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/classLoaderDataGraph.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 "gc/shared/workerPolicy.hpp" 65 #include "logging/log.hpp" 66 #include "logging/logStream.hpp" 67 #include "memory/allocation.hpp" 68 #include "memory/binaryTreeDictionary.inline.hpp" 69 #include "memory/iterator.inline.hpp" 70 #include "memory/padded.hpp" 71 #include "memory/resourceArea.hpp" 72 #include "oops/access.inline.hpp" 73 #include "oops/oop.inline.hpp" 74 #include "prims/jvmtiExport.hpp" 75 #include "runtime/atomic.hpp" 76 #include "runtime/flags/flagSetting.hpp" 77 #include "runtime/globals_extension.hpp" 78 #include "runtime/handles.inline.hpp" 79 #include "runtime/java.hpp" 80 #include "runtime/orderAccess.hpp" 81 #include "runtime/timer.hpp" 82 #include "runtime/vmThread.hpp" 83 #include "services/memoryService.hpp" 84 #include "services/runtimeService.hpp" 85 #include "utilities/align.hpp" 86 #include "utilities/stack.inline.hpp" 87 88 // statics 89 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 90 bool CMSCollector::_full_gc_requested = false; 91 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 92 93 ////////////////////////////////////////////////////////////////// 94 // In support of CMS/VM thread synchronization 95 ////////////////////////////////////////////////////////////////// 96 // We split use of the CGC_lock into 2 "levels". 97 // The low-level locking is of the usual CGC_lock monitor. We introduce 98 // a higher level "token" (hereafter "CMS token") built on top of the 99 // low level monitor (hereafter "CGC lock"). 100 // The token-passing protocol gives priority to the VM thread. The 101 // CMS-lock doesn't provide any fairness guarantees, but clients 102 // should ensure that it is only held for very short, bounded 103 // durations. 104 // 105 // When either of the CMS thread or the VM thread is involved in 106 // collection operations during which it does not want the other 107 // thread to interfere, it obtains the CMS token. 108 // 109 // If either thread tries to get the token while the other has 110 // it, that thread waits. However, if the VM thread and CMS thread 111 // both want the token, then the VM thread gets priority while the 112 // CMS thread waits. This ensures, for instance, that the "concurrent" 113 // phases of the CMS thread's work do not block out the VM thread 114 // for long periods of time as the CMS thread continues to hog 115 // the token. (See bug 4616232). 116 // 117 // The baton-passing functions are, however, controlled by the 118 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 119 // and here the low-level CMS lock, not the high level token, 120 // ensures mutual exclusion. 121 // 122 // Two important conditions that we have to satisfy: 123 // 1. if a thread does a low-level wait on the CMS lock, then it 124 // relinquishes the CMS token if it were holding that token 125 // when it acquired the low-level CMS lock. 126 // 2. any low-level notifications on the low-level lock 127 // should only be sent when a thread has relinquished the token. 128 // 129 // In the absence of either property, we'd have potential deadlock. 130 // 131 // We protect each of the CMS (concurrent and sequential) phases 132 // with the CMS _token_, not the CMS _lock_. 133 // 134 // The only code protected by CMS lock is the token acquisition code 135 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 136 // baton-passing code. 137 // 138 // Unfortunately, i couldn't come up with a good abstraction to factor and 139 // hide the naked CGC_lock manipulation in the baton-passing code 140 // further below. That's something we should try to do. Also, the proof 141 // of correctness of this 2-level locking scheme is far from obvious, 142 // and potentially quite slippery. We have an uneasy suspicion, for instance, 143 // that there may be a theoretical possibility of delay/starvation in the 144 // low-level lock/wait/notify scheme used for the baton-passing because of 145 // potential interference with the priority scheme embodied in the 146 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 147 // invocation further below and marked with "XXX 20011219YSR". 148 // Indeed, as we note elsewhere, this may become yet more slippery 149 // in the presence of multiple CMS and/or multiple VM threads. XXX 150 151 class CMSTokenSync: public StackObj { 152 private: 153 bool _is_cms_thread; 154 public: 155 CMSTokenSync(bool is_cms_thread): 156 _is_cms_thread(is_cms_thread) { 157 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 158 "Incorrect argument to constructor"); 159 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 160 } 161 162 ~CMSTokenSync() { 163 assert(_is_cms_thread ? 164 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 165 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 166 "Incorrect state"); 167 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 168 } 169 }; 170 171 // Convenience class that does a CMSTokenSync, and then acquires 172 // upto three locks. 173 class CMSTokenSyncWithLocks: public CMSTokenSync { 174 private: 175 // Note: locks are acquired in textual declaration order 176 // and released in the opposite order 177 MutexLockerEx _locker1, _locker2, _locker3; 178 public: 179 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 180 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 181 CMSTokenSync(is_cms_thread), 182 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 183 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 184 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 185 { } 186 }; 187 188 189 ////////////////////////////////////////////////////////////////// 190 // Concurrent Mark-Sweep Generation ///////////////////////////// 191 ////////////////////////////////////////////////////////////////// 192 193 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 194 195 // This struct contains per-thread things necessary to support parallel 196 // young-gen collection. 