPolicyCounters.hpp" 51 #include "gc/shared/gcTimer.hpp" 52 #include "gc/shared/gcTrace.hpp" 53 #include "gc/shared/gcTraceTime.inline.hpp" 54 #include "gc/shared/genCollectedHeap.hpp" 55 #include "gc/shared/genOopClosures.inline.hpp" 56 #include "gc/shared/isGCActiveMark.hpp" 57 #include "gc/shared/oopStorageParState.hpp" 58 #include "gc/shared/referencePolicy.hpp" 59 #include "gc/shared/referenceProcessorPhaseTimes.hpp" 60 #include "gc/shared/space.inline.hpp" 61 #include "gc/shared/strongRootsScope.hpp" 62 #include "gc/shared/taskqueue.inline.hpp" 63 #include "gc/shared/weakProcessor.hpp" 64 #include "logging/log.hpp" 65 #include "logging/logStream.hpp" 66 #include "memory/allocation.hpp" 67 #include "memory/binaryTreeDictionary.inline.hpp" 68 #include "memory/iterator.inline.hpp" 69 #include "memory/padded.hpp" 70 #include "memory/resourceArea.hpp" 71 #include "oops/access.inline.hpp" 72 #include "oops/oop.inline.hpp" 73 #include "prims/jvmtiExport.hpp" 74 #include "runtime/atomic.hpp" 75 #include "runtime/flags/flagSetting.hpp" 76 #include "runtime/globals_extension.hpp" 77 #include "runtime/handles.inline.hpp" 78 #include "runtime/java.hpp" 79 #include "runtime/orderAccess.hpp" 80 #include "runtime/timer.hpp" 81 #include "runtime/vmThread.hpp" 82 #include "services/memoryService.hpp" 83 #include "services/runtimeService.hpp" 84 #include "utilities/align.hpp" 85 #include "utilities/stack.inline.hpp" 86 87 // statics 88 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 89 bool CMSCollector::_full_gc_requested = false; 90 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 91 92 ////////////////////////////////////////////////////////////////// 93 // In support of CMS/VM thread synchronization 94 ////////////////////////////////////////////////////////////////// 95 // We split use of the CGC_lock into 2 "levels". 96 // The low-level locking is of the usual CGC_lock monitor. We introduce 97 // a higher level "token" (hereafter "CMS token") built on top of the 98 // low level monitor (hereafter "CGC lock"). 99 // The token-passing protocol gives priority to the VM thread. The 100 // CMS-lock doesn't provide any fairness guarantees, but clients 101 // should ensure that it is only held for very short, bounded 102 // durations. 103 // 104 // When either of the CMS thread or the VM thread is involved in 105 // collection operations during which it does not want the other 106 // thread to interfere, it obtains the CMS token. 107 // 108 // If either thread tries to get the token while the other has 109 // it, that thread waits. However, if the VM thread and CMS thread 110 // both want the token, then the VM thread gets priority while the 111 // CMS thread waits. This ensures, for instance, that the "concurrent" 112 // phases of the CMS thread's work do not block out the VM thread 113 // for long periods of time as the CMS thread continues to hog 114 // the token. (See bug 4616232). 115 // 116 // The baton-passing functions are, however, controlled by the 117 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 118 // and here the low-level CMS lock, not the high level token, 119 // ensures mutual exclusion. 120 // 121 // Two important conditions that we have to satisfy: 122 // 1. if a thread does a low-level wait on the CMS lock, then it 123 // relinquishes the CMS token if it were holding that token 124 // when it acquired the low-level CMS lock. 125 // 2. any low-level notifications on the low-level lock 126 // should only be sent when a thread has relinquished the token. 127 // 128 // In the absence of either property, we'd have potential deadlock. 129 // 130 // We protect each of the CMS (concurrent and sequential) phases 131 // with the CMS _token_, not the CMS _lock_. 132 // 133 // The only code protected by CMS lock is the token acquisition code 134 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 135 // baton-passing code. 136 // 137 // Unfortunately, i couldn't come up with a good abstraction to factor and 138 // hide the naked CGC_lock manipulation in the baton-passing code 139 // further below. That's something we should try to do. Also, the proof 140 // of correctness of this 2-level locking scheme is far from obvious, 141 // and potentially quite slippery. We have an uneasy suspicion, for instance, 142 // that there may be a theoretical possibility of delay/starvation in the 143 // low-level lock/wait/notify scheme used for the baton-passing because of 144 // potential interference with the priority scheme embodied in the 145 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 146 // invocation further below and marked with "XXX 20011219YSR". 147 // Indeed, as we note elsewhere, this may become yet more slippery 148 // in the presence of multiple CMS and/or multiple VM threads. XXX 149 150 class CMSTokenSync: public StackObj { 151 private: 152 bool _is_cms_thread; 153 public: 154 CMSTokenSync(bool is_cms_thread): 155 _is_cms_thread(is_cms_thread) { 156 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 157 "Incorrect argument to constructor"); 158 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 159 } 160 161 ~CMSTokenSync() { 162 assert(_is_cms_thread ? 163 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 164 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 165 "Incorrect state"); 166 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 167 } 168 }; 169 170 // Convenience class that does a CMSTokenSync, and then acquires 171 // upto three locks. 172 class CMSTokenSyncWithLocks: public CMSTokenSync { 173 private: 174 // Note: locks are acquired in textual declaration order 175 // and released in the opposite order 176 MutexLockerEx _locker1, _locker2, _locker3; 177 public: 178 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 179 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 180 CMSTokenSync(is_cms_thread), 181 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 182 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 183 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 184 { } 185 }; 186 187 188 ////////////////////////////////////////////////////////////////// 189 // Concurrent Mark-Sweep Generation ///////////////////////////// 190 ////////////////////////////////////////////////////////////////// 191 192 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 193 194 // This struct contains per-thread things necessary to support parallel 195 // young-gen collection. 196 class CMSParGCThreadState: public CHeapObj<mtGC> { 197 public: 198 CompactibleFreeListSpaceLAB lab; 199 PromotionInfo promo; 200 201 // Constructor. 202 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 203 promo.setSpace(cfls); 204 } 205 }; 206 207 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 208 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 209 CardGeneration(rs, initial_byte_size, ct), 210 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 211 _did_compact(false) 212 { 213 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 214 HeapWord* end = (HeapWord*) _virtual_space.high(); 215 216 _direct_allocated_words = 0; 217 NOT_PRODUCT( 218 _numObjectsPromoted = 0; 219 _numWordsPromoted = 0; 220 _numObjectsAllocated = 0; 221 _numWordsAllocated = 0; 222 ) 223 224 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 225 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 226 _cmsSpace->_old_gen = this; 227 228 _gc_stats = new CMSGCStats(); 229 230 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 231 // offsets match. The ability to tell free chunks from objects 232 // depends on this property. 233 debug_only( 234 FreeChunk* junk = NULL; 235 assert(UseCompressedClassPointers || 236 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 237 "Offset of FreeChunk::_prev within FreeChunk must match" 238 " that of OopDesc::_klass within OopDesc"); 239 ) 240 241 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 242 for (uint i = 0; i < ParallelGCThreads; i++) { 243 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 244 } 245 246 _incremental_collection_failed = false; 247 // The "dilatation_factor" is the expansion that can occur on 248 // account of the fact that the minimum object size in the CMS 249 // generation may be larger than that in, say, a contiguous young 250 // generation. 251 // Ideally, in the calculation below, we'd compute the dilatation 252 // factor as: MinChunkSize/(promoting_gen's min object size) 253 // Since we do not have such a general query interface for the 254 // promoting generation, we'll instead just use the minimum 255 // object size (which today is a header's worth of space); 256 // note that all arithmetic is in units of HeapWords. 257 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 258 assert(_dilatation_factor >= 1.0, "from previous assert"); 259 } 260 261 262 // The field "_initiating_occupancy" represents the occupancy percentage 263 // at which we trigger a new collection cycle. Unless explicitly specified 264 // via CMSInitiatingOccupancyFraction (argument "io" below), it 265 // is calculated by: 266 // 267 // Let "f" be MinHeapFreeRatio in 268 // 269 // _initiating_occupancy = 100-f + 270 // f * (CMSTriggerRatio/100) 271 // where CMSTriggerRatio is the argument "tr" below. 272 // 273 // That is, if we assume the heap is at its desired maximum occupancy at the 274 // end of a collection, we let CMSTriggerRatio of the (purported) free 275 // space be allocated before initiating a new collection cycle. 276 // 277 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 278 assert(io <= 100 && tr <= 100, "Check the arguments"); 279 if (io >= 0) { 280 _initiating_occupancy = (double)io / 100.0; 281 } else { 282 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 283 (double)(tr * MinHeapFreeRatio) / 100.0) 284 / 100.0; 285 } 286 } 287 288 void ConcurrentMarkSweepGeneration::ref_processor_init() { 289 assert(collector() != NULL, "no collector"); 290 collector()->ref_processor_init(); 291 } 292 293 void CMSCollector::ref_processor_init() { 294 if (_ref_processor == NULL) { 295 // Allocate and initialize a reference processor 296 _ref_processor = 297 new ReferenceProcessor(&_span_based_discoverer, 298 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 299 ParallelGCThreads, // mt processing degree 300 _cmsGen->refs_discovery_is_mt(), // mt discovery 301 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 302 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 303 &_is_alive_closure, // closure for liveness info 304 false); // disable adjusting number of processing threads 305 // Initialize the _ref_processor field of CMSGen 306 _cmsGen->set_ref_processor(_ref_processor); 307 308 } 309 } 310 311 AdaptiveSizePolicy* CMSCollector::size_policy() { 312 return CMSHeap::heap()->size_policy(); 313 } 314 315 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 316 317 const char* gen_name = "old"; 318 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 319 // Generation Counters - generation 1, 1 subspace 320 _gen_counters = new GenerationCounters(gen_name, 1, 1, 321 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 322 323 _space_counters = new GSpaceCounters(gen_name, 0, 324 _virtual_space.reserved_size(), 325 this, _gen_counters); 326 } 327 328 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 329 _cms_gen(cms_gen) 330 { 331 assert(alpha <= 100, "bad value"); 332 _saved_alpha = alpha; 333 334 // Initialize the alphas to the bootstrap value of 100. 335 _gc0_alpha = _cms_alpha = 100; 336 337 _cms_begin_time.update(); 338 _cms_end_time.update(); 339 340 _gc0_duration = 0.0; 341 _gc0_period = 0.0; 342 _gc0_promoted = 0; 343 344 _cms_duration = 0.0; 345 _cms_period = 0.0; 346 _cms_allocated = 0; 347 348 _cms_used_at_gc0_begin = 0; 349 _cms_used_at_gc0_end = 0; 350 _allow_duty_cycle_reduction = false; 351 _valid_bits = 0; 352 } 353 354 double CMSStats::cms_free_adjustment_factor(size_t free) const { 355 // TBD: CR 6909490 356 return 1.0; 357 } 358 359 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 360 } 361 362 // If promotion failure handling is on use 363 // the padded average size of the promotion for each 364 // young generation collection. 365 double CMSStats::time_until_cms_gen_full() const { 366 size_t cms_free = _cms_gen->cmsSpace()->free(); 367 CMSHeap* heap = CMSHeap::heap(); 368 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 369 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 370 if (cms_free > expected_promotion) { 371 // Start a cms collection if there isn't enough space to promote 372 // for the next young collection. Use the padded average as 373 // a safety factor. 374 cms_free -= expected_promotion; 375 376 // Adjust by the safety factor. 377 double cms_free_dbl = (double)cms_free; 378 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 379 // Apply a further correction factor which tries to adjust 380 // for recent occurance of concurrent mode failures. 381 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 382 cms_free_dbl = cms_free_dbl * cms_adjustment; 383 384 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 385 cms_free, expected_promotion); 386 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 387 // Add 1 in case the consumption rate goes to zero. 388 return cms_free_dbl / (cms_consumption_rate() + 1.0); 389 } 390 return 0.0; 391 } 392 393 // Compare the duration of the cms collection to the 394 // time remaining before the cms generation is empty. 395 // Note that the time from the start of the cms collection 396 // to the start of the cms sweep (less than the total 397 // duration of the cms collection) can be used. This 398 // has been tried and some applications experienced 399 // promotion failures early in execution. This was 400 // possibly because the averages were not accurate 401 // enough at the beginning. 402 double CMSStats::time_until_cms_start() const { 403 // We add "gc0_period" to the "work" calculation 404 // below because this query is done (mostly) at the 405 // end of a scavenge, so we need to conservatively 406 // account for that much possible delay 407 // in the query so as to avoid concurrent mode failures 408 // due to starting the collection just a wee bit too 409 // late. 410 double work = cms_duration() + gc0_period(); 411 double deadline = time_until_cms_gen_full(); 412 // If a concurrent mode failure occurred recently, we want to be 413 // more conservative and halve our expected time_until_cms_gen_full() 414 if (work > deadline) { 415 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 416 cms_duration(), gc0_period(), time_until_cms_gen_full()); 417 return 0.0; 418 } 419 return work - deadline; 420 } 421 422 #ifndef PRODUCT 423 void CMSStats::print_on(outputStream *st) const { 424 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 425 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 426 gc0_duration(), gc0_period(), gc0_promoted()); 427 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 428 cms_duration(), cms_period(), cms_allocated()); 429 st->print(",cms_since_beg=%g,cms_since_end=%g", 430 cms_time_since_begin(), cms_time_since_end()); 431 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 432 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 433 434 if (valid()) { 435 st->print(",promo_rate=%g,cms_alloc_rate=%g", 436 promotion_rate(), cms_allocation_rate()); 437 st->print(",cms_consumption_rate=%g,time_until_full=%g", 438 cms_consumption_rate(), time_until_cms_gen_full()); 439 } 440 st->cr(); 441 } 442 #endif // #ifndef PRODUCT 443 444 CMSCollector::CollectorState CMSCollector::_collectorState = 445 CMSCollector::Idling; 446 bool CMSCollector::_foregroundGCIsActive = false; 447 bool CMSCollector::_foregroundGCShouldWait = false; 448 449 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 450 CardTableRS* ct, 451 ConcurrentMarkSweepPolicy* cp): 452 _cmsGen(cmsGen), 453 // Adjust span to cover old (cms) gen 454 _span(cmsGen->reserved()), 455 _ct(ct), 456 _span_based_discoverer(_span), 457 _ref_processor(NULL), // will be set later 458 _conc_workers(NULL), // may be set later 459 _abort_preclean(false), 460 _start_sampling(false), 461 _between_prologue_and_epilogue(false), 462 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 463 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 464 -1 /* lock-free */, "No_lock" /* dummy */), 465 _modUnionClosurePar(&_modUnionTable), 466 // Construct the is_alive_closure with _span & markBitMap 467 _is_alive_closure(_span, &_markBitMap), 468 _restart_addr(NULL), 469 _overflow_list(NULL), 470 _stats(cmsGen), 471 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 472 //verify that this lock should be acquired with safepoint check. 473 Monitor::_safepoint_check_sometimes)), 474 _eden_chunk_array(NULL), // may be set in ctor body 475 _eden_chunk_capacity(0), // -- ditto -- 476 _eden_chunk_index(0), // -- ditto -- 477 _survivor_plab_array(NULL), // -- ditto -- 478 _survivor_chunk_array(NULL), // -- ditto -- 479 _survivor_chunk_capacity(0), // -- ditto -- 480 _survivor_chunk_index(0), // -- ditto -- 481 _ser_pmc_preclean_ovflw(0), 482 _ser_kac_preclean_ovflw(0), 483 _ser_pmc_remark_ovflw(0), 484 _par_pmc_remark_ovflw(0), 485 _ser_kac_ovflw(0), 486 _par_kac_ovflw(0), 487 #ifndef PRODUCT 488 _num_par_pushes(0), 489 #endif 490 _collection_count_start(0), 491 _verifying(false), 492 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 493 _completed_initialization(false), 494 _collector_policy(cp), 495 _should_unload_classes(CMSClassUnloadingEnabled), 496 _concurrent_cycles_since_last_unload(0), 497 _roots_scanning_options(GenCollectedHeap::SO_None), 498 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 499 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 500 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 501 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 502 _cms_start_registered(false) 503 { 504 // Now expand the span and allocate the collection support structures 505 // (MUT, marking bit map etc.) to cover both generations subject to 506 // collection. 507 508 // For use by dirty card to oop closures. 509 _cmsGen->cmsSpace()->set_collector(this); 510 511 // Allocate MUT and marking bit map 512 { 513 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 514 if (!_markBitMap.allocate(_span)) { 515 log_warning(gc)("Failed to allocate CMS Bit Map"); 516 return; 517 } 518 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 519 } 520 { 521 _modUnionTable.allocate(_span); 522 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 523 } 524 525 if (!_markStack.allocate(MarkStackSize)) { 526 log_warning(gc)("Failed to allocate CMS Marking Stack"); 527 return; 528 } 529 530 // Support for multi-threaded concurrent phases 531 if (CMSConcurrentMTEnabled) { 532 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 533 // just for now 534 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 535 } 536 if (ConcGCThreads > 1) { 537 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 538 ConcGCThreads, true); 539 if (_conc_workers == NULL) { 540 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 541 CMSConcurrentMTEnabled = false; 542 } else { 543 _conc_workers->initialize_workers(); 544 } 545 } else { 546 CMSConcurrentMTEnabled = false; 547 } 548 } 549 if (!CMSConcurrentMTEnabled) { 550 ConcGCThreads = 0; 551 } else { 552 // Turn off CMSCleanOnEnter optimization temporarily for 553 // the MT case where it's not fixed yet; see 6178663. 554 CMSCleanOnEnter = false; 555 } 556 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 557 "Inconsistency"); 558 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 559 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 560 561 // Parallel task queues; these are shared for the 562 // concurrent and stop-world phases of CMS, but 563 // are not shared with parallel scavenge (ParNew). 564 { 565 uint i; 566 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 567 568 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 569 || ParallelRefProcEnabled) 570 && num_queues > 0) { 571 _task_queues = new OopTaskQueueSet(num_queues); 572 if (_task_queues == NULL) { 573 log_warning(gc)("task_queues allocation failure."); 574 return; 575 } 576 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC); 577 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 578 for (i = 0; i < num_queues; i++) { 579 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 580 if (q == NULL) { 581 log_warning(gc)("work_queue allocation failure."); 582 return; 583 } 584 _task_queues->register_queue(i, q); 585 } 586 for (i = 0; i < num_queues; i++) { 587 _task_queues->queue(i)->initialize(); 588 _hash_seed[i] = 17; // copied from ParNew 589 } 590 } 591 } 592 593 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 594 595 // Clip CMSBootstrapOccupancy between 0 and 100. 596 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 597 598 // Now tell CMS generations the identity of their collector 599 ConcurrentMarkSweepGeneration::set_collector(this); 600 601 // Create & start a CMS thread for this CMS collector 602 _cmsThread = ConcurrentMarkSweepThread::start(this); 603 assert(cmsThread() != NULL, "CMS Thread should have been created"); 604 assert(cmsThread()->collector() == this, 605 "CMS Thread should refer to this gen"); 606 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 607 608 // Support for parallelizing young gen rescan 609 CMSHeap* heap = CMSHeap::heap(); 610 _young_gen = heap->young_gen(); 611 if (heap->supports_inline_contig_alloc()) { 612 _top_addr = heap->top_addr(); 613 _end_addr = heap->end_addr(); 614 assert(_young_gen != NULL, "no _young_gen"); 615 _eden_chunk_index = 0; 616 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 617 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 618 } 619 620 // Support for parallelizing survivor space rescan 621 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 622 const size_t max_plab_samples = 623 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 624 625 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 626 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 627 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 628 _survivor_chunk_capacity = max_plab_samples; 629 for (uint i = 0; i < ParallelGCThreads; i++) { 630 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 631 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 632 assert(cur->end() == 0, "Should be 0"); 633 assert(cur->array() == vec, "Should be vec"); 634 assert(cur->capacity() == max_plab_samples, "Error"); 635 } 636 } 637 638 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 639 _gc_counters = new CollectorCounters("CMS", 1); 640 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 641 _completed_initialization = true; 642 _inter_sweep_timer.start(); // start of time 643 } 644 645 const char* ConcurrentMarkSweepGeneration::name() const { 646 return "concurrent mark-sweep generation"; 647 } 648 void ConcurrentMarkSweepGeneration::update_counters() { 649 if (UsePerfData) { 650 _space_counters->update_all(); 651 _gen_counters->update_all(); 652 } 653 } 654 655 // this is an optimized version of update_counters(). it takes the 656 // used value as a parameter rather than computing it. 657 // 658 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 659 if (UsePerfData) { 660 _space_counters->update_used(used); 661 _space_counters->update_capacity(); 662 _gen_counters->update_all(); 663 } 664 } 665 666 void ConcurrentMarkSweepGeneration::print() const { 667 Generation::print(); 668 cmsSpace()->print(); 669 } 670 671 #ifndef PRODUCT 672 void ConcurrentMarkSweepGeneration::print_statistics() { 673 cmsSpace()->printFLCensus(0); 674 } 675 #endif 676 677 size_t 678 ConcurrentMarkSweepGeneration::contiguous_available() const { 679 // dld proposes an improvement in precision here. If the committed 680 // part of the space ends in a free block we should add that to 681 // uncommitted size in the calculation below. Will make this 682 // change later, staying with the approximation below for the 683 // time being. -- ysr. 684 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 685 } 686 687 size_t 688 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 689 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 690 } 691 692 size_t ConcurrentMarkSweepGeneration::max_available() const { 693 return free() + _virtual_space.uncommitted_size(); 694 } 695 696 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 697 size_t available = max_available(); 698 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 699 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 700 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 701 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 702 return res; 703 } 704 705 // At a promotion failure dump information on block layout in heap 706 // (cms old generation). 707 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 708 Log(gc, promotion) log; 709 if (log.is_trace()) { 710 LogStream ls(log.trace()); 711 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 712 } 713 } 714 715 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 716 // Clear the promotion information. These pointers can be adjusted 717 // along with all the other pointers into the heap but 718 // compaction is expected to be a rare event with 719 // a heap using cms so don't do it without seeing the need. 720 for (uint i = 0; i < ParallelGCThreads; i++) { 721 _par_gc_thread_states[i]->promo.reset(); 722 } 723 } 724 725 void ConcurrentMarkSweepGeneration::compute_new_size() { 726 assert_locked_or_safepoint(Heap_lock); 727 728 // If incremental collection failed, we just want to expand 729 // to the limit. 730 if (incremental_collection_failed()) { 731 clear_incremental_collection_failed(); 732 grow_to_reserved(); 733 return; 734 } 735 736 // The heap has been compacted but not reset yet. 737 // Any metric such as free() or used() will be incorrect. 738 739 CardGeneration::compute_new_size(); 740 741 // Reset again after a possible resizing 742 if (did_compact()) { 743 cmsSpace()->reset_after_compaction(); 744 } 745 } 746 747 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 748 assert_locked_or_safepoint(Heap_lock); 749 750 // If incremental collection failed, we just want to expand 751 // to the limit. 752 if (incremental_collection_failed()) { 753 clear_incremental_collection_failed(); 754 grow_to_reserved(); 755 return; 756 } 757 758 double free_percentage = ((double) free()) / capacity(); 759 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 760 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 761 762 // compute expansion delta needed for reaching desired free percentage 763 if (free_percentage < desired_free_percentage) { 764 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 765 assert(desired_capacity >= capacity(), "invalid expansion size"); 766 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 767 Log(gc) log; 768 if (log.is_trace()) { 769 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 770 log.trace("From compute_new_size: "); 771 log.trace(" Free fraction %f", free_percentage); 772 log.trace(" Desired free fraction %f", desired_free_percentage); 773 log.trace(" Maximum free fraction %f", maximum_free_percentage); 774 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 775 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 776 CMSHeap* heap = CMSHeap::heap(); 777 size_t young_size = heap->young_gen()->capacity(); 778 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 779 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 780 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 781 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 782 } 783 // safe if expansion fails 784 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 785 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 786 } else { 787 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 788 assert(desired_capacity <= capacity(), "invalid expansion size"); 789 size_t shrink_bytes = capacity() - desired_capacity; 790 // Don't shrink unless the delta is greater than the minimum shrink we want 791 if (shrink_bytes >= MinHeapDeltaBytes) { 792 shrink_free_list_by(shrink_bytes); 793 } 794 } 795 } 796 797 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 798 return cmsSpace()->freelistLock(); 799 } 800 801 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 802 CMSSynchronousYieldRequest yr; 803 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 804 return have_lock_and_allocate(size, tlab); 805 } 806 807 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 808 bool tlab /* ignored */) { 809 assert_lock_strong(freelistLock()); 810 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 811 HeapWord* res = cmsSpace()->allocate(adjustedSize); 812 // Allocate the object live (grey) if the background collector has 813 // started marking. This is necessary because the marker may 814 // have passed this address and consequently this object will 815 // not otherwise be greyed and would be incorrectly swept up. 816 // Note that if this object contains references, the writing 817 // of those references will dirty the card containing this object 818 // allowing the object to be blackened (and its references scanned) 819 // either during a preclean phase or at the final checkpoint. 820 if (res != NULL) { 821 // We may block here with an uninitialized object with 822 // its mark-bit or P-bits not yet set. Such objects need 823 // to be safely navigable by block_start(). 824 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 825 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 826 collector()->direct_allocated(res, adjustedSize); 827 _direct_allocated_words += adjustedSize; 828 // allocation counters 829 NOT_PRODUCT( 830 _numObjectsAllocated++; 831 _numWordsAllocated += (int)adjustedSize; 832 ) 833 } 834 return res; 835 } 836 837 // In the case of direct allocation by mutators in a generation that 838 // is being concurrently collected, the object must be allocated 839 // live (grey) if the background collector has started marking. 840 // This is necessary because the marker may 841 // have passed this address and consequently this object will 842 // not otherwise be greyed and would be incorrectly swept up. 843 // Note that if this object contains references, the writing 844 // of those references will dirty the card containing this object 845 // allowing the object to be blackened (and its references scanned) 846 // either during a preclean phase or at the final checkpoint. 847 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 848 assert(_markBitMap.covers(start, size), "Out of bounds"); 849 if (_collectorState >= Marking) { 850 MutexLockerEx y(_markBitMap.lock(), 851 Mutex::_no_safepoint_check_flag); 852 // [see comments preceding SweepClosure::do_blk() below for details] 853 // 854 // Can the P-bits be deleted now? JJJ 855 // 856 // 1. need to mark the object as live so it isn't collected 857 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 858 // 3. need to mark the end of the object so marking, precleaning or sweeping 859 // can skip over uninitialized or unparsable objects. An allocated 860 // object is considered uninitialized for our purposes as long as 861 // its klass word is NULL. All old gen objects are parsable 862 // as soon as they are initialized.) 863 _markBitMap.mark(start); // object is live 864 _markBitMap.mark(start + 1); // object is potentially uninitialized? 865 _markBitMap.mark(start + size - 1); 866 // mark end of object 867 } 868 // check that oop looks uninitialized 869 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 870 } 871 872 void CMSCollector::promoted(bool par, HeapWord* start, 873 bool is_obj_array, size_t obj_size) { 874 assert(_markBitMap.covers(start), "Out of bounds"); 875 // See comment in direct_allocated() about when objects should 876 // be allocated live. 877 if (_collectorState >= Marking) { 878 // we already hold the marking bit map lock, taken in 879 // the prologue 880 if (par) { 881 _markBitMap.par_mark(start); 882 } else { 883 _markBitMap.mark(start); 884 } 885 // We don't need to mark the object as uninitialized (as 886 // in direct_allocated above) because this is being done with the 887 // world stopped and the object will be initialized by the 888 // time the marking, precleaning or sweeping get to look at it. 889 // But see the code for copying objects into the CMS generation, 890 // where we need to ensure that concurrent readers of the 891 // block offset table are able to safely navigate a block that 892 // is in flux from being free to being allocated (and in 893 // transition while being copied into) and subsequently 894 // becoming a bona-fide object when the copy/promotion is complete. 895 assert(SafepointSynchronize::is_at_safepoint(), 896 "expect promotion only at safepoints"); 897 898 if (_collectorState < Sweeping) { 899 // Mark the appropriate cards in the modUnionTable, so that 900 // this object gets scanned before the sweep. If this is 901 // not done, CMS generation references in the object might 902 // not get marked. 903 // For the case of arrays, which are otherwise precisely 904 // marked, we need to dirty the entire array, not just its head. 905 if (is_obj_array) { 906 // The [par_]mark_range() method expects mr.end() below to 907 // be aligned to the granularity of a bit's representation 908 // in the heap. In the case of the MUT below, that's a 909 // card size. 910 MemRegion mr(start, 911 align_up(start + obj_size, 912 CardTable::card_size /* bytes */)); 913 if (par) { 914 _modUnionTable.par_mark_range(mr); 915 } else { 916 _modUnionTable.mark_range(mr); 917 } 918 } else { // not an obj array; we can just mark the head 919 if (par) { 920 _modUnionTable.par_mark(start); 921 } else { 922 _modUnionTable.mark(start); 923 } 924 } 925 } 926 } 927 } 928 929 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 930 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 931 // allocate, copy and if necessary update promoinfo -- 932 // delegate to underlying space. 933 assert_lock_strong(freelistLock()); 934 935 #ifndef PRODUCT 936 if (CMSHeap::heap()->promotion_should_fail()) { 937 return NULL; 938 } 939 #endif // #ifndef PRODUCT 940 941 oop res = _cmsSpace->promote(obj, obj_size); 942 if (res == NULL) { 943 // expand and retry 944 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 945 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 946 // Since this is the old generation, we don't try to promote 947 // into a more senior generation. 948 res = _cmsSpace->promote(obj, obj_size); 949 } 950 if (res != NULL) { 951 // See comment in allocate() about when objects should 952 // be allocated live. 953 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 954 collector()->promoted(false, // Not parallel 955 (HeapWord*)res, obj->is_objArray(), obj_size); 956 // promotion counters 957 NOT_PRODUCT( 958 _numObjectsPromoted++; 959 _numWordsPromoted += 960 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 961 ) 962 } 963 return res; 964 } 965 966 967 // IMPORTANT: Notes on object size recognition in CMS. 968 // --------------------------------------------------- 969 // A block of storage in the CMS generation is always in 970 // one of three states. A free block (FREE), an allocated 971 // object (OBJECT) whose size() method reports the correct size, 972 // and an intermediate state (TRANSIENT) in which its size cannot 973 // be accurately determined. 974 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 975 // ----------------------------------------------------- 976 // FREE: klass_word & 1 == 1; mark_word holds block size 977 // 978 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 979 // obj->size() computes correct size 980 // 981 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 982 // 983 // STATE IDENTIFICATION: (64 bit+COOPS) 984 // ------------------------------------ 985 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 986 // 987 // OBJECT: klass_word installed; klass_word != 0; 988 // obj->size() computes correct size 989 // 990 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 991 // 992 // 993 // STATE TRANSITION DIAGRAM 994 // 995 // mut / parnew mut / parnew 996 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 997 // ^ | 998 // |------------------------ DEAD <------------------------------------| 999 // sweep mut 1000 // 1001 // While a block is in TRANSIENT state its size cannot be determined 1002 // so readers will either need to come back later or stall until 1003 // the size can be determined. Note that for the case of direct 1004 // allocation, P-bits, when available, may be used to determine the 1005 // size of an object that may not yet have been initialized. 1006 1007 // Things to support parallel young-gen collection. 1008 oop 1009 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1010 oop old, markOop m, 1011 size_t word_sz) { 1012 #ifndef PRODUCT 1013 if (CMSHeap::heap()->promotion_should_fail()) { 1014 return NULL; 1015 } 1016 #endif // #ifndef PRODUCT 1017 1018 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1019 PromotionInfo* promoInfo = &ps->promo; 1020 // if we are tracking promotions, then first ensure space for 1021 // promotion (including spooling space for saving header if necessary). 1022 // then allocate and copy, then track promoted info if needed. 1023 // When tracking (see PromotionInfo::track()), the mark word may 1024 // be displaced and in this case restoration of the mark word 1025 // occurs in the (oop_since_save_marks_)iterate phase. 1026 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1027 // Out of space for allocating spooling buffers; 1028 // try expanding and allocating spooling buffers. 1029 if (!expand_and_ensure_spooling_space(promoInfo)) { 1030 return NULL; 1031 } 1032 } 1033 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1034 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1035 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1036 if (obj_ptr == NULL) { 1037 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1038 if (obj_ptr == NULL) { 1039 return NULL; 1040 } 1041 } 1042 oop obj = oop(obj_ptr); 1043 OrderAccess::storestore(); 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 // IMPORTANT: See note on object initialization for CMS above. 1047 // Otherwise, copy the object. Here we must be careful to insert the 1048 // klass pointer last, since this marks the block as an allocated object. 1049 // Except with compressed oops it's the mark word. 1050 HeapWord* old_ptr = (HeapWord*)old; 1051 // Restore the mark word copied above. 1052 obj->set_mark_raw(m); 1053 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1054 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1055 OrderAccess::storestore(); 1056 1057 if (UseCompressedClassPointers) { 1058 // Copy gap missed by (aligned) header size calculation below 1059 obj->set_klass_gap(old->klass_gap()); 1060 } 1061 if (word_sz > (size_t)oopDesc::header_size()) { 1062 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1063 obj_ptr + oopDesc::header_size(), 1064 word_sz - oopDesc::header_size()); 1065 } 1066 1067 // Now we can track the promoted object, if necessary. We take care 1068 // to delay the transition from uninitialized to full object 1069 // (i.e., insertion of klass pointer) until after, so that it 1070 // atomically becomes a promoted object. 