197 class CMSParGCThreadState: public CHeapObj<mtGC> { 198 public: 199 CompactibleFreeListSpaceLAB lab; 200 PromotionInfo promo; 201 202 // Constructor. 203 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 204 promo.setSpace(cfls); 205 } 206 }; 207 208 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 209 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 210 CardGeneration(rs, initial_byte_size, ct), 211 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 212 _did_compact(false) 213 { 214 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 215 HeapWord* end = (HeapWord*) _virtual_space.high(); 216 217 _direct_allocated_words = 0; 218 NOT_PRODUCT( 219 _numObjectsPromoted = 0; 220 _numWordsPromoted = 0; 221 _numObjectsAllocated = 0; 222 _numWordsAllocated = 0; 223 ) 224 225 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 226 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 227 _cmsSpace->_old_gen = this; 228 229 _gc_stats = new CMSGCStats(); 230 231 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 232 // offsets match. The ability to tell free chunks from objects 233 // depends on this property. 234 debug_only( 235 FreeChunk* junk = NULL; 236 assert(UseCompressedClassPointers || 237 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 238 "Offset of FreeChunk::_prev within FreeChunk must match" 239 " that of OopDesc::_klass within OopDesc"); 240 ) 241 242 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 243 for (uint i = 0; i < ParallelGCThreads; i++) { 244 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 245 } 246 247 _incremental_collection_failed = false; 248 // The "dilatation_factor" is the expansion that can occur on 249 // account of the fact that the minimum object size in the CMS 250 // generation may be larger than that in, say, a contiguous young 251 // generation. 252 // Ideally, in the calculation below, we'd compute the dilatation 253 // factor as: MinChunkSize/(promoting_gen's min object size) 254 // Since we do not have such a general query interface for the 255 // promoting generation, we'll instead just use the minimum 256 // object size (which today is a header's worth of space); 257 // note that all arithmetic is in units of HeapWords. 258 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 259 assert(_dilatation_factor >= 1.0, "from previous assert"); 260 } 261 262 263 // The field "_initiating_occupancy" represents the occupancy percentage 264 // at which we trigger a new collection cycle. Unless explicitly specified 265 // via CMSInitiatingOccupancyFraction (argument "io" below), it 266 // is calculated by: 267 // 268 // Let "f" be MinHeapFreeRatio in 269 // 270 // _initiating_occupancy = 100-f + 271 // f * (CMSTriggerRatio/100) 272 // where CMSTriggerRatio is the argument "tr" below. 273 // 274 // That is, if we assume the heap is at its desired maximum occupancy at the 275 // end of a collection, we let CMSTriggerRatio of the (purported) free 276 // space be allocated before initiating a new collection cycle. 277 // 278 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 279 assert(io <= 100 && tr <= 100, "Check the arguments"); 280 if (io >= 0) { 281 _initiating_occupancy = (double)io / 100.0; 282 } else { 283 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 284 (double)(tr * MinHeapFreeRatio) / 100.0) 285 / 100.0; 286 } 287 } 288 289 void ConcurrentMarkSweepGeneration::ref_processor_init() { 290 assert(collector() != NULL, "no collector"); 291 collector()->ref_processor_init(); 292 } 293 294 void CMSCollector::ref_processor_init() { 295 if (_ref_processor == NULL) { 296 // Allocate and initialize a reference processor 297 _ref_processor = 298 new ReferenceProcessor(&_span_based_discoverer, 299 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 300 ParallelGCThreads, // mt processing degree 301 _cmsGen->refs_discovery_is_mt(), // mt discovery 302 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 303 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 304 &_is_alive_closure, // closure for liveness info 305 false); // disable adjusting number of processing threads 306 // Initialize the _ref_processor field of CMSGen 307 _cmsGen->set_ref_processor(_ref_processor); 308 309 } 310 } 311 312 AdaptiveSizePolicy* CMSCollector::size_policy() { 313 return CMSHeap::heap()->size_policy(); 314 } 315 316 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 317 318 const char* gen_name = "old"; 319 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 320 // Generation Counters - generation 1, 1 subspace 321 _gen_counters = new GenerationCounters(gen_name, 1, 1, 322 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 323 324 _space_counters = new GSpaceCounters(gen_name, 0, 325 _virtual_space.reserved_size(), 326 this, _gen_counters); 327 } 328 329 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 330 _cms_gen(cms_gen) 331 { 332 assert(alpha <= 100, "bad value"); 333 _saved_alpha = alpha; 334 335 // Initialize the alphas to the bootstrap value of 100. 336 _gc0_alpha = _cms_alpha = 100; 337 338 _cms_begin_time.update(); 339 _cms_end_time.update(); 340 341 _gc0_duration = 0.0; 342 _gc0_period = 0.0; 343 _gc0_promoted = 0; 344 345 _cms_duration = 0.0; 346 _cms_period = 0.0; 347 _cms_allocated = 0; 348 349 _cms_used_at_gc0_begin = 0; 350 _cms_used_at_gc0_end = 0; 351 _allow_duty_cycle_reduction = false; 352 _valid_bits = 0; 353 } 354 355 double CMSStats::cms_free_adjustment_factor(size_t free) const { 356 // TBD: CR 6909490 357 return 1.0; 358 } 359 360 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 361 } 362 363 // If promotion failure handling is on use 364 // the padded average size of the promotion for each 365 // young generation collection. 366 double CMSStats::time_until_cms_gen_full() const { 367 size_t cms_free = _cms_gen->cmsSpace()->free(); 368 CMSHeap* heap = CMSHeap::heap(); 369 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 370 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 371 if (cms_free > expected_promotion) { 372 // Start a cms collection if there isn't enough space to promote 373 // for the next young collection. Use the padded average as 374 // a safety factor. 375 cms_free -= expected_promotion; 376 377 // Adjust by the safety factor. 378 double cms_free_dbl = (double)cms_free; 379 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 380 // Apply a further correction factor which tries to adjust 381 // for recent occurance of concurrent mode failures. 382 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 383 cms_free_dbl = cms_free_dbl * cms_adjustment; 384 385 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 386 cms_free, expected_promotion); 387 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 388 // Add 1 in case the consumption rate goes to zero. 389 return cms_free_dbl / (cms_consumption_rate() + 1.0); 390 } 391 return 0.0; 392 } 393 394 // Compare the duration of the cms collection to the 395 // time remaining before the cms generation is empty. 396 // Note that the time from the start of the cms collection 397 // to the start of the cms sweep (less than the total 398 // duration of the cms collection) can be used. This 399 // has been tried and some applications experienced 400 // promotion failures early in execution. This was 401 // possibly because the averages were not accurate 402 // enough at the beginning. 403 double CMSStats::time_until_cms_start() const { 404 // We add "gc0_period" to the "work" calculation 405 // below because this query is done (mostly) at the 406 // end of a scavenge, so we need to conservatively 407 // account for that much possible delay 408 // in the query so as to avoid concurrent mode failures 409 // due to starting the collection just a wee bit too 410 // late. 411 double work = cms_duration() + gc0_period(); 412 double deadline = time_until_cms_gen_full(); 413 // If a concurrent mode failure occurred recently, we want to be 414 // more conservative and halve our expected time_until_cms_gen_full() 415 if (work > deadline) { 416 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 417 cms_duration(), gc0_period(), time_until_cms_gen_full()); 418 return 0.0; 419 } 420 return work - deadline; 421 } 422 423 #ifndef PRODUCT 424 void CMSStats::print_on(outputStream *st) const { 425 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 426 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 427 gc0_duration(), gc0_period(), gc0_promoted()); 428 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 429 cms_duration(), cms_period(), cms_allocated()); 430 st->print(",cms_since_beg=%g,cms_since_end=%g", 431 cms_time_since_begin(), cms_time_since_end()); 432 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 433 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 434 435 if (valid()) { 436 st->print(",promo_rate=%g,cms_alloc_rate=%g", 437 promotion_rate(), cms_allocation_rate()); 438 st->print(",cms_consumption_rate=%g,time_until_full=%g", 439 cms_consumption_rate(), time_until_cms_gen_full()); 440 } 441 st->cr(); 442 } 443 #endif // #ifndef PRODUCT 444 445 CMSCollector::CollectorState CMSCollector::_collectorState = 446 CMSCollector::Idling; 447 bool CMSCollector::_foregroundGCIsActive = false; 448 bool CMSCollector::_foregroundGCShouldWait = false; 449 450 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 451 CardTableRS* ct, 452 ConcurrentMarkSweepPolicy* cp): 453 _overflow_list(NULL), 454 _conc_workers(NULL), // may be set later 455 _completed_initialization(false), 456 _collection_count_start(0), 457 _should_unload_classes(CMSClassUnloadingEnabled), 458 _concurrent_cycles_since_last_unload(0), 459 _roots_scanning_options(GenCollectedHeap::SO_None), 460 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 461 _verifying(false), 462 _collector_policy(cp), 463 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 464 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 465 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 466 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 467 _cms_start_registered(false), 468 _cmsGen(cmsGen), 469 // Adjust span to cover old (cms) gen 470 _span(cmsGen->reserved()), 471 _ct(ct), 472 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 473 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 474 -1 /* lock-free */, "No_lock" /* dummy */), 475 _restart_addr(NULL), 476 _ser_pmc_preclean_ovflw(0), 477 _ser_pmc_remark_ovflw(0), 478 _par_pmc_remark_ovflw(0), 479 _ser_kac_preclean_ovflw(0), 480 _ser_kac_ovflw(0), 481 _par_kac_ovflw(0), 482 #ifndef PRODUCT 483 _num_par_pushes(0), 484 #endif 485 _span_based_discoverer(_span), 486 _ref_processor(NULL), // will be set later 487 // Construct the is_alive_closure with _span & markBitMap 488 _is_alive_closure(_span, &_markBitMap), 489 _modUnionClosurePar(&_modUnionTable), 490 _between_prologue_and_epilogue(false), 491 _abort_preclean(false), 492 _start_sampling(false), 493 _stats(cmsGen), 494 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 495 //verify that this lock should be acquired with safepoint check. 