1071 if (promoInfo->tracking()) { 1072 promoInfo->track((PromotedObject*)obj, old->klass()); 1073 } 1074 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1075 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1076 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1077 1078 // Finally, install the klass pointer (this should be volatile). 1079 OrderAccess::storestore(); 1080 obj->set_klass(old->klass()); 1081 // We should now be able to calculate the right size for this object 1082 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1083 1084 collector()->promoted(true, // parallel 1085 obj_ptr, old->is_objArray(), word_sz); 1086 1087 NOT_PRODUCT( 1088 Atomic::inc(&_numObjectsPromoted); 1089 Atomic::add(alloc_sz, &_numWordsPromoted); 1090 ) 1091 1092 return obj; 1093 } 1094 1095 void 1096 ConcurrentMarkSweepGeneration:: 1097 par_promote_alloc_done(int thread_num) { 1098 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1099 ps->lab.retire(thread_num); 1100 } 1101 1102 void 1103 ConcurrentMarkSweepGeneration:: 1104 par_oop_since_save_marks_iterate_done(int thread_num) { 1105 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1106 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1107 ps->promo.promoted_oops_iterate(dummy_cl); 1108 1109 // Because card-scanning has been completed, subsequent phases 1110 // (e.g., reference processing) will not need to recognize which 1111 // objects have been promoted during this GC. So, we can now disable 1112 // promotion tracking. 1113 ps->promo.stopTrackingPromotions(); 1114 } 1115 1116 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1117 size_t size, 1118 bool tlab) 1119 { 1120 // We allow a STW collection only if a full 1121 // collection was requested. 1122 return full || should_allocate(size, tlab); // FIX ME !!! 1123 // This and promotion failure handling are connected at the 1124 // hip and should be fixed by untying them. 1125 } 1126 1127 bool CMSCollector::shouldConcurrentCollect() { 1128 LogTarget(Trace, gc) log; 1129 1130 if (_full_gc_requested) { 1131 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1132 return true; 1133 } 1134 1135 FreelistLocker x(this); 1136 // ------------------------------------------------------------------ 1137 // Print out lots of information which affects the initiation of 1138 // a collection. 1139 if (log.is_enabled() && stats().valid()) { 1140 log.print("CMSCollector shouldConcurrentCollect: "); 1141 1142 LogStream out(log); 1143 stats().print_on(&out); 1144 1145 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1146 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1147 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1148 log.print("promotion_rate=%g", stats().promotion_rate()); 1149 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1150 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1151 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1152 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1153 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1154 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1155 } 1156 // ------------------------------------------------------------------ 1157 1158 // If the estimated time to complete a cms collection (cms_duration()) 1159 // is less than the estimated time remaining until the cms generation 1160 // is full, start a collection. 1161 if (!UseCMSInitiatingOccupancyOnly) { 1162 if (stats().valid()) { 1163 if (stats().time_until_cms_start() == 0.0) { 1164 return true; 1165 } 1166 } else { 1167 // We want to conservatively collect somewhat early in order 1168 // to try and "bootstrap" our CMS/promotion statistics; 1169 // this branch will not fire after the first successful CMS 1170 // collection because the stats should then be valid. 1171 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1172 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1173 _cmsGen->occupancy(), _bootstrap_occupancy); 1174 return true; 1175 } 1176 } 1177 } 1178 1179 // Otherwise, we start a collection cycle if 1180 // old gen want a collection cycle started. Each may use 1181 // an appropriate criterion for making this decision. 1182 // XXX We need to make sure that the gen expansion 1183 // criterion dovetails well with this. XXX NEED TO FIX THIS 1184 if (_cmsGen->should_concurrent_collect()) { 1185 log.print("CMS old gen initiated"); 1186 return true; 1187 } 1188 1189 // We start a collection if we believe an incremental collection may fail; 1190 // this is not likely to be productive in practice because it's probably too 1191 // late anyway. 1192 CMSHeap* heap = CMSHeap::heap(); 1193 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1194 log.print("CMSCollector: collect because incremental collection will fail "); 1195 return true; 1196 } 1197 1198 if (MetaspaceGC::should_concurrent_collect()) { 1199 log.print("CMSCollector: collect for metadata allocation "); 1200 return true; 1201 } 1202 1203 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1204 if (CMSTriggerInterval >= 0) { 1205 if (CMSTriggerInterval == 0) { 1206 // Trigger always 1207 return true; 1208 } 1209 1210 // Check the CMS time since begin (we do not check the stats validity 1211 // as we want to be able to trigger the first CMS cycle as well) 1212 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1213 if (stats().valid()) { 1214 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1215 stats().cms_time_since_begin()); 1216 } else { 1217 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1218 } 1219 return true; 1220 } 1221 } 1222 1223 return false; 1224 } 1225 1226 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1227 1228 // Clear _expansion_cause fields of constituent generations 1229 void CMSCollector::clear_expansion_cause() { 1230 _cmsGen->clear_expansion_cause(); 1231 } 1232 1233 // We should be conservative in starting a collection cycle. To 1234 // start too eagerly runs the risk of collecting too often in the 1235 // extreme. To collect too rarely falls back on full collections, 1236 // which works, even if not optimum in terms of concurrent work. 1237 // As a work around for too eagerly collecting, use the flag 1238 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1239 // giving the user an easily understandable way of controlling the 1240 // collections. 1241 // We want to start a new collection cycle if any of the following 1242 // conditions hold: 1243 // . our current occupancy exceeds the configured initiating occupancy 1244 // for this generation, or 1245 // . we recently needed to expand this space and have not, since that 1246 // expansion, done a collection of this generation, or 1247 // . the underlying space believes that it may be a good idea to initiate 1248 // a concurrent collection (this may be based on criteria such as the 1249 // following: the space uses linear allocation and linear allocation is 1250 // going to fail, or there is believed to be excessive fragmentation in 1251 // the generation, etc... or ... 1252 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1253 // the case of the old generation; see CR 6543076): 1254 // we may be approaching a point at which allocation requests may fail because 1255 // we will be out of sufficient free space given allocation rate estimates.] 1256 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1257 1258 assert_lock_strong(freelistLock()); 1259 if (occupancy() > initiating_occupancy()) { 1260 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1261 short_name(), occupancy(), initiating_occupancy()); 1262 return true; 1263 } 1264 if (UseCMSInitiatingOccupancyOnly) { 1265 return false; 1266 } 1267 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1268 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1269 return true; 1270 } 1271 return false; 1272 } 1273 1274 void ConcurrentMarkSweepGeneration::collect(bool full, 1275 bool clear_all_soft_refs, 1276 size_t size, 1277 bool tlab) 1278 { 1279 collector()->collect(full, clear_all_soft_refs, size, tlab); 1280 } 1281 1282 void CMSCollector::collect(bool full, 1283 bool clear_all_soft_refs, 1284 size_t size, 1285 bool tlab) 1286 { 1287 // The following "if" branch is present for defensive reasons. 1288 // In the current uses of this interface, it can be replaced with: 1289 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1290 // But I am not placing that assert here to allow future 1291 // generality in invoking this interface. 1292 if (GCLocker::is_active()) { 1293 // A consistency test for GCLocker 1294 assert(GCLocker::needs_gc(), "Should have been set already"); 1295 // Skip this foreground collection, instead 1296 // expanding the heap if necessary. 1297 // Need the free list locks for the call to free() in compute_new_size() 1298 compute_new_size(); 1299 return; 1300 } 1301 acquire_control_and_collect(full, clear_all_soft_refs); 1302 } 1303 1304 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1305 CMSHeap* heap = CMSHeap::heap(); 1306 unsigned int gc_count = heap->total_full_collections(); 1307 if (gc_count == full_gc_count) { 1308 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1309 _full_gc_requested = true; 1310 _full_gc_cause = cause; 1311 CGC_lock->notify(); // nudge CMS thread 1312 } else { 1313 assert(gc_count > full_gc_count, "Error: causal loop"); 1314 } 1315 } 1316 1317 bool CMSCollector::is_external_interruption() { 1318 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1319 return GCCause::is_user_requested_gc(cause) || 1320 GCCause::is_serviceability_requested_gc(cause); 1321 } 1322 1323 void CMSCollector::report_concurrent_mode_interruption() { 1324 if (is_external_interruption()) { 1325 log_debug(gc)("Concurrent mode interrupted"); 1326 } else { 1327 log_debug(gc)("Concurrent mode failure"); 1328 _gc_tracer_cm->report_concurrent_mode_failure(); 1329 } 1330 } 1331 1332 1333 // The foreground and background collectors need to coordinate in order 1334 // to make sure that they do not mutually interfere with CMS collections. 1335 // When a background collection is active, 1336 // the foreground collector may need to take over (preempt) and 1337 // synchronously complete an ongoing collection. Depending on the 1338 // frequency of the background collections and the heap usage 1339 // of the application, this preemption can be seldom or frequent. 1340 // There are only certain 1341 // points in the background collection that the "collection-baton" 1342 // can be passed to the foreground collector. 1343 // 1344 // The foreground collector will wait for the baton before 1345 // starting any part of the collection. The foreground collector 1346 // will only wait at one location. 1347 // 1348 // The background collector will yield the baton before starting a new 1349 // phase of the collection (e.g., before initial marking, marking from roots, 1350 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1351 // of the loop which switches the phases. The background collector does some 1352 // of the phases (initial mark, final re-mark) with the world stopped. 1353 // Because of locking involved in stopping the world, 1354 // the foreground collector should not block waiting for the background 1355 // collector when it is doing a stop-the-world phase. The background 1356 // collector will yield the baton at an additional point just before 1357 // it enters a stop-the-world phase. Once the world is stopped, the 1358 // background collector checks the phase of the collection. If the 1359 // phase has not changed, it proceeds with the collection. If the 1360 // phase has changed, it skips that phase of the collection. See 1361 // the comments on the use of the Heap_lock in collect_in_background(). 1362 // 1363 // Variable used in baton passing. 1364 // _foregroundGCIsActive - Set to true by the foreground collector when 1365 // it wants the baton. The foreground clears it when it has finished 1366 // the collection. 1367 // _foregroundGCShouldWait - Set to true by the background collector 1368 // when it is running. The foreground collector waits while 1369 // _foregroundGCShouldWait is true. 1370 // CGC_lock - monitor used to protect access to the above variables 1371 // and to notify the foreground and background collectors. 1372 // _collectorState - current state of the CMS collection. 1373 // 1374 // The foreground collector 1375 // acquires the CGC_lock 1376 // sets _foregroundGCIsActive 1377 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1378 // various locks acquired in preparation for the collection 1379 // are released so as not to block the background collector 1380 // that is in the midst of a collection 1381 // proceeds with the collection 1382 // clears _foregroundGCIsActive 1383 // returns 1384 // 1385 // The background collector in a loop iterating on the phases of the 1386 // collection 1387 // acquires the CGC_lock 1388 // sets _foregroundGCShouldWait 1389 // if _foregroundGCIsActive is set 1390 // clears _foregroundGCShouldWait, notifies _CGC_lock 1391 // waits on _CGC_lock for _foregroundGCIsActive to become false 1392 // and exits the loop. 1393 // otherwise 1394 // proceed with that phase of the collection 1395 // if the phase is a stop-the-world phase, 1396 // yield the baton once more just before enqueueing 1397 // the stop-world CMS operation (executed by the VM thread). 1398 // returns after all phases of the collection are done 1399 // 1400 1401 void CMSCollector::acquire_control_and_collect(bool full, 1402 bool clear_all_soft_refs) { 1403 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1404 assert(!Thread::current()->is_ConcurrentGC_thread(), 1405 "shouldn't try to acquire control from self!"); 1406 1407 // Start the protocol for acquiring control of the 1408 // collection from the background collector (aka CMS thread). 1409 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1410 "VM thread should have CMS token"); 1411 // Remember the possibly interrupted state of an ongoing 1412 // concurrent collection 1413 CollectorState first_state = _collectorState; 1414 1415 // Signal to a possibly ongoing concurrent collection that 1416 // we want to do a foreground collection. 1417 _foregroundGCIsActive = true; 1418 1419 // release locks and wait for a notify from the background collector 1420 // releasing the locks in only necessary for phases which 1421 // do yields to improve the granularity of the collection. 1422 assert_lock_strong(bitMapLock()); 1423 // We need to lock the Free list lock for the space that we are 1424 // currently collecting. 1425 assert(haveFreelistLocks(), "Must be holding free list locks"); 1426 bitMapLock()->unlock(); 1427 releaseFreelistLocks(); 1428 { 1429 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1430 if (_foregroundGCShouldWait) { 1431 // We are going to be waiting for action for the CMS thread; 1432 // it had better not be gone (for instance at shutdown)! 1433 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1434 "CMS thread must be running"); 1435 // Wait here until the background collector gives us the go-ahead 1436 ConcurrentMarkSweepThread::clear_CMS_flag( 1437 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1438 // Get a possibly blocked CMS thread going: 1439 // Note that we set _foregroundGCIsActive true above, 1440 // without protection of the CGC_lock. 1441 CGC_lock->notify(); 1442 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1443 "Possible deadlock"); 1444 while (_foregroundGCShouldWait) { 1445 // wait for notification 1446 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1447 // Possibility of delay/starvation here, since CMS token does 1448 // not know to give priority to VM thread? Actually, i think 1449 // there wouldn't be any delay/starvation, but the proof of 1450 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1451 } 1452 ConcurrentMarkSweepThread::set_CMS_flag( 1453 ConcurrentMarkSweepThread::CMS_vm_has_token); 1454 } 1455 } 1456 // The CMS_token is already held. Get back the other locks. 1457 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1458 "VM thread should have CMS token"); 1459 getFreelistLocks(); 1460 bitMapLock()->lock_without_safepoint_check(); 1461 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1462 p2i(Thread::current()), first_state); 1463 log_debug(gc, state)(" gets control with state %d", _collectorState); 1464 1465 // Inform cms gen if this was due to partial collection failing. 1466 // The CMS gen may use this fact to determine its expansion policy. 1467 CMSHeap* heap = CMSHeap::heap(); 1468 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1469 assert(!_cmsGen->incremental_collection_failed(), 1470 "Should have been noticed, reacted to and cleared"); 1471 _cmsGen->set_incremental_collection_failed(); 1472 } 1473 1474 if (first_state > Idling) { 1475 report_concurrent_mode_interruption(); 1476 } 1477 1478 set_did_compact(true); 1479 1480 // If the collection is being acquired from the background 1481 // collector, there may be references on the discovered 1482 // references lists. Abandon those references, since some 1483 // of them may have become unreachable after concurrent 1484 // discovery; the STW compacting collector will redo discovery 1485 // more precisely, without being subject to floating garbage. 1486 // Leaving otherwise unreachable references in the discovered 1487 // lists would require special handling. 1488 ref_processor()->disable_discovery(); 1489 ref_processor()->abandon_partial_discovery(); 1490 ref_processor()->verify_no_references_recorded(); 1491 1492 if (first_state > Idling) { 1493 save_heap_summary(); 1494 } 1495 1496 do_compaction_work(clear_all_soft_refs); 1497 1498 // Has the GC time limit been exceeded? 1499 size_t max_eden_size = _young_gen->max_eden_size(); 1500 GCCause::Cause gc_cause = heap->gc_cause(); 1501 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1502 _young_gen->eden()->used(), 1503 _cmsGen->max_capacity(), 1504 max_eden_size, 1505 full, 1506 gc_cause, 1507 heap->soft_ref_policy()); 1508 1509 // Reset the expansion cause, now that we just completed 1510 // a collection cycle. 1511 clear_expansion_cause(); 1512 _foregroundGCIsActive = false; 1513 return; 1514 } 1515 1516 // Resize the tenured generation 1517 // after obtaining the free list locks for the 1518 // two generations. 1519 void CMSCollector::compute_new_size() { 1520 assert_locked_or_safepoint(Heap_lock); 1521 FreelistLocker z(this); 1522 MetaspaceGC::compute_new_size(); 1523 _cmsGen->compute_new_size_free_list(); 1524 } 1525 1526 // A work method used by the foreground collector to do 1527 // a mark-sweep-compact. 1528 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1529 CMSHeap* heap = CMSHeap::heap(); 1530 1531 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1532 gc_timer->register_gc_start(); 1533 1534 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1535 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1536 1537 heap->pre_full_gc_dump(gc_timer); 1538 1539 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1540 1541 // Temporarily widen the span of the weak reference processing to 1542 // the entire heap. 1543 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1544 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1545 // Temporarily, clear the "is_alive_non_header" field of the 1546 // reference processor. 1547 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1548 // Temporarily make reference _processing_ single threaded (non-MT). 1549 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1550 // Temporarily make refs discovery atomic 1551 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1552 // Temporarily make reference _discovery_ single threaded (non-MT) 1553 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1554 1555 ref_processor()->set_enqueuing_is_done(false); 1556 ref_processor()->enable_discovery(); 1557 ref_processor()->setup_policy(clear_all_soft_refs); 1558 // If an asynchronous collection finishes, the _modUnionTable is 1559 // all clear. If we are assuming the collection from an asynchronous 1560 // collection, clear the _modUnionTable. 1561 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1562 "_modUnionTable should be clear if the baton was not passed"); 1563 _modUnionTable.clear_all(); 1564 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1565 "mod union for klasses should be clear if the baton was passed"); 1566 _ct->cld_rem_set()->clear_mod_union(); 1567 1568 1569 // We must adjust the allocation statistics being maintained 1570 // in the free list space. We do so by reading and clearing 1571 // the sweep timer and updating the block flux rate estimates below. 1572 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1573 if (_inter_sweep_timer.is_active()) { 1574 _inter_sweep_timer.stop(); 1575 // Note that we do not use this sample to update the _inter_sweep_estimate. 1576 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1577 _inter_sweep_estimate.padded_average(), 1578 _intra_sweep_estimate.padded_average()); 1579 } 1580 1581 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1582 #ifdef ASSERT 1583 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1584 size_t free_size = cms_space->free(); 1585 assert(free_size == 1586 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1587 * HeapWordSize, 1588 "All the free space should be compacted into one chunk at top"); 1589 assert(cms_space->dictionary()->total_chunk_size( 1590 debug_only(cms_space->freelistLock())) == 0 || 1591 cms_space->totalSizeInIndexedFreeLists() == 0, 1592 "All the free space should be in a single chunk"); 1593 size_t num = cms_space->totalCount(); 1594 assert((free_size == 0 && num == 0) || 1595 (free_size > 0 && (num == 1 || num == 2)), 1596 "There should be at most 2 free chunks after compaction"); 1597 #endif // ASSERT 1598 _collectorState = Resetting; 1599 assert(_restart_addr == NULL, 1600 "Should have been NULL'd before baton was passed"); 1601 reset_stw(); 1602 _cmsGen->reset_after_compaction(); 1603 _concurrent_cycles_since_last_unload = 0; 1604 1605 // Clear any data recorded in the PLAB chunk arrays. 1606 if (_survivor_plab_array != NULL) { 1607 reset_survivor_plab_arrays(); 1608 } 1609 1610 // Adjust the per-size allocation stats for the next epoch. 1611 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1612 // Restart the "inter sweep timer" for the next epoch. 1613 _inter_sweep_timer.reset(); 1614 _inter_sweep_timer.start(); 1615 1616 // No longer a need to do a concurrent collection for Metaspace. 1617 MetaspaceGC::set_should_concurrent_collect(false); 1618 1619 heap->post_full_gc_dump(gc_timer); 1620 1621 gc_timer->register_gc_end(); 1622 1623 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1624 1625 // For a mark-sweep-compact, compute_new_size() will be called 1626 // in the heap's do_collection() method. 1627 } 1628 1629 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1630 Log(gc, heap) log; 1631 if (!log.is_trace()) { 1632 return; 1633 } 1634 1635 ContiguousSpace* eden_space = _young_gen->eden(); 1636 ContiguousSpace* from_space = _young_gen->from(); 1637 ContiguousSpace* to_space = _young_gen->to(); 1638 // Eden 1639 if (_eden_chunk_array != NULL) { 1640 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1641 p2i(eden_space->bottom()), p2i(eden_space->top()), 1642 p2i(eden_space->end()), eden_space->capacity()); 1643 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1644 _eden_chunk_index, _eden_chunk_capacity); 1645 for (size_t i = 0; i < _eden_chunk_index; i++) { 1646 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1647 } 1648 } 1649 // Survivor 1650 if (_survivor_chunk_array != NULL) { 1651 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1652 p2i(from_space->bottom()), p2i(from_space->top()), 1653 p2i(from_space->end()), from_space->capacity()); 1654 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1655 _survivor_chunk_index, _survivor_chunk_capacity); 1656 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1657 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1658 } 1659 } 1660 } 1661 1662 void CMSCollector::getFreelistLocks() const { 1663 // Get locks for all free lists in all generations that this 1664 // collector is responsible for 1665 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1666 } 1667 1668 void CMSCollector::releaseFreelistLocks() const { 1669 // Release locks for all free lists in all generations that this 1670 // collector is responsible for 1671 _cmsGen->freelistLock()->unlock(); 1672 } 1673 1674 bool CMSCollector::haveFreelistLocks() const { 1675 // Check locks for all free lists in all generations that this 1676 // collector is responsible for 1677 assert_lock_strong(_cmsGen->freelistLock()); 1678 PRODUCT_ONLY(ShouldNotReachHere()); 1679 return true; 1680 } 1681 1682 // A utility class that is used by the CMS collector to 1683 // temporarily "release" the foreground collector from its 1684 // usual obligation to wait for the background collector to 1685 // complete an ongoing phase before proceeding. 1686 class ReleaseForegroundGC: public StackObj { 1687 private: 1688 CMSCollector* _c; 1689 public: 1690 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1691 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1692 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1693 // allow a potentially blocked foreground collector to proceed 1694 _c->_foregroundGCShouldWait = false; 1695 if (_c->_foregroundGCIsActive) { 1696 CGC_lock->notify(); 1697 } 1698 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1699 "Possible deadlock"); 1700 } 1701 1702 ~ReleaseForegroundGC() { 1703 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1704 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1705 _c->_foregroundGCShouldWait = true; 1706 } 1707 }; 1708 1709 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1710 assert(Thread::current()->is_ConcurrentGC_thread(), 1711 "A CMS asynchronous collection is only allowed on a CMS thread."); 1712 1713 CMSHeap* heap = CMSHeap::heap(); 1714 { 1715 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1716 MutexLockerEx hl(Heap_lock, safepoint_check); 1717 FreelistLocker fll(this); 1718 MutexLockerEx x(CGC_lock, safepoint_check); 1719 if (_foregroundGCIsActive) { 1720 // The foreground collector is. Skip this 1721 // background collection. 1722 assert(!_foregroundGCShouldWait, "Should be clear"); 1723 return; 1724 } else { 1725 assert(_collectorState == Idling, "Should be idling before start."); 1726 _collectorState = InitialMarking; 1727 register_gc_start(cause); 1728 // Reset the expansion cause, now that we are about to begin 1729 // a new cycle. 1730 clear_expansion_cause(); 1731 1732 // Clear the MetaspaceGC flag since a concurrent collection 1733 // is starting but also clear it after the collection. 1734 MetaspaceGC::set_should_concurrent_collect(false); 1735 } 1736 // Decide if we want to enable class unloading as part of the 1737 // ensuing concurrent GC cycle. 1738 update_should_unload_classes(); 1739 _full_gc_requested = false; // acks all outstanding full gc requests 1740 _full_gc_cause = GCCause::_no_gc; 1741 // Signal that we are about to start a collection 1742 heap->increment_total_full_collections(); // ... starting a collection cycle 1743 _collection_count_start = heap->total_full_collections(); 1744 } 1745 1746 size_t prev_used = _cmsGen->used(); 1747 1748 // The change of the collection state is normally done at this level; 1749 // the exceptions are phases that are executed while the world is 1750 // stopped. For those phases the change of state is done while the 1751 // world is stopped. For baton passing purposes this allows the 1752 // background collector to finish the phase and change state atomically. 1753 // The foreground collector cannot wait on a phase that is done 1754 // while the world is stopped because the foreground collector already 1755 // has the world stopped and would deadlock. 1756 while (_collectorState != Idling) { 1757 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1758 p2i(Thread::current()), _collectorState); 1759 // The foreground collector 1760 // holds the Heap_lock throughout its collection. 1761 // holds the CMS token (but not the lock) 1762 // except while it is waiting for the background collector to yield. 1763 // 1764 // The foreground collector should be blocked (not for long) 1765 // if the background collector is about to start a phase 1766 // executed with world stopped. If the background 1767 // collector has already started such a phase, the 1768 // foreground collector is blocked waiting for the 1769 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1770 // are executed in the VM thread. 1771 // 1772 // The locking order is 1773 // PendingListLock (PLL) -- if applicable (FinalMarking) 1774 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1775 // CMS token (claimed in 1776 // stop_world_and_do() --> 1777 // safepoint_synchronize() --> 1778 // CMSThread::synchronize()) 1779 1780 { 1781 // Check if the FG collector wants us to yield. 1782 CMSTokenSync x(true); // is cms thread 1783 if (waitForForegroundGC()) { 1784 // We yielded to a foreground GC, nothing more to be 1785 // done this round. 1786 assert(_foregroundGCShouldWait == false, "We set it to false in " 1787 "waitForForegroundGC()"); 1788 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1789 p2i(Thread::current()), _collectorState); 1790 return; 1791 } else { 1792 // The background collector can run but check to see if the 1793 // foreground collector has done a collection while the 1794 // background collector was waiting to get the CGC_lock 1795 // above. If yes, break so that _foregroundGCShouldWait 1796 // is cleared before returning. 1797 if (_collectorState == Idling) { 1798 break; 1799 } 1800 } 1801 } 1802 1803 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1804 "should be waiting"); 1805 1806 switch (_collectorState) { 1807 case InitialMarking: 1808 { 1809 ReleaseForegroundGC x(this); 1810 stats().record_cms_begin(); 1811 VM_CMS_Initial_Mark initial_mark_op(this); 1812 VMThread::execute(&initial_mark_op); 1813 } 1814 // The collector state may be any legal state at this point 1815 // since the background collector may have yielded to the 1816 // foreground collector. 1817 break; 1818 case Marking: 1819 // initial marking in checkpointRootsInitialWork has been completed 1820 if (markFromRoots()) { // we were successful 1821 assert(_collectorState == Precleaning, "Collector state should " 1822 "have changed"); 1823 } else { 1824 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1825 } 1826 break; 1827 case Precleaning: 1828 // marking from roots in markFromRoots has been completed 1829 preclean(); 1830 assert(_collectorState == AbortablePreclean || 1831 _collectorState == FinalMarking, 1832 "Collector state should have changed"); 1833 break; 1834 case AbortablePreclean: 1835 abortable_preclean(); 1836 assert(_collectorState == FinalMarking, "Collector state should " 1837 "have changed"); 1838 break; 1839 case FinalMarking: 1840 { 1841 ReleaseForegroundGC x(this); 1842 1843 VM_CMS_Final_Remark final_remark_op(this); 1844 VMThread::execute(&final_remark_op); 1845 } 1846 assert(_foregroundGCShouldWait, "block post-condition"); 1847 break; 1848 case Sweeping: 1849 // final marking in checkpointRootsFinal has been completed 1850 sweep(); 1851 assert(_collectorState == Resizing, "Collector state change " 1852 "to Resizing must be done under the free_list_lock"); 1853 1854 case Resizing: { 1855 // Sweeping has been completed... 1856 // At this point the background collection has completed. 1857 // Don't move the call to compute_new_size() down 1858 // into code that might be executed if the background 1859 // collection was preempted. 1860 { 1861 ReleaseForegroundGC x(this); // unblock FG collection 1862 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1863 CMSTokenSync z(true); // not strictly needed. 1864 if (_collectorState == Resizing) { 1865 compute_new_size(); 1866 save_heap_summary(); 1867 _collectorState = Resetting; 1868 } else { 1869 assert(_collectorState == Idling, "The state should only change" 1870 " because the foreground collector has finished the collection"); 1871 } 1872 } 1873 break; 1874 } 1875 case Resetting: 1876 // CMS heap resizing has been completed 1877 reset_concurrent(); 1878 assert(_collectorState == Idling, "Collector state should " 1879 "have changed"); 1880 1881 MetaspaceGC::set_should_concurrent_collect(false); 1882 1883 stats().record_cms_end(); 1884 // Don't move the concurrent_phases_end() and compute_new_size() 1885 // calls to here because a preempted background collection 1886 // has it's state set to "Resetting". 1887 break; 1888 case Idling: 1889 default: 1890 ShouldNotReachHere(); 1891 break; 1892 } 1893 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1894 p2i(Thread::current()), _collectorState); 1895 assert(_foregroundGCShouldWait, "block post-condition"); 1896 } 1897 1898 // Should this be in gc_epilogue? 1899 heap->counters()->update_counters(); 1900 1901 { 1902 // Clear _foregroundGCShouldWait and, in the event that the 1903 // foreground collector is waiting, notify it, before 1904 // returning. 1905 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1906 _foregroundGCShouldWait = false; 1907 if (_foregroundGCIsActive) { 1908 CGC_lock->notify(); 1909 } 1910 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1911 "Possible deadlock"); 1912 } 1913 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1914 p2i(Thread::current()), _collectorState); 1915 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1916 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1917 } 1918 1919 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1920 _cms_start_registered = true; 1921 _gc_timer_cm->register_gc_start(); 1922 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1923 } 1924 1925 void CMSCollector::register_gc_end() { 1926 if (_cms_start_registered) { 1927 report_heap_summary(GCWhen::AfterGC); 1928 1929 _gc_timer_cm->register_gc_end(); 1930 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1931 _cms_start_registered = false; 1932 } 1933 } 1934 1935 void CMSCollector::save_heap_summary() { 1936 CMSHeap* heap = CMSHeap::heap(); 1937 _last_heap_summary = heap->create_heap_summary(); 1938 _last_metaspace_summary = heap->create_metaspace_summary(); 1939 } 1940 1941 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1942 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1943 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1944 } 1945 1946 bool CMSCollector::waitForForegroundGC() { 1947 bool res = false; 1948 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1949 "CMS thread should have CMS token"); 1950 // Block the foreground collector until the 1951 // background collectors decides whether to 1952 // yield. 1953 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1954 _foregroundGCShouldWait = true; 1955 if (_foregroundGCIsActive) { 1956 // The background collector yields to the 1957 // foreground collector and returns a value 1958 // indicating that it has yielded. The foreground 1959 // collector can proceed. 1960 res = true; 1961 _foregroundGCShouldWait = false; 1962 ConcurrentMarkSweepThread::clear_CMS_flag( 1963 ConcurrentMarkSweepThread::CMS_cms_has_token); 1964 ConcurrentMarkSweepThread::set_CMS_flag( 1965 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1966 // Get a possibly blocked foreground thread going 1967 CGC_lock->notify(); 1968 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1969 p2i(Thread::current()), _collectorState); 1970 while (_foregroundGCIsActive) { 1971 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1972 } 1973 ConcurrentMarkSweepThread::set_CMS_flag( 1974 ConcurrentMarkSweepThread::CMS_cms_has_token); 1975 ConcurrentMarkSweepThread::clear_CMS_flag( 1976 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1977 } 1978 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1979 p2i(Thread::current()), _collectorState); 1980 return res; 1981 } 1982 1983 // Because of the need to lock the free lists and other structures in 1984 // the collector, common to all the generations that the collector is 1985 // collecting, we need the gc_prologues of individual CMS generations 1986 // delegate to their collector. It may have been simpler had the 1987 // current infrastructure allowed one to call a prologue on a 1988 // collector. In the absence of that we have the generation's 1989 // prologue delegate to the collector, which delegates back 1990 // some "local" work to a worker method in the individual generations 1991 // that it's responsible for collecting, while itself doing any 1992 // work common to all generations it's responsible for. A similar 1993 // comment applies to the gc_epilogue()'s. 1994 // The role of the variable _between_prologue_and_epilogue is to 1995 // enforce the invocation protocol. 1996 void CMSCollector::gc_prologue(bool full) { 1997 // Call gc_prologue_work() for the CMSGen 1998 // we are responsible for. 1999 2000 // The following locking discipline assumes that we are only called 2001 // when the world is stopped. 2002 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2003 2004 // The CMSCollector prologue must call the gc_prologues for the 2005 // "generations" that it's responsible 2006 // for. 2007 2008 assert( Thread::current()->is_VM_thread() 2009 || ( CMSScavengeBeforeRemark 2010 && Thread::current()->is_ConcurrentGC_thread()), 2011 "Incorrect thread type for prologue execution"); 2012 2013 if (_between_prologue_and_epilogue) { 2014 // We have already been invoked; this is a gc_prologue delegation 2015 // from yet another CMS generation that we are responsible for, just 2016 // ignore it since all relevant work has already been done. 2017 return; 2018 } 2019 2020 // set a bit saying prologue has been called; cleared in epilogue 2021 _between_prologue_and_epilogue = true; 2022 // Claim locks for common data structures, then call gc_prologue_work() 2023 // for each CMSGen. 2024 2025 getFreelistLocks(); // gets free list locks on constituent spaces 2026 bitMapLock()->lock_without_safepoint_check(); 2027 2028 // Should call gc_prologue_work() for all cms gens we are responsible for 2029 bool duringMarking = _collectorState >= Marking 2030 && _collectorState < Sweeping; 2031 2032 // The young collections clear the modified oops state, which tells if 2033 // there are any modified oops in the class. The remark phase also needs 2034 // that information. Tell the young collection to save the union of all 2035 // modified klasses. 2036 if (duringMarking) { 2037 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2038 } 2039 2040 bool registerClosure = duringMarking; 2041 2042 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2043 2044 if (!full) { 2045 stats().record_gc0_begin(); 2046 } 2047 } 2048 2049 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2050 2051 _capacity_at_prologue = capacity(); 2052 _used_at_prologue = used(); 2053 2054 // We enable promotion tracking so that card-scanning can recognize 2055 // which objects have been promoted during this GC and skip them. 2056 for (uint i = 0; i < ParallelGCThreads; i++) { 2057 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2058 } 2059 2060 // Delegate to CMScollector which knows how to coordinate between 2061 // this and any other CMS generations that it is responsible for 2062 // collecting. 