496 Monitor::_safepoint_check_sometimes)), 497 _eden_chunk_array(NULL), // may be set in ctor body 498 _eden_chunk_index(0), // -- ditto -- 499 _eden_chunk_capacity(0), // -- ditto -- 500 _survivor_chunk_array(NULL), // -- ditto -- 501 _survivor_chunk_index(0), // -- ditto -- 502 _survivor_chunk_capacity(0), // -- ditto -- 503 _survivor_plab_array(NULL) // -- ditto -- 504 { 505 // Now expand the span and allocate the collection support structures 506 // (MUT, marking bit map etc.) to cover both generations subject to 507 // collection. 508 509 // For use by dirty card to oop closures. 510 _cmsGen->cmsSpace()->set_collector(this); 511 512 // Allocate MUT and marking bit map 513 { 514 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 515 if (!_markBitMap.allocate(_span)) { 516 log_warning(gc)("Failed to allocate CMS Bit Map"); 517 return; 518 } 519 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 520 } 521 { 522 _modUnionTable.allocate(_span); 523 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 524 } 525 526 if (!_markStack.allocate(MarkStackSize)) { 527 log_warning(gc)("Failed to allocate CMS Marking Stack"); 528 return; 529 } 530 531 // Support for multi-threaded concurrent phases 532 if (CMSConcurrentMTEnabled) { 533 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 534 // just for now 535 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 536 } 537 if (ConcGCThreads > 1) { 538 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 539 ConcGCThreads, true); 540 if (_conc_workers == NULL) { 541 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 542 CMSConcurrentMTEnabled = false; 543 } else { 544 _conc_workers->initialize_workers(); 545 } 546 } else { 547 CMSConcurrentMTEnabled = false; 548 } 549 } 550 if (!CMSConcurrentMTEnabled) { 551 ConcGCThreads = 0; 552 } else { 553 // Turn off CMSCleanOnEnter optimization temporarily for 554 // the MT case where it's not fixed yet; see 6178663. 555 CMSCleanOnEnter = false; 556 } 557 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 558 "Inconsistency"); 559 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 560 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 561 562 // Parallel task queues; these are shared for the 563 // concurrent and stop-world phases of CMS, but 564 // are not shared with parallel scavenge (ParNew). 565 { 566 uint i; 567 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 568 569 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 570 || ParallelRefProcEnabled) 571 && num_queues > 0) { 572 _task_queues = new OopTaskQueueSet(num_queues); 573 if (_task_queues == NULL) { 574 log_warning(gc)("task_queues allocation failure."); 575 return; 576 } 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 } 589 } 590 } 591 592 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 593 594 // Clip CMSBootstrapOccupancy between 0 and 100. 595 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 596 597 // Now tell CMS generations the identity of their collector 598 ConcurrentMarkSweepGeneration::set_collector(this); 599 600 // Create & start a CMS thread for this CMS collector 601 _cmsThread = ConcurrentMarkSweepThread::start(this); 602 assert(cmsThread() != NULL, "CMS Thread should have been created"); 603 assert(cmsThread()->collector() == this, 604 "CMS Thread should refer to this gen"); 605 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 606 607 // Support for parallelizing young gen rescan 608 CMSHeap* heap = CMSHeap::heap(); 609 _young_gen = heap->young_gen(); 610 if (heap->supports_inline_contig_alloc()) { 611 _top_addr = heap->top_addr(); 612 _end_addr = heap->end_addr(); 613 assert(_young_gen != NULL, "no _young_gen"); 614 _eden_chunk_index = 0; 615 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 616 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 617 } 618 619 // Support for parallelizing survivor space rescan 620 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 621 const size_t max_plab_samples = 622 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 623 624 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 625 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 626 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 627 _survivor_chunk_capacity = max_plab_samples; 628 for (uint i = 0; i < ParallelGCThreads; i++) { 629 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 630 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 631 assert(cur->end() == 0, "Should be 0"); 632 assert(cur->array() == vec, "Should be vec"); 633 assert(cur->capacity() == max_plab_samples, "Error"); 634 } 635 } 636 637 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 638 _gc_counters = new CollectorCounters("CMS", 1); 639 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 640 _completed_initialization = true; 641 _inter_sweep_timer.start(); // start of time 642 } 643 644 const char* ConcurrentMarkSweepGeneration::name() const { 645 return "concurrent mark-sweep generation"; 646 } 647 void ConcurrentMarkSweepGeneration::update_counters() { 648 if (UsePerfData) { 649 _space_counters->update_all(); 650 _gen_counters->update_all(); 651 } 652 } 653 654 // this is an optimized version of update_counters(). it takes the 655 // used value as a parameter rather than computing it. 656 // 657 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 658 if (UsePerfData) { 659 _space_counters->update_used(used); 660 _space_counters->update_capacity(); 