2063 collector()->gc_prologue(full); 2064 } 2065 2066 // This is a "private" interface for use by this generation's CMSCollector. 2067 // Not to be called directly by any other entity (for instance, 2068 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2069 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2070 bool registerClosure, ModUnionClosure* modUnionClosure) { 2071 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2072 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2073 "Should be NULL"); 2074 if (registerClosure) { 2075 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2076 } 2077 cmsSpace()->gc_prologue(); 2078 // Clear stat counters 2079 NOT_PRODUCT( 2080 assert(_numObjectsPromoted == 0, "check"); 2081 assert(_numWordsPromoted == 0, "check"); 2082 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2083 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2084 _numObjectsAllocated = 0; 2085 _numWordsAllocated = 0; 2086 ) 2087 } 2088 2089 void CMSCollector::gc_epilogue(bool full) { 2090 // The following locking discipline assumes that we are only called 2091 // when the world is stopped. 2092 assert(SafepointSynchronize::is_at_safepoint(), 2093 "world is stopped assumption"); 2094 2095 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2096 // if linear allocation blocks need to be appropriately marked to allow the 2097 // the blocks to be parsable. We also check here whether we need to nudge the 2098 // CMS collector thread to start a new cycle (if it's not already active). 2099 assert( Thread::current()->is_VM_thread() 2100 || ( CMSScavengeBeforeRemark 2101 && Thread::current()->is_ConcurrentGC_thread()), 2102 "Incorrect thread type for epilogue execution"); 2103 2104 if (!_between_prologue_and_epilogue) { 2105 // We have already been invoked; this is a gc_epilogue delegation 2106 // from yet another CMS generation that we are responsible for, just 2107 // ignore it since all relevant work has already been done. 2108 return; 2109 } 2110 assert(haveFreelistLocks(), "must have freelist locks"); 2111 assert_lock_strong(bitMapLock()); 2112 2113 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2114 2115 _cmsGen->gc_epilogue_work(full); 2116 2117 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2118 // in case sampling was not already enabled, enable it 2119 _start_sampling = true; 2120 } 2121 // reset _eden_chunk_array so sampling starts afresh 2122 _eden_chunk_index = 0; 2123 2124 size_t cms_used = _cmsGen->cmsSpace()->used(); 2125 2126 // update performance counters - this uses a special version of 2127 // update_counters() that allows the utilization to be passed as a 2128 // parameter, avoiding multiple calls to used(). 2129 // 2130 _cmsGen->update_counters(cms_used); 2131 2132 bitMapLock()->unlock(); 2133 releaseFreelistLocks(); 2134 2135 if (!CleanChunkPoolAsync) { 2136 Chunk::clean_chunk_pool(); 2137 } 2138 2139 set_did_compact(false); 2140 _between_prologue_and_epilogue = false; // ready for next cycle 2141 } 2142 2143 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2144 collector()->gc_epilogue(full); 2145 2146 // When using ParNew, promotion tracking should have already been 2147 // disabled. However, the prologue (which enables promotion 2148 // tracking) and epilogue are called irrespective of the type of 2149 // GC. So they will also be called before and after Full GCs, during 2150 // which promotion tracking will not be explicitly disabled. So, 2151 // it's safer to also disable it here too (to be symmetric with 2152 // enabling it in the prologue). 2153 for (uint i = 0; i < ParallelGCThreads; i++) { 2154 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2155 } 2156 } 2157 2158 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2159 assert(!incremental_collection_failed(), "Should have been cleared"); 2160 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2161 cmsSpace()->gc_epilogue(); 2162 // Print stat counters 2163 NOT_PRODUCT( 2164 assert(_numObjectsAllocated == 0, "check"); 2165 assert(_numWordsAllocated == 0, "check"); 2166 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2167 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2168 _numObjectsPromoted = 0; 2169 _numWordsPromoted = 0; 2170 ) 2171 2172 // Call down the chain in contiguous_available needs the freelistLock 2173 // so print this out before releasing the freeListLock. 2174 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2175 } 2176 2177 #ifndef PRODUCT 2178 bool CMSCollector::have_cms_token() { 2179 Thread* thr = Thread::current(); 2180 if (thr->is_VM_thread()) { 2181 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2182 } else if (thr->is_ConcurrentGC_thread()) { 2183 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2184 } else if (thr->is_GC_task_thread()) { 2185 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2186 ParGCRareEvent_lock->owned_by_self(); 2187 } 2188 return false; 2189 } 2190 2191 // Check reachability of the given heap address in CMS generation, 2192 // treating all other generations as roots. 2193 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2194 // We could "guarantee" below, rather than assert, but I'll 2195 // leave these as "asserts" so that an adventurous debugger 2196 // could try this in the product build provided some subset of 2197 // the conditions were met, provided they were interested in the 2198 // results and knew that the computation below wouldn't interfere 2199 // with other concurrent computations mutating the structures 2200 // being read or written. 2201 assert(SafepointSynchronize::is_at_safepoint(), 2202 "Else mutations in object graph will make answer suspect"); 2203 assert(have_cms_token(), "Should hold cms token"); 2204 assert(haveFreelistLocks(), "must hold free list locks"); 2205 assert_lock_strong(bitMapLock()); 2206 2207 // Clear the marking bit map array before starting, but, just 2208 // for kicks, first report if the given address is already marked 2209 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2210 _markBitMap.isMarked(addr) ? "" : " not"); 2211 2212 if (verify_after_remark()) { 2213 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2214 bool result = verification_mark_bm()->isMarked(addr); 2215 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2216 result ? "IS" : "is NOT"); 2217 return result; 2218 } else { 2219 tty->print_cr("Could not compute result"); 2220 return false; 2221 } 2222 } 2223 #endif 2224 2225 void 2226 CMSCollector::print_on_error(outputStream* st) { 2227 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2228 if (collector != NULL) { 2229 CMSBitMap* bitmap = &collector->_markBitMap; 2230 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2231 bitmap->print_on_error(st, " Bits: "); 2232 2233 st->cr(); 2234 2235 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2236 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2237 mut_bitmap->print_on_error(st, " Bits: "); 2238 } 2239 } 2240 2241 //////////////////////////////////////////////////////// 2242 // CMS Verification Support 2243 //////////////////////////////////////////////////////// 2244 // Following the remark phase, the following invariant 2245 // should hold -- each object in the CMS heap which is 2246 // marked in markBitMap() should be marked in the verification_mark_bm(). 2247 2248 class VerifyMarkedClosure: public BitMapClosure { 2249 CMSBitMap* _marks; 2250 bool _failed; 2251 2252 public: 2253 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2254 2255 bool do_bit(size_t offset) { 2256 HeapWord* addr = _marks->offsetToHeapWord(offset); 2257 if (!_marks->isMarked(addr)) { 2258 Log(gc, verify) log; 2259 ResourceMark rm; 2260 LogStream ls(log.error()); 2261 oop(addr)->print_on(&ls); 2262 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2263 _failed = true; 2264 } 2265 return true; 2266 } 2267 2268 bool failed() { return _failed; } 2269 }; 2270 2271 bool CMSCollector::verify_after_remark() { 2272 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2273 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2274 static bool init = false; 2275 2276 assert(SafepointSynchronize::is_at_safepoint(), 2277 "Else mutations in object graph will make answer suspect"); 2278 assert(have_cms_token(), 2279 "Else there may be mutual interference in use of " 2280 " verification data structures"); 2281 assert(_collectorState > Marking && _collectorState <= Sweeping, 2282 "Else marking info checked here may be obsolete"); 2283 assert(haveFreelistLocks(), "must hold free list locks"); 2284 assert_lock_strong(bitMapLock()); 2285 2286 2287 // Allocate marking bit map if not already allocated 2288 if (!init) { // first time 2289 if (!verification_mark_bm()->allocate(_span)) { 2290 return false; 2291 } 2292 init = true; 2293 } 2294 2295 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2296 2297 // Turn off refs discovery -- so we will be tracing through refs. 2298 // This is as intended, because by this time 2299 // GC must already have cleared any refs that need to be cleared, 2300 // and traced those that need to be marked; moreover, 2301 // the marking done here is not going to interfere in any 2302 // way with the marking information used by GC. 2303 NoRefDiscovery no_discovery(ref_processor()); 2304 2305 #if COMPILER2_OR_JVMCI 2306 DerivedPointerTableDeactivate dpt_deact; 2307 #endif 2308 2309 // Clear any marks from a previous round 2310 verification_mark_bm()->clear_all(); 2311 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2312 verify_work_stacks_empty(); 2313 2314 CMSHeap* heap = CMSHeap::heap(); 2315 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2316 // Update the saved marks which may affect the root scans. 2317 heap->save_marks(); 2318 2319 if (CMSRemarkVerifyVariant == 1) { 2320 // In this first variant of verification, we complete 2321 // all marking, then check if the new marks-vector is 2322 // a subset of the CMS marks-vector. 2323 verify_after_remark_work_1(); 2324 } else { 2325 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2326 // In this second variant of verification, we flag an error 2327 // (i.e. an object reachable in the new marks-vector not reachable 2328 // in the CMS marks-vector) immediately, also indicating the 2329 // identify of an object (A) that references the unmarked object (B) -- 2330 // presumably, a mutation to A failed to be picked up by preclean/remark? 2331 verify_after_remark_work_2(); 2332 } 2333 2334 return true; 2335 } 2336 2337 void CMSCollector::verify_after_remark_work_1() { 2338 ResourceMark rm; 2339 HandleMark hm; 2340 CMSHeap* heap = CMSHeap::heap(); 2341 2342 // Get a clear set of claim bits for the roots processing to work with. 2343 ClassLoaderDataGraph::clear_claimed_marks(); 2344 2345 // Mark from roots one level into CMS 2346 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2347 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2348 2349 { 2350 StrongRootsScope srs(1); 2351 2352 heap->cms_process_roots(&srs, 2353 true, // young gen as roots 2354 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2355 should_unload_classes(), 2356 ¬Older, 2357 NULL); 2358 } 2359 2360 // Now mark from the roots 2361 MarkFromRootsClosure markFromRootsClosure(this, _span, 2362 verification_mark_bm(), verification_mark_stack(), 2363 false /* don't yield */, true /* verifying */); 2364 assert(_restart_addr == NULL, "Expected pre-condition"); 2365 verification_mark_bm()->iterate(&markFromRootsClosure); 2366 while (_restart_addr != NULL) { 2367 // Deal with stack overflow: by restarting at the indicated 2368 // address. 2369 HeapWord* ra = _restart_addr; 2370 markFromRootsClosure.reset(ra); 2371 _restart_addr = NULL; 2372 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2373 } 2374 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2375 verify_work_stacks_empty(); 2376 2377 // Marking completed -- now verify that each bit marked in 2378 // verification_mark_bm() is also marked in markBitMap(); flag all 2379 // errors by printing corresponding objects. 2380 VerifyMarkedClosure vcl(markBitMap()); 2381 verification_mark_bm()->iterate(&vcl); 2382 if (vcl.failed()) { 2383 Log(gc, verify) log; 2384 log.error("Failed marking verification after remark"); 2385 ResourceMark rm; 2386 LogStream ls(log.error()); 2387 heap->print_on(&ls); 2388 fatal("CMS: failed marking verification after remark"); 2389 } 2390 } 2391 2392 class VerifyCLDOopsCLDClosure : public CLDClosure { 2393 class VerifyCLDOopsClosure : public OopClosure { 2394 CMSBitMap* _bitmap; 2395 public: 2396 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2397 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2398 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2399 } _oop_closure; 2400 public: 2401 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2402 void do_cld(ClassLoaderData* cld) { 2403 cld->oops_do(&_oop_closure, false, false); 2404 } 2405 }; 2406 2407 void CMSCollector::verify_after_remark_work_2() { 2408 ResourceMark rm; 2409 HandleMark hm; 2410 CMSHeap* heap = CMSHeap::heap(); 2411 2412 // Get a clear set of claim bits for the roots processing to work with. 2413 ClassLoaderDataGraph::clear_claimed_marks(); 2414 2415 // Mark from roots one level into CMS 2416 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2417 markBitMap()); 2418 CLDToOopClosure cld_closure(¬Older, true); 2419 2420 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2421 2422 { 2423 StrongRootsScope srs(1); 2424 2425 heap->cms_process_roots(&srs, 2426 true, // young gen as roots 2427 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2428 should_unload_classes(), 2429 ¬Older, 2430 &cld_closure); 2431 } 2432 2433 // Now mark from the roots 2434 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2435 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2436 assert(_restart_addr == NULL, "Expected pre-condition"); 2437 verification_mark_bm()->iterate(&markFromRootsClosure); 2438 while (_restart_addr != NULL) { 2439 // Deal with stack overflow: by restarting at the indicated 2440 // address. 2441 HeapWord* ra = _restart_addr; 2442 markFromRootsClosure.reset(ra); 2443 _restart_addr = NULL; 2444 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2445 } 2446 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2447 verify_work_stacks_empty(); 2448 2449 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2450 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2451 2452 // Marking completed -- now verify that each bit marked in 2453 // verification_mark_bm() is also marked in markBitMap(); flag all 2454 // errors by printing corresponding objects. 2455 VerifyMarkedClosure vcl(markBitMap()); 2456 verification_mark_bm()->iterate(&vcl); 2457 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2458 } 2459 2460 void ConcurrentMarkSweepGeneration::save_marks() { 2461 // delegate to CMS space 2462 cmsSpace()->save_marks(); 2463 } 2464 2465 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2466 return cmsSpace()->no_allocs_since_save_marks(); 2467 } 2468 2469 void 2470 ConcurrentMarkSweepGeneration::oop_iterate(ExtendedOopClosure* cl) { 2471 if (freelistLock()->owned_by_self()) { 2472 Generation::oop_iterate(cl); 2473 } else { 2474 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2475 Generation::oop_iterate(cl); 2476 } 2477 } 2478 2479 void 2480 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2481 if (freelistLock()->owned_by_self()) { 2482 Generation::object_iterate(cl); 2483 } else { 2484 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2485 Generation::object_iterate(cl); 2486 } 2487 } 2488 2489 void 2490 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2491 if (freelistLock()->owned_by_self()) { 2492 Generation::safe_object_iterate(cl); 2493 } else { 2494 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2495 Generation::safe_object_iterate(cl); 2496 } 2497 } 2498 2499 void 2500 ConcurrentMarkSweepGeneration::post_compact() { 2501 } 2502 2503 void 2504 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2505 // Fix the linear allocation blocks to look like free blocks. 2506 2507 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2508 // are not called when the heap is verified during universe initialization and 2509 // at vm shutdown. 2510 if (freelistLock()->owned_by_self()) { 2511 cmsSpace()->prepare_for_verify(); 2512 } else { 2513 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2514 cmsSpace()->prepare_for_verify(); 2515 } 2516 } 2517 2518 void 2519 ConcurrentMarkSweepGeneration::verify() { 2520 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2521 // are not called when the heap is verified during universe initialization and 2522 // at vm shutdown. 2523 if (freelistLock()->owned_by_self()) { 2524 cmsSpace()->verify(); 2525 } else { 2526 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2527 cmsSpace()->verify(); 2528 } 2529 } 2530 2531 void CMSCollector::verify() { 2532 _cmsGen->verify(); 2533 } 2534 2535 #ifndef PRODUCT 2536 bool CMSCollector::overflow_list_is_empty() const { 2537 assert(_num_par_pushes >= 0, "Inconsistency"); 2538 if (_overflow_list == NULL) { 2539 assert(_num_par_pushes == 0, "Inconsistency"); 2540 } 2541 return _overflow_list == NULL; 2542 } 2543 2544 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2545 // merely consolidate assertion checks that appear to occur together frequently. 2546 void CMSCollector::verify_work_stacks_empty() const { 2547 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2548 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2549 } 2550 2551 void CMSCollector::verify_overflow_empty() const { 2552 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2553 assert(no_preserved_marks(), "No preserved marks"); 2554 } 2555 #endif // PRODUCT 2556 2557 // Decide if we want to enable class unloading as part of the 2558 // ensuing concurrent GC cycle. We will collect and 2559 // unload classes if it's the case that: 2560 // (a) class unloading is enabled at the command line, and 2561 // (b) old gen is getting really full 2562 // NOTE: Provided there is no change in the state of the heap between 2563 // calls to this method, it should have idempotent results. Moreover, 2564 // its results should be monotonically increasing (i.e. going from 0 to 1, 2565 // but not 1 to 0) between successive calls between which the heap was 2566 // not collected. For the implementation below, it must thus rely on 2567 // the property that concurrent_cycles_since_last_unload() 2568 // will not decrease unless a collection cycle happened and that 2569 // _cmsGen->is_too_full() are 2570 // themselves also monotonic in that sense. See check_monotonicity() 2571 // below. 2572 void CMSCollector::update_should_unload_classes() { 2573 _should_unload_classes = false; 2574 if (CMSClassUnloadingEnabled) { 2575 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2576 CMSClassUnloadingMaxInterval) 2577 || _cmsGen->is_too_full(); 2578 } 2579 } 2580 2581 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2582 bool res = should_concurrent_collect(); 2583 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2584 return res; 2585 } 2586 2587 void CMSCollector::setup_cms_unloading_and_verification_state() { 2588 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2589 || VerifyBeforeExit; 2590 const int rso = GenCollectedHeap::SO_AllCodeCache; 2591 2592 // We set the proper root for this CMS cycle here. 2593 if (should_unload_classes()) { // Should unload classes this cycle 2594 remove_root_scanning_option(rso); // Shrink the root set appropriately 2595 set_verifying(should_verify); // Set verification state for this cycle 2596 return; // Nothing else needs to be done at this time 2597 } 2598 2599 // Not unloading classes this cycle 2600 assert(!should_unload_classes(), "Inconsistency!"); 2601 2602 // If we are not unloading classes then add SO_AllCodeCache to root 2603 // scanning options. 2604 add_root_scanning_option(rso); 2605 2606 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2607 set_verifying(true); 2608 } else if (verifying() && !should_verify) { 2609 // We were verifying, but some verification flags got disabled. 2610 set_verifying(false); 2611 // Exclude symbols, strings and code cache elements from root scanning to 2612 // reduce IM and RM pauses. 2613 remove_root_scanning_option(rso); 2614 } 2615 } 2616 2617 2618 #ifndef PRODUCT 2619 HeapWord* CMSCollector::block_start(const void* p) const { 2620 const HeapWord* addr = (HeapWord*)p; 2621 if (_span.contains(p)) { 2622 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2623 return _cmsGen->cmsSpace()->block_start(p); 2624 } 2625 } 2626 return NULL; 2627 } 2628 #endif 2629 2630 HeapWord* 2631 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2632 bool tlab, 2633 bool parallel) { 2634 CMSSynchronousYieldRequest yr; 2635 assert(!tlab, "Can't deal with TLAB allocation"); 2636 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2637 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2638 if (GCExpandToAllocateDelayMillis > 0) { 2639 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2640 } 2641 return have_lock_and_allocate(word_size, tlab); 2642 } 2643 2644 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2645 size_t bytes, 2646 size_t expand_bytes, 2647 CMSExpansionCause::Cause cause) 2648 { 2649 2650 bool success = expand(bytes, expand_bytes); 2651 2652 // remember why we expanded; this information is used 2653 // by shouldConcurrentCollect() when making decisions on whether to start 2654 // a new CMS cycle. 2655 if (success) { 2656 set_expansion_cause(cause); 2657 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2658 } 2659 } 2660 2661 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2662 HeapWord* res = NULL; 2663 MutexLocker x(ParGCRareEvent_lock); 2664 while (true) { 2665 // Expansion by some other thread might make alloc OK now: 2666 res = ps->lab.alloc(word_sz); 2667 if (res != NULL) return res; 2668 // If there's not enough expansion space available, give up. 2669 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2670 return NULL; 2671 } 2672 // Otherwise, we try expansion. 2673 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2674 // Now go around the loop and try alloc again; 2675 // A competing par_promote might beat us to the expansion space, 2676 // so we may go around the loop again if promotion fails again. 2677 if (GCExpandToAllocateDelayMillis > 0) { 2678 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2679 } 2680 } 2681 } 2682 2683 2684 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2685 PromotionInfo* promo) { 2686 MutexLocker x(ParGCRareEvent_lock); 2687 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2688 while (true) { 2689 // Expansion by some other thread might make alloc OK now: 2690 if (promo->ensure_spooling_space()) { 2691 assert(promo->has_spooling_space(), 2692 "Post-condition of successful ensure_spooling_space()"); 2693 return true; 2694 } 2695 // If there's not enough expansion space available, give up. 2696 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2697 return false; 2698 } 2699 // Otherwise, we try expansion. 2700 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2701 // Now go around the loop and try alloc again; 2702 // A competing allocation might beat us to the expansion space, 2703 // so we may go around the loop again if allocation fails again. 2704 if (GCExpandToAllocateDelayMillis > 0) { 2705 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2706 } 2707 } 2708 } 2709 2710 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2711 // Only shrink if a compaction was done so that all the free space 2712 // in the generation is in a contiguous block at the end. 2713 if (did_compact()) { 2714 CardGeneration::shrink(bytes); 2715 } 2716 } 2717 2718 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2719 assert_locked_or_safepoint(Heap_lock); 2720 } 2721 2722 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2723 assert_locked_or_safepoint(Heap_lock); 2724 assert_lock_strong(freelistLock()); 2725 log_trace(gc)("Shrinking of CMS not yet implemented"); 2726 return; 2727 } 2728 2729 2730 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2731 // phases. 2732 class CMSPhaseAccounting: public StackObj { 2733 public: 2734 CMSPhaseAccounting(CMSCollector *collector, 2735 const char *title); 2736 ~CMSPhaseAccounting(); 2737 2738 private: 2739 CMSCollector *_collector; 2740 const char *_title; 2741 GCTraceConcTime(Info, gc) _trace_time; 2742 2743 public: 2744 // Not MT-safe; so do not pass around these StackObj's 2745 // where they may be accessed by other threads. 2746 double wallclock_millis() { 2747 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2748 } 2749 }; 2750 2751 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2752 const char *title) : 2753 _collector(collector), _title(title), _trace_time(title) { 2754 2755 _collector->resetYields(); 2756 _collector->resetTimer(); 2757 _collector->startTimer(); 2758 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2759 } 2760 2761 CMSPhaseAccounting::~CMSPhaseAccounting() { 2762 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2763 _collector->stopTimer(); 2764 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2765 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2766 } 2767 2768 // CMS work 2769 2770 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2771 class CMSParMarkTask : public AbstractGangTask { 2772 protected: 2773 CMSCollector* _collector; 2774 uint _n_workers; 2775 OopStorage::ParState<false, false> _par_state_string; 2776 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2777 AbstractGangTask(name), 2778 _collector(collector), 2779 _n_workers(n_workers), 2780 _par_state_string(StringTable::weak_storage()) {} 2781 // Work method in support of parallel rescan ... of young gen spaces 2782 void do_young_space_rescan(OopsInGenClosure* cl, 2783 ContiguousSpace* space, 2784 HeapWord** chunk_array, size_t chunk_top); 2785 void work_on_young_gen_roots(OopsInGenClosure* cl); 2786 }; 2787 2788 // Parallel initial mark task 2789 class CMSParInitialMarkTask: public CMSParMarkTask { 2790 StrongRootsScope* _strong_roots_scope; 2791 public: 2792 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2793 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2794 _strong_roots_scope(strong_roots_scope) {} 2795 void work(uint worker_id); 2796 }; 2797 2798 // Checkpoint the roots into this generation from outside 2799 // this generation. [Note this initial checkpoint need only 2800 // be approximate -- we'll do a catch up phase subsequently.] 2801 void CMSCollector::checkpointRootsInitial() { 2802 assert(_collectorState == InitialMarking, "Wrong collector state"); 2803 check_correct_thread_executing(); 2804 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2805 2806 save_heap_summary(); 2807 report_heap_summary(GCWhen::BeforeGC); 2808 2809 ReferenceProcessor* rp = ref_processor(); 2810 assert(_restart_addr == NULL, "Control point invariant"); 2811 { 2812 // acquire locks for subsequent manipulations 2813 MutexLockerEx x(bitMapLock(), 2814 Mutex::_no_safepoint_check_flag); 2815 checkpointRootsInitialWork(); 2816 // enable ("weak") refs discovery 2817 rp->enable_discovery(); 2818 _collectorState = Marking; 2819 } 2820 } 2821 2822 void CMSCollector::checkpointRootsInitialWork() { 2823 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2824 assert(_collectorState == InitialMarking, "just checking"); 2825 2826 // Already have locks. 2827 assert_lock_strong(bitMapLock()); 2828 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2829 2830 // Setup the verification and class unloading state for this 2831 // CMS collection cycle. 2832 setup_cms_unloading_and_verification_state(); 2833 2834 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2835 2836 // Reset all the PLAB chunk arrays if necessary. 2837 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2838 reset_survivor_plab_arrays(); 2839 } 2840 2841 ResourceMark rm; 2842 HandleMark hm; 2843 2844 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2845 CMSHeap* heap = CMSHeap::heap(); 2846 2847 verify_work_stacks_empty(); 2848 verify_overflow_empty(); 2849 2850 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2851 // Update the saved marks which may affect the root scans. 2852 heap->save_marks(); 2853 2854 // weak reference processing has not started yet. 2855 ref_processor()->set_enqueuing_is_done(false); 2856 2857 // Need to remember all newly created CLDs, 2858 // so that we can guarantee that the remark finds them. 2859 ClassLoaderDataGraph::remember_new_clds(true); 2860 2861 // Whenever a CLD is found, it will be claimed before proceeding to mark 2862 // the klasses. The claimed marks need to be cleared before marking starts. 2863 ClassLoaderDataGraph::clear_claimed_marks(); 2864 2865 print_eden_and_survivor_chunk_arrays(); 2866 2867 { 2868 #if COMPILER2_OR_JVMCI 2869 DerivedPointerTableDeactivate dpt_deact; 2870 #endif 2871 if (CMSParallelInitialMarkEnabled) { 2872 // The parallel version. 2873 WorkGang* workers = heap->workers(); 2874 assert(workers != NULL, "Need parallel worker threads."); 2875 uint n_workers = workers->active_workers(); 2876 2877 StrongRootsScope srs(n_workers); 2878 2879 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2880 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2881 // If the total workers is greater than 1, then multiple workers 2882 // may be used at some time and the initialization has been set 2883 // such that the single threaded path cannot be used. 2884 if (workers->total_workers() > 1) { 2885 workers->run_task(&tsk); 2886 } else { 2887 tsk.work(0); 2888 } 2889 } else { 2890 // The serial version. 2891 CLDToOopClosure cld_closure(¬Older, true); 2892 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2893 2894 StrongRootsScope srs(1); 2895 2896 heap->cms_process_roots(&srs, 2897 true, // young gen as roots 2898 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2899 should_unload_classes(), 2900 ¬Older, 2901 &cld_closure); 2902 } 2903 } 2904 2905 // Clear mod-union table; it will be dirtied in the prologue of 2906 // CMS generation per each young generation collection. 2907 2908 assert(_modUnionTable.isAllClear(), 2909 "Was cleared in most recent final checkpoint phase" 2910 " or no bits are set in the gc_prologue before the start of the next " 2911 "subsequent marking phase."); 2912 2913 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2914 2915 // Save the end of the used_region of the constituent generations 2916 // to be used to limit the extent of sweep in each generation. 2917 save_sweep_limits(); 2918 verify_overflow_empty(); 2919 } 2920 2921 bool CMSCollector::markFromRoots() { 2922 // we might be tempted to assert that: 2923 // assert(!SafepointSynchronize::is_at_safepoint(), 2924 // "inconsistent argument?"); 2925 // However that wouldn't be right, because it's possible that 2926 // a safepoint is indeed in progress as a young generation 2927 // stop-the-world GC happens even as we mark in this generation. 2928 assert(_collectorState == Marking, "inconsistent state?"); 2929 check_correct_thread_executing(); 2930 verify_overflow_empty(); 2931 2932 // Weak ref discovery note: We may be discovering weak 2933 // refs in this generation concurrent (but interleaved) with 2934 // weak ref discovery by the young generation collector. 2935 2936 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2937 GCTraceCPUTime tcpu; 2938 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2939 bool res = markFromRootsWork(); 2940 if (res) { 2941 _collectorState = Precleaning; 2942 } else { // We failed and a foreground collection wants to take over 2943 assert(_foregroundGCIsActive, "internal state inconsistency"); 2944 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2945 log_debug(gc)("bailing out to foreground collection"); 2946 } 2947 verify_overflow_empty(); 2948 return res; 2949 } 2950 2951 bool CMSCollector::markFromRootsWork() { 2952 // iterate over marked bits in bit map, doing a full scan and mark 2953 // from these roots using the following algorithm: 2954 // . if oop is to the right of the current scan pointer, 2955 // mark corresponding bit (we'll process it later) 2956 // . else (oop is to left of current scan pointer) 2957 // push oop on marking stack 2958 // . drain the marking stack 2959 2960 // Note that when we do a marking step we need to hold the 2961 // bit map lock -- recall that direct allocation (by mutators) 2962 // and promotion (by the young generation collector) is also 2963 // marking the bit map. [the so-called allocate live policy.] 2964 // Because the implementation of bit map marking is not 2965 // robust wrt simultaneous marking of bits in the same word, 2966 // we need to make sure that there is no such interference 2967 // between concurrent such updates. 2968 2969 // already have locks 2970 assert_lock_strong(bitMapLock()); 2971 2972 verify_work_stacks_empty(); 2973 verify_overflow_empty(); 2974 bool result = false; 2975 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2976 result = do_marking_mt(); 2977 } else { 2978 result = do_marking_st(); 2979 } 2980 return result; 2981 } 2982 2983 // Forward decl 2984 class CMSConcMarkingTask; 2985 2986 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2987 CMSCollector* _collector; 2988 CMSConcMarkingTask* _task; 2989 public: 2990 virtual void yield(); 2991 2992 // "n_threads" is the number of threads to be terminated. 2993 // "queue_set" is a set of work queues of other threads. 2994 // "collector" is the CMS collector associated with this task terminator. 2995 // "yield" indicates whether we need the gang as a whole to yield. 2996 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2997 ParallelTaskTerminator(n_threads, queue_set), 2998 _collector(collector) { } 2999 3000 void set_task(CMSConcMarkingTask* task) { 3001 _task = task; 3002 } 3003 }; 3004 3005 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3006 CMSConcMarkingTask* _task; 3007 public: 3008 bool should_exit_termination(); 3009 void set_task(CMSConcMarkingTask* task) { 3010 _task = task; 3011 } 3012 }; 3013 3014 // MT Concurrent Marking Task 3015 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3016 CMSCollector* _collector; 3017 uint _n_workers; // requested/desired # workers 3018 bool _result; 3019 CompactibleFreeListSpace* _cms_space; 3020 char _pad_front[64]; // padding to ... 3021 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3022 char _pad_back[64]; 3023 HeapWord* _restart_addr; 3024 3025 // Exposed here for yielding support 3026 Mutex* const _bit_map_lock; 3027 3028 // The per thread work queues, available here for stealing 3029 OopTaskQueueSet* _task_queues; 3030 3031 // Termination (and yielding) support 3032 CMSConcMarkingTerminator _term; 3033 CMSConcMarkingTerminatorTerminator _term_term; 3034 3035 public: 3036 CMSConcMarkingTask(CMSCollector* collector, 3037 CompactibleFreeListSpace* cms_space, 3038 YieldingFlexibleWorkGang* workers, 3039 OopTaskQueueSet* task_queues): 3040 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3041 _collector(collector), 3042 _cms_space(cms_space), 3043 _n_workers(0), _result(true), 3044 _task_queues(task_queues), 3045 _term(_n_workers, task_queues, _collector), 3046 _bit_map_lock(collector->bitMapLock()) 3047 { 3048 _requested_size = _n_workers; 3049 _term.set_task(this); 3050 _term_term.set_task(this); 3051 _restart_addr = _global_finger = _cms_space->bottom(); 3052 } 3053 3054 3055 OopTaskQueueSet* task_queues() { return _task_queues; } 3056 3057 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3058 3059 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3060 3061 CMSConcMarkingTerminator* terminator() { return &_term; } 3062 3063 virtual void set_for_termination(uint active_workers) { 3064 terminator()->reset_for_reuse(active_workers); 3065 } 3066 3067 void work(uint worker_id); 3068 bool should_yield() { 3069 return ConcurrentMarkSweepThread::should_yield() 3070 && !_collector->foregroundGCIsActive(); 3071 } 3072 3073 virtual void coordinator_yield(); // stuff done by coordinator 3074 bool result() { return _result; } 3075 3076 void reset(HeapWord* ra) { 3077 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3078 _restart_addr = _global_finger = ra; 3079 _term.reset_for_reuse(); 3080 } 3081 3082 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3083 OopTaskQueue* work_q); 3084 3085 private: 3086 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3087 void do_work_steal(int i); 3088 void bump_global_finger(HeapWord* f); 3089 }; 3090 3091 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3092 assert(_task != NULL, "Error"); 3093 return _task->yielding(); 3094 // Note that we do not need the disjunct || _task->should_yield() above 3095 // because we want terminating threads to yield only if the task 3096 // is already in the midst of yielding, which happens only after at least one 3097 // thread has yielded. 3098 } 3099 3100 void CMSConcMarkingTerminator::yield() { 3101 if (_task->should_yield()) { 3102 _task->yield(); 3103 } else { 3104 ParallelTaskTerminator::yield(); 3105 } 3106 } 3107 3108 //////////////////////////////////////////////////////////////// 3109 // Concurrent Marking Algorithm Sketch 3110 //////////////////////////////////////////////////////////////// 3111 // Until all tasks exhausted (both spaces): 3112 // -- claim next available chunk 3113 // -- bump global finger via CAS 3114 // -- find first object that starts in this chunk 3115 // and start scanning bitmap from that position 3116 // -- scan marked objects for oops 3117 // -- CAS-mark target, and if successful: 3118 // . if target oop is above global finger (volatile read) 3119 // nothing to do 3120 // . if target oop is in chunk and above local finger 3121 // then nothing to do 3122 // . else push on work-queue 3123 // -- Deal with possible overflow issues: 3124 // . local work-queue overflow causes stuff to be pushed on 3125 // global (common) overflow queue 3126 // . always first empty local work queue 3127 // . then get a batch of oops from global work queue if any 3128 // . then do work stealing 3129 // -- When all tasks claimed (both spaces) 3130 // and local work queue empty, 3131 // then in a loop do: 3132 // . check global overflow stack; steal a batch of oops and trace 3133 // . try to steal from other threads oif GOS is empty 3134 // . if neither is available, offer termination 3135 // -- Terminate and return result 3136 // 3137 void CMSConcMarkingTask::work(uint worker_id) { 3138 elapsedTimer _timer; 3139 ResourceMark rm; 3140 HandleMark hm; 3141 3142 DEBUG_ONLY(_collector->verify_overflow_empty();) 3143 3144 // Before we begin work, our work queue should be empty 3145 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3146 // Scan the bitmap covering _cms_space, tracing through grey objects. 