661 _gen_counters->update_all(); 662 } 663 } 664 665 void ConcurrentMarkSweepGeneration::print() const { 666 Generation::print(); 667 cmsSpace()->print(); 668 } 669 670 #ifndef PRODUCT 671 void ConcurrentMarkSweepGeneration::print_statistics() { 672 cmsSpace()->printFLCensus(0); 673 } 674 #endif 675 676 size_t 677 ConcurrentMarkSweepGeneration::contiguous_available() const { 678 // dld proposes an improvement in precision here. If the committed 679 // part of the space ends in a free block we should add that to 680 // uncommitted size in the calculation below. Will make this 681 // change later, staying with the approximation below for the 682 // time being. -- ysr. 683 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 684 } 685 686 size_t 687 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 688 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 689 } 690 691 size_t ConcurrentMarkSweepGeneration::max_available() const { 692 return free() + _virtual_space.uncommitted_size(); 693 } 694 695 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 696 size_t available = max_available(); 697 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 698 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 699 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 700 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 701 return res; 702 } 703 704 // At a promotion failure dump information on block layout in heap 705 // (cms old generation). 706 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 707 Log(gc, promotion) log; 708 if (log.is_trace()) { 709 LogStream ls(log.trace()); 710 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 711 } 712 } 713 714 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 715 // Clear the promotion information. These pointers can be adjusted 716 // along with all the other pointers into the heap but 717 // compaction is expected to be a rare event with 718 // a heap using cms so don't do it without seeing the need. 719 for (uint i = 0; i < ParallelGCThreads; i++) { 720 _par_gc_thread_states[i]->promo.reset(); 721 } 722 } 723 724 void ConcurrentMarkSweepGeneration::compute_new_size() { 725 assert_locked_or_safepoint(Heap_lock); 726 727 // If incremental collection failed, we just want to expand 728 // to the limit. 729 if (incremental_collection_failed()) { 730 clear_incremental_collection_failed(); 731 grow_to_reserved(); 732 return; 733 } 734 735 // The heap has been compacted but not reset yet. 736 // Any metric such as free() or used() will be incorrect. 737 738 CardGeneration::compute_new_size(); 739 740 // Reset again after a possible resizing 741 if (did_compact()) { 742 cmsSpace()->reset_after_compaction(); 743 } 744 } 745 746 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 747 assert_locked_or_safepoint(Heap_lock); 748 749 // If incremental collection failed, we just want to expand 750 // to the limit. 751 if (incremental_collection_failed()) { 752 clear_incremental_collection_failed(); 753 grow_to_reserved(); 754 return; 755 } 756 757 double free_percentage = ((double) free()) / capacity(); 758 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 759 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 760 761 // compute expansion delta needed for reaching desired free percentage 762 if (free_percentage < desired_free_percentage) { 763 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 764 assert(desired_capacity >= capacity(), "invalid expansion size"); 765 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 766 Log(gc) log; 767 if (log.is_trace()) { 768 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 769 log.trace("From compute_new_size: "); 770 log.trace(" Free fraction %f", free_percentage); 771 log.trace(" Desired free fraction %f", desired_free_percentage); 772 log.trace(" Maximum free fraction %f", maximum_free_percentage); 773 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 774 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 775 CMSHeap* heap = CMSHeap::heap(); 776 size_t young_size = heap->young_gen()->capacity(); 777 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 778 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 779 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 780 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 781 } 782 // safe if expansion fails 783 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 784 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 785 } else { 786 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 787 assert(desired_capacity <= capacity(), "invalid expansion size"); 788 size_t shrink_bytes = capacity() - desired_capacity; 789 // Don't shrink unless the delta is greater than the minimum shrink we want 790 if (shrink_bytes >= MinHeapDeltaBytes) { 791 shrink_free_list_by(shrink_bytes); 792 } 793 } 794 } 795 796 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 797 return cmsSpace()->freelistLock(); 798 } 799 800 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 801 CMSSynchronousYieldRequest yr; 802 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 803 return have_lock_and_allocate(size, tlab); 804 } 805 806 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 807 bool tlab /* ignored */) { 808 assert_lock_strong(freelistLock()); 809 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 810 HeapWord* res = cmsSpace()->allocate(adjustedSize); 811 // Allocate the object live (grey) if the background collector has 812 // started marking. This is necessary because the marker may 813 // have passed this address and consequently this object will 814 // not otherwise be greyed and would be incorrectly swept up. 815 // Note that if this object contains references, the writing 816 // of those references will dirty the card containing this object 817 // allowing the object to be blackened (and its references scanned) 818 // either during a preclean phase or at the final checkpoint. 819 if (res != NULL) { 820 // We may block here with an uninitialized object with 821 // its mark-bit or P-bits not yet set. Such objects need 822 // to be safely navigable by block_start(). 823 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 824 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 825 collector()->direct_allocated(res, adjustedSize); 826 _direct_allocated_words += adjustedSize; 827 // allocation counters 828 NOT_PRODUCT( 829 _numObjectsAllocated++; 830 _numWordsAllocated += (int)adjustedSize; 831 ) 832 } 833 return res; 834 } 835 836 // In the case of direct allocation by mutators in a generation that 837 // is being concurrently collected, the object must be allocated 838 // live (grey) if the background collector has started marking. 839 // This is necessary because the marker may 840 // have passed this address and consequently this object will 841 // not otherwise be greyed and would be incorrectly swept up. 842 // Note that if this object contains references, the writing 843 // of those references will dirty the card containing this object 844 // allowing the object to be blackened (and its references scanned) 845 // either during a preclean phase or at the final checkpoint. 846 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 847 assert(_markBitMap.covers(start, size), "Out of bounds"); 848 if (_collectorState >= Marking) { 849 MutexLockerEx y(_markBitMap.lock(), 850 Mutex::_no_safepoint_check_flag); 851 // [see comments preceding SweepClosure::do_blk() below for details] 852 // 853 // Can the P-bits be deleted now? JJJ 854 // 855 // 1. need to mark the object as live so it isn't collected 856 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 857 // 3. need to mark the end of the object so marking, precleaning or sweeping 858 // can skip over uninitialized or unparsable objects. An allocated 859 // object is considered uninitialized for our purposes as long as 860 // its klass word is NULL. All old gen objects are parsable 861 // as soon as they are initialized.) 862 _markBitMap.mark(start); // object is live 863 _markBitMap.mark(start + 1); // object is potentially uninitialized? 864 _markBitMap.mark(start + size - 1); 865 // mark end of object 866 } 867 // check that oop looks uninitialized 868 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 869 } 870 871 void CMSCollector::promoted(bool par, HeapWord* start, 872 bool is_obj_array, size_t obj_size) { 873 assert(_markBitMap.covers(start), "Out of bounds"); 874 // See comment in direct_allocated() about when objects should 875 // be allocated live. 876 if (_collectorState >= Marking) { 877 // we already hold the marking bit map lock, taken in 878 // the prologue 879 if (par) { 880 _markBitMap.par_mark(start); 881 } else { 882 _markBitMap.mark(start); 883 } 884 // We don't need to mark the object as uninitialized (as 885 // in direct_allocated above) because this is being done with the 886 // world stopped and the object will be initialized by the 887 // time the marking, precleaning or sweeping get to look at it. 888 // But see the code for copying objects into the CMS generation, 889 // where we need to ensure that concurrent readers of the 890 // block offset table are able to safely navigate a block that 891 // is in flux from being free to being allocated (and in 892 // transition while being copied into) and subsequently 893 // becoming a bona-fide object when the copy/promotion is complete. 894 assert(SafepointSynchronize::is_at_safepoint(), 895 "expect promotion only at safepoints"); 896 897 if (_collectorState < Sweeping) { 898 // Mark the appropriate cards in the modUnionTable, so that 899 // this object gets scanned before the sweep. If this is 900 // not done, CMS generation references in the object might 901 // not get marked. 902 // For the case of arrays, which are otherwise precisely 903 // marked, we need to dirty the entire array, not just its head. 904 if (is_obj_array) { 905 // The [par_]mark_range() method expects mr.end() below to 906 // be aligned to the granularity of a bit's representation 907 // in the heap. In the case of the MUT below, that's a 908 // card size. 909 MemRegion mr(start, 910 align_up(start + obj_size, 911 CardTable::card_size /* bytes */)); 912 if (par) { 913 _modUnionTable.par_mark_range(mr); 914 } else { 915 _modUnionTable.mark_range(mr); 916 } 917 } else { // not an obj array; we can just mark the head 918 if (par) { 919 _modUnionTable.par_mark(start); 920 } else { 921 _modUnionTable.mark(start); 922 } 923 } 924 } 925 } 926 } 927 928 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 929 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 930 // allocate, copy and if necessary update promoinfo -- 931 // delegate to underlying space. 932 assert_lock_strong(freelistLock()); 933 934 #ifndef PRODUCT 935 if (CMSHeap::heap()->promotion_should_fail()) { 936 return NULL; 937 } 938 #endif // #ifndef PRODUCT 939 940 oop res = _cmsSpace->promote(obj, obj_size); 941 if (res == NULL) { 942 // expand and retry 943 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 944 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 945 // Since this is the old generation, we don't try to promote 946 // into a more senior generation. 