3147 _timer.start(); 3148 do_scan_and_mark(worker_id, _cms_space); 3149 _timer.stop(); 3150 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3151 3152 // ... do work stealing 3153 _timer.reset(); 3154 _timer.start(); 3155 do_work_steal(worker_id); 3156 _timer.stop(); 3157 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3158 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3159 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3160 // Note that under the current task protocol, the 3161 // following assertion is true even of the spaces 3162 // expanded since the completion of the concurrent 3163 // marking. XXX This will likely change under a strict 3164 // ABORT semantics. 3165 // After perm removal the comparison was changed to 3166 // greater than or equal to from strictly greater than. 3167 // Before perm removal the highest address sweep would 3168 // have been at the end of perm gen but now is at the 3169 // end of the tenured gen. 3170 assert(_global_finger >= _cms_space->end(), 3171 "All tasks have been completed"); 3172 DEBUG_ONLY(_collector->verify_overflow_empty();) 3173 } 3174 3175 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3176 HeapWord* read = _global_finger; 3177 HeapWord* cur = read; 3178 while (f > read) { 3179 cur = read; 3180 read = Atomic::cmpxchg(f, &_global_finger, cur); 3181 if (cur == read) { 3182 // our cas succeeded 3183 assert(_global_finger >= f, "protocol consistency"); 3184 break; 3185 } 3186 } 3187 } 3188 3189 // This is really inefficient, and should be redone by 3190 // using (not yet available) block-read and -write interfaces to the 3191 // stack and the work_queue. XXX FIX ME !!! 3192 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3193 OopTaskQueue* work_q) { 3194 // Fast lock-free check 3195 if (ovflw_stk->length() == 0) { 3196 return false; 3197 } 3198 assert(work_q->size() == 0, "Shouldn't steal"); 3199 MutexLockerEx ml(ovflw_stk->par_lock(), 3200 Mutex::_no_safepoint_check_flag); 3201 // Grab up to 1/4 the size of the work queue 3202 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3203 (size_t)ParGCDesiredObjsFromOverflowList); 3204 num = MIN2(num, ovflw_stk->length()); 3205 for (int i = (int) num; i > 0; i--) { 3206 oop cur = ovflw_stk->pop(); 3207 assert(cur != NULL, "Counted wrong?"); 3208 work_q->push(cur); 3209 } 3210 return num > 0; 3211 } 3212 3213 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3214 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3215 int n_tasks = pst->n_tasks(); 3216 // We allow that there may be no tasks to do here because 3217 // we are restarting after a stack overflow. 3218 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3219 uint nth_task = 0; 3220 3221 HeapWord* aligned_start = sp->bottom(); 3222 if (sp->used_region().contains(_restart_addr)) { 3223 // Align down to a card boundary for the start of 0th task 3224 // for this space. 3225 aligned_start = align_down(_restart_addr, CardTable::card_size); 3226 } 3227 3228 size_t chunk_size = sp->marking_task_size(); 3229 while (!pst->is_task_claimed(/* reference */ nth_task)) { 3230 // Having claimed the nth task in this space, 3231 // compute the chunk that it corresponds to: 3232 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3233 aligned_start + (nth_task+1)*chunk_size); 3234 // Try and bump the global finger via a CAS; 3235 // note that we need to do the global finger bump 3236 // _before_ taking the intersection below, because 3237 // the task corresponding to that region will be 3238 // deemed done even if the used_region() expands 3239 // because of allocation -- as it almost certainly will 3240 // during start-up while the threads yield in the 3241 // closure below. 3242 HeapWord* finger = span.end(); 3243 bump_global_finger(finger); // atomically 3244 // There are null tasks here corresponding to chunks 3245 // beyond the "top" address of the space. 3246 span = span.intersection(sp->used_region()); 3247 if (!span.is_empty()) { // Non-null task 3248 HeapWord* prev_obj; 3249 assert(!span.contains(_restart_addr) || nth_task == 0, 3250 "Inconsistency"); 3251 if (nth_task == 0) { 3252 // For the 0th task, we'll not need to compute a block_start. 3253 if (span.contains(_restart_addr)) { 3254 // In the case of a restart because of stack overflow, 3255 // we might additionally skip a chunk prefix. 3256 prev_obj = _restart_addr; 3257 } else { 3258 prev_obj = span.start(); 3259 } 3260 } else { 3261 // We want to skip the first object because 3262 // the protocol is to scan any object in its entirety 3263 // that _starts_ in this span; a fortiori, any 3264 // object starting in an earlier span is scanned 3265 // as part of an earlier claimed task. 3266 // Below we use the "careful" version of block_start 3267 // so we do not try to navigate uninitialized objects. 3268 prev_obj = sp->block_start_careful(span.start()); 3269 // Below we use a variant of block_size that uses the 3270 // Printezis bits to avoid waiting for allocated 3271 // objects to become initialized/parsable. 3272 while (prev_obj < span.start()) { 3273 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3274 if (sz > 0) { 3275 prev_obj += sz; 3276 } else { 3277 // In this case we may end up doing a bit of redundant 3278 // scanning, but that appears unavoidable, short of 3279 // locking the free list locks; see bug 6324141. 3280 break; 3281 } 3282 } 3283 } 3284 if (prev_obj < span.end()) { 3285 MemRegion my_span = MemRegion(prev_obj, span.end()); 3286 // Do the marking work within a non-empty span -- 3287 // the last argument to the constructor indicates whether the 3288 // iteration should be incremental with periodic yields. 3289 ParMarkFromRootsClosure cl(this, _collector, my_span, 3290 &_collector->_markBitMap, 3291 work_queue(i), 3292 &_collector->_markStack); 3293 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3294 } // else nothing to do for this task 3295 } // else nothing to do for this task 3296 } 3297 // We'd be tempted to assert here that since there are no 3298 // more tasks left to claim in this space, the global_finger 3299 // must exceed space->top() and a fortiori space->end(). However, 3300 // that would not quite be correct because the bumping of 3301 // global_finger occurs strictly after the claiming of a task, 3302 // so by the time we reach here the global finger may not yet 3303 // have been bumped up by the thread that claimed the last 3304 // task. 3305 pst->all_tasks_completed(); 3306 } 3307 3308 class ParConcMarkingClosure: public MetadataAwareOopClosure { 3309 private: 3310 CMSCollector* _collector; 3311 CMSConcMarkingTask* _task; 3312 MemRegion _span; 3313 CMSBitMap* _bit_map; 3314 CMSMarkStack* _overflow_stack; 3315 OopTaskQueue* _work_queue; 3316 protected: 3317 DO_OOP_WORK_DEFN 3318 public: 3319 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3320 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3321 MetadataAwareOopClosure(collector->ref_processor()), 3322 _collector(collector), 3323 _task(task), 3324 _span(collector->_span), 3325 _work_queue(work_queue), 3326 _bit_map(bit_map), 3327 _overflow_stack(overflow_stack) 3328 { } 3329 virtual void do_oop(oop* p); 3330 virtual void do_oop(narrowOop* p); 3331 3332 void trim_queue(size_t max); 3333 void handle_stack_overflow(HeapWord* lost); 3334 void do_yield_check() { 3335 if (_task->should_yield()) { 3336 _task->yield(); 3337 } 3338 } 3339 }; 3340 3341 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3342 3343 // Grey object scanning during work stealing phase -- 3344 // the salient assumption here is that any references 3345 // that are in these stolen objects being scanned must 3346 // already have been initialized (else they would not have 3347 // been published), so we do not need to check for 3348 // uninitialized objects before pushing here. 3349 void ParConcMarkingClosure::do_oop(oop obj) { 3350 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3351 HeapWord* addr = (HeapWord*)obj; 3352 // Check if oop points into the CMS generation 3353 // and is not marked 3354 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3355 // a white object ... 3356 // If we manage to "claim" the object, by being the 3357 // first thread to mark it, then we push it on our 3358 // marking stack 3359 if (_bit_map->par_mark(addr)) { // ... now grey 3360 // push on work queue (grey set) 3361 bool simulate_overflow = false; 3362 NOT_PRODUCT( 3363 if (CMSMarkStackOverflowALot && 3364 _collector->simulate_overflow()) { 3365 // simulate a stack overflow 3366 simulate_overflow = true; 3367 } 3368 ) 3369 if (simulate_overflow || 3370 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3371 // stack overflow 3372 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3373 // We cannot assert that the overflow stack is full because 3374 // it may have been emptied since. 3375 assert(simulate_overflow || 3376 _work_queue->size() == _work_queue->max_elems(), 3377 "Else push should have succeeded"); 3378 handle_stack_overflow(addr); 3379 } 3380 } // Else, some other thread got there first 3381 do_yield_check(); 3382 } 3383 } 3384 3385 void ParConcMarkingClosure::do_oop(oop* p) { ParConcMarkingClosure::do_oop_work(p); } 3386 void ParConcMarkingClosure::do_oop(narrowOop* p) { ParConcMarkingClosure::do_oop_work(p); } 3387 3388 void ParConcMarkingClosure::trim_queue(size_t max) { 3389 while (_work_queue->size() > max) { 3390 oop new_oop; 3391 if (_work_queue->pop_local(new_oop)) { 3392 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3393 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3394 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3395 new_oop->oop_iterate(this); // do_oop() above 3396 do_yield_check(); 3397 } 3398 } 3399 } 3400 3401 // Upon stack overflow, we discard (part of) the stack, 3402 // remembering the least address amongst those discarded 3403 // in CMSCollector's _restart_address. 3404 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3405 // We need to do this under a mutex to prevent other 3406 // workers from interfering with the work done below. 3407 MutexLockerEx ml(_overflow_stack->par_lock(), 3408 Mutex::_no_safepoint_check_flag); 3409 // Remember the least grey address discarded 3410 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3411 _collector->lower_restart_addr(ra); 3412 _overflow_stack->reset(); // discard stack contents 3413 _overflow_stack->expand(); // expand the stack if possible 3414 } 3415 3416 3417 void CMSConcMarkingTask::do_work_steal(int i) { 3418 OopTaskQueue* work_q = work_queue(i); 3419 oop obj_to_scan; 3420 CMSBitMap* bm = &(_collector->_markBitMap); 3421 CMSMarkStack* ovflw = &(_collector->_markStack); 3422 int* seed = _collector->hash_seed(i); 3423 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3424 while (true) { 3425 cl.trim_queue(0); 3426 assert(work_q->size() == 0, "Should have been emptied above"); 3427 if (get_work_from_overflow_stack(ovflw, work_q)) { 3428 // Can't assert below because the work obtained from the 3429 // overflow stack may already have been stolen from us. 3430 // assert(work_q->size() > 0, "Work from overflow stack"); 3431 continue; 3432 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 3433 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3434 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3435 obj_to_scan->oop_iterate(&cl); 3436 } else if (terminator()->offer_termination(&_term_term)) { 3437 assert(work_q->size() == 0, "Impossible!"); 3438 break; 3439 } else if (yielding() || should_yield()) { 3440 yield(); 3441 } 3442 } 3443 } 3444 3445 // This is run by the CMS (coordinator) thread. 3446 void CMSConcMarkingTask::coordinator_yield() { 3447 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3448 "CMS thread should hold CMS token"); 3449 // First give up the locks, then yield, then re-lock 3450 // We should probably use a constructor/destructor idiom to 3451 // do this unlock/lock or modify the MutexUnlocker class to 3452 // serve our purpose. XXX 3453 assert_lock_strong(_bit_map_lock); 3454 _bit_map_lock->unlock(); 3455 ConcurrentMarkSweepThread::desynchronize(true); 3456 _collector->stopTimer(); 3457 _collector->incrementYields(); 3458 3459 // It is possible for whichever thread initiated the yield request 3460 // not to get a chance to wake up and take the bitmap lock between 3461 // this thread releasing it and reacquiring it. So, while the 3462 // should_yield() flag is on, let's sleep for a bit to give the 3463 // other thread a chance to wake up. The limit imposed on the number 3464 // of iterations is defensive, to avoid any unforseen circumstances 3465 // putting us into an infinite loop. Since it's always been this 3466 // (coordinator_yield()) method that was observed to cause the 3467 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3468 // which is by default non-zero. For the other seven methods that 3469 // also perform the yield operation, as are using a different 3470 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3471 // can enable the sleeping for those methods too, if necessary. 3472 // See 6442774. 3473 // 3474 // We really need to reconsider the synchronization between the GC 3475 // thread and the yield-requesting threads in the future and we 3476 // should really use wait/notify, which is the recommended 3477 // way of doing this type of interaction. Additionally, we should 3478 // consolidate the eight methods that do the yield operation and they 3479 // are almost identical into one for better maintainability and 3480 // readability. See 6445193. 3481 // 3482 // Tony 2006.06.29 3483 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3484 ConcurrentMarkSweepThread::should_yield() && 3485 !CMSCollector::foregroundGCIsActive(); ++i) { 3486 os::sleep(Thread::current(), 1, false); 3487 } 3488 3489 ConcurrentMarkSweepThread::synchronize(true); 3490 _bit_map_lock->lock_without_safepoint_check(); 3491 _collector->startTimer(); 3492 } 3493 3494 bool CMSCollector::do_marking_mt() { 3495 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3496 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3497 conc_workers()->active_workers(), 3498 Threads::number_of_non_daemon_threads()); 3499 num_workers = conc_workers()->update_active_workers(num_workers); 3500 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3501 3502 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3503 3504 CMSConcMarkingTask tsk(this, 3505 cms_space, 3506 conc_workers(), 3507 task_queues()); 3508 3509 // Since the actual number of workers we get may be different 3510 // from the number we requested above, do we need to do anything different 3511 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3512 // class?? XXX 3513 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3514 3515 // Refs discovery is already non-atomic. 3516 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3517 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3518 conc_workers()->start_task(&tsk); 3519 while (tsk.yielded()) { 3520 tsk.coordinator_yield(); 3521 conc_workers()->continue_task(&tsk); 3522 } 3523 // If the task was aborted, _restart_addr will be non-NULL 3524 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3525 while (_restart_addr != NULL) { 3526 // XXX For now we do not make use of ABORTED state and have not 3527 // yet implemented the right abort semantics (even in the original 3528 // single-threaded CMS case). That needs some more investigation 3529 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3530 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3531 // If _restart_addr is non-NULL, a marking stack overflow 3532 // occurred; we need to do a fresh marking iteration from the 3533 // indicated restart address. 3534 if (_foregroundGCIsActive) { 3535 // We may be running into repeated stack overflows, having 3536 // reached the limit of the stack size, while making very 3537 // slow forward progress. It may be best to bail out and 3538 // let the foreground collector do its job. 3539 // Clear _restart_addr, so that foreground GC 3540 // works from scratch. This avoids the headache of 3541 // a "rescan" which would otherwise be needed because 3542 // of the dirty mod union table & card table. 3543 _restart_addr = NULL; 3544 return false; 3545 } 3546 // Adjust the task to restart from _restart_addr 3547 tsk.reset(_restart_addr); 3548 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3549 _restart_addr); 3550 _restart_addr = NULL; 3551 // Get the workers going again 3552 conc_workers()->start_task(&tsk); 3553 while (tsk.yielded()) { 3554 tsk.coordinator_yield(); 3555 conc_workers()->continue_task(&tsk); 3556 } 3557 } 3558 assert(tsk.completed(), "Inconsistency"); 3559 assert(tsk.result() == true, "Inconsistency"); 3560 return true; 3561 } 3562 3563 bool CMSCollector::do_marking_st() { 3564 ResourceMark rm; 3565 HandleMark hm; 3566 3567 // Temporarily make refs discovery single threaded (non-MT) 3568 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3569 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3570 &_markStack, CMSYield); 3571 // the last argument to iterate indicates whether the iteration 3572 // should be incremental with periodic yields. 3573 _markBitMap.iterate(&markFromRootsClosure); 3574 // If _restart_addr is non-NULL, a marking stack overflow 3575 // occurred; we need to do a fresh iteration from the 3576 // indicated restart address. 3577 while (_restart_addr != NULL) { 3578 if (_foregroundGCIsActive) { 3579 // We may be running into repeated stack overflows, having 3580 // reached the limit of the stack size, while making very 3581 // slow forward progress. It may be best to bail out and 3582 // let the foreground collector do its job. 3583 // Clear _restart_addr, so that foreground GC 3584 // works from scratch. This avoids the headache of 3585 // a "rescan" which would otherwise be needed because 3586 // of the dirty mod union table & card table. 3587 _restart_addr = NULL; 3588 return false; // indicating failure to complete marking 3589 } 3590 // Deal with stack overflow: 3591 // we restart marking from _restart_addr 3592 HeapWord* ra = _restart_addr; 3593 markFromRootsClosure.reset(ra); 3594 _restart_addr = NULL; 3595 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3596 } 3597 return true; 3598 } 3599 3600 void CMSCollector::preclean() { 3601 check_correct_thread_executing(); 3602 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3603 verify_work_stacks_empty(); 3604 verify_overflow_empty(); 3605 _abort_preclean = false; 3606 if (CMSPrecleaningEnabled) { 3607 if (!CMSEdenChunksRecordAlways) { 3608 _eden_chunk_index = 0; 3609 } 3610 size_t used = get_eden_used(); 3611 size_t capacity = get_eden_capacity(); 3612 // Don't start sampling unless we will get sufficiently 3613 // many samples. 3614 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3615 * CMSScheduleRemarkEdenPenetration)) { 3616 _start_sampling = true; 3617 } else { 3618 _start_sampling = false; 3619 } 3620 GCTraceCPUTime tcpu; 3621 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3622 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3623 } 3624 CMSTokenSync x(true); // is cms thread 3625 if (CMSPrecleaningEnabled) { 3626 sample_eden(); 3627 _collectorState = AbortablePreclean; 3628 } else { 3629 _collectorState = FinalMarking; 3630 } 3631 verify_work_stacks_empty(); 3632 verify_overflow_empty(); 3633 } 3634 3635 // Try and schedule the remark such that young gen 3636 // occupancy is CMSScheduleRemarkEdenPenetration %. 3637 void CMSCollector::abortable_preclean() { 3638 check_correct_thread_executing(); 3639 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3640 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3641 3642 // If Eden's current occupancy is below this threshold, 3643 // immediately schedule the remark; else preclean 3644 // past the next scavenge in an effort to 3645 // schedule the pause as described above. By choosing 3646 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3647 // we will never do an actual abortable preclean cycle. 3648 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3649 GCTraceCPUTime tcpu; 3650 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3651 // We need more smarts in the abortable preclean 3652 // loop below to deal with cases where allocation 3653 // in young gen is very very slow, and our precleaning 3654 // is running a losing race against a horde of 3655 // mutators intent on flooding us with CMS updates 3656 // (dirty cards). 3657 // One, admittedly dumb, strategy is to give up 3658 // after a certain number of abortable precleaning loops 3659 // or after a certain maximum time. We want to make 3660 // this smarter in the next iteration. 3661 // XXX FIX ME!!! YSR 3662 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3663 while (!(should_abort_preclean() || 3664 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3665 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3666 cumworkdone += workdone; 3667 loops++; 3668 // Voluntarily terminate abortable preclean phase if we have 3669 // been at it for too long. 3670 if ((CMSMaxAbortablePrecleanLoops != 0) && 3671 loops >= CMSMaxAbortablePrecleanLoops) { 3672 log_debug(gc)(" CMS: abort preclean due to loops "); 3673 break; 3674 } 3675 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3676 log_debug(gc)(" CMS: abort preclean due to time "); 3677 break; 3678 } 3679 // If we are doing little work each iteration, we should 3680 // take a short break. 3681 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3682 // Sleep for some time, waiting for work to accumulate 3683 stopTimer(); 3684 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3685 startTimer(); 3686 waited++; 3687 } 3688 } 3689 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3690 loops, waited, cumworkdone); 3691 } 3692 CMSTokenSync x(true); // is cms thread 3693 if (_collectorState != Idling) { 3694 assert(_collectorState == AbortablePreclean, 3695 "Spontaneous state transition?"); 3696 _collectorState = FinalMarking; 3697 } // Else, a foreground collection completed this CMS cycle. 3698 return; 3699 } 3700 3701 // Respond to an Eden sampling opportunity 3702 void CMSCollector::sample_eden() { 3703 // Make sure a young gc cannot sneak in between our 3704 // reading and recording of a sample. 3705 assert(Thread::current()->is_ConcurrentGC_thread(), 3706 "Only the cms thread may collect Eden samples"); 3707 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3708 "Should collect samples while holding CMS token"); 3709 if (!_start_sampling) { 3710 return; 3711 } 3712 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3713 // is populated by the young generation. 3714 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3715 if (_eden_chunk_index < _eden_chunk_capacity) { 3716 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3717 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3718 "Unexpected state of Eden"); 3719 // We'd like to check that what we just sampled is an oop-start address; 3720 // however, we cannot do that here since the object may not yet have been 3721 // initialized. So we'll instead do the check when we _use_ this sample 3722 // later. 3723 if (_eden_chunk_index == 0 || 3724 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3725 _eden_chunk_array[_eden_chunk_index-1]) 3726 >= CMSSamplingGrain)) { 3727 _eden_chunk_index++; // commit sample 3728 } 3729 } 3730 } 3731 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3732 size_t used = get_eden_used(); 3733 size_t capacity = get_eden_capacity(); 3734 assert(used <= capacity, "Unexpected state of Eden"); 3735 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3736 _abort_preclean = true; 3737 } 3738 } 3739 } 3740 3741 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3742 assert(_collectorState == Precleaning || 3743 _collectorState == AbortablePreclean, "incorrect state"); 3744 ResourceMark rm; 3745 HandleMark hm; 3746 3747 // Precleaning is currently not MT but the reference processor 3748 // may be set for MT. Disable it temporarily here. 3749 ReferenceProcessor* rp = ref_processor(); 3750 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3751 3752 // Do one pass of scrubbing the discovered reference lists 3753 // to remove any reference objects with strongly-reachable 3754 // referents. 3755 if (clean_refs) { 3756 CMSPrecleanRefsYieldClosure yield_cl(this); 3757 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3758 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3759 &_markStack, true /* preclean */); 3760 CMSDrainMarkingStackClosure complete_trace(this, 3761 _span, &_markBitMap, &_markStack, 3762 &keep_alive, true /* preclean */); 3763 3764 // We don't want this step to interfere with a young 3765 // collection because we don't want to take CPU 3766 // or memory bandwidth away from the young GC threads 3767 // (which may be as many as there are CPUs). 3768 // Note that we don't need to protect ourselves from 3769 // interference with mutators because they can't 3770 // manipulate the discovered reference lists nor affect 3771 // the computed reachability of the referents, the 3772 // only properties manipulated by the precleaning 3773 // of these reference lists. 3774 stopTimer(); 3775 CMSTokenSyncWithLocks x(true /* is cms thread */, 3776 bitMapLock()); 3777 startTimer(); 3778 sample_eden(); 3779 3780 // The following will yield to allow foreground 3781 // collection to proceed promptly. XXX YSR: 3782 // The code in this method may need further 3783 // tweaking for better performance and some restructuring 3784 // for cleaner interfaces. 3785 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3786 rp->preclean_discovered_references( 3787 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3788 gc_timer); 3789 } 3790 3791 if (clean_survivor) { // preclean the active survivor space(s) 3792 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3793 &_markBitMap, &_modUnionTable, 3794 &_markStack, true /* precleaning phase */); 3795 stopTimer(); 3796 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3797 bitMapLock()); 3798 startTimer(); 3799 unsigned int before_count = 3800 CMSHeap::heap()->total_collections(); 3801 SurvivorSpacePrecleanClosure 3802 sss_cl(this, _span, &_markBitMap, &_markStack, 3803 &pam_cl, before_count, CMSYield); 3804 _young_gen->from()->object_iterate_careful(&sss_cl); 3805 _young_gen->to()->object_iterate_careful(&sss_cl); 3806 } 3807 MarkRefsIntoAndScanClosure 3808 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3809 &_markStack, this, CMSYield, 3810 true /* precleaning phase */); 3811 // CAUTION: The following closure has persistent state that may need to 3812 // be reset upon a decrease in the sequence of addresses it 3813 // processes. 3814 ScanMarkedObjectsAgainCarefullyClosure 3815 smoac_cl(this, _span, 3816 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3817 3818 // Preclean dirty cards in ModUnionTable and CardTable using 3819 // appropriate convergence criterion; 3820 // repeat CMSPrecleanIter times unless we find that 3821 // we are losing. 3822 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3823 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3824 "Bad convergence multiplier"); 3825 assert(CMSPrecleanThreshold >= 100, 3826 "Unreasonably low CMSPrecleanThreshold"); 3827 3828 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3829 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3830 numIter < CMSPrecleanIter; 3831 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3832 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3833 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3834 // Either there are very few dirty cards, so re-mark 3835 // pause will be small anyway, or our pre-cleaning isn't 3836 // that much faster than the rate at which cards are being 3837 // dirtied, so we might as well stop and re-mark since 3838 // precleaning won't improve our re-mark time by much. 3839 if (curNumCards <= CMSPrecleanThreshold || 3840 (numIter > 0 && 3841 (curNumCards * CMSPrecleanDenominator > 3842 lastNumCards * CMSPrecleanNumerator))) { 3843 numIter++; 3844 cumNumCards += curNumCards; 3845 break; 3846 } 3847 } 3848 3849 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3850 3851 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3852 cumNumCards += curNumCards; 3853 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3854 curNumCards, cumNumCards, numIter); 3855 return cumNumCards; // as a measure of useful work done 3856 } 3857 3858 // PRECLEANING NOTES: 3859 // Precleaning involves: 3860 // . reading the bits of the modUnionTable and clearing the set bits. 3861 // . For the cards corresponding to the set bits, we scan the 3862 // objects on those cards. This means we need the free_list_lock 3863 // so that we can safely iterate over the CMS space when scanning 3864 // for oops. 3865 // . When we scan the objects, we'll be both reading and setting 3866 // marks in the marking bit map, so we'll need the marking bit map. 3867 // . For protecting _collector_state transitions, we take the CGC_lock. 3868 // Note that any races in the reading of of card table entries by the 3869 // CMS thread on the one hand and the clearing of those entries by the 3870 // VM thread or the setting of those entries by the mutator threads on the 3871 // other are quite benign. However, for efficiency it makes sense to keep 3872 // the VM thread from racing with the CMS thread while the latter is 3873 // dirty card info to the modUnionTable. We therefore also use the 3874 // CGC_lock to protect the reading of the card table and the mod union 3875 // table by the CM thread. 3876 // . We run concurrently with mutator updates, so scanning 3877 // needs to be done carefully -- we should not try to scan 3878 // potentially uninitialized objects. 3879 // 3880 // Locking strategy: While holding the CGC_lock, we scan over and 3881 // reset a maximal dirty range of the mod union / card tables, then lock 3882 // the free_list_lock and bitmap lock to do a full marking, then 3883 // release these locks; and repeat the cycle. This allows for a 3884 // certain amount of fairness in the sharing of these locks between 3885 // the CMS collector on the one hand, and the VM thread and the 3886 // mutators on the other. 3887 3888 // NOTE: preclean_mod_union_table() and preclean_card_table() 3889 // further below are largely identical; if you need to modify 3890 // one of these methods, please check the other method too. 3891 3892 size_t CMSCollector::preclean_mod_union_table( 3893 ConcurrentMarkSweepGeneration* old_gen, 3894 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3895 verify_work_stacks_empty(); 3896 verify_overflow_empty(); 3897 3898 // strategy: starting with the first card, accumulate contiguous 3899 // ranges of dirty cards; clear these cards, then scan the region 3900 // covered by these cards. 3901 3902 // Since all of the MUT is committed ahead, we can just use 3903 // that, in case the generations expand while we are precleaning. 3904 // It might also be fine to just use the committed part of the 3905 // generation, but we might potentially miss cards when the 3906 // generation is rapidly expanding while we are in the midst 3907 // of precleaning. 3908 HeapWord* startAddr = old_gen->reserved().start(); 3909 HeapWord* endAddr = old_gen->reserved().end(); 3910 3911 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3912 3913 size_t numDirtyCards, cumNumDirtyCards; 3914 HeapWord *nextAddr, *lastAddr; 3915 for (cumNumDirtyCards = numDirtyCards = 0, 3916 nextAddr = lastAddr = startAddr; 3917 nextAddr < endAddr; 3918 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3919 3920 ResourceMark rm; 3921 HandleMark hm; 3922 3923 MemRegion dirtyRegion; 3924 { 3925 stopTimer(); 3926 // Potential yield point 3927 CMSTokenSync ts(true); 3928 startTimer(); 3929 sample_eden(); 3930 // Get dirty region starting at nextOffset (inclusive), 3931 // simultaneously clearing it. 3932 dirtyRegion = 3933 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3934 assert(dirtyRegion.start() >= nextAddr, 3935 "returned region inconsistent?"); 3936 } 3937 // Remember where the next search should begin. 3938 // The returned region (if non-empty) is a right open interval, 3939 // so lastOffset is obtained from the right end of that 3940 // interval. 3941 lastAddr = dirtyRegion.end(); 3942 // Should do something more transparent and less hacky XXX 3943 numDirtyCards = 3944 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3945 3946 // We'll scan the cards in the dirty region (with periodic 3947 // yields for foreground GC as needed). 3948 if (!dirtyRegion.is_empty()) { 3949 assert(numDirtyCards > 0, "consistency check"); 3950 HeapWord* stop_point = NULL; 3951 stopTimer(); 3952 // Potential yield point 3953 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3954 bitMapLock()); 3955 startTimer(); 3956 { 3957 verify_work_stacks_empty(); 3958 verify_overflow_empty(); 3959 sample_eden(); 3960 stop_point = 3961 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3962 } 3963 if (stop_point != NULL) { 3964 // The careful iteration stopped early either because it found an 3965 // uninitialized object, or because we were in the midst of an 3966 // "abortable preclean", which should now be aborted. Redirty 3967 // the bits corresponding to the partially-scanned or unscanned 3968 // cards. We'll either restart at the next block boundary or 3969 // abort the preclean. 3970 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3971 "Should only be AbortablePreclean."); 3972 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3973 if (should_abort_preclean()) { 3974 break; // out of preclean loop 3975 } else { 3976 // Compute the next address at which preclean should pick up; 3977 // might need bitMapLock in order to read P-bits. 3978 lastAddr = next_card_start_after_block(stop_point); 3979 } 3980 } 3981 } else { 3982 assert(lastAddr == endAddr, "consistency check"); 3983 assert(numDirtyCards == 0, "consistency check"); 3984 break; 3985 } 3986 } 3987 verify_work_stacks_empty(); 3988 verify_overflow_empty(); 3989 return cumNumDirtyCards; 3990 } 3991 3992 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3993 // below are largely identical; if you need to modify 3994 // one of these methods, please check the other method too. 3995 3996 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3997 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3998 // strategy: it's similar to precleamModUnionTable above, in that 3999 // we accumulate contiguous ranges of dirty cards, mark these cards 4000 // precleaned, then scan the region covered by these cards. 4001 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 4002 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 4003 4004 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4005 4006 size_t numDirtyCards, cumNumDirtyCards; 4007 HeapWord *lastAddr, *nextAddr; 4008 4009 for (cumNumDirtyCards = numDirtyCards = 0, 4010 nextAddr = lastAddr = startAddr; 4011 nextAddr < endAddr; 4012 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4013 4014 ResourceMark rm; 4015 HandleMark hm; 4016 4017 MemRegion dirtyRegion; 4018 { 4019 // See comments in "Precleaning notes" above on why we 4020 // do this locking. XXX Could the locking overheads be 4021 // too high when dirty cards are sparse? [I don't think so.] 4022 stopTimer(); 4023 CMSTokenSync x(true); // is cms thread 4024 startTimer(); 4025 sample_eden(); 4026 // Get and clear dirty region from card table 4027 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4028 true, 4029 CardTable::precleaned_card_val()); 4030 4031 assert(dirtyRegion.start() >= nextAddr, 4032 "returned region inconsistent?"); 4033 } 4034 lastAddr = dirtyRegion.end(); 4035 numDirtyCards = 4036 dirtyRegion.word_size()/CardTable::card_size_in_words; 4037 4038 if (!dirtyRegion.is_empty()) { 4039 stopTimer(); 4040 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4041 startTimer(); 4042 sample_eden(); 4043 verify_work_stacks_empty(); 4044 verify_overflow_empty(); 4045 HeapWord* stop_point = 4046 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4047 if (stop_point != NULL) { 4048 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4049 "Should only be AbortablePreclean."); 4050 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4051 if (should_abort_preclean()) { 4052 break; // out of preclean loop 4053 } else { 4054 // Compute the next address at which preclean should pick up. 4055 lastAddr = next_card_start_after_block(stop_point); 4056 } 4057 } 4058 } else { 4059 break; 4060 } 4061 } 4062 verify_work_stacks_empty(); 4063 verify_overflow_empty(); 4064 return cumNumDirtyCards; 4065 } 4066 4067 class PrecleanCLDClosure : public CLDClosure { 4068 MetadataAwareOopsInGenClosure* _cm_closure; 4069 public: 4070 PrecleanCLDClosure(MetadataAwareOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4071 void do_cld(ClassLoaderData* cld) { 4072 if (cld->has_accumulated_modified_oops()) { 4073 cld->clear_accumulated_modified_oops(); 4074 4075 _cm_closure->do_cld(cld); 4076 } 4077 } 4078 }; 4079 4080 // The freelist lock is needed to prevent asserts, is it really needed? 4081 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4082 4083 cl->set_freelistLock(freelistLock); 4084 4085 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4086 4087 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4088 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4089 PrecleanCLDClosure preclean_closure(cl); 4090 ClassLoaderDataGraph::cld_do(&preclean_closure); 4091 4092 verify_work_stacks_empty(); 4093 verify_overflow_empty(); 4094 } 4095 4096 void CMSCollector::checkpointRootsFinal() { 4097 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4098 check_correct_thread_executing(); 4099 // world is stopped at this checkpoint 4100 assert(SafepointSynchronize::is_at_safepoint(), 4101 "world should be stopped"); 4102 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4103 4104 verify_work_stacks_empty(); 4105 verify_overflow_empty(); 4106 4107 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4108 _young_gen->used() / K, _young_gen->capacity() / K); 4109 { 4110 if (CMSScavengeBeforeRemark) { 4111 CMSHeap* heap = CMSHeap::heap(); 4112 // Temporarily set flag to false, GCH->do_collection will 4113 // expect it to be false and set to true 4114 FlagSetting fl(heap->_is_gc_active, false); 4115 4116 heap->do_collection(true, // full (i.e. force, see below) 4117 false, // !clear_all_soft_refs 4118 0, // size 4119 false, // is_tlab 4120 GenCollectedHeap::YoungGen // type 4121 ); 4122 } 4123 FreelistLocker x(this); 4124 MutexLockerEx y(bitMapLock(), 4125 Mutex::_no_safepoint_check_flag); 4126 checkpointRootsFinalWork(); 4127 } 4128 verify_work_stacks_empty(); 4129 verify_overflow_empty(); 4130 } 4131 4132 void CMSCollector::checkpointRootsFinalWork() { 4133 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4134 4135 assert(haveFreelistLocks(), "must have free list locks"); 4136 assert_lock_strong(bitMapLock()); 4137 4138 ResourceMark rm; 4139 HandleMark hm; 4140 4141 CMSHeap* heap = CMSHeap::heap(); 4142 4143 if (should_unload_classes()) { 4144 CodeCache::gc_prologue(); 4145 } 4146 assert(haveFreelistLocks(), "must have free list locks"); 4147 assert_lock_strong(bitMapLock()); 4148 4149 // We might assume that we need not fill TLAB's when 4150 // CMSScavengeBeforeRemark is set, because we may have just done 4151 // a scavenge which would have filled all TLAB's -- and besides 4152 // Eden would be empty. This however may not always be the case -- 4153 // for instance although we asked for a scavenge, it may not have 4154 // happened because of a JNI critical section. We probably need 4155 // a policy for deciding whether we can in that case wait until 4156 // the critical section releases and then do the remark following 4157 // the scavenge, and skip it here. In the absence of that policy, 4158 // or of an indication of whether the scavenge did indeed occur, 4159 // we cannot rely on TLAB's having been filled and must do 4160 // so here just in case a scavenge did not happen. 