947 res = _cmsSpace->promote(obj, obj_size); 948 } 949 if (res != NULL) { 950 // See comment in allocate() about when objects should 951 // be allocated live. 952 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 953 collector()->promoted(false, // Not parallel 954 (HeapWord*)res, obj->is_objArray(), obj_size); 955 // promotion counters 956 NOT_PRODUCT( 957 _numObjectsPromoted++; 958 _numWordsPromoted += 959 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 960 ) 961 } 962 return res; 963 } 964 965 966 // IMPORTANT: Notes on object size recognition in CMS. 967 // --------------------------------------------------- 968 // A block of storage in the CMS generation is always in 969 // one of three states. A free block (FREE), an allocated 970 // object (OBJECT) whose size() method reports the correct size, 971 // and an intermediate state (TRANSIENT) in which its size cannot 972 // be accurately determined. 973 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 974 // ----------------------------------------------------- 975 // FREE: klass_word & 1 == 1; mark_word holds block size 976 // 977 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 978 // obj->size() computes correct size 979 // 980 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 981 // 982 // STATE IDENTIFICATION: (64 bit+COOPS) 983 // ------------------------------------ 984 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 985 // 986 // OBJECT: klass_word installed; klass_word != 0; 987 // obj->size() computes correct size 988 // 989 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 990 // 991 // 992 // STATE TRANSITION DIAGRAM 993 // 994 // mut / parnew mut / parnew 995 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 996 // ^ | 997 // |------------------------ DEAD <------------------------------------| 998 // sweep mut 999 // 1000 // While a block is in TRANSIENT state its size cannot be determined 1001 // so readers will either need to come back later or stall until 1002 // the size can be determined. Note that for the case of direct 1003 // allocation, P-bits, when available, may be used to determine the 1004 // size of an object that may not yet have been initialized. 1005 1006 // Things to support parallel young-gen collection. 1007 oop 1008 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1009 oop old, markOop m, 1010 size_t word_sz) { 1011 #ifndef PRODUCT 1012 if (CMSHeap::heap()->promotion_should_fail()) { 1013 return NULL; 1014 } 1015 #endif // #ifndef PRODUCT 1016 1017 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1018 PromotionInfo* promoInfo = &ps->promo; 1019 // if we are tracking promotions, then first ensure space for 1020 // promotion (including spooling space for saving header if necessary). 1021 // then allocate and copy, then track promoted info if needed. 1022 // When tracking (see PromotionInfo::track()), the mark word may 1023 // be displaced and in this case restoration of the mark word 1024 // occurs in the (oop_since_save_marks_)iterate phase. 1025 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1026 // Out of space for allocating spooling buffers; 1027 // try expanding and allocating spooling buffers. 1028 if (!expand_and_ensure_spooling_space(promoInfo)) { 1029 return NULL; 1030 } 1031 } 1032 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1033 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1034 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1035 if (obj_ptr == NULL) { 1036 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1037 if (obj_ptr == NULL) { 1038 return NULL; 1039 } 1040 } 1041 oop obj = oop(obj_ptr); 1042 OrderAccess::storestore(); 1043 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1044 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1045 // IMPORTANT: See note on object initialization for CMS above. 1046 // Otherwise, copy the object. Here we must be careful to insert the 1047 // klass pointer last, since this marks the block as an allocated object. 1048 // Except with compressed oops it's the mark word. 1049 HeapWord* old_ptr = (HeapWord*)old; 1050 // Restore the mark word copied above. 1051 obj->set_mark_raw(m); 1052 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1053 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1054 OrderAccess::storestore(); 1055 1056 if (UseCompressedClassPointers) { 1057 // Copy gap missed by (aligned) header size calculation below 1058 obj->set_klass_gap(old->klass_gap()); 1059 } 1060 if (word_sz > (size_t)oopDesc::header_size()) { 1061 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1062 obj_ptr + oopDesc::header_size(), 1063 word_sz - oopDesc::header_size()); 1064 } 1065 1066 // Now we can track the promoted object, if necessary. We take care 1067 // to delay the transition from uninitialized to full object 1068 // (i.e., insertion of klass pointer) until after, so that it 1069 // atomically becomes a promoted object. 1070 if (promoInfo->tracking()) { 1071 promoInfo->track((PromotedObject*)obj, old->klass()); 1072 } 1073 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1074 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1075 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1076 1077 // Finally, install the klass pointer (this should be volatile). 1078 OrderAccess::storestore(); 1079 obj->set_klass(old->klass()); 1080 // We should now be able to calculate the right size for this object 1081 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1082 1083 collector()->promoted(true, // parallel 1084 obj_ptr, old->is_objArray(), word_sz); 1085 1086 NOT_PRODUCT( 1087 Atomic::inc(&_numObjectsPromoted); 1088 Atomic::add(alloc_sz, &_numWordsPromoted); 1089 ) 1090 1091 return obj; 1092 } 1093 1094 void 1095 ConcurrentMarkSweepGeneration:: 1096 par_promote_alloc_done(int thread_num) { 1097 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1098 ps->lab.retire(thread_num); 1099 } 1100 1101 void 1102 ConcurrentMarkSweepGeneration:: 1103 par_oop_since_save_marks_iterate_done(int thread_num) { 1104 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1105 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1106 ps->promo.promoted_oops_iterate(dummy_cl); 1107 1108 // Because card-scanning has been completed, subsequent phases 1109 // (e.g., reference processing) will not need to recognize which 1110 // objects have been promoted during this GC. So, we can now disable 1111 // promotion tracking. 1112 ps->promo.stopTrackingPromotions(); 1113 } 1114 1115 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1116 size_t size, 1117 bool tlab) 1118 { 1119 // We allow a STW collection only if a full 1120 // collection was requested. 1121 return full || should_allocate(size, tlab); // FIX ME !!! 1122 // This and promotion failure handling are connected at the 1123 // hip and should be fixed by untying them. 1124 } 1125 1126 bool CMSCollector::shouldConcurrentCollect() { 1127 LogTarget(Trace, gc) log; 1128 1129 if (_full_gc_requested) { 1130 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1131 return true; 1132 } 1133 1134 FreelistLocker x(this); 1135 // ------------------------------------------------------------------ 1136 // Print out lots of information which affects the initiation of 1137 // a collection. 1138 if (log.is_enabled() && stats().valid()) { 1139 log.print("CMSCollector shouldConcurrentCollect: "); 1140 1141 LogStream out(log); 1142 stats().print_on(&out); 1143 1144 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1145 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1146 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1147 log.print("promotion_rate=%g", stats().promotion_rate()); 1148 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1149 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1150 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1151 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1152 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1153 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1154 } 1155 // ------------------------------------------------------------------ 1156 1157 // If the estimated time to complete a cms collection (cms_duration()) 1158 // is less than the estimated time remaining until the cms generation 1159 // is full, start a collection. 1160 if (!UseCMSInitiatingOccupancyOnly) { 1161 if (stats().valid()) { 1162 if (stats().time_until_cms_start() == 0.0) { 1163 return true; 1164 } 1165 } else { 1166 // We want to conservatively collect somewhat early in order 1167 // to try and "bootstrap" our CMS/promotion statistics; 1168 // this branch will not fire after the first successful CMS 1169 // collection because the stats should then be valid. 1170 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1171 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1172 _cmsGen->occupancy(), _bootstrap_occupancy); 1173 return true; 1174 } 1175 } 1176 } 1177 1178 // Otherwise, we start a collection cycle if 1179 // old gen want a collection cycle started. Each may use 1180 // an appropriate criterion for making this decision. 1181 // XXX We need to make sure that the gen expansion 1182 // criterion dovetails well with this. XXX NEED TO FIX THIS 1183 if (_cmsGen->should_concurrent_collect()) { 1184 log.print("CMS old gen initiated"); 1185 return true; 1186 } 1187 1188 // We start a collection if we believe an incremental collection may fail; 1189 // this is not likely to be productive in practice because it's probably too 1190 // late anyway. 1191 CMSHeap* heap = CMSHeap::heap(); 1192 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1193 log.print("CMSCollector: collect because incremental collection will fail "); 1194 return true; 1195 } 1196 1197 if (MetaspaceGC::should_concurrent_collect()) { 1198 log.print("CMSCollector: collect for metadata allocation "); 1199 return true; 1200 } 1201 1202 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1203 if (CMSTriggerInterval >= 0) { 1204 if (CMSTriggerInterval == 0) { 1205 // Trigger always 1206 return true; 1207 } 1208 1209 // Check the CMS time since begin (we do not check the stats validity 1210 // as we want to be able to trigger the first CMS cycle as well) 1211 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1212 if (stats().valid()) { 1213 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1214 stats().cms_time_since_begin()); 1215 } else { 1216 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1217 } 1218 return true; 1219 } 1220 } 1221 1222 return false; 1223 } 1224 1225 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1226 1227 // Clear _expansion_cause fields of constituent generations 1228 void CMSCollector::clear_expansion_cause() { 1229 _cmsGen->clear_expansion_cause(); 1230 } 1231 1232 // We should be conservative in starting a collection cycle. To 1233 // start too eagerly runs the risk of collecting too often in the 1234 // extreme. To collect too rarely falls back on full collections, 1235 // which works, even if not optimum in terms of concurrent work. 1236 // As a work around for too eagerly collecting, use the flag 1237 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1238 // giving the user an easily understandable way of controlling the 1239 // collections. 1240 // We want to start a new collection cycle if any of the following 1241