4161 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4162 // Update the saved marks which may affect the root scans. 4163 heap->save_marks(); 4164 4165 print_eden_and_survivor_chunk_arrays(); 4166 4167 { 4168 #if COMPILER2_OR_JVMCI 4169 DerivedPointerTableDeactivate dpt_deact; 4170 #endif 4171 4172 // Note on the role of the mod union table: 4173 // Since the marker in "markFromRoots" marks concurrently with 4174 // mutators, it is possible for some reachable objects not to have been 4175 // scanned. For instance, an only reference to an object A was 4176 // placed in object B after the marker scanned B. Unless B is rescanned, 4177 // A would be collected. Such updates to references in marked objects 4178 // are detected via the mod union table which is the set of all cards 4179 // dirtied since the first checkpoint in this GC cycle and prior to 4180 // the most recent young generation GC, minus those cleaned up by the 4181 // concurrent precleaning. 4182 if (CMSParallelRemarkEnabled) { 4183 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4184 do_remark_parallel(); 4185 } else { 4186 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4187 do_remark_non_parallel(); 4188 } 4189 } 4190 verify_work_stacks_empty(); 4191 verify_overflow_empty(); 4192 4193 { 4194 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4195 refProcessingWork(); 4196 } 4197 verify_work_stacks_empty(); 4198 verify_overflow_empty(); 4199 4200 if (should_unload_classes()) { 4201 CodeCache::gc_epilogue(); 4202 } 4203 JvmtiExport::gc_epilogue(); 4204 4205 // If we encountered any (marking stack / work queue) overflow 4206 // events during the current CMS cycle, take appropriate 4207 // remedial measures, where possible, so as to try and avoid 4208 // recurrence of that condition. 4209 assert(_markStack.isEmpty(), "No grey objects"); 4210 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4211 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4212 if (ser_ovflw > 0) { 4213 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4214 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4215 _markStack.expand(); 4216 _ser_pmc_remark_ovflw = 0; 4217 _ser_pmc_preclean_ovflw = 0; 4218 _ser_kac_preclean_ovflw = 0; 4219 _ser_kac_ovflw = 0; 4220 } 4221 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4222 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4223 _par_pmc_remark_ovflw, _par_kac_ovflw); 4224 _par_pmc_remark_ovflw = 0; 4225 _par_kac_ovflw = 0; 4226 } 4227 if (_markStack._hit_limit > 0) { 4228 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4229 _markStack._hit_limit); 4230 } 4231 if (_markStack._failed_double > 0) { 4232 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4233 _markStack._failed_double, _markStack.capacity()); 4234 } 4235 _markStack._hit_limit = 0; 4236 _markStack._failed_double = 0; 4237 4238 if ((VerifyAfterGC || VerifyDuringGC) && 4239 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4240 verify_after_remark(); 4241 } 4242 4243 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4244 4245 // Change under the freelistLocks. 4246 _collectorState = Sweeping; 4247 // Call isAllClear() under bitMapLock 4248 assert(_modUnionTable.isAllClear(), 4249 "Should be clear by end of the final marking"); 4250 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4251 "Should be clear by end of the final marking"); 4252 } 4253 4254 void CMSParInitialMarkTask::work(uint worker_id) { 4255 elapsedTimer _timer; 4256 ResourceMark rm; 4257 HandleMark hm; 4258 4259 // ---------- scan from roots -------------- 4260 _timer.start(); 4261 CMSHeap* heap = CMSHeap::heap(); 4262 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4263 4264 // ---------- young gen roots -------------- 4265 { 4266 work_on_young_gen_roots(&par_mri_cl); 4267 _timer.stop(); 4268 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4269 } 4270 4271 // ---------- remaining roots -------------- 4272 _timer.reset(); 4273 _timer.start(); 4274 4275 CLDToOopClosure cld_closure(&par_mri_cl, true); 4276 4277 heap->cms_process_roots(_strong_roots_scope, 4278 false, // yg was scanned above 4279 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4280 _collector->should_unload_classes(), 4281 &par_mri_cl, 4282 &cld_closure, 4283 &_par_state_string); 4284 4285 assert(_collector->should_unload_classes() 4286 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4287 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4288 _timer.stop(); 4289 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4290 } 4291 4292 // Parallel remark task 4293 class CMSParRemarkTask: public CMSParMarkTask { 4294 CompactibleFreeListSpace* _cms_space; 4295 4296 // The per-thread work queues, available here for stealing. 4297 OopTaskQueueSet* _task_queues; 4298 ParallelTaskTerminator _term; 4299 StrongRootsScope* _strong_roots_scope; 4300 4301 public: 4302 // A value of 0 passed to n_workers will cause the number of 4303 // workers to be taken from the active workers in the work gang. 4304 CMSParRemarkTask(CMSCollector* collector, 4305 CompactibleFreeListSpace* cms_space, 4306 uint n_workers, WorkGang* workers, 4307 OopTaskQueueSet* task_queues, 4308 StrongRootsScope* strong_roots_scope): 4309 CMSParMarkTask("Rescan roots and grey objects in parallel", 4310 collector, n_workers), 4311 _cms_space(cms_space), 4312 _task_queues(task_queues), 4313 _term(n_workers, task_queues), 4314 _strong_roots_scope(strong_roots_scope) { } 4315 4316 OopTaskQueueSet* task_queues() { return _task_queues; } 4317 4318 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4319 4320 ParallelTaskTerminator* terminator() { return &_term; } 4321 uint n_workers() { return _n_workers; } 4322 4323 void work(uint worker_id); 4324 4325 private: 4326 // ... of dirty cards in old space 4327 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4328 ParMarkRefsIntoAndScanClosure* cl); 4329 4330 // ... work stealing for the above 4331 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, int* seed); 4332 }; 4333 4334 class RemarkCLDClosure : public CLDClosure { 4335 CLDToOopClosure _cm_closure; 4336 public: 4337 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure) {} 4338 void do_cld(ClassLoaderData* cld) { 4339 // Check if we have modified any oops in the CLD during the concurrent marking. 4340 if (cld->has_accumulated_modified_oops()) { 4341 cld->clear_accumulated_modified_oops(); 4342 4343 // We could have transfered the current modified marks to the accumulated marks, 4344 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4345 } else if (cld->has_modified_oops()) { 4346 // Don't clear anything, this info is needed by the next young collection. 4347 } else { 4348 // No modified oops in the ClassLoaderData. 4349 return; 4350 } 4351 4352 // The klass has modified fields, need to scan the klass. 4353 _cm_closure.do_cld(cld); 4354 } 4355 }; 4356 4357 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4358 ParNewGeneration* young_gen = _collector->_young_gen; 4359 ContiguousSpace* eden_space = young_gen->eden(); 4360 ContiguousSpace* from_space = young_gen->from(); 4361 ContiguousSpace* to_space = young_gen->to(); 4362 4363 HeapWord** eca = _collector->_eden_chunk_array; 4364 size_t ect = _collector->_eden_chunk_index; 4365 HeapWord** sca = _collector->_survivor_chunk_array; 4366 size_t sct = _collector->_survivor_chunk_index; 4367 4368 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4369 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4370 4371 do_young_space_rescan(cl, to_space, NULL, 0); 4372 do_young_space_rescan(cl, from_space, sca, sct); 4373 do_young_space_rescan(cl, eden_space, eca, ect); 4374 } 4375 4376 // work_queue(i) is passed to the closure 4377 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4378 // also is passed to do_dirty_card_rescan_tasks() and to 4379 // do_work_steal() to select the i-th task_queue. 4380 4381 void CMSParRemarkTask::work(uint worker_id) { 4382 elapsedTimer _timer; 4383 ResourceMark rm; 4384 HandleMark hm; 4385 4386 // ---------- rescan from roots -------------- 4387 _timer.start(); 4388 CMSHeap* heap = CMSHeap::heap(); 4389 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4390 _collector->_span, _collector->ref_processor(), 4391 &(_collector->_markBitMap), 4392 work_queue(worker_id)); 4393 4394 // Rescan young gen roots first since these are likely 4395 // coarsely partitioned and may, on that account, constitute 4396 // the critical path; thus, it's best to start off that 4397 // work first. 4398 // ---------- young gen roots -------------- 4399 { 4400 work_on_young_gen_roots(&par_mrias_cl); 4401 _timer.stop(); 4402 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4403 } 4404 4405 // ---------- remaining roots -------------- 4406 _timer.reset(); 4407 _timer.start(); 4408 heap->cms_process_roots(_strong_roots_scope, 4409 false, // yg was scanned above 4410 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4411 _collector->should_unload_classes(), 4412 &par_mrias_cl, 4413 NULL, // The dirty klasses will be handled below 4414 &_par_state_string); 4415 4416 assert(_collector->should_unload_classes() 4417 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4418 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4419 _timer.stop(); 4420 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4421 4422 // ---------- unhandled CLD scanning ---------- 4423 if (worker_id == 0) { // Single threaded at the moment. 4424 _timer.reset(); 4425 _timer.start(); 4426 4427 // Scan all new class loader data objects and new dependencies that were 4428 // introduced during concurrent marking. 4429 ResourceMark rm; 4430 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4431 for (int i = 0; i < array->length(); i++) { 4432 par_mrias_cl.do_cld_nv(array->at(i)); 4433 } 4434 4435 // We don't need to keep track of new CLDs anymore. 4436 ClassLoaderDataGraph::remember_new_clds(false); 4437 4438 _timer.stop(); 4439 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4440 } 4441 4442 // We might have added oops to ClassLoaderData::_handles during the 4443 // concurrent marking phase. These oops do not always point to newly allocated objects 4444 // that are guaranteed to be kept alive. Hence, 4445 // we do have to revisit the _handles block during the remark phase. 4446 4447 // ---------- dirty CLD scanning ---------- 4448 if (worker_id == 0) { // Single threaded at the moment. 4449 _timer.reset(); 4450 _timer.start(); 4451 4452 // Scan all classes that was dirtied during the concurrent marking phase. 4453 RemarkCLDClosure remark_closure(&par_mrias_cl); 4454 ClassLoaderDataGraph::cld_do(&remark_closure); 4455 4456 _timer.stop(); 4457 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4458 } 4459 4460 4461 // ---------- rescan dirty cards ------------ 4462 _timer.reset(); 4463 _timer.start(); 4464 4465 // Do the rescan tasks for each of the two spaces 4466 // (cms_space) in turn. 4467 // "worker_id" is passed to select the task_queue for "worker_id" 4468 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4469 _timer.stop(); 4470 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4471 4472 // ---------- steal work from other threads ... 4473 // ---------- ... and drain overflow list. 4474 _timer.reset(); 4475 _timer.start(); 4476 do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id)); 4477 _timer.stop(); 4478 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4479 } 4480 4481 void 4482 CMSParMarkTask::do_young_space_rescan( 4483 OopsInGenClosure* cl, ContiguousSpace* space, 4484 HeapWord** chunk_array, size_t chunk_top) { 4485 // Until all tasks completed: 4486 // . claim an unclaimed task 4487 // . compute region boundaries corresponding to task claimed 4488 // using chunk_array 4489 // . par_oop_iterate(cl) over that region 4490 4491 ResourceMark rm; 4492 HandleMark hm; 4493 4494 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4495 4496 uint nth_task = 0; 4497 uint n_tasks = pst->n_tasks(); 4498 4499 if (n_tasks > 0) { 4500 assert(pst->valid(), "Uninitialized use?"); 4501 HeapWord *start, *end; 4502 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4503 // We claimed task # nth_task; compute its boundaries. 4504 if (chunk_top == 0) { // no samples were taken 4505 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4506 start = space->bottom(); 4507 end = space->top(); 4508 } else if (nth_task == 0) { 4509 start = space->bottom(); 4510 end = chunk_array[nth_task]; 4511 } else if (nth_task < (uint)chunk_top) { 4512 assert(nth_task >= 1, "Control point invariant"); 4513 start = chunk_array[nth_task - 1]; 4514 end = chunk_array[nth_task]; 4515 } else { 4516 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4517 start = chunk_array[chunk_top - 1]; 4518 end = space->top(); 4519 } 4520 MemRegion mr(start, end); 4521 // Verify that mr is in space 4522 assert(mr.is_empty() || space->used_region().contains(mr), 4523 "Should be in space"); 4524 // Verify that "start" is an object boundary 4525 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4526 "Should be an oop"); 4527 space->par_oop_iterate(mr, cl); 4528 } 4529 pst->all_tasks_completed(); 4530 } 4531 } 4532 4533 void 4534 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4535 CompactibleFreeListSpace* sp, int i, 4536 ParMarkRefsIntoAndScanClosure* cl) { 4537 // Until all tasks completed: 4538 // . claim an unclaimed task 4539 // . compute region boundaries corresponding to task claimed 4540 // . transfer dirty bits ct->mut for that region 4541 // . apply rescanclosure to dirty mut bits for that region 4542 4543 ResourceMark rm; 4544 HandleMark hm; 4545 4546 OopTaskQueue* work_q = work_queue(i); 4547 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4548 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4549 // CAUTION: This closure has state that persists across calls to 4550 // the work method dirty_range_iterate_clear() in that it has 4551 // embedded in it a (subtype of) UpwardsObjectClosure. The 4552 // use of that state in the embedded UpwardsObjectClosure instance 4553 // assumes that the cards are always iterated (even if in parallel 4554 // by several threads) in monotonically increasing order per each 4555 // thread. This is true of the implementation below which picks 4556 // card ranges (chunks) in monotonically increasing order globally 4557 // and, a-fortiori, in monotonically increasing order per thread 4558 // (the latter order being a subsequence of the former). 4559 // If the work code below is ever reorganized into a more chaotic 4560 // work-partitioning form than the current "sequential tasks" 4561 // paradigm, the use of that persistent state will have to be 4562 // revisited and modified appropriately. See also related 4563 // bug 4756801 work on which should examine this code to make 4564 // sure that the changes there do not run counter to the 4565 // assumptions made here and necessary for correctness and 4566 // efficiency. Note also that this code might yield inefficient 4567 // behavior in the case of very large objects that span one or 4568 // more work chunks. Such objects would potentially be scanned 4569 // several times redundantly. Work on 4756801 should try and 4570 // address that performance anomaly if at all possible. XXX 4571 MemRegion full_span = _collector->_span; 4572 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4573 MarkFromDirtyCardsClosure 4574 greyRescanClosure(_collector, full_span, // entire span of interest 4575 sp, bm, work_q, cl); 4576 4577 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4578 assert(pst->valid(), "Uninitialized use?"); 4579 uint nth_task = 0; 4580 const int alignment = CardTable::card_size * BitsPerWord; 4581 MemRegion span = sp->used_region(); 4582 HeapWord* start_addr = span.start(); 4583 HeapWord* end_addr = align_up(span.end(), alignment); 4584 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4585 assert(is_aligned(start_addr, alignment), "Check alignment"); 4586 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4587 4588 while (!pst->is_task_claimed(/* reference */ nth_task)) { 4589 // Having claimed the nth_task, compute corresponding mem-region, 4590 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4591 // The alignment restriction ensures that we do not need any 4592 // synchronization with other gang-workers while setting or 4593 // clearing bits in thus chunk of the MUT. 4594 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4595 start_addr + (nth_task+1)*chunk_size); 4596 // The last chunk's end might be way beyond end of the 4597 // used region. In that case pull back appropriately. 4598 if (this_span.end() > end_addr) { 4599 this_span.set_end(end_addr); 4600 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4601 } 4602 // Iterate over the dirty cards covering this chunk, marking them 4603 // precleaned, and setting the corresponding bits in the mod union 4604 // table. Since we have been careful to partition at Card and MUT-word 4605 // boundaries no synchronization is needed between parallel threads. 4606 _collector->_ct->dirty_card_iterate(this_span, 4607 &modUnionClosure); 4608 4609 // Having transferred these marks into the modUnionTable, 4610 // rescan the marked objects on the dirty cards in the modUnionTable. 4611 // Even if this is at a synchronous collection, the initial marking 4612 // may have been done during an asynchronous collection so there 4613 // may be dirty bits in the mod-union table. 4614 _collector->_modUnionTable.dirty_range_iterate_clear( 4615 this_span, &greyRescanClosure); 4616 _collector->_modUnionTable.verifyNoOneBitsInRange( 4617 this_span.start(), 4618 this_span.end()); 4619 } 4620 pst->all_tasks_completed(); // declare that i am done 4621 } 4622 4623 // . see if we can share work_queues with ParNew? XXX 4624 void 4625 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl, 4626 int* seed) { 4627 OopTaskQueue* work_q = work_queue(i); 4628 NOT_PRODUCT(int num_steals = 0;) 4629 oop obj_to_scan; 4630 CMSBitMap* bm = &(_collector->_markBitMap); 4631 4632 while (true) { 4633 // Completely finish any left over work from (an) earlier round(s) 4634 cl->trim_queue(0); 4635 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4636 (size_t)ParGCDesiredObjsFromOverflowList); 4637 // Now check if there's any work in the overflow list 4638 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4639 // only affects the number of attempts made to get work from the 4640 // overflow list and does not affect the number of workers. Just 4641 // pass ParallelGCThreads so this behavior is unchanged. 4642 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4643 work_q, 4644 ParallelGCThreads)) { 4645 // found something in global overflow list; 4646 // not yet ready to go stealing work from others. 4647 // We'd like to assert(work_q->size() != 0, ...) 4648 // because we just took work from the overflow list, 4649 // but of course we can't since all of that could have 4650 // been already stolen from us. 4651 // "He giveth and He taketh away." 4652 continue; 4653 } 4654 // Verify that we have no work before we resort to stealing 4655 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4656 // Try to steal from other queues that have work 4657 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 4658 NOT_PRODUCT(num_steals++;) 4659 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4660 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4661 // Do scanning work 4662 obj_to_scan->oop_iterate(cl); 4663 // Loop around, finish this work, and try to steal some more 4664 } else if (terminator()->offer_termination()) { 4665 break; // nirvana from the infinite cycle 4666 } 4667 } 4668 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4669 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4670 "Else our work is not yet done"); 4671 } 4672 4673 // Record object boundaries in _eden_chunk_array by sampling the eden 4674 // top in the slow-path eden object allocation code path and record 4675 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4676 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4677 // sampling in sample_eden() that activates during the part of the 4678 // preclean phase. 4679 void CMSCollector::sample_eden_chunk() { 4680 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4681 if (_eden_chunk_lock->try_lock()) { 4682 // Record a sample. This is the critical section. The contents 4683 // of the _eden_chunk_array have to be non-decreasing in the 4684 // address order. 4685 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4686 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4687 "Unexpected state of Eden"); 4688 if (_eden_chunk_index == 0 || 4689 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4690 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4691 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4692 _eden_chunk_index++; // commit sample 4693 } 4694 _eden_chunk_lock->unlock(); 4695 } 4696 } 4697 } 4698 4699 // Return a thread-local PLAB recording array, as appropriate. 4700 void* CMSCollector::get_data_recorder(int thr_num) { 4701 if (_survivor_plab_array != NULL && 4702 (CMSPLABRecordAlways || 4703 (_collectorState > Marking && _collectorState < FinalMarking))) { 4704 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4705 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4706 ca->reset(); // clear it so that fresh data is recorded 4707 return (void*) ca; 4708 } else { 4709 return NULL; 4710 } 4711 } 4712 4713 // Reset all the thread-local PLAB recording arrays 4714 void CMSCollector::reset_survivor_plab_arrays() { 4715 for (uint i = 0; i < ParallelGCThreads; i++) { 4716 _survivor_plab_array[i].reset(); 4717 } 4718 } 4719 4720 // Merge the per-thread plab arrays into the global survivor chunk 4721 // array which will provide the partitioning of the survivor space 4722 // for CMS initial scan and rescan. 4723 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4724 int no_of_gc_threads) { 4725 assert(_survivor_plab_array != NULL, "Error"); 4726 assert(_survivor_chunk_array != NULL, "Error"); 4727 assert(_collectorState == FinalMarking || 4728 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4729 for (int j = 0; j < no_of_gc_threads; j++) { 4730 _cursor[j] = 0; 4731 } 4732 HeapWord* top = surv->top(); 4733 size_t i; 4734 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4735 HeapWord* min_val = top; // Higher than any PLAB address 4736 uint min_tid = 0; // position of min_val this round 4737 for (int j = 0; j < no_of_gc_threads; j++) { 4738 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4739 if (_cursor[j] == cur_sca->end()) { 4740 continue; 4741 } 4742 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4743 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4744 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4745 if (cur_val < min_val) { 4746 min_tid = j; 4747 min_val = cur_val; 4748 } else { 4749 assert(cur_val < top, "All recorded addresses should be less"); 4750 } 4751 } 4752 // At this point min_val and min_tid are respectively 4753 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4754 // and the thread (j) that witnesses that address. 4755 // We record this address in the _survivor_chunk_array[i] 4756 // and increment _cursor[min_tid] prior to the next round i. 4757 if (min_val == top) { 4758 break; 4759 } 4760 _survivor_chunk_array[i] = min_val; 4761 _cursor[min_tid]++; 4762 } 4763 // We are all done; record the size of the _survivor_chunk_array 4764 _survivor_chunk_index = i; // exclusive: [0, i) 4765 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4766 // Verify that we used up all the recorded entries 4767 #ifdef ASSERT 4768 size_t total = 0; 4769 for (int j = 0; j < no_of_gc_threads; j++) { 4770 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4771 total += _cursor[j]; 4772 } 4773 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4774 // Check that the merged array is in sorted order 4775 if (total > 0) { 4776 for (size_t i = 0; i < total - 1; i++) { 4777 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4778 i, p2i(_survivor_chunk_array[i])); 4779 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4780 "Not sorted"); 4781 } 4782 } 4783 #endif // ASSERT 4784 } 4785 4786 // Set up the space's par_seq_tasks structure for work claiming 4787 // for parallel initial scan and rescan of young gen. 4788 // See ParRescanTask where this is currently used. 4789 void 4790 CMSCollector:: 4791 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4792 assert(n_threads > 0, "Unexpected n_threads argument"); 4793 4794 // Eden space 4795 if (!_young_gen->eden()->is_empty()) { 4796 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4797 assert(!pst->valid(), "Clobbering existing data?"); 4798 // Each valid entry in [0, _eden_chunk_index) represents a task. 4799 size_t n_tasks = _eden_chunk_index + 1; 4800 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4801 // Sets the condition for completion of the subtask (how many threads 4802 // need to finish in order to be done). 4803 pst->set_n_threads(n_threads); 4804 pst->set_n_tasks((int)n_tasks); 4805 } 4806 4807 // Merge the survivor plab arrays into _survivor_chunk_array 4808 if (_survivor_plab_array != NULL) { 4809 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4810 } else { 4811 assert(_survivor_chunk_index == 0, "Error"); 4812 } 4813 4814 // To space 4815 { 4816 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4817 assert(!pst->valid(), "Clobbering existing data?"); 4818 // Sets the condition for completion of the subtask (how many threads 4819 // need to finish in order to be done). 4820 pst->set_n_threads(n_threads); 4821 pst->set_n_tasks(1); 4822 assert(pst->valid(), "Error"); 4823 } 4824 4825 // From space 4826 { 4827 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4828 assert(!pst->valid(), "Clobbering existing data?"); 4829 size_t n_tasks = _survivor_chunk_index + 1; 4830 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4831 // Sets the condition for completion of the subtask (how many threads 4832 // need to finish in order to be done). 4833 pst->set_n_threads(n_threads); 4834 pst->set_n_tasks((int)n_tasks); 4835 assert(pst->valid(), "Error"); 4836 } 4837 } 4838 4839 // Parallel version of remark 4840 void CMSCollector::do_remark_parallel() { 4841 CMSHeap* heap = CMSHeap::heap(); 4842 WorkGang* workers = heap->workers(); 4843 assert(workers != NULL, "Need parallel worker threads."); 4844 // Choose to use the number of GC workers most recently set 4845 // into "active_workers". 4846 uint n_workers = workers->active_workers(); 4847 4848 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4849 4850 StrongRootsScope srs(n_workers); 4851 4852 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4853 4854 // We won't be iterating over the cards in the card table updating 4855 // the younger_gen cards, so we shouldn't call the following else 4856 // the verification code as well as subsequent younger_refs_iterate 4857 // code would get confused. XXX 4858 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4859 4860 // The young gen rescan work will not be done as part of 4861 // process_roots (which currently doesn't know how to 4862 // parallelize such a scan), but rather will be broken up into 4863 // a set of parallel tasks (via the sampling that the [abortable] 4864 // preclean phase did of eden, plus the [two] tasks of 4865 // scanning the [two] survivor spaces. Further fine-grain 4866 // parallelization of the scanning of the survivor spaces 4867 // themselves, and of precleaning of the young gen itself 4868 // is deferred to the future. 4869 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4870 4871 // The dirty card rescan work is broken up into a "sequence" 4872 // of parallel tasks (per constituent space) that are dynamically 4873 // claimed by the parallel threads. 4874 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4875 4876 // It turns out that even when we're using 1 thread, doing the work in a 4877 // separate thread causes wide variance in run times. We can't help this 4878 // in the multi-threaded case, but we special-case n=1 here to get 4879 // repeatable measurements of the 1-thread overhead of the parallel code. 4880 if (n_workers > 1) { 4881 // Make refs discovery MT-safe, if it isn't already: it may not 4882 // necessarily be so, since it's possible that we are doing 4883 // ST marking. 4884 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4885 workers->run_task(&tsk); 4886 } else { 4887 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4888 tsk.work(0); 4889 } 4890 4891 // restore, single-threaded for now, any preserved marks 4892 // as a result of work_q overflow 4893 restore_preserved_marks_if_any(); 4894 } 4895 4896 // Non-parallel version of remark 4897 void CMSCollector::do_remark_non_parallel() { 4898 ResourceMark rm; 4899 HandleMark hm; 4900 CMSHeap* heap = CMSHeap::heap(); 4901 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4902 4903 MarkRefsIntoAndScanClosure 4904 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4905 &_markStack, this, 4906 false /* should_yield */, false /* not precleaning */); 4907 MarkFromDirtyCardsClosure 4908 markFromDirtyCardsClosure(this, _span, 4909 NULL, // space is set further below 4910 &_markBitMap, &_markStack, &mrias_cl); 4911 { 4912 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4913 // Iterate over the dirty cards, setting the corresponding bits in the 4914 // mod union table. 4915 { 4916 ModUnionClosure modUnionClosure(&_modUnionTable); 4917 _ct->dirty_card_iterate(_cmsGen->used_region(), 4918 &modUnionClosure); 4919 } 4920 // Having transferred these marks into the modUnionTable, we just need 4921 // to rescan the marked objects on the dirty cards in the modUnionTable. 4922 // The initial marking may have been done during an asynchronous 4923 // collection so there may be dirty bits in the mod-union table. 4924 const int alignment = CardTable::card_size * BitsPerWord; 4925 { 4926 // ... First handle dirty cards in CMS gen 4927 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4928 MemRegion ur = _cmsGen->used_region(); 4929 HeapWord* lb = ur.start(); 4930 HeapWord* ub = align_up(ur.end(), alignment); 4931 MemRegion cms_span(lb, ub); 4932 _modUnionTable.dirty_range_iterate_clear(cms_span, 4933 &markFromDirtyCardsClosure); 4934 verify_work_stacks_empty(); 4935 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4936 } 4937 } 4938 if (VerifyDuringGC && 4939 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4940 HandleMark hm; // Discard invalid handles created during verification 4941 Universe::verify(); 4942 } 4943 { 4944 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4945 4946 verify_work_stacks_empty(); 4947 4948 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4949 StrongRootsScope srs(1); 4950 4951 heap->cms_process_roots(&srs, 4952 true, // young gen as roots 4953 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4954 should_unload_classes(), 4955 &mrias_cl, 4956 NULL); // The dirty klasses will be handled below 4957 4958 assert(should_unload_classes() 4959 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4960 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4961 } 4962 4963 { 4964 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4965 4966 verify_work_stacks_empty(); 4967 4968 // Scan all class loader data objects that might have been introduced 4969 // during concurrent marking. 4970 ResourceMark rm; 4971 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4972 for (int i = 0; i < array->length(); i++) { 4973 mrias_cl.do_cld_nv(array->at(i)); 4974 } 4975 4976 // We don't need to keep track of new CLDs anymore. 4977 ClassLoaderDataGraph::remember_new_clds(false); 4978 4979 verify_work_stacks_empty(); 4980 } 4981 4982 // We might have added oops to ClassLoaderData::_handles during the 4983 // concurrent marking phase. These oops do not point to newly allocated objects 4984 // that are guaranteed to be kept alive. Hence, 4985 // we do have to revisit the _handles block during the remark phase. 4986 { 4987 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4988 4989 verify_work_stacks_empty(); 4990 4991 RemarkCLDClosure remark_closure(&mrias_cl); 4992 ClassLoaderDataGraph::cld_do(&remark_closure); 4993 4994 verify_work_stacks_empty(); 4995 } 4996 4997 verify_work_stacks_empty(); 4998 // Restore evacuated mark words, if any, used for overflow list links 4999 restore_preserved_marks_if_any(); 5000 5001 verify_overflow_empty(); 5002 } 5003 5004 //////////////////////////////////////////////////////// 5005 // Parallel Reference Processing Task Proxy Class 5006 //////////////////////////////////////////////////////// 5007 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5008 OopTaskQueueSet* _queues; 5009 ParallelTaskTerminator _terminator; 5010 public: 5011 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5012 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5013 ParallelTaskTerminator* terminator() { return &_terminator; } 5014 OopTaskQueueSet* queues() { return _queues; } 5015 }; 5016 5017 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5018 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5019 CMSCollector* _collector; 5020 CMSBitMap* _mark_bit_map; 5021 const MemRegion _span; 5022 ProcessTask& _task; 5023 5024 public: 5025 CMSRefProcTaskProxy(ProcessTask& task, 5026 CMSCollector* collector, 5027 const MemRegion& span, 5028 CMSBitMap* mark_bit_map, 5029 AbstractWorkGang* workers, 5030 OopTaskQueueSet* task_queues): 5031 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5032 task_queues, 5033 workers->active_workers()), 5034 _task(task), 5035 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) 5036 { 5037 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5038 "Inconsistency in _span"); 5039 } 5040 5041 OopTaskQueueSet* task_queues() { return queues(); } 5042 5043 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5044 5045 void do_work_steal(int i, 5046 CMSParDrainMarkingStackClosure* drain, 5047 CMSParKeepAliveClosure* keep_alive, 5048 int* seed); 5049 5050 virtual void work(uint worker_id); 5051 }; 5052 5053 void CMSRefProcTaskProxy::work(uint worker_id) { 5054 ResourceMark rm; 5055 HandleMark hm; 5056 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5057 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5058 _mark_bit_map, 5059 work_queue(worker_id)); 5060 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5061 _mark_bit_map, 5062 work_queue(worker_id)); 5063 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5064 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5065 if (_task.marks_oops_alive()) { 5066 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive, 5067 _collector->hash_seed(worker_id)); 5068 } 5069 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5070 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5071 } 5072 5073 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5074 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5075 _span(span), 5076 _bit_map(bit_map), 5077 _work_queue(work_queue), 5078 _mark_and_push(collector, span, bit_map, work_queue), 5079 _low_water_mark(MIN2((work_queue->max_elems()/4), 5080 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5081 { } 5082 5083 // . see if we can share work_queues with ParNew? XXX 5084 void CMSRefProcTaskProxy::do_work_steal(int i, 5085 CMSParDrainMarkingStackClosure* drain, 5086 CMSParKeepAliveClosure* keep_alive, 5087 int* seed) { 5088 OopTaskQueue* work_q = work_queue(i); 5089 NOT_PRODUCT(int num_steals = 0;) 5090 oop obj_to_scan; 5091 5092 while (true) { 5093 // Completely finish any left over work from (an) earlier round(s) 5094 drain->trim_queue(0); 5095 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5096 (size_t)ParGCDesiredObjsFromOverflowList); 5097 // Now check if there's any work in the overflow list 5098 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5099 // only affects the number of attempts made to get work from the 5100 // overflow list and does not affect the number of workers. Just 5101 // pass ParallelGCThreads so this behavior is unchanged. 5102 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5103 work_q, 5104 ParallelGCThreads)) { 5105 // Found something in global overflow list; 5106 // not yet ready to go stealing work from others. 5107 // We'd like to assert(work_q->size() != 0, ...) 5108 // because we just took work from the overflow list, 5109 // but of course we can't, since all of that might have 5110 // been already stolen from us. 5111 continue; 5112 } 5113 // Verify that we have no work before we resort to stealing 5114 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5115 // Try to steal from other queues that have work 5116 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { 5117 NOT_PRODUCT(num_steals++;) 5118 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5119 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5120 // Do scanning work 5121 obj_to_scan->oop_iterate(keep_alive); 5122 // Loop around, finish this work, and try to steal some more 5123 } else if (terminator()->offer_termination()) { 5124 break; // nirvana from the infinite cycle 5125 } 5126 } 5127 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5128 } 5129 5130 void CMSRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 5131 CMSHeap* heap = CMSHeap::heap(); 5132 WorkGang* workers = heap->workers(); 5133 assert(workers != NULL, "Need parallel worker threads."); 5134 assert(workers->active_workers() == ergo_workers, 5135 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 5136 ergo_workers, workers->active_workers()); 5137 CMSRefProcTaskProxy rp_task(task, &_collector, 5138 _collector.ref_processor_span(), 5139 _collector.markBitMap(), 5140 workers, _collector.task_queues()); 5141 workers->run_task(&rp_task, workers->active_workers()); 5142 } 5143 5144 void CMSCollector::refProcessingWork() { 5145 ResourceMark rm; 5146 HandleMark hm; 5147 5148 ReferenceProcessor* rp = ref_processor(); 5149 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5150 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5151 // Process weak references. 5152 rp->setup_policy(false); 5153 verify_work_stacks_empty(); 5154 5155 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 5156 { 5157 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5158 5159 // Setup keep_alive and complete closures. 5160 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5161 &_markStack, false /* !preclean */); 5162 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5163 _span, &_markBitMap, &_markStack, 5164 &cmsKeepAliveClosure, false /* !preclean */); 5165 5166 ReferenceProcessorStats stats; 5167 if (rp->processing_is_mt()) { 5168 // Set the degree of MT here. If the discovery is done MT, there 5169 // may have been a different number of threads doing the discovery 5170 // and a different number of discovered lists may have Ref objects. 5171 // That is OK as long as the Reference lists are balanced (see 5172 // balance_all_queues() and balance_queues()). 5173 CMSHeap* heap = CMSHeap::heap(); 5174 uint active_workers = ParallelGCThreads; 5175 WorkGang* workers = heap->workers(); 5176 if (workers != NULL) { 5177 active_workers = workers->active_workers(); 5178 // The expectation is that active_workers will have already 5179 // been set to a reasonable value. If it has not been set, 5180 // investigate. 5181 assert(active_workers > 0, "Should have been set during scavenge"); 5182 } 5183 rp->set_active_mt_degree(active_workers); 5184 CMSRefProcTaskExecutor task_executor(*this); 5185 stats = rp->process_discovered_references(&_is_alive_closure, 5186 &cmsKeepAliveClosure, 5187 &cmsDrainMarkingStackClosure, 5188 &task_executor, 5189 &pt); 5190 } else { 5191 stats = rp->process_discovered_references(&_is_alive_closure, 5192 &cmsKeepAliveClosure, 5193 &cmsDrainMarkingStackClosure, 5194 NULL, 5195 &pt); 5196 } 5197 _gc_tracer_cm->report_gc_reference_stats(stats); 5198 pt.print_all_references(); 5199 } 5200 5201 // This is the point where the entire marking should have completed. 5202 verify_work_stacks_empty(); 5203 5204 { 5205 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5206 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5207 } 5208 5209 if (should_unload_classes()) { 5210 { 5211 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5212 5213 // Unload classes and purge the SystemDictionary. 5214 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm); 5215 5216 // Unload nmethods. 5217 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5218 5219 // Prune dead klasses from subklass/sibling/implementor lists. 5220 Klass::clean_weak_klass_links(purged_class); 5221 } 5222 5223 { 5224 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5225 // Clean up unreferenced symbols in symbol table. 5226 SymbolTable::unlink(); 5227 } 5228 5229 { 5230 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5231 // Delete entries for dead interned strings. 5232 StringTable::unlink(&_is_alive_closure); 5233 } 5234 } 5235 5236 // Restore any preserved marks as a result of mark stack or 5237 // work queue overflow 5238 restore_preserved_marks_if_any(); // done single-threaded for now 5239 5240 rp->set_enqueuing_is_done(true); 5241 rp->verify_no_references_recorded(); 5242 } 5243 5244 #ifndef PRODUCT 5245 void CMSCollector::check_correct_thread_executing() { 5246 Thread* t = Thread::current(); 5247 // Only the VM thread or the CMS thread should be here. 5248 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5249 "Unexpected thread type"); 5250 // If this is the vm thread, the foreground process 5251 // should not be waiting. Note that _foregroundGCIsActive is 5252 // true while the foreground collector is waiting. 5253 if (_foregroundGCShouldWait) { 5254 // We cannot be the VM thread 5255 assert(t->is_ConcurrentGC_thread(), 5256 "Should be CMS thread"); 5257 } else { 5258 // We can be the CMS thread only if we are in a stop-world 5259 // phase of CMS collection. 5260 if (t->is_ConcurrentGC_thread()) { 5261 assert(_collectorState == InitialMarking || 5262 _collectorState == FinalMarking, 5263 "Should be a stop-world phase"); 5264 // The CMS thread should be holding the CMS_token. 5265 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5266 "Potential interference with concurrently " 5267 "executing VM thread"); 5268 } 5269 } 5270 } 5271 #endif 5272 5273 void CMSCollector::sweep() { 5274 assert(_collectorState == Sweeping, "just checking"); 5275 check_correct_thread_executing(); 5276 verify_work_stacks_empty(); 5277 verify_overflow_empty(); 5278 increment_sweep_count(); 5279 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5280 5281 _inter_sweep_timer.stop(); 5282 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5283 5284 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5285 _intra_sweep_timer.reset(); 5286 _intra_sweep_timer.start(); 5287 { 5288 GCTraceCPUTime tcpu; 5289 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5290 // First sweep the old gen 5291 { 5292 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5293 bitMapLock()); 5294 sweepWork(_cmsGen); 5295 } 5296 5297 // Update Universe::_heap_*_at_gc figures. 5298 // We need all the free list locks to make the abstract state 5299 // transition from Sweeping to Resetting. See detailed note 5300 // further below. 5301 { 5302 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5303 // Update heap occupancy information which is used as 5304 // input to soft ref clearing policy at the next gc. 5305 Universe::update_heap_info_at_gc(); 5306 _collectorState = Resizing; 5307 } 5308 } 5309 verify_work_stacks_empty(); 5310 verify_overflow_empty(); 5311 5312 if (should_unload_classes()) { 5313 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5314 // requires that the virtual spaces are stable and not deleted. 5315 ClassLoaderDataGraph::set_should_purge(true); 5316 } 5317 5318 _intra_sweep_timer.stop(); 5319 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5320 5321 _inter_sweep_timer.reset(); 5322 _inter_sweep_timer.start(); 5323 5324 // We need to use a monotonically non-decreasing time in ms 5325 // or we will see time-warp warnings and os::javaTimeMillis() 5326 // does not guarantee monotonicity. 5327 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5328 update_time_of_last_gc(now); 5329 5330 // NOTE on abstract state transitions: 5331 // Mutators allocate-live and/or mark the mod-union table dirty 5332 // based on the state of the collection. The former is done in 5333 // the interval [Marking, Sweeping] and the latter in the interval 5334 // [Marking, Sweeping). Thus the transitions into the Marking state 5335 // and out of the Sweeping state must be synchronously visible 5336 // globally to the mutators. 5337 // The transition into the Marking state happens with the world 5338 // stopped so the mutators will globally see it. Sweeping is 5339 // done asynchronously by the background collector so the transition 5340 // from the Sweeping state to the Resizing state must be done 5341 // under the freelistLock (as is the check for whether to 5342 // allocate-live and whether to dirty the mod-union table). 5343 assert(_collectorState == Resizing, "Change of collector state to" 5344 " Resizing must be done under the freelistLocks (plural)"); 5345 5346 // Now that sweeping has been completed, we clear 5347 // the incremental_collection_failed flag, 5348 // thus inviting a younger gen collection to promote into 5349 // this generation. If such a promotion may still fail, 5350 // the flag will be set again when a young collection is 5351 // attempted. 5352 CMSHeap* heap = CMSHeap::heap(); 5353 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5354 heap->update_full_collections_completed(_collection_count_start); 5355 } 5356 5357 // FIX ME!!! Looks like this belongs in CFLSpace, with 5358 // CMSGen merely delegating to it. 5359 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5360 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5361 HeapWord* minAddr = _cmsSpace->bottom(); 5362 HeapWord* largestAddr = 5363 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5364 if (largestAddr == NULL) { 5365 // The dictionary appears to be empty. In this case 5366 // try to coalesce at the end of the heap. 5367 largestAddr = _cmsSpace->end(); 5368 } 5369 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5370 size_t nearLargestOffset = 5371 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5372 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5373 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5374 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5375 } 5376 5377 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5378 return addr >= _cmsSpace->nearLargestChunk(); 5379 } 5380 5381 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5382 return _cmsSpace->find_chunk_at_end(); 5383 } 5384 5385 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5386 bool full) { 5387 // If the young generation has been collected, gather any statistics 5388 // that are of interest at this point. 5389 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5390 if (!full && current_is_young) { 5391 // Gather statistics on the young generation collection. 5392 collector()->stats().record_gc0_end(used()); 5393 } 5394 } 5395 5396 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5397 // We iterate over the space(s) underlying this generation, 5398 // checking the mark bit map to see if the bits corresponding 5399 // to specific blocks are marked or not. Blocks that are 5400 // marked are live and are not swept up. All remaining blocks 5401 // are swept up, with coalescing on-the-fly as we sweep up 5402 // contiguous free and/or garbage blocks: 5403 // We need to ensure that the sweeper synchronizes with allocators 5404 // and stop-the-world collectors. In particular, the following 5405 // locks are used: 5406 // . CMS token: if this is held, a stop the world collection cannot occur 5407 // . freelistLock: if this is held no allocation can occur from this 5408 // generation by another thread 5409 // . bitMapLock: if this is held, no other thread can access or update 5410 // 5411 5412 // Note that we need to hold the freelistLock if we use 5413 // block iterate below; else the iterator might go awry if 5414 // a mutator (or promotion) causes block contents to change 5415 // (for instance if the allocator divvies up a block). 5416 // If we hold the free list lock, for all practical purposes 5417 // young generation GC's can't occur (they'll usually need to 5418 // promote), so we might as well prevent all young generation 5419 // GC's while we do a sweeping step. For the same reason, we might 5420 // as well take the bit map lock for the entire duration 5421 5422 // check that we hold the requisite locks 5423 assert(have_cms_token(), "Should hold cms token"); 5424 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5425 assert_lock_strong(old_gen->freelistLock()); 5426 assert_lock_strong(bitMapLock()); 5427 5428 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5429 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5430 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5431 _inter_sweep_estimate.padded_average(), 5432 _intra_sweep_estimate.padded_average()); 5433 old_gen->setNearLargestChunk(); 5434 5435 { 5436 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5437 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5438 // We need to free-up/coalesce garbage/blocks from a 5439 // co-terminal free run. This is done in the SweepClosure 5440 // destructor; so, do not remove this scope, else the 5441 // end-of-sweep-census below will be off by a little bit. 5442 } 5443 old_gen->cmsSpace()->sweep_completed(); 5444 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5445 if (should_unload_classes()) { // unloaded classes this cycle, 5446 _concurrent_cycles_since_last_unload = 0; // ... reset count 5447 } else { // did not unload classes, 5448 _concurrent_cycles_since_last_unload++; // ... increment count 5449 } 5450 } 5451 5452 // Reset CMS data structures (for now just the marking bit map) 5453 // preparatory for the next cycle. 5454 void CMSCollector::reset_concurrent() { 5455 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5456 5457 // If the state is not "Resetting", the foreground thread 5458 // has done a collection and the resetting. 5459 if (_collectorState != Resetting) { 5460 assert(_collectorState == Idling, "The state should only change" 5461 " because the foreground collector has finished the collection"); 5462 return; 5463 } 5464 5465 { 5466 // Clear the mark bitmap (no grey objects to start with) 5467 // for the next cycle. 5468 GCTraceCPUTime tcpu; 5469 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5470 5471 HeapWord* curAddr = _markBitMap.startWord(); 5472 while (curAddr < _markBitMap.endWord()) { 5473 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5474 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5475 _markBitMap.clear_large_range(chunk); 5476 if (ConcurrentMarkSweepThread::should_yield() && 5477 !foregroundGCIsActive() && 5478 CMSYield) { 5479 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5480 "CMS thread should hold CMS token"); 5481 assert_lock_strong(bitMapLock()); 5482 bitMapLock()->unlock(); 5483 ConcurrentMarkSweepThread::desynchronize(true); 5484 stopTimer(); 5485 incrementYields(); 5486 5487 // See the comment in coordinator_yield() 5488 for (unsigned i = 0; i < CMSYieldSleepCount && 5489 ConcurrentMarkSweepThread::should_yield() && 5490 !CMSCollector::foregroundGCIsActive(); ++i) { 5491 os::sleep(Thread::current(), 1, false); 5492 } 5493 5494 ConcurrentMarkSweepThread::synchronize(true); 5495 bitMapLock()->lock_without_safepoint_check(); 5496 startTimer(); 5497 } 5498 curAddr = chunk.end(); 5499 } 5500 // A successful mostly concurrent collection has been done. 5501 // Because only the full (i.e., concurrent mode failure) collections 5502 // are being measured for gc overhead limits, clean the "near" flag 5503 // and count. 5504 size_policy()->reset_gc_overhead_limit_count(); 5505 _collectorState = Idling; 5506 } 5507 5508 register_gc_end(); 5509 } 5510 5511 // Same as above but for STW paths 5512 void CMSCollector::reset_stw() { 5513 // already have the lock 5514 assert(_collectorState == Resetting, "just checking"); 5515 assert_lock_strong(bitMapLock()); 5516 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5517 _markBitMap.clear_all(); 5518 _collectorState = Idling; 5519 register_gc_end(); 5520 } 5521 5522 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5523 GCTraceCPUTime tcpu; 5524 TraceCollectorStats tcs_cgc(cgc_counters()); 5525 5526 switch (op) { 5527 case CMS_op_checkpointRootsInitial: { 5528 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5529 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5530 checkpointRootsInitial(); 5531 break; 5532 } 5533 case CMS_op_checkpointRootsFinal: { 5534 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5535 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5536 checkpointRootsFinal(); 5537 break; 5538 } 5539 default: 5540 fatal("No such CMS_op"); 5541 } 5542 } 5543 5544 #ifndef PRODUCT 5545 size_t const CMSCollector::skip_header_HeapWords() { 5546 return FreeChunk::header_size(); 5547 } 5548 5549 // Try and collect here conditions that should hold when 5550 // CMS thread is exiting. The idea is that the foreground GC 5551 // thread should not be blocked if it wants to terminate 5552 // the CMS thread and yet continue to run the VM for a while 5553 // after that. 5554 void CMSCollector::verify_ok_to_terminate() const { 5555 assert(Thread::current()->is_ConcurrentGC_thread(), 5556 "should be called by CMS thread"); 5557 assert(!_foregroundGCShouldWait, "should be false"); 5558 // We could check here that all the various low-level locks 5559 // are not held by the CMS thread, but that is overkill; see 5560 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5561 // is checked. 5562 } 5563 #endif 5564 5565 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5566 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5567 "missing Printezis mark?"); 5568 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5569 size_t size = pointer_delta(nextOneAddr + 1, addr); 5570 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5571 "alignment problem"); 5572 assert(size >= 3, "Necessary for Printezis marks to work"); 5573 return size; 5574 } 5575 5576 // A variant of the above (block_size_using_printezis_bits()) except 5577 // that we return 0 if the P-bits are not yet set. 5578 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5579 if (_markBitMap.isMarked(addr + 1)) { 5580 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5581 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5582 size_t size = pointer_delta(nextOneAddr + 1, addr); 5583 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5584 "alignment problem"); 5585 assert(size >= 3, "Necessary for Printezis marks to work"); 5586 return size; 5587 } 5588 return 0; 5589 } 5590 5591 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5592 size_t sz = 0; 5593 oop p = (oop)addr; 5594 if (p->klass_or_null_acquire() != NULL) { 5595 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5596 } else { 5597 sz = block_size_using_printezis_bits(addr); 5598 } 5599 assert(sz > 0, "size must be nonzero"); 5600 HeapWord* next_block = addr + sz; 5601 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5602 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5603 align_down((uintptr_t)next_card, CardTable::card_size), 5604 "must be different cards"); 5605 return next_card; 5606 } 5607 5608 5609 // CMS Bit Map Wrapper ///////////////////////////////////////// 5610 5611 // Construct a CMS bit map infrastructure, but don't create the 5612 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5613 // further below. 5614 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5615 _bm(), 5616 _shifter(shifter), 5617 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5618 Monitor::_safepoint_check_sometimes) : NULL) 5619 { 5620 _bmStartWord = 0; 5621 _bmWordSize = 0; 5622 } 5623 5624 bool CMSBitMap::allocate(MemRegion mr) { 5625 _bmStartWord = mr.start(); 5626 _bmWordSize = mr.word_size(); 5627 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5628 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5629 if (!brs.is_reserved()) { 5630 log_warning(gc)("CMS bit map allocation failure"); 5631 return false; 5632 } 5633 // For now we'll just commit all of the bit map up front. 5634 // Later on we'll try to be more parsimonious with swap. 5635 if (!_virtual_space.initialize(brs, brs.size())) { 5636 log_warning(gc)("CMS bit map backing store failure"); 5637 return false; 5638 } 5639 assert(_virtual_space.committed_size() == brs.size(), 5640 "didn't reserve backing store for all of CMS bit map?"); 5641 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5642 _bmWordSize, "inconsistency in bit map sizing"); 5643 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5644 5645 // bm.clear(); // can we rely on getting zero'd memory? verify below 5646 assert(isAllClear(), 5647 "Expected zero'd memory from ReservedSpace constructor"); 5648 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5649 "consistency check"); 5650 return true; 5651 } 5652 5653 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5654 HeapWord *next_addr, *end_addr, *last_addr; 5655 assert_locked(); 5656 assert(covers(mr), "out-of-range error"); 5657 // XXX assert that start and end are appropriately aligned 5658 for (next_addr = mr.start(), end_addr = mr.end(); 5659 next_addr < end_addr; next_addr = last_addr) { 5660 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5661 last_addr = dirty_region.end(); 5662 if (!dirty_region.is_empty()) { 5663 cl->do_MemRegion(dirty_region); 5664 } else { 5665 assert(last_addr == end_addr, "program logic"); 5666 return; 5667 } 5668 } 5669 } 5670 5671 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5672 _bm.print_on_error(st, prefix); 5673 } 5674 5675 #ifndef PRODUCT 5676 void CMSBitMap::assert_locked() const { 5677 CMSLockVerifier::assert_locked(lock()); 5678 } 5679 5680 bool CMSBitMap::covers(MemRegion mr) const { 5681 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5682 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5683 "size inconsistency"); 5684 return (mr.start() >= _bmStartWord) && 5685 (mr.end() <= endWord()); 5686 } 5687 5688 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5689 return (start >= _bmStartWord && (start + size) <= endWord()); 5690 } 5691 5692 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5693 // verify that there are no 1 bits in the interval [left, right) 5694 FalseBitMapClosure falseBitMapClosure; 5695 iterate(&falseBitMapClosure, left, right); 5696 } 5697 5698 void CMSBitMap::region_invariant(MemRegion mr) 5699 { 5700 assert_locked(); 5701 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5702 assert(!mr.is_empty(), "unexpected empty region"); 5703 assert(covers(mr), "mr should be covered by bit map"); 5704 // convert address range into offset range 5705 size_t start_ofs = heapWordToOffset(mr.start()); 5706 // Make sure that end() is appropriately aligned 5707 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5708 "Misaligned mr.end()"); 5709 size_t end_ofs = heapWordToOffset(mr.end()); 5710 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5711 } 5712 5713 #endif 5714 5715 bool CMSMarkStack::allocate(size_t size) { 5716 // allocate a stack of the requisite depth 5717 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5718 size * sizeof(oop))); 5719 if (!rs.is_reserved()) { 5720 log_warning(gc)("CMSMarkStack allocation failure"); 5721 return false; 5722 } 5723 if (!_virtual_space.initialize(rs, rs.size())) { 5724 log_warning(gc)("CMSMarkStack backing store failure"); 5725 return false; 5726 } 5727 assert(_virtual_space.committed_size() == rs.size(), 5728 "didn't reserve backing store for all of CMS stack?"); 5729 _base = (oop*)(_virtual_space.low()); 5730 _index = 0; 5731 _capacity = size; 5732 NOT_PRODUCT(_max_depth = 0); 5733 return true; 5734 } 5735 5736 // XXX FIX ME !!! In the MT case we come in here holding a 5737 // leaf lock. For printing we need to take a further lock 5738 // which has lower rank. We need to recalibrate the two 5739 // lock-ranks involved in order to be able to print the 5740 // messages below. (Or defer the printing to the caller. 5741 // For now we take the expedient path of just disabling the 5742 // messages for the problematic case.) 5743 void CMSMarkStack::expand() { 5744 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5745 if (_capacity == MarkStackSizeMax) { 5746 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5747 // We print a warning message only once per CMS cycle. 5748 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5749 } 5750 return; 5751 } 5752 // Double capacity if possible 5753 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5754 // Do not give up existing stack until we have managed to 5755 // get the double capacity that we desired. 5756 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5757 new_capacity * sizeof(oop))); 5758 if (rs.is_reserved()) { 5759 // Release the backing store associated with old stack 5760 _virtual_space.release(); 5761 // Reinitialize virtual space for new stack 5762 if (!_virtual_space.initialize(rs, rs.size())) { 5763 fatal("Not enough swap for expanded marking stack"); 5764 } 5765 _base = (oop*)(_virtual_space.low()); 5766 _index = 0; 5767 _capacity = new_capacity; 5768 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5769 // Failed to double capacity, continue; 5770 // we print a detail message only once per CMS cycle. 5771 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5772 _capacity / K, new_capacity / K); 5773 } 5774 } 5775 5776 5777 // Closures 5778 // XXX: there seems to be a lot of code duplication here; 5779 // should refactor and consolidate common code. 5780 5781 // This closure is used to mark refs into the CMS generation in 5782 // the CMS bit map. Called at the first checkpoint. This closure 5783 // assumes that we do not need to re-mark dirty cards; if the CMS 5784 // generation on which this is used is not an oldest 5785 // generation then this will lose younger_gen cards! 5786 5787 MarkRefsIntoClosure::MarkRefsIntoClosure( 5788 MemRegion span, CMSBitMap* bitMap): 5789 _span(span), 5790 _bitMap(bitMap) 5791 { 5792 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5793 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5794 } 5795 5796 void MarkRefsIntoClosure::do_oop(oop obj) { 5797 // if p points into _span, then mark corresponding bit in _markBitMap 5798 assert(oopDesc::is_oop(obj), "expected an oop"); 5799 HeapWord* addr = (HeapWord*)obj; 5800 if (_span.contains(addr)) { 5801 // this should be made more efficient 5802 _bitMap->mark(addr); 5803 } 5804 } 5805 5806 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5807 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } 5808 5809 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5810 MemRegion span, CMSBitMap* bitMap): 5811 _span(span), 5812 _bitMap(bitMap) 5813 { 5814 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5815 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5816 } 5817 5818 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5819 // if p points into _span, then mark corresponding bit in _markBitMap 5820 assert(oopDesc::is_oop(obj), "expected an oop"); 5821 HeapWord* addr = (HeapWord*)obj; 5822 if (_span.contains(addr)) { 5823 // this should be made more efficient 5824 _bitMap->par_mark(addr); 5825 } 5826 } 5827 5828 void ParMarkRefsIntoClosure::do_oop(oop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5829 void ParMarkRefsIntoClosure::do_oop(narrowOop* p) { ParMarkRefsIntoClosure::do_oop_work(p); } 5830 5831 // A variant of the above, used for CMS marking verification. 5832 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5833 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5834 _span(span), 5835 _verification_bm(verification_bm), 5836 _cms_bm(cms_bm) 5837 { 5838 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5839 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5840 } 5841 5842 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5843 // if p points into _span, then mark corresponding bit in _markBitMap 5844 assert(oopDesc::is_oop(obj), "expected an oop"); 5845 HeapWord* addr = (HeapWord*)obj; 5846 if (_span.contains(addr)) { 5847 _verification_bm->mark(addr); 5848 if (!_cms_bm->isMarked(addr)) { 5849 Log(gc, verify) log; 5850 ResourceMark rm; 5851 LogStream ls(log.error()); 5852 oop(addr)->print_on(&ls); 5853 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5854 fatal("... aborting"); 5855 } 5856 } 5857 } 5858 5859 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5860 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } 5861 5862 ////////////////////////////////////////////////// 5863 // MarkRefsIntoAndScanClosure 5864 ////////////////////////////////////////////////// 5865 5866 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5867 ReferenceDiscoverer* rd, 5868 CMSBitMap* bit_map, 5869 CMSBitMap* mod_union_table, 5870 CMSMarkStack* mark_stack, 5871 CMSCollector* collector, 5872 bool should_yield, 5873 bool concurrent_precleaning): 5874 _collector(collector), 5875 _span(span), 5876 _bit_map(bit_map), 5877 _mark_stack(mark_stack), 5878 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5879 mark_stack, concurrent_precleaning), 5880 _yield(should_yield), 5881 _concurrent_precleaning(concurrent_precleaning), 5882 _freelistLock(NULL) 5883 { 5884 // FIXME: Should initialize in base class constructor. 5885 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5886 set_ref_discoverer_internal(rd); 5887 } 5888 5889 // This closure is used to mark refs into the CMS generation at the 5890 // second (final) checkpoint, and to scan and transitively follow 5891 // the unmarked oops. It is also used during the concurrent precleaning 5892 // phase while scanning objects on dirty cards in the CMS generation. 5893 // The marks are made in the marking bit map and the marking stack is 5894 // used for keeping the (newly) grey objects during the scan. 5895 // The parallel version (Par_...) appears further below. 5896 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5897 if (obj != NULL) { 5898 assert(oopDesc::is_oop(obj), "expected an oop"); 5899 HeapWord* addr = (HeapWord*)obj; 5900 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5901 assert(_collector->overflow_list_is_empty(), 5902 "overflow list should be empty"); 5903 if (_span.contains(addr) && 5904 !_bit_map->isMarked(addr)) { 5905 // mark bit map (object is now grey) 5906 _bit_map->mark(addr); 5907 // push on marking stack (stack should be empty), and drain the 5908 // stack by applying this closure to the oops in the oops popped 5909 // from the stack (i.e. blacken the grey objects) 5910 bool res = _mark_stack->push(obj); 5911 assert(res, "Should have space to push on empty stack"); 5912 do { 5913 oop new_oop = _mark_stack->pop(); 5914 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5915 assert(_bit_map->isMarked((HeapWord*)new_oop), 5916 "only grey objects on this stack"); 5917 // iterate over the oops in this oop, marking and pushing 5918 // the ones in CMS heap (i.e. in _span). 5919 new_oop->oop_iterate(&_pushAndMarkClosure); 5920 // check if it's time to yield 5921 do_yield_check(); 5922 } while (!_mark_stack->isEmpty() || 5923 (!_concurrent_precleaning && take_from_overflow_list())); 5924 // if marking stack is empty, and we are not doing this 5925 // during precleaning, then check the overflow list 5926 } 5927 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5928 assert(_collector->overflow_list_is_empty(), 5929 "overflow list was drained above"); 5930 5931 assert(_collector->no_preserved_marks(), 5932 "All preserved marks should have been restored above"); 5933 } 5934 } 5935 5936 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5937 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } 5938 5939 void MarkRefsIntoAndScanClosure::do_yield_work() { 5940 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5941 "CMS thread should hold CMS token"); 5942 assert_lock_strong(_freelistLock); 5943 assert_lock_strong(_bit_map->lock()); 5944 // relinquish the free_list_lock and bitMaplock() 5945 _bit_map->lock()->unlock(); 5946 _freelistLock->unlock(); 5947 ConcurrentMarkSweepThread::desynchronize(true); 5948 _collector->stopTimer(); 5949 _collector->incrementYields(); 5950 5951 // See the comment in coordinator_yield() 5952 for (unsigned i = 0; 5953 i < CMSYieldSleepCount && 5954 ConcurrentMarkSweepThread::should_yield() && 5955 !CMSCollector::foregroundGCIsActive(); 5956 ++i) { 5957 os::sleep(Thread::current(), 1, false); 5958 } 5959 5960 ConcurrentMarkSweepThread::synchronize(true); 5961 _freelistLock->lock_without_safepoint_check(); 5962 _bit_map->lock()->lock_without_safepoint_check(); 5963 _collector->startTimer(); 5964 } 5965 5966 /////////////////////////////////////////////////////////// 5967 // ParMarkRefsIntoAndScanClosure: a parallel version of 5968 // MarkRefsIntoAndScanClosure 5969 /////////////////////////////////////////////////////////// 5970 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5971 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 5972 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5973 _span(span), 5974 _bit_map(bit_map), 5975 _work_queue(work_queue), 5976 _low_water_mark(MIN2((work_queue->max_elems()/4), 5977 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 5978 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 5979 { 5980 // FIXME: Should initialize in base class constructor. 5981 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5982 set_ref_discoverer_internal(rd); 5983 } 5984 5985 // This closure is used to mark refs into the CMS generation at the 5986 // second (final) checkpoint, and to scan and transitively follow 5987 // the unmarked oops. The marks are made in the marking bit map and 5988 // the work_queue is used for keeping the (newly) grey objects during 5989 // the scan phase whence they are also available for stealing by parallel 5990 // threads. Since the marking bit map is shared, updates are 5991 // synchronized (via CAS). 5992 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 5993 if (obj != NULL) { 5994 // Ignore mark word because this could be an already marked oop 5995 // that may be chained at the end of the overflow list. 5996 assert(oopDesc::is_oop(obj, true), "expected an oop"); 5997 HeapWord* addr = (HeapWord*)obj; 5998 if (_span.contains(addr) && 5999 !_bit_map->isMarked(addr)) { 6000 // mark bit map (object will become grey): 6001 // It is possible for several threads to be 6002 // trying to "claim" this object concurrently; 6003 // the unique thread that succeeds in marking the 6004 // object first will do the subsequent push on 6005 // to the work queue (or overflow list). 6006 if (_bit_map->par_mark(addr)) { 6007 // push on work_queue (which may not be empty), and trim the 6008 // queue to an appropriate length by applying this closure to 6009 // the oops in the oops popped from the stack (i.e. blacken the 6010 // grey objects) 6011 bool res = _work_queue->push(obj); 6012 assert(res, "Low water mark should be less than capacity?"); 6013 trim_queue(_low_water_mark); 6014 } // Else, another thread claimed the object 6015 } 6016 } 6017 } 6018 6019 void ParMarkRefsIntoAndScanClosure::do_oop(oop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6020 void ParMarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { ParMarkRefsIntoAndScanClosure::do_oop_work(p); } 6021 6022 // This closure is used to rescan the marked objects on the dirty cards 6023 // in the mod union table and the card table proper. 6024 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6025 oop p, MemRegion mr) { 6026 6027 size_t size = 0; 6028 HeapWord* addr = (HeapWord*)p; 6029 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6030 assert(_span.contains(addr), "we are scanning the CMS generation"); 6031 // check if it's time to yield 6032 if (do_yield_check()) { 6033 // We yielded for some foreground stop-world work, 6034 // and we have been asked to abort this ongoing preclean cycle. 6035 return 0; 6036 } 6037 if (_bitMap->isMarked(addr)) { 6038 // it's marked; is it potentially uninitialized? 6039 if (p->klass_or_null_acquire() != NULL) { 6040 // an initialized object; ignore mark word in verification below 6041 // since we are running concurrent with mutators 6042 assert(oopDesc::is_oop(p, true), "should be an oop"); 6043 if (p->is_objArray()) { 6044 // objArrays are precisely marked; restrict scanning 6045 // to dirty cards only. 6046 size = CompactibleFreeListSpace::adjustObjectSize( 6047 p->oop_iterate_size(_scanningClosure, mr)); 6048 } else { 6049 // A non-array may have been imprecisely marked; we need 6050 // to scan object in its entirety. 6051 size = CompactibleFreeListSpace::adjustObjectSize( 6052 p->oop_iterate_size(_scanningClosure)); 6053 } 6054 #ifdef ASSERT 6055 size_t direct_size = 6056 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6057 assert(size == direct_size, "Inconsistency in size"); 6058 assert(size >= 3, "Necessary for Printezis marks to work"); 6059 HeapWord* start_pbit = addr + 1; 6060 HeapWord* end_pbit = addr + size - 1; 6061 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6062 "inconsistent Printezis mark"); 6063 // Verify inner mark bits (between Printezis bits) are clear, 6064 // but don't repeat if there are multiple dirty regions for 6065 // the same object, to avoid potential O(N^2) performance. 6066 if (addr != _last_scanned_object) { 6067 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6068 _last_scanned_object = addr; 6069 } 6070 #endif // ASSERT 6071 } else { 6072 // An uninitialized object. 6073 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6074 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6075 size = pointer_delta(nextOneAddr + 1, addr); 6076 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6077 "alignment problem"); 6078 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6079 // will dirty the card when the klass pointer is installed in the 6080 // object (signaling the completion of initialization). 6081 } 6082 } else { 6083 // Either a not yet marked object or an uninitialized object 6084 if (p->klass_or_null_acquire() == NULL) { 6085 // An uninitialized object, skip to the next card, since 6086 // we may not be able to read its P-bits yet. 6087 assert(size == 0, "Initial value"); 6088 } else { 6089 // An object not (yet) reached by marking: we merely need to 6090 // compute its size so as to go look at the next block. 6091 assert(oopDesc::is_oop(p, true), "should be an oop"); 6092 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6093 } 6094 } 6095 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6096 return size; 6097 } 6098 6099 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6100 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6101 "CMS thread should hold CMS token"); 6102 assert_lock_strong(_freelistLock); 6103 assert_lock_strong(_bitMap->lock()); 6104 // relinquish the free_list_lock and bitMaplock() 6105 _bitMap->lock()->unlock(); 6106 _freelistLock->unlock(); 6107 ConcurrentMarkSweepThread::desynchronize(true); 6108 _collector->stopTimer(); 6109 _collector->incrementYields(); 6110 6111 // See the comment in coordinator_yield() 6112 for (unsigned i = 0; i < CMSYieldSleepCount && 6113 ConcurrentMarkSweepThread::should_yield() && 6114 !CMSCollector::foregroundGCIsActive(); ++i) { 6115 os::sleep(Thread::current(), 1, false); 6116 } 6117 6118 ConcurrentMarkSweepThread::synchronize(true); 6119 _freelistLock->lock_without_safepoint_check(); 6120 _bitMap->lock()->lock_without_safepoint_check(); 6121 _collector->startTimer(); 6122 } 6123 6124 6125 ////////////////////////////////////////////////////////////////// 6126 // SurvivorSpacePrecleanClosure 6127 ////////////////////////////////////////////////////////////////// 6128 // This (single-threaded) closure is used to preclean the oops in 6129 // the survivor spaces. 6130 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6131 6132 HeapWord* addr = (HeapWord*)p; 6133 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6134 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6135 assert(p->klass_or_null() != NULL, "object should be initialized"); 6136 // an initialized object; ignore mark word in verification below 6137 // since we are running concurrent with mutators 6138 assert(oopDesc::is_oop(p, true), "should be an oop"); 6139 // Note that we do not yield while we iterate over 6140 // the interior oops of p, pushing the relevant ones 6141 // on our marking stack. 6142 size_t size = p->oop_iterate_size(_scanning_closure); 6143 do_yield_check(); 6144 // Observe that below, we do not abandon the preclean 6145 // phase as soon as we should; rather we empty the 6146 // marking stack before returning. This is to satisfy 6147 // some existing assertions. In general, it may be a 6148 // good idea to abort immediately and complete the marking 6149 // from the grey objects at a later time. 6150 while (!_mark_stack->isEmpty()) { 6151 oop new_oop = _mark_stack->pop(); 6152 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6153 assert(_bit_map->isMarked((HeapWord*)new_oop), 6154 "only grey objects on this stack"); 6155 // iterate over the oops in this oop, marking and pushing 6156 // the ones in CMS heap (i.e. in _span). 6157 new_oop->oop_iterate(_scanning_closure); 6158 // check if it's time to yield 6159 do_yield_check(); 6160 } 6161 unsigned int after_count = 6162 CMSHeap::heap()->total_collections(); 6163 bool abort = (_before_count != after_count) || 6164 _collector->should_abort_preclean(); 6165 return abort ? 0 : size; 6166 } 6167 6168 void SurvivorSpacePrecleanClosure::do_yield_work() { 6169 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6170 "CMS thread should hold CMS token"); 6171 assert_lock_strong(_bit_map->lock()); 6172 // Relinquish the bit map lock 6173 _bit_map->lock()->unlock(); 6174 ConcurrentMarkSweepThread::desynchronize(true); 6175 _collector->stopTimer(); 6176 _collector->incrementYields(); 6177 6178 // See the comment in coordinator_yield() 6179 for (unsigned i = 0; i < CMSYieldSleepCount && 6180 ConcurrentMarkSweepThread::should_yield() && 6181 !CMSCollector::foregroundGCIsActive(); ++i) { 6182 os::sleep(Thread::current(), 1, false); 6183 } 6184 6185 ConcurrentMarkSweepThread::synchronize(true); 6186 _bit_map->lock()->lock_without_safepoint_check(); 6187 _collector->startTimer(); 6188 } 6189 6190 // This closure is used to rescan the marked objects on the dirty cards 6191 // in the mod union table and the card table proper. In the parallel 6192 // case, although the bitMap is shared, we do a single read so the 6193 // isMarked() query is "safe". 6194 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6195 // Ignore mark word because we are running concurrent with mutators 6196 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6197 HeapWord* addr = (HeapWord*)p; 6198 assert(_span.contains(addr), "we are scanning the CMS generation"); 6199 bool is_obj_array = false; 6200 #ifdef ASSERT 6201 if (!_parallel) { 6202 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6203 assert(_collector->overflow_list_is_empty(), 6204 "overflow list should be empty"); 6205 6206 } 6207 #endif // ASSERT 6208 if (_bit_map->isMarked(addr)) { 6209 // Obj arrays are precisely marked, non-arrays are not; 6210 // so we scan objArrays precisely and non-arrays in their 6211 // entirety. 6212 if (p->is_objArray()) { 6213 is_obj_array = true; 6214 if (_parallel) { 6215 p->oop_iterate(_par_scan_closure, mr); 6216 } else { 6217 p->oop_iterate(_scan_closure, mr); 6218 } 6219 } else { 6220 if (_parallel) { 6221 p->oop_iterate(_par_scan_closure); 6222 } else { 6223 p->oop_iterate(_scan_closure); 6224 } 6225 } 6226 } 6227 #ifdef ASSERT 6228 if (!_parallel) { 6229 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6230 assert(_collector->overflow_list_is_empty(), 6231 "overflow list should be empty"); 6232 6233 } 6234 #endif // ASSERT 6235 return is_obj_array; 6236 } 6237 6238 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6239 MemRegion span, 6240 CMSBitMap* bitMap, CMSMarkStack* markStack, 6241 bool should_yield, bool verifying): 6242 _collector(collector), 6243 _span(span), 6244 _bitMap(bitMap), 6245 _mut(&collector->_modUnionTable), 6246 _markStack(markStack), 6247 _yield(should_yield), 6248 _skipBits(0) 6249 { 6250 assert(_markStack->isEmpty(), "stack should be empty"); 6251 _finger = _bitMap->startWord(); 6252 _threshold = _finger; 6253 assert(_collector->_restart_addr == NULL, "Sanity check"); 6254 assert(_span.contains(_finger), "Out of bounds _finger?"); 6255 DEBUG_ONLY(_verifying = verifying;) 6256 } 6257 6258 void MarkFromRootsClosure::reset(HeapWord* addr) { 6259 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6260 assert(_span.contains(addr), "Out of bounds _finger?"); 6261 _finger = addr; 6262 _threshold = align_up(_finger, CardTable::card_size); 6263 } 6264 6265 // Should revisit to see if this should be restructured for 6266 // greater efficiency. 6267 bool MarkFromRootsClosure::do_bit(size_t offset) { 6268 if (_skipBits > 0) { 6269 _skipBits--; 6270 return true; 6271 } 6272 // convert offset into a HeapWord* 6273 HeapWord* addr = _bitMap->startWord() + offset; 6274 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6275 "address out of range"); 6276 assert(_bitMap->isMarked(addr), "tautology"); 6277 if (_bitMap->isMarked(addr+1)) { 6278 // this is an allocated but not yet initialized object 6279 assert(_skipBits == 0, "tautology"); 6280 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6281 oop p = oop(addr); 6282 if (p->klass_or_null_acquire() == NULL) { 6283 DEBUG_ONLY(if (!_verifying) {) 6284 // We re-dirty the cards on which this object lies and increase 6285 // the _threshold so that we'll come back to scan this object 6286 // during the preclean or remark phase. (CMSCleanOnEnter) 6287 if (CMSCleanOnEnter) { 6288 size_t sz = _collector->block_size_using_printezis_bits(addr); 6289 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6290 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6291 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6292 // Bump _threshold to end_card_addr; note that 6293 // _threshold cannot possibly exceed end_card_addr, anyhow. 6294 // This prevents future clearing of the card as the scan proceeds 6295 // to the right. 6296 assert(_threshold <= end_card_addr, 6297 "Because we are just scanning into this object"); 6298 if (_threshold < end_card_addr) { 6299 _threshold = end_card_addr; 6300 } 6301 if (p->klass_or_null_acquire() != NULL) { 6302 // Redirty the range of cards... 6303 _mut->mark_range(redirty_range); 6304 } // ...else the setting of klass will dirty the card anyway. 6305 } 6306 DEBUG_ONLY(}) 6307 return true; 6308 } 6309 } 6310 scanOopsInOop(addr); 6311 return true; 6312 } 6313 6314 // We take a break if we've been at this for a while, 6315 // so as to avoid monopolizing the locks involved. 6316 void MarkFromRootsClosure::do_yield_work() { 6317 // First give up the locks, then yield, then re-lock 6318 // We should probably use a constructor/destructor idiom to 6319 // do this unlock/lock or modify the MutexUnlocker class to 6320 // serve our purpose. XXX 6321 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6322 "CMS thread should hold CMS token"); 6323 assert_lock_strong(_bitMap->lock()); 6324 _bitMap->lock()->unlock(); 6325 ConcurrentMarkSweepThread::desynchronize(true); 6326 _collector->stopTimer(); 6327 _collector->incrementYields(); 6328 6329 // See the comment in coordinator_yield() 6330 for (unsigned i = 0; i < CMSYieldSleepCount && 6331 ConcurrentMarkSweepThread::should_yield() && 6332 !CMSCollector::foregroundGCIsActive(); ++i) { 6333 os::sleep(Thread::current(), 1, false); 6334 } 6335 6336 ConcurrentMarkSweepThread::synchronize(true); 6337 _bitMap->lock()->lock_without_safepoint_check(); 6338 _collector->startTimer(); 6339 } 6340 6341 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6342 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6343 assert(_markStack->isEmpty(), 6344 "should drain stack to limit stack usage"); 6345 // convert ptr to an oop preparatory to scanning 6346 oop obj = oop(ptr); 6347 // Ignore mark word in verification below, since we 6348 // may be running concurrent with mutators. 6349 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6350 assert(_finger <= ptr, "_finger runneth ahead"); 6351 // advance the finger to right end of this object 6352 _finger = ptr + obj->size(); 6353 assert(_finger > ptr, "we just incremented it above"); 6354 // On large heaps, it may take us some time to get through 6355 // the marking phase. During 6356 // this time it's possible that a lot of mutations have 6357 // accumulated in the card table and the mod union table -- 6358 // these mutation records are redundant until we have 6359 // actually traced into the corresponding card. 6360 // Here, we check whether advancing the finger would make 6361 // us cross into a new card, and if so clear corresponding 6362 // cards in the MUT (preclean them in the card-table in the 6363 // future). 6364 6365 DEBUG_ONLY(if (!_verifying) {) 6366 // The clean-on-enter optimization is disabled by default, 6367 // until we fix 6178663. 6368 if (CMSCleanOnEnter && (_finger > _threshold)) { 6369 // [_threshold, _finger) represents the interval 6370 // of cards to be cleared in MUT (or precleaned in card table). 6371 // The set of cards to be cleared is all those that overlap 6372 // with the interval [_threshold, _finger); note that 6373 // _threshold is always kept card-aligned but _finger isn't 6374 // always card-aligned. 6375 HeapWord* old_threshold = _threshold; 6376 assert(is_aligned(old_threshold, CardTable::card_size), 6377 "_threshold should always be card-aligned"); 6378 _threshold = align_up(_finger, CardTable::card_size); 6379 MemRegion mr(old_threshold, _threshold); 6380 assert(!mr.is_empty(), "Control point invariant"); 6381 assert(_span.contains(mr), "Should clear within span"); 6382 _mut->clear_range(mr); 6383 } 6384 DEBUG_ONLY(}) 6385 // Note: the finger doesn't advance while we drain 6386 // the stack below. 6387 PushOrMarkClosure pushOrMarkClosure(_collector, 6388 _span, _bitMap, _markStack, 6389 _finger, this); 6390 bool res = _markStack->push(obj); 6391 assert(res, "Empty non-zero size stack should have space for single push"); 6392 while (!_markStack->isEmpty()) { 6393 oop new_oop = _markStack->pop(); 6394 // Skip verifying header mark word below because we are 6395 // running concurrent with mutators. 6396 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6397 // now scan this oop's oops 6398 new_oop->oop_iterate(&pushOrMarkClosure); 6399 do_yield_check(); 6400 } 6401 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6402 } 6403 6404 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6405 CMSCollector* collector, MemRegion span, 6406 CMSBitMap* bit_map, 6407 OopTaskQueue* work_queue, 6408 CMSMarkStack* overflow_stack): 6409 _collector(collector), 6410 _whole_span(collector->_span), 6411 _span(span), 6412 _bit_map(bit_map), 6413 _mut(&collector->_modUnionTable), 6414 _work_queue(work_queue), 6415 _overflow_stack(overflow_stack), 6416 _skip_bits(0), 6417 _task(task) 6418 { 6419 assert(_work_queue->size() == 0, "work_queue should be empty"); 6420 _finger = span.start(); 6421 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6422 assert(_span.contains(_finger), "Out of bounds _finger?"); 6423 } 6424 6425 // Should revisit to see if this should be restructured for 6426 // greater efficiency. 6427 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6428 if (_skip_bits > 0) { 6429 _skip_bits--; 6430 return true; 6431 } 6432 // convert offset into a HeapWord* 6433 HeapWord* addr = _bit_map->startWord() + offset; 6434 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6435 "address out of range"); 6436 assert(_bit_map->isMarked(addr), "tautology"); 6437 if (_bit_map->isMarked(addr+1)) { 6438 // this is an allocated object that might not yet be initialized 6439 assert(_skip_bits == 0, "tautology"); 6440 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6441 oop p = oop(addr); 6442 if (p->klass_or_null_acquire() == NULL) { 6443 // in the case of Clean-on-Enter optimization, redirty card 6444 // and avoid clearing card by increasing the threshold. 6445 return true; 6446 } 6447 } 6448 scan_oops_in_oop(addr); 6449 return true; 6450 } 6451 6452 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6453 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6454 // Should we assert that our work queue is empty or 6455 // below some drain limit? 6456 assert(_work_queue->size() == 0, 6457 "should drain stack to limit stack usage"); 6458 // convert ptr to an oop preparatory to scanning 6459 oop obj = oop(ptr); 6460 // Ignore mark word in verification below, since we 6461 // may be running concurrent with mutators. 6462 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6463 assert(_finger <= ptr, "_finger runneth ahead"); 6464 // advance the finger to right end of this object 6465 _finger = ptr + obj->size(); 6466 assert(_finger > ptr, "we just incremented it above"); 6467 // On large heaps, it may take us some time to get through 6468 // the marking phase. During 6469 // this time it's possible that a lot of mutations have 6470 // accumulated in the card table and the mod union table -- 6471 // these mutation records are redundant until we have 6472 // actually traced into the corresponding card. 6473 // Here, we check whether advancing the finger would make 6474 // us cross into a new card, and if so clear corresponding 6475 // cards in the MUT (preclean them in the card-table in the 6476 // future). 6477 6478 // The clean-on-enter optimization is disabled by default, 6479 // until we fix 6178663. 6480 if (CMSCleanOnEnter && (_finger > _threshold)) { 6481 // [_threshold, _finger) represents the interval 6482 // of cards to be cleared in MUT (or precleaned in card table). 6483 // The set of cards to be cleared is all those that overlap 6484 // with the interval [_threshold, _finger); note that 6485 // _threshold is always kept card-aligned but _finger isn't 6486 // always card-aligned. 6487 HeapWord* old_threshold = _threshold; 6488 assert(is_aligned(old_threshold, CardTable::card_size), 6489 "_threshold should always be card-aligned"); 6490 _threshold = align_up(_finger, CardTable::card_size); 6491 MemRegion mr(old_threshold, _threshold); 6492 assert(!mr.is_empty(), "Control point invariant"); 6493 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6494 _mut->clear_range(mr); 6495 } 6496 6497 // Note: the local finger doesn't advance while we drain 6498 // the stack below, but the global finger sure can and will. 6499 HeapWord* volatile* gfa = _task->global_finger_addr(); 6500 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6501 _span, _bit_map, 6502 _work_queue, 6503 _overflow_stack, 6504 _finger, 6505 gfa, this); 6506 bool res = _work_queue->push(obj); // overflow could occur here 6507 assert(res, "Will hold once we use workqueues"); 6508 while (true) { 6509 oop new_oop; 6510 if (!_work_queue->pop_local(new_oop)) { 6511 // We emptied our work_queue; check if there's stuff that can 6512 // be gotten from the overflow stack. 6513 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6514 _overflow_stack, _work_queue)) { 6515 do_yield_check(); 6516 continue; 6517 } else { // done 6518 break; 6519 } 6520 } 6521 // Skip verifying header mark word below because we are 6522 // running concurrent with mutators. 6523 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6524 // now scan this oop's oops 6525 new_oop->oop_iterate(&pushOrMarkClosure); 6526 do_yield_check(); 6527 } 6528 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6529 } 6530 6531 // Yield in response to a request from VM Thread or 6532 // from mutators. 6533 void ParMarkFromRootsClosure::do_yield_work() { 6534 assert(_task != NULL, "sanity"); 6535 _task->yield(); 6536 } 6537 6538 // A variant of the above used for verifying CMS marking work. 6539 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6540 MemRegion span, 6541 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6542 CMSMarkStack* mark_stack): 6543 _collector(collector), 6544 _span(span), 6545 _verification_bm(verification_bm), 6546 _cms_bm(cms_bm), 6547 _mark_stack(mark_stack), 6548 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6549 mark_stack) 6550 { 6551 assert(_mark_stack->isEmpty(), "stack should be empty"); 6552 _finger = _verification_bm->startWord(); 6553 assert(_collector->_restart_addr == NULL, "Sanity check"); 6554 assert(_span.contains(_finger), "Out of bounds _finger?"); 6555 } 6556 6557 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6558 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6559 assert(_span.contains(addr), "Out of bounds _finger?"); 6560 _finger = addr; 6561 } 6562 6563 // Should revisit to see if this should be restructured for 6564 // greater efficiency. 6565 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6566 // convert offset into a HeapWord* 6567 HeapWord* addr = _verification_bm->startWord() + offset; 6568 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6569 "address out of range"); 6570 assert(_verification_bm->isMarked(addr), "tautology"); 6571 assert(_cms_bm->isMarked(addr), "tautology"); 6572 6573 assert(_mark_stack->isEmpty(), 6574 "should drain stack to limit stack usage"); 6575 // convert addr to an oop preparatory to scanning 6576 oop obj = oop(addr); 6577 assert(oopDesc::is_oop(obj), "should be an oop"); 6578 assert(_finger <= addr, "_finger runneth ahead"); 6579 // advance the finger to right end of this object 6580 _finger = addr + obj->size(); 6581 assert(_finger > addr, "we just incremented it above"); 6582 // Note: the finger doesn't advance while we drain 6583 // the stack below. 6584 bool res = _mark_stack->push(obj); 6585 assert(res, "Empty non-zero size stack should have space for single push"); 6586 while (!_mark_stack->isEmpty()) { 6587 oop new_oop = _mark_stack->pop(); 6588 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6589 // now scan this oop's oops 6590 new_oop->oop_iterate(&_pam_verify_closure); 6591 } 6592 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6593 return true; 6594 } 6595 6596 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6597 CMSCollector* collector, MemRegion span, 6598 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6599 CMSMarkStack* mark_stack): 6600 MetadataAwareOopClosure(collector->ref_processor()), 6601 _collector(collector), 6602 _span(span), 6603 _verification_bm(verification_bm), 6604 _cms_bm(cms_bm), 6605 _mark_stack(mark_stack) 6606 { } 6607 6608 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6609 oop obj = RawAccess<>::oop_load(p); 6610 do_oop(obj); 6611 } 6612 6613 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6614 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6615 6616 // Upon stack overflow, we discard (part of) the stack, 6617 // remembering the least address amongst those discarded 6618 // in CMSCollector's _restart_address. 6619 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6620 // Remember the least grey address discarded 6621 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6622 _collector->lower_restart_addr(ra); 6623 _mark_stack->reset(); // discard stack contents 6624 _mark_stack->expand(); // expand the stack if possible 6625 } 6626 6627 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6628 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6629 HeapWord* addr = (HeapWord*)obj; 6630 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6631 // Oop lies in _span and isn't yet grey or black 6632 _verification_bm->mark(addr); // now grey 6633 if (!_cms_bm->isMarked(addr)) { 6634 Log(gc, verify) log; 6635 ResourceMark rm; 6636 LogStream ls(log.error()); 6637 oop(addr)->print_on(&ls); 6638 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6639 fatal("... aborting"); 6640 } 6641 6642 if (!_mark_stack->push(obj)) { // stack overflow 6643 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6644 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6645 handle_stack_overflow(addr); 6646 } 6647 // anything including and to the right of _finger 6648 // will be scanned as we iterate over the remainder of the 6649 // bit map 6650 } 6651 } 6652 6653 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6654 MemRegion span, 6655 CMSBitMap* bitMap, CMSMarkStack* markStack, 6656 HeapWord* finger, MarkFromRootsClosure* parent) : 6657 MetadataAwareOopClosure(collector->ref_processor()), 6658 _collector(collector), 6659 _span(span), 6660 _bitMap(bitMap), 6661 _markStack(markStack), 6662 _finger(finger), 6663 _parent(parent) 6664 { } 6665 6666 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6667 MemRegion span, 6668 CMSBitMap* bit_map, 6669 OopTaskQueue* work_queue, 6670 CMSMarkStack* overflow_stack, 6671 HeapWord* finger, 6672 HeapWord* volatile* global_finger_addr, 6673 ParMarkFromRootsClosure* parent) : 6674 MetadataAwareOopClosure(collector->ref_processor()), 6675 _collector(collector), 6676 _whole_span(collector->_span), 6677 _span(span), 6678 _bit_map(bit_map), 6679 _work_queue(work_queue), 6680 _overflow_stack(overflow_stack), 6681 _finger(finger), 6682 _global_finger_addr(global_finger_addr), 6683 _parent(parent) 6684 { } 6685 6686 // Assumes thread-safe access by callers, who are 6687 // responsible for mutual exclusion. 6688 void CMSCollector::lower_restart_addr(HeapWord* low) { 6689 assert(_span.contains(low), "Out of bounds addr"); 6690 if (_restart_addr == NULL) { 6691 _restart_addr = low; 6692 } else { 6693 _restart_addr = MIN2(_restart_addr, low); 6694 } 6695 } 6696 6697 // Upon stack overflow, we discard (part of) the stack, 6698 // remembering the least address amongst those discarded 6699 // in CMSCollector's _restart_address. 6700 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6701 // Remember the least grey address discarded 6702 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6703 _collector->lower_restart_addr(ra); 6704 _markStack->reset(); // discard stack contents 6705 _markStack->expand(); // expand the stack if possible 6706 } 6707 6708 // Upon stack overflow, we discard (part of) the stack, 6709 // remembering the least address amongst those discarded 6710 // in CMSCollector's _restart_address. 6711 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6712 // We need to do this under a mutex to prevent other 6713 // workers from interfering with the work done below. 6714 MutexLockerEx ml(_overflow_stack->par_lock(), 6715 Mutex::_no_safepoint_check_flag); 6716 // Remember the least grey address discarded 6717 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6718 _collector->lower_restart_addr(ra); 6719 _overflow_stack->reset(); // discard stack contents 6720 _overflow_stack->expand(); // expand the stack if possible 6721 } 6722 6723 void PushOrMarkClosure::do_oop(oop obj) { 6724 // Ignore mark word because we are running concurrent with mutators. 6725 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6726 HeapWord* addr = (HeapWord*)obj; 6727 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6728 // Oop lies in _span and isn't yet grey or black 6729 _bitMap->mark(addr); // now grey 6730 if (addr < _finger) { 6731 // the bit map iteration has already either passed, or 6732 // sampled, this bit in the bit map; we'll need to 6733 // use the marking stack to scan this oop's oops. 6734 bool simulate_overflow = false; 6735 NOT_PRODUCT( 6736 if (CMSMarkStackOverflowALot && 6737 _collector->simulate_overflow()) { 6738 // simulate a stack overflow 6739 simulate_overflow = true; 6740 } 6741 ) 6742 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6743 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6744 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6745 handle_stack_overflow(addr); 6746 } 6747 } 6748 // anything including and to the right of _finger 6749 // will be scanned as we iterate over the remainder of the 6750 // bit map 6751 do_yield_check(); 6752 } 6753 } 6754 6755 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } 6756 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } 6757 6758 void ParPushOrMarkClosure::do_oop(oop obj) { 6759 // Ignore mark word because we are running concurrent with mutators. 6760 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6761 HeapWord* addr = (HeapWord*)obj; 6762 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6763 // Oop lies in _span and isn't yet grey or black 6764 // We read the global_finger (volatile read) strictly after marking oop 6765 bool res = _bit_map->par_mark(addr); // now grey 6766 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6767 // Should we push this marked oop on our stack? 6768 // -- if someone else marked it, nothing to do 6769 // -- if target oop is above global finger nothing to do 6770 // -- if target oop is in chunk and above local finger 6771 // then nothing to do 6772 // -- else push on work queue 6773 if ( !res // someone else marked it, they will deal with it 6774 || (addr >= *gfa) // will be scanned in a later task 6775 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6776 return; 6777 } 6778 // the bit map iteration has already either passed, or 6779 // sampled, this bit in the bit map; we'll need to 6780 // use the marking stack to scan this oop's oops. 6781 bool simulate_overflow = false; 6782 NOT_PRODUCT( 6783 if (CMSMarkStackOverflowALot && 6784 _collector->simulate_overflow()) { 6785 // simulate a stack overflow 6786 simulate_overflow = true; 6787 } 6788 ) 6789 if (simulate_overflow || 6790 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6791 // stack overflow 6792 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6793 // We cannot assert that the overflow stack is full because 6794 // it may have been emptied since. 6795 assert(simulate_overflow || 6796 _work_queue->size() == _work_queue->max_elems(), 6797 "Else push should have succeeded"); 6798 handle_stack_overflow(addr); 6799 } 6800 do_yield_check(); 6801 } 6802 } 6803 6804 void ParPushOrMarkClosure::do_oop(oop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6805 void ParPushOrMarkClosure::do_oop(narrowOop* p) { ParPushOrMarkClosure::do_oop_work(p); } 6806 6807 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6808 MemRegion span, 6809 ReferenceDiscoverer* rd, 6810 CMSBitMap* bit_map, 6811 CMSBitMap* mod_union_table, 6812 CMSMarkStack* mark_stack, 6813 bool concurrent_precleaning): 6814 MetadataAwareOopClosure(rd), 6815 _collector(collector), 6816 _span(span), 6817 _bit_map(bit_map), 6818 _mod_union_table(mod_union_table), 6819 _mark_stack(mark_stack), 6820 _concurrent_precleaning(concurrent_precleaning) 6821 { 6822 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6823 } 6824 6825 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6826 // the non-parallel version (the parallel version appears further below.) 6827 void PushAndMarkClosure::do_oop(oop obj) { 6828 // Ignore mark word verification. If during concurrent precleaning, 6829 // the object monitor may be locked. If during the checkpoint 6830 // phases, the object may already have been reached by a different 6831 // path and may be at the end of the global overflow list (so 6832 // the mark word may be NULL). 6833 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6834 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6835 HeapWord* addr = (HeapWord*)obj; 6836 // Check if oop points into the CMS generation 6837 // and is not marked 6838 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6839 // a white object ... 6840 _bit_map->mark(addr); // ... now grey 6841 // push on the marking stack (grey set) 6842 bool simulate_overflow = false; 6843 NOT_PRODUCT( 6844 if (CMSMarkStackOverflowALot && 6845 _collector->simulate_overflow()) { 6846 // simulate a stack overflow 6847 simulate_overflow = true; 6848 } 6849 ) 6850 if (simulate_overflow || !_mark_stack->push(obj)) { 6851 if (_concurrent_precleaning) { 6852 // During precleaning we can just dirty the appropriate card(s) 6853 // in the mod union table, thus ensuring that the object remains 6854 // in the grey set and continue. In the case of object arrays 6855 // we need to dirty all of the cards that the object spans, 6856 // since the rescan of object arrays will be limited to the 6857 // dirty cards. 6858 // Note that no one can be interfering with us in this action 6859 // of dirtying the mod union table, so no locking or atomics 6860 // are required. 6861 if (obj->is_objArray()) { 6862 size_t sz = obj->size(); 6863 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6864 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6865 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6866 _mod_union_table->mark_range(redirty_range); 6867 } else { 6868 _mod_union_table->mark(addr); 6869 } 6870 _collector->_ser_pmc_preclean_ovflw++; 6871 } else { 6872 // During the remark phase, we need to remember this oop 6873 // in the overflow list. 6874 _collector->push_on_overflow_list(obj); 6875 _collector->_ser_pmc_remark_ovflw++; 6876 } 6877 } 6878 } 6879 } 6880 6881 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6882 MemRegion span, 6883 ReferenceDiscoverer* rd, 6884 CMSBitMap* bit_map, 6885 OopTaskQueue* work_queue): 6886 MetadataAwareOopClosure(rd), 6887 _collector(collector), 6888 _span(span), 6889 _bit_map(bit_map), 6890 _work_queue(work_queue) 6891 { 6892 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6893 } 6894 6895 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } 6896 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } 6897 6898 // Grey object rescan during second checkpoint phase -- 6899 // the parallel version. 6900 void ParPushAndMarkClosure::do_oop(oop obj) { 6901 // In the assert below, we ignore the mark word because 6902 // this oop may point to an already visited object that is 6903 // on the overflow stack (in which case the mark word has 6904 // been hijacked for chaining into the overflow stack -- 6905 // if this is the last object in the overflow stack then 6906 // its mark word will be NULL). Because this object may 6907 // have been subsequently popped off the global overflow 6908 // stack, and the mark word possibly restored to the prototypical 6909 // value, by the time we get to examined this failing assert in 6910 // the debugger, is_oop_or_null(false) may subsequently start 6911 // to hold. 6912 assert(oopDesc::is_oop_or_null(obj, true), 6913 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6914 HeapWord* addr = (HeapWord*)obj; 6915 // Check if oop points into the CMS generation 6916 // and is not marked 6917 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6918 // a white object ... 6919 // If we manage to "claim" the object, by being the 6920 // first thread to mark it, then we push it on our 6921 // marking stack 6922 if (_bit_map->par_mark(addr)) { // ... now grey 6923 // push on work queue (grey set) 6924 bool simulate_overflow = false; 6925 NOT_PRODUCT( 6926 if (CMSMarkStackOverflowALot && 6927 _collector->par_simulate_overflow()) { 6928 // simulate a stack overflow 6929 simulate_overflow = true; 6930 } 6931 ) 6932 if (simulate_overflow || !_work_queue->push(obj)) { 6933 _collector->par_push_on_overflow_list(obj); 6934 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6935 } 6936 } // Else, some other thread got there first 6937 } 6938 } 6939 6940 void ParPushAndMarkClosure::do_oop(oop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6941 void ParPushAndMarkClosure::do_oop(narrowOop* p) { ParPushAndMarkClosure::do_oop_work(p); } 6942 6943 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6944 Mutex* bml = _collector->bitMapLock(); 6945 assert_lock_strong(bml); 6946 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6947 "CMS thread should hold CMS token"); 6948 6949 bml->unlock(); 6950 ConcurrentMarkSweepThread::desynchronize(true); 6951 6952 _collector->stopTimer(); 6953 _collector->incrementYields(); 6954 6955 // See the comment in coordinator_yield() 6956 for (unsigned i = 0; i < CMSYieldSleepCount && 6957 ConcurrentMarkSweepThread::should_yield() && 6958 !CMSCollector::foregroundGCIsActive(); ++i) { 6959 os::sleep(Thread::current(), 1, false); 6960 } 6961 6962 ConcurrentMarkSweepThread::synchronize(true); 6963 bml->lock(); 6964 6965 _collector->startTimer(); 6966 } 6967 6968 bool CMSPrecleanRefsYieldClosure::should_return() { 6969 if (ConcurrentMarkSweepThread::should_yield()) { 6970 do_yield_work(); 6971 } 6972 return _collector->foregroundGCIsActive(); 6973 } 6974 6975 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6976 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 6977 "mr should be aligned to start at a card boundary"); 6978 // We'd like to assert: 6979 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 6980 // "mr should be a range of cards"); 6981 // However, that would be too strong in one case -- the last 6982 // partition ends at _unallocated_block which, in general, can be 6983 // an arbitrary boundary, not necessarily card aligned. 6984 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 6985 _space->object_iterate_mem(mr, &_scan_cl); 6986 } 6987 6988 SweepClosure::SweepClosure(CMSCollector* collector, 6989 ConcurrentMarkSweepGeneration* g, 6990 CMSBitMap* bitMap, bool should_yield) : 6991 _collector(collector), 6992 _g(g), 6993 _sp(g->cmsSpace()), 6994 _limit(_sp->sweep_limit()), 6995 _freelistLock(_sp->freelistLock()), 6996 _bitMap(bitMap), 6997 _yield(should_yield), 6998 _inFreeRange(false), // No free range at beginning of sweep 6999 _freeRangeInFreeLists(false), // No free range at beginning of sweep 7000 _lastFreeRangeCoalesced(false), 7001 _freeFinger(g->used_region().start()) 7002 { 7003 NOT_PRODUCT( 7004 _numObjectsFreed = 0; 7005 _numWordsFreed = 0; 7006 _numObjectsLive = 0; 7007 _numWordsLive = 0; 7008 _numObjectsAlreadyFree = 0; 7009 _numWordsAlreadyFree = 0; 7010 _last_fc = NULL; 7011 7012 _sp->initializeIndexedFreeListArrayReturnedBytes(); 7013 _sp->dictionary()->initialize_dict_returned_bytes(); 7014 ) 7015 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7016 "sweep _limit out of bounds"); 7017 log_develop_trace(gc, sweep)("===================="); 7018 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 7019 } 7020 7021 void SweepClosure::print_on(outputStream* st) const { 7022 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 7023 p2i(_sp->bottom()), p2i(_sp->end())); 7024 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 7025 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 7026 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 7027 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 7028 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 7029 } 7030 7031 #ifndef PRODUCT 7032 // Assertion checking only: no useful work in product mode -- 7033 // however, if any of the flags below become product flags, 7034 // you may need to review this code to see if it needs to be 7035 // enabled in product mode. 7036 SweepClosure::~SweepClosure() { 7037 assert_lock_strong(_freelistLock); 7038 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7039 "sweep _limit out of bounds"); 7040 if (inFreeRange()) { 7041 Log(gc, sweep) log; 7042 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7043 ResourceMark rm; 7044 LogStream ls(log.error()); 7045 print_on(&ls); 7046 ShouldNotReachHere(); 7047 } 7048 7049 if (log_is_enabled(Debug, gc, sweep)) { 7050 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7051 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7052 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7053 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7054 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7055 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7056 } 7057 7058 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7059 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7060 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7061 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7062 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7063 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7064 } 7065 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7066 log_develop_trace(gc, sweep)("================"); 7067 } 7068 #endif // PRODUCT 7069 7070 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7071 bool freeRangeInFreeLists) { 7072 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7073 p2i(freeFinger), freeRangeInFreeLists); 7074 assert(!inFreeRange(), "Trampling existing free range"); 7075 set_inFreeRange(true); 7076 set_lastFreeRangeCoalesced(false); 7077 7078 set_freeFinger(freeFinger); 7079 set_freeRangeInFreeLists(freeRangeInFreeLists); 7080 if (CMSTestInFreeList) { 7081 if (freeRangeInFreeLists) { 7082 FreeChunk* fc = (FreeChunk*) freeFinger; 7083 assert(fc->is_free(), "A chunk on the free list should be free."); 7084 assert(fc->size() > 0, "Free range should have a size"); 7085 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7086 } 7087 } 7088 } 7089 7090 // Note that the sweeper runs concurrently with mutators. Thus, 7091 // it is possible for direct allocation in this generation to happen 7092 // in the middle of the sweep. Note that the sweeper also coalesces 7093 // contiguous free blocks. Thus, unless the sweeper and the allocator 7094 // synchronize appropriately freshly allocated blocks may get swept up. 7095 // This is accomplished by the sweeper locking the free lists while 7096 // it is sweeping. Thus blocks that are determined to be free are 7097 // indeed free. There is however one additional complication: 7098 // blocks that have been allocated since the final checkpoint and 7099 // mark, will not have been marked and so would be treated as 7100 // unreachable and swept up. To prevent this, the allocator marks 7101 // the bit map when allocating during the sweep phase. This leads, 7102 // however, to a further complication -- objects may have been allocated 7103 // but not yet initialized -- in the sense that the header isn't yet 7104 // installed. The sweeper can not then determine the size of the block 7105 // in order to skip over it. To deal with this case, we use a technique 7106 // (due to Printezis) to encode such uninitialized block sizes in the 7107 // bit map. Since the bit map uses a bit per every HeapWord, but the 7108 // CMS generation has a minimum object size of 3 HeapWords, it follows 7109 // that "normal marks" won't be adjacent in the bit map (there will 7110 // always be at least two 0 bits between successive 1 bits). We make use 7111 // of these "unused" bits to represent uninitialized blocks -- the bit 7112 // corresponding to the start of the uninitialized object and the next 7113 // bit are both set. Finally, a 1 bit marks the end of the object that 7114 // started with the two consecutive 1 bits to indicate its potentially 7115 // uninitialized state. 7116 7117 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7118 FreeChunk* fc = (FreeChunk*)addr; 7119 size_t res; 7120 7121 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7122 // than "addr == _limit" because although _limit was a block boundary when 7123 // we started the sweep, it may no longer be one because heap expansion 7124 // may have caused us to coalesce the block ending at the address _limit 7125 // with a newly expanded chunk (this happens when _limit was set to the 7126 // previous _end of the space), so we may have stepped past _limit: 7127 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7128 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7129 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7130 "sweep _limit out of bounds"); 7131 assert(addr < _sp->end(), "addr out of bounds"); 7132 // Flush any free range we might be holding as a single 7133 // coalesced chunk to the appropriate free list. 7134 if (inFreeRange()) { 7135 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7136 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7137 flush_cur_free_chunk(freeFinger(), 7138 pointer_delta(addr, freeFinger())); 7139 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7140 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7141 lastFreeRangeCoalesced() ? 1 : 0); 7142 } 7143 7144 // help the iterator loop finish 7145 return pointer_delta(_sp->end(), addr); 7146 } 7147 7148 assert(addr < _limit, "sweep invariant"); 7149 // check if we should yield 7150 do_yield_check(addr); 7151 if (fc->is_free()) { 7152 // Chunk that is already free 7153 res = fc->size(); 7154 do_already_free_chunk(fc); 7155 debug_only(_sp->verifyFreeLists()); 7156 // If we flush the chunk at hand in lookahead_and_flush() 7157 // and it's coalesced with a preceding chunk, then the 7158 // process of "mangling" the payload of the coalesced block 7159 // will cause erasure of the size information from the 7160 // (erstwhile) header of all the coalesced blocks but the 7161 // first, so the first disjunct in the assert will not hold 7162 // in that specific case (in which case the second disjunct 7163 // will hold). 7164 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7165 "Otherwise the size info doesn't change at this step"); 7166 NOT_PRODUCT( 7167 _numObjectsAlreadyFree++; 7168 _numWordsAlreadyFree += res; 7169 ) 7170 NOT_PRODUCT(_last_fc = fc;) 7171 } else if (!_bitMap->isMarked(addr)) { 7172 // Chunk is fresh garbage 7173 res = do_garbage_chunk(fc); 7174 debug_only(_sp->verifyFreeLists()); 7175 NOT_PRODUCT( 7176 _numObjectsFreed++; 7177 _numWordsFreed += res; 7178 ) 7179 } else { 7180 // Chunk that is alive. 7181 res = do_live_chunk(fc); 7182 debug_only(_sp->verifyFreeLists()); 7183 NOT_PRODUCT( 7184 _numObjectsLive++; 7185 _numWordsLive += res; 7186 ) 7187 } 7188 return res; 7189 } 7190 7191 // For the smart allocation, record following 7192 // split deaths - a free chunk is removed from its free list because 7193 // it is being split into two or more chunks. 7194 // split birth - a free chunk is being added to its free list because 7195 // a larger free chunk has been split and resulted in this free chunk. 7196 // coal death - a free chunk is being removed from its free list because 7197 // it is being coalesced into a large free chunk. 7198 // coal birth - a free chunk is being added to its free list because 7199 // it was created when two or more free chunks where coalesced into 7200 // this free chunk. 7201 // 7202 // These statistics are used to determine the desired number of free 7203 // chunks of a given size. The desired number is chosen to be relative 7204 // to the end of a CMS sweep. The desired number at the end of a sweep 7205 // is the 7206 // count-at-end-of-previous-sweep (an amount that was enough) 7207 // - count-at-beginning-of-current-sweep (the excess) 7208 // + split-births (gains in this size during interval) 7209 // - split-deaths (demands on this size during interval) 7210 // where the interval is from the end of one sweep to the end of the 7211 // next. 7212 // 7213 // When sweeping the sweeper maintains an accumulated chunk which is 7214 // the chunk that is made up of chunks that have been coalesced. That 7215 // will be termed the left-hand chunk. A new chunk of garbage that 7216 // is being considered for coalescing will be referred to as the 7217 // right-hand chunk. 7218 // 7219 // When making a decision on whether to coalesce a right-hand chunk with 7220 // the current left-hand chunk, the current count vs. the desired count 7221 // of the left-hand chunk is considered. Also if the right-hand chunk 7222 // is near the large chunk at the end of the heap (see 7223 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7224 // left-hand chunk is coalesced. 7225 // 7226 // When making a decision about whether to split a chunk, the desired count 7227 // vs. the current count of the candidate to be split is also considered. 7228 // If the candidate is underpopulated (currently fewer chunks than desired) 7229 // a chunk of an overpopulated (currently more chunks than desired) size may 7230 // be chosen. The "hint" associated with a free list, if non-null, points 7231 // to a free list which may be overpopulated. 7232 // 7233 7234 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7235 const size_t size = fc->size(); 7236 // Chunks that cannot be coalesced are not in the 7237 // free lists. 7238 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7239 assert(_sp->verify_chunk_in_free_list(fc), 7240 "free chunk should be in free lists"); 7241 } 7242 // a chunk that is already free, should not have been 7243 // marked in the bit map 7244 HeapWord* const addr = (HeapWord*) fc; 7245 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7246 // Verify that the bit map has no bits marked between 7247 // addr and purported end of this block. 7248 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7249 7250 // Some chunks cannot be coalesced under any circumstances. 7251 // See the definition of cantCoalesce(). 7252 if (!fc->cantCoalesce()) { 7253 // This chunk can potentially be coalesced. 7254 // All the work is done in 7255 do_post_free_or_garbage_chunk(fc, size); 7256 // Note that if the chunk is not coalescable (the else arm 7257 // below), we unconditionally flush, without needing to do 7258 // a "lookahead," as we do below. 7259 if (inFreeRange()) lookahead_and_flush(fc, size); 7260 } else { 7261 // Code path common to both original and adaptive free lists. 7262 7263 // cant coalesce with previous block; this should be treated 7264 // as the end of a free run if any 7265 if (inFreeRange()) { 7266 // we kicked some butt; time to pick up the garbage 7267 assert(freeFinger() < addr, "freeFinger points too high"); 7268 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7269 } 7270 // else, nothing to do, just continue 7271 } 7272 } 7273 7274 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7275 // This is a chunk of garbage. It is not in any free list. 7276 // Add it to a free list or let it possibly be coalesced into 7277 // a larger chunk. 7278 HeapWord* const addr = (HeapWord*) fc; 7279 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7280 7281 // Verify that the bit map has no bits marked between 7282 // addr and purported end of just dead object. 7283 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7284 do_post_free_or_garbage_chunk(fc, size); 7285 7286 assert(_limit >= addr + size, 7287 "A freshly garbage chunk can't possibly straddle over _limit"); 7288 if (inFreeRange()) lookahead_and_flush(fc, size); 7289 return size; 7290 } 7291 7292 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7293 HeapWord* addr = (HeapWord*) fc; 7294 // The sweeper has just found a live object. Return any accumulated 7295 // left hand chunk to the free lists. 7296 if (inFreeRange()) { 7297 assert(freeFinger() < addr, "freeFinger points too high"); 7298 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7299 } 7300 7301 // This object is live: we'd normally expect this to be 7302 // an oop, and like to assert the following: 7303 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7304 // However, as we commented above, this may be an object whose 7305 // header hasn't yet been initialized. 7306 size_t size; 7307 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7308 if (_bitMap->isMarked(addr + 1)) { 7309 // Determine the size from the bit map, rather than trying to 7310 // compute it from the object header. 7311 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7312 size = pointer_delta(nextOneAddr + 1, addr); 7313 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7314 "alignment problem"); 7315 7316 #ifdef ASSERT 7317 if (oop(addr)->klass_or_null_acquire() != NULL) { 7318 // Ignore mark word because we are running concurrent with mutators 7319 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7320 assert(size == 7321 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7322 "P-mark and computed size do not agree"); 7323 } 7324 #endif 7325 7326 } else { 7327 // This should be an initialized object that's alive. 7328 assert(oop(addr)->klass_or_null_acquire() != NULL, 7329 "Should be an initialized object"); 7330 // Ignore mark word because we are running concurrent with mutators 7331 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7332 // Verify that the bit map has no bits marked between 7333 // addr and purported end of this block. 7334 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7335 assert(size >= 3, "Necessary for Printezis marks to work"); 7336 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7337 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7338 } 7339 return size; 7340 } 7341 7342 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7343 size_t chunkSize) { 7344 // do_post_free_or_garbage_chunk() should only be called in the case 7345 // of the adaptive free list allocator. 7346 const bool fcInFreeLists = fc->is_free(); 7347 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7348 if (CMSTestInFreeList && fcInFreeLists) { 7349 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7350 } 7351 7352 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7353 7354 HeapWord* const fc_addr = (HeapWord*) fc; 7355 7356 bool coalesce = false; 7357 const size_t left = pointer_delta(fc_addr, freeFinger()); 7358 const size_t right = chunkSize; 7359 switch (FLSCoalescePolicy) { 7360 // numeric value forms a coalition aggressiveness metric 7361 case 0: { // never coalesce 7362 coalesce = false; 7363 break; 7364 } 7365 case 1: { // coalesce if left & right chunks on overpopulated lists 7366 coalesce = _sp->coalOverPopulated(left) && 7367 _sp->coalOverPopulated(right); 7368 break; 7369 } 7370 case 2: { // coalesce if left chunk on overpopulated list (default) 7371 coalesce = _sp->coalOverPopulated(left); 7372 break; 7373 } 7374 case 3: { // coalesce if left OR right chunk on overpopulated list 7375 coalesce = _sp->coalOverPopulated(left) || 7376 _sp->coalOverPopulated(right); 7377 break; 7378 } 7379 case 4: { // always coalesce 7380 coalesce = true; 7381 break; 7382 } 7383 default: 7384 ShouldNotReachHere(); 7385 } 7386 7387 // Should the current free range be coalesced? 7388 // If the chunk is in a free range and either we decided to coalesce above 7389 // or the chunk is near the large block at the end of the heap 7390 // (isNearLargestChunk() returns true), then coalesce this chunk. 7391 const bool doCoalesce = inFreeRange() 7392 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7393 if (doCoalesce) { 7394 // Coalesce the current free range on the left with the new 7395 // chunk on the right. If either is on a free list, 7396 // it must be removed from the list and stashed in the closure. 7397 if (freeRangeInFreeLists()) { 7398 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7399 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7400 "Size of free range is inconsistent with chunk size."); 7401 if (CMSTestInFreeList) { 7402 assert(_sp->verify_chunk_in_free_list(ffc), 7403 "Chunk is not in free lists"); 7404 } 7405 _sp->coalDeath(ffc->size()); 7406 _sp->removeFreeChunkFromFreeLists(ffc); 7407 set_freeRangeInFreeLists(false); 7408 } 7409 if (fcInFreeLists) { 7410 _sp->coalDeath(chunkSize); 7411 assert(fc->size() == chunkSize, 7412 "The chunk has the wrong size or is not in the free lists"); 7413 _sp->removeFreeChunkFromFreeLists(fc); 7414 } 7415 set_lastFreeRangeCoalesced(true); 7416 print_free_block_coalesced(fc); 7417 } else { // not in a free range and/or should not coalesce 7418 // Return the current free range and start a new one. 7419 if (inFreeRange()) { 7420 // In a free range but cannot coalesce with the right hand chunk. 7421 // Put the current free range into the free lists. 7422 flush_cur_free_chunk(freeFinger(), 7423 pointer_delta(fc_addr, freeFinger())); 7424 } 7425 // Set up for new free range. Pass along whether the right hand 7426 // chunk is in the free lists. 7427 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7428 } 7429 } 7430 7431 // Lookahead flush: 7432 // If we are tracking a free range, and this is the last chunk that 7433 // we'll look at because its end crosses past _limit, we'll preemptively 7434 // flush it along with any free range we may be holding on to. Note that 7435 // this can be the case only for an already free or freshly garbage 7436 // chunk. If this block is an object, it can never straddle 7437 // over _limit. The "straddling" occurs when _limit is set at 7438 // the previous end of the space when this cycle started, and 7439 // a subsequent heap expansion caused the previously co-terminal 7440 // free block to be coalesced with the newly expanded portion, 7441 // thus rendering _limit a non-block-boundary making it dangerous 7442 // for the sweeper to step over and examine. 7443 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7444 assert(inFreeRange(), "Should only be called if currently in a free range."); 7445 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7446 assert(_sp->used_region().contains(eob - 1), 7447 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7448 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7449 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7450 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7451 if (eob >= _limit) { 7452 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7453 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7454 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7455 "[" PTR_FORMAT "," PTR_FORMAT ")", 7456 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7457 // Return the storage we are tracking back into the free lists. 7458 log_develop_trace(gc, sweep)("Flushing ... "); 7459 assert(freeFinger() < eob, "Error"); 7460 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7461 } 7462 } 7463 7464 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7465 assert(inFreeRange(), "Should only be called if currently in a free range."); 7466 assert(size > 0, 7467 "A zero sized chunk cannot be added to the free lists."); 7468 if (!freeRangeInFreeLists()) { 7469 if (CMSTestInFreeList) { 7470 FreeChunk* fc = (FreeChunk*) chunk; 7471 fc->set_size(size); 7472 assert(!_sp->verify_chunk_in_free_list(fc), 7473 "chunk should not be in free lists yet"); 7474 } 7475 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7476 // A new free range is going to be starting. The current 7477 // free range has not been added to the free lists yet or 7478 // was removed so add it back. 7479 // If the current free range was coalesced, then the death 7480 // of the free range was recorded. Record a birth now. 7481 if (lastFreeRangeCoalesced()) { 7482 _sp->coalBirth(size); 7483 } 7484 _sp->addChunkAndRepairOffsetTable(chunk, size, 7485 lastFreeRangeCoalesced()); 7486 } else { 7487 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7488 } 7489 set_inFreeRange(false); 7490 set_freeRangeInFreeLists(false); 7491 } 7492 7493 // We take a break if we've been at this for a while, 7494 // so as to avoid monopolizing the locks involved. 7495 void SweepClosure::do_yield_work(HeapWord* addr) { 7496 // Return current free chunk being used for coalescing (if any) 7497 // to the appropriate freelist. After yielding, the next 7498 // free block encountered will start a coalescing range of 7499 // free blocks. If the next free block is adjacent to the 7500 // chunk just flushed, they will need to wait for the next 7501 // sweep to be coalesced. 7502 if (inFreeRange()) { 7503 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7504 } 7505 7506 // First give up the locks, then yield, then re-lock. 7507 // We should probably use a constructor/destructor idiom to 7508 // do this unlock/lock or modify the MutexUnlocker class to 7509 // serve our purpose. XXX 7510 assert_lock_strong(_bitMap->lock()); 7511 assert_lock_strong(_freelistLock); 7512 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7513 "CMS thread should hold CMS token"); 7514 _bitMap->lock()->unlock(); 7515 _freelistLock->unlock(); 7516 ConcurrentMarkSweepThread::desynchronize(true); 7517 _collector->stopTimer(); 7518 _collector->incrementYields(); 7519 7520 // See the comment in coordinator_yield() 7521 for (unsigned i = 0; i < CMSYieldSleepCount && 7522 ConcurrentMarkSweepThread::should_yield() && 7523 !CMSCollector::foregroundGCIsActive(); ++i) { 7524 os::sleep(Thread::current(), 1, false); 7525 } 7526 7527 ConcurrentMarkSweepThread::synchronize(true); 7528 _freelistLock->lock(); 7529 _bitMap->lock()->lock_without_safepoint_check(); 7530 _collector->startTimer(); 7531 } 7532 7533 #ifndef PRODUCT 7534 // This is actually very useful in a product build if it can 7535 // be called from the debugger. Compile it into the product 7536 // as needed. 7537 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7538 return debug_cms_space->verify_chunk_in_free_list(fc); 7539 } 7540 #endif 7541 7542 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7543 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7544 p2i(fc), fc->size()); 7545 } 7546 7547 // CMSIsAliveClosure 7548 bool CMSIsAliveClosure::do_object_b(oop obj) { 7549 HeapWord* addr = (HeapWord*)obj; 7550 return addr != NULL && 7551 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7552 } 7553 7554 7555 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7556 MemRegion span, 7557 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7558 bool cpc): 7559 _collector(collector), 7560 _span(span), 7561 _bit_map(bit_map), 7562 _mark_stack(mark_stack), 7563 _concurrent_precleaning(cpc) { 7564 assert(!_span.is_empty(), "Empty span could spell trouble"); 7565 } 7566 7567 7568 // CMSKeepAliveClosure: the serial version 7569 void CMSKeepAliveClosure::do_oop(oop obj) { 7570 HeapWord* addr = (HeapWord*)obj; 7571 if (_span.contains(addr) && 7572 !_bit_map->isMarked(addr)) { 7573 _bit_map->mark(addr); 7574 bool simulate_overflow = false; 7575 NOT_PRODUCT( 7576 if (CMSMarkStackOverflowALot && 7577 _collector->simulate_overflow()) { 7578 // simulate a stack overflow 7579 simulate_overflow = true; 7580 } 7581 ) 7582 if (simulate_overflow || !_mark_stack->push(obj)) { 7583 if (_concurrent_precleaning) { 7584 // We dirty the overflown object and let the remark 7585 // phase deal with it. 7586 assert(_collector->overflow_list_is_empty(), "Error"); 7587 // In the case of object arrays, we need to dirty all of 7588 // the cards that the object spans. No locking or atomics 7589 // are needed since no one else can be mutating the mod union 7590 // table. 7591 if (obj->is_objArray()) { 7592 size_t sz = obj->size(); 7593 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7594 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7595 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7596 _collector->_modUnionTable.mark_range(redirty_range); 7597 } else { 7598 _collector->_modUnionTable.mark(addr); 7599 } 7600 _collector->_ser_kac_preclean_ovflw++; 7601 } else { 7602 _collector->push_on_overflow_list(obj); 7603 _collector->_ser_kac_ovflw++; 7604 } 7605 } 7606 } 7607 } 7608 7609 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7610 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } 7611 7612 // CMSParKeepAliveClosure: a parallel version of the above. 7613 // The work queues are private to each closure (thread), 7614 // but (may be) available for stealing by other threads. 7615 void CMSParKeepAliveClosure::do_oop(oop obj) { 7616 HeapWord* addr = (HeapWord*)obj; 7617 if (_span.contains(addr) && 7618 !_bit_map->isMarked(addr)) { 7619 // In general, during recursive tracing, several threads 7620 // may be concurrently getting here; the first one to 7621 // "tag" it, claims it. 7622 if (_bit_map->par_mark(addr)) { 7623 bool res = _work_queue->push(obj); 7624 assert(res, "Low water mark should be much less than capacity"); 7625 // Do a recursive trim in the hope that this will keep 7626 // stack usage lower, but leave some oops for potential stealers 7627 trim_queue(_low_water_mark); 7628 } // Else, another thread got there first 7629 } 7630 } 7631 7632 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7633 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } 7634 7635 void CMSParKeepAliveClosure::trim_queue(uint max) { 7636 while (_work_queue->size() > max) { 7637 oop new_oop; 7638 if (_work_queue->pop_local(new_oop)) { 7639 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7640 assert(_bit_map->isMarked((HeapWord*)new_oop), 7641 "no white objects on this stack!"); 7642 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7643 // iterate over the oops in this oop, marking and pushing 7644 // the ones in CMS heap (i.e. in _span). 7645 new_oop->oop_iterate(&_mark_and_push); 7646 } 7647 } 7648 } 7649 7650 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7651 CMSCollector* collector, 7652 MemRegion span, CMSBitMap* bit_map, 7653 OopTaskQueue* work_queue): 7654 _collector(collector), 7655 _span(span), 7656 _bit_map(bit_map), 7657 _work_queue(work_queue) { } 7658 7659 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7660 HeapWord* addr = (HeapWord*)obj; 7661 if (_span.contains(addr) && 7662 !_bit_map->isMarked(addr)) { 7663 if (_bit_map->par_mark(addr)) { 7664 bool simulate_overflow = false; 7665 NOT_PRODUCT( 7666 if (CMSMarkStackOverflowALot && 7667 _collector->par_simulate_overflow()) { 7668 // simulate a stack overflow 7669 simulate_overflow = true; 7670 } 7671 ) 7672 if (simulate_overflow || !_work_queue->push(obj)) { 7673 _collector->par_push_on_overflow_list(obj); 7674 _collector->_par_kac_ovflw++; 7675 } 7676 } // Else another thread got there already 7677 } 7678 } 7679 7680 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7681 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } 7682 7683 ////////////////////////////////////////////////////////////////// 7684 // CMSExpansionCause ///////////////////////////// 7685 ////////////////////////////////////////////////////////////////// 7686 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7687 switch (cause) { 7688 case _no_expansion: 7689 return "No expansion"; 7690 case _satisfy_free_ratio: 7691 return "Free ratio"; 7692 case _satisfy_promotion: 7693 return "Satisfy promotion"; 7694 case _satisfy_allocation: 7695 return "allocation"; 7696 case _allocate_par_lab: 7697 return "Par LAB"; 7698 case _allocate_par_spooling_space: 7699 return "Par Spooling Space"; 7700 case _adaptive_size_policy: 7701 return "Ergonomics"; 7702 default: 7703 return "unknown"; 7704 } 7705 } 7706 7707 void CMSDrainMarkingStackClosure::do_void() { 7708 // the max number to take from overflow list at a time 7709 const size_t num = _mark_stack->capacity()/4; 7710 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7711 "Overflow list should be NULL during concurrent phases"); 7712 while (!_mark_stack->isEmpty() || 7713 // if stack is empty, check the overflow list 7714 _collector->take_from_overflow_list(num, _mark_stack)) { 7715 oop obj = _mark_stack->pop(); 7716 HeapWord* addr = (HeapWord*)obj; 7717 assert(_span.contains(addr), "Should be within span"); 7718 assert(_bit_map->isMarked(addr), "Should be marked"); 7719 assert(oopDesc::is_oop(obj), "Should be an oop"); 7720 obj->oop_iterate(_keep_alive); 7721 } 7722 } 7723 7724 void CMSParDrainMarkingStackClosure::do_void() { 7725 // drain queue 7726 trim_queue(0); 7727 } 7728 7729 // Trim our work_queue so its length is below max at return 7730 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7731 while (_work_queue->size() > max) { 7732 oop new_oop; 7733 if (_work_queue->pop_local(new_oop)) { 7734 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7735 assert(_bit_map->isMarked((HeapWord*)new_oop), 7736 "no white objects on this stack!"); 7737 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7738 // iterate over the oops in this oop, marking and pushing 7739 // the ones in CMS heap (i.e. in _span). 7740 new_oop->oop_iterate(&_mark_and_push); 7741 } 7742 } 7743 } 7744 7745 //////////////////////////////////////////////////////////////////// 7746 // Support for Marking Stack Overflow list handling and related code 7747 //////////////////////////////////////////////////////////////////// 7748 // Much of the following code is similar in shape and spirit to the 7749 // code used in ParNewGC. We should try and share that code 7750 // as much as possible in the future. 7751 7752 #ifndef PRODUCT 7753 // Debugging support for CMSStackOverflowALot 7754 7755 // It's OK to call this multi-threaded; the worst thing 7756 // that can happen is that we'll get a bunch of closely 7757 // spaced simulated overflows, but that's OK, in fact 7758 // probably good as it would exercise the overflow code 7759 // under contention. 7760 bool CMSCollector::simulate_overflow() { 7761 if (_overflow_counter-- <= 0) { // just being defensive 7762 _overflow_counter = CMSMarkStackOverflowInterval; 7763 return true; 7764 } else { 7765 return false; 7766 } 7767 } 7768 7769 bool CMSCollector::par_simulate_overflow() { 7770 return simulate_overflow(); 7771 } 7772 #endif 7773 7774 // Single-threaded 7775 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7776 assert(stack->isEmpty(), "Expected precondition"); 7777 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7778 size_t i = num; 7779 oop cur = _overflow_list; 7780 const markOop proto = markOopDesc::prototype(); 7781 NOT_PRODUCT(ssize_t n = 0;) 7782 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7783 next = oop(cur->mark_raw()); 7784 cur->set_mark_raw(proto); // until proven otherwise 7785 assert(oopDesc::is_oop(cur), "Should be an oop"); 7786 bool res = stack->push(cur); 7787 assert(res, "Bit off more than can chew?"); 7788 NOT_PRODUCT(n++;) 7789 } 7790 _overflow_list = cur; 7791 #ifndef PRODUCT 7792 assert(_num_par_pushes >= n, "Too many pops?"); 7793 _num_par_pushes -=n; 7794 #endif 7795 return !stack->isEmpty(); 7796 } 7797 7798 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7799 // (MT-safe) Get a prefix of at most "num" from the list. 7800 // The overflow list is chained through the mark word of 7801 // each object in the list. We fetch the entire list, 7802 // break off a prefix of the right size and return the 7803 // remainder. If other threads try to take objects from 7804 // the overflow list at that time, they will wait for 7805 // some time to see if data becomes available. If (and 7806 // only if) another thread places one or more object(s) 7807 // on the global list before we have returned the suffix 7808 // to the global list, we will walk down our local list 7809 // to find its end and append the global list to 7810 // our suffix before returning it. This suffix walk can 7811 // prove to be expensive (quadratic in the amount of traffic) 7812 // when there are many objects in the overflow list and 7813 // there is much producer-consumer contention on the list. 7814 // *NOTE*: The overflow list manipulation code here and 7815 // in ParNewGeneration:: are very similar in shape, 7816 // except that in the ParNew case we use the old (from/eden) 7817 // copy of the object to thread the list via its klass word. 7818 // Because of the common code, if you make any changes in 7819 // the code below, please check the ParNew version to see if 7820 // similar changes might be needed. 7821 // CR 6797058 has been filed to consolidate the common code. 7822 bool CMSCollector::par_take_from_overflow_list(size_t num, 7823 OopTaskQueue* work_q, 7824 int no_of_gc_threads) { 7825 assert(work_q->size() == 0, "First empty local work queue"); 7826 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7827 if (_overflow_list == NULL) { 7828 return false; 7829 } 7830 // Grab the entire list; we'll put back a suffix 7831 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7832 Thread* tid = Thread::current(); 7833 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7834 // set to ParallelGCThreads. 7835 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7836 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7837 // If the list is busy, we spin for a short while, 7838 // sleeping between attempts to get the list. 7839 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7840 os::sleep(tid, sleep_time_millis, false); 7841 if (_overflow_list == NULL) { 7842 // Nothing left to take 7843 return false; 7844 } else if (_overflow_list != BUSY) { 7845 // Try and grab the prefix 7846 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7847 } 7848 } 7849 // If the list was found to be empty, or we spun long 7850 // enough, we give up and return empty-handed. If we leave 7851 // the list in the BUSY state below, it must be the case that 7852 // some other thread holds the overflow list and will set it 7853 // to a non-BUSY state in the future. 7854 if (prefix == NULL || prefix == BUSY) { 7855 // Nothing to take or waited long enough 7856 if (prefix == NULL) { 7857 // Write back the NULL in case we overwrote it with BUSY above 7858 // and it is still the same value. 7859 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7860 } 7861 return false; 7862 } 7863 assert(prefix != NULL && prefix != BUSY, "Error"); 7864 size_t i = num; 7865 oop cur = prefix; 7866 // Walk down the first "num" objects, unless we reach the end. 7867 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7868 if (cur->mark_raw() == NULL) { 7869 // We have "num" or fewer elements in the list, so there 7870 // is nothing to return to the global list. 7871 // Write back the NULL in lieu of the BUSY we wrote 7872 // above, if it is still the same value. 7873 if (_overflow_list == BUSY) { 7874 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7875 } 7876 } else { 7877 // Chop off the suffix and return it to the global list. 7878 assert(cur->mark_raw() != BUSY, "Error"); 7879 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7880 cur->set_mark_raw(NULL); // break off suffix 7881 // It's possible that the list is still in the empty(busy) state 7882 // we left it in a short while ago; in that case we may be 7883 // able to place back the suffix without incurring the cost 7884 // of a walk down the list. 7885 oop observed_overflow_list = _overflow_list; 7886 oop cur_overflow_list = observed_overflow_list; 7887 bool attached = false; 7888 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7889 observed_overflow_list = 7890 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7891 if (cur_overflow_list == observed_overflow_list) { 7892 attached = true; 7893 break; 7894 } else cur_overflow_list = observed_overflow_list; 7895 } 7896 if (!attached) { 7897 // Too bad, someone else sneaked in (at least) an element; we'll need 7898 // to do a splice. Find tail of suffix so we can prepend suffix to global 7899 // list. 7900 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7901 oop suffix_tail = cur; 7902 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7903 "Tautology"); 7904 observed_overflow_list = _overflow_list; 7905 do { 7906 cur_overflow_list = observed_overflow_list; 7907 if (cur_overflow_list != BUSY) { 7908 // Do the splice ... 7909 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7910 } else { // cur_overflow_list == BUSY 7911 suffix_tail->set_mark_raw(NULL); 7912 } 7913 // ... and try to place spliced list back on overflow_list ... 7914 observed_overflow_list = 7915 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7916 } while (cur_overflow_list != observed_overflow_list); 7917 // ... until we have succeeded in doing so. 7918 } 7919 } 7920 7921 // Push the prefix elements on work_q 7922 assert(prefix != NULL, "control point invariant"); 7923 const markOop proto = markOopDesc::prototype(); 7924 oop next; 7925 NOT_PRODUCT(ssize_t n = 0;) 7926 for (cur = prefix; cur != NULL; cur = next) { 7927 next = oop(cur->mark_raw()); 7928 cur->set_mark_raw(proto); // until proven otherwise 7929 assert(oopDesc::is_oop(cur), "Should be an oop"); 7930 bool res = work_q->push(cur); 7931 assert(res, "Bit off more than we can chew?"); 7932 NOT_PRODUCT(n++;) 7933 } 7934 #ifndef PRODUCT 7935 assert(_num_par_pushes >= n, "Too many pops?"); 7936 Atomic::sub(n, &_num_par_pushes); 7937 #endif 7938 return true; 7939 } 7940 7941 // Single-threaded 7942 void CMSCollector::push_on_overflow_list(oop p) { 7943 NOT_PRODUCT(_num_par_pushes++;) 7944 assert(oopDesc::is_oop(p), "Not an oop"); 7945 preserve_mark_if_necessary(p); 7946 p->set_mark_raw((markOop)_overflow_list); 7947 _overflow_list = p; 7948 } 7949 7950 // Multi-threaded; use CAS to prepend to overflow list 7951 void CMSCollector::par_push_on_overflow_list(oop p) { 7952 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7953 assert(oopDesc::is_oop(p), "Not an oop"); 7954 par_preserve_mark_if_necessary(p); 7955 oop observed_overflow_list = _overflow_list; 7956 oop cur_overflow_list; 7957 do { 7958 cur_overflow_list = observed_overflow_list; 7959 if (cur_overflow_list != BUSY) { 7960 p->set_mark_raw(markOop(cur_overflow_list)); 7961 } else { 7962 p->set_mark_raw(NULL); 7963 } 7964 observed_overflow_list = 7965 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7966 } while (cur_overflow_list != observed_overflow_list); 7967 } 7968 #undef BUSY 7969 7970 // Single threaded 7971 // General Note on GrowableArray: pushes may silently fail 7972 // because we are (temporarily) out of C-heap for expanding 7973 // the stack. The problem is quite ubiquitous and affects 7974 // a lot of code in the JVM. The prudent thing for GrowableArray 7975 // to do (for now) is to exit with an error. However, that may 7976 // be too draconian in some cases because the caller may be 7977 // able to recover without much harm. For such cases, we 7978 // should probably introduce a "soft_push" method which returns 7979 // an indication of success or failure with the assumption that 7980 // the caller may be able to recover from a failure; code in 7981 // the VM can then be changed, incrementally, to deal with such 7982 // failures where possible, thus, incrementally hardening the VM 7983 // in such low resource situations. 7984 void CMSCollector::preserve_mark_work(oop p, markOop m) { 7985 _preserved_oop_stack.push(p); 7986 _preserved_mark_stack.push(m); 7987 assert(m == p->mark_raw(), "Mark word changed"); 7988 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7989 "bijection"); 7990 } 7991 7992 // Single threaded 7993 void CMSCollector::preserve_mark_if_necessary(oop p) { 7994 markOop m = p->mark_raw(); 7995 if (m->must_be_preserved(p)) { 7996 preserve_mark_work(p, m); 7997 } 7998 } 7999 8000 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 8001 markOop m = p->mark_raw(); 8002 if (m->must_be_preserved(p)) { 8003 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 8004 // Even though we read the mark word without holding 8005 // the lock, we are assured that it will not change 8006 // because we "own" this oop, so no other thread can 8007 // be trying to push it on the overflow list; see 8008 // the assertion in preserve_mark_work() that checks 8009 // that m == p->mark_raw(). 8010 preserve_mark_work(p, m); 8011 } 8012 } 8013 8014 // We should be able to do this multi-threaded, 8015 // a chunk of stack being a task (this is 8016 // correct because each oop only ever appears 8017 // once in the overflow list. However, it's 8018 // not very easy to completely overlap this with 8019 // other operations, so will generally not be done 8020 // until all work's been completed. Because we 8021 // expect the preserved oop stack (set) to be small, 8022 // it's probably fine to do this single-threaded. 8023 // We can explore cleverer concurrent/overlapped/parallel 8024 // processing of preserved marks if we feel the 8025 // need for this in the future. Stack overflow should 8026 // be so rare in practice and, when it happens, its 8027 // effect on performance so great that this will 8028 // likely just be in the noise anyway. 8029 void CMSCollector::restore_preserved_marks_if_any() { 8030 assert(SafepointSynchronize::is_at_safepoint(), 8031 "world should be stopped"); 8032 assert(Thread::current()->is_ConcurrentGC_thread() || 8033 Thread::current()->is_VM_thread(), 8034 "should be single-threaded"); 8035 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 8036 "bijection"); 8037 8038 while (!_preserved_oop_stack.is_empty()) { 8039 oop p = _preserved_oop_stack.pop(); 8040 assert(oopDesc::is_oop(p), "Should be an oop"); 8041 assert(_span.contains(p), "oop should be in _span"); 8042 assert(p->mark_raw() == markOopDesc::prototype(), 8043 "Set when taken from overflow list"); 8044 markOop m = _preserved_mark_stack.pop(); 8045 p->set_mark_raw(m); 8046 } 8047 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8048 "stacks were cleared above"); 8049 } 8050 8051 #ifndef PRODUCT 8052 bool CMSCollector::no_preserved_marks() const { 8053 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8054 } 8055 #endif 8056 8057 // Transfer some number of overflown objects to usual marking 8058 // stack. Return true if some objects were transferred. 8059 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8060 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8061 (size_t)ParGCDesiredObjsFromOverflowList); 8062 8063 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8064 assert(_collector->overflow_list_is_empty() || res, 8065 "If list is not empty, we should have taken something"); 8066 assert(!res || !_mark_stack->isEmpty(), 8067 "If we took something, it should now be on our stack"); 8068 return res; 8069 } 8070 8071 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8072 size_t res = _sp->block_size_no_stall(addr, _collector); 8073 if (_sp->block_is_obj(addr)) { 8074 if (_live_bit_map->isMarked(addr)) { 8075 // It can't have been dead in a previous cycle 8076 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8077 } else { 8078 _dead_bit_map->mark(addr); // mark the dead object 8079 } 8080 } 8081 // Could be 0, if the block size could not be computed without stalling. 8082 return res; 8083 } 8084 8085 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8086 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8087 switch (phase) { 8088 case CMSCollector::InitialMarking: 8089 initialize(manager /* GC manager */ , 8090 cause /* cause of the GC */, 8091 true /* allMemoryPoolsAffected */, 8092 true /* recordGCBeginTime */, 8093 true /* recordPreGCUsage */, 8094 false /* recordPeakUsage */, 8095 false /* recordPostGCusage */, 8096 true /* recordAccumulatedGCTime */, 8097 false /* recordGCEndTime */, 8098 false /* countCollection */ ); 8099 break; 8100 8101 case CMSCollector::FinalMarking: 8102 initialize(manager /* GC manager */ , 8103 cause /* cause of the GC */, 8104 true /* allMemoryPoolsAffected */, 8105 false /* recordGCBeginTime */, 8106 false /* recordPreGCUsage */, 8107 false /* recordPeakUsage */, 8108 false /* recordPostGCusage */, 8109 true /* recordAccumulatedGCTime */, 8110 false /* recordGCEndTime */, 8111 false /* countCollection */ ); 8112 break; 8113 8114 case CMSCollector::Sweeping: 8115 initialize(manager /* GC manager */ , 8116 cause /* cause of the GC */, 8117 true /* allMemoryPoolsAffected */, 8118 false /* recordGCBeginTime */, 8119 false /* recordPreGCUsage */, 8120 true /* recordPeakUsage */, 8121 true /* recordPostGCusage */, 8122 false /* recordAccumulatedGCTime */, 8123 true /* recordGCEndTime */, 8124 true /* countCollection */ ); 8125 break; 8126 8127 default: 8128 ShouldNotReachHere(); 8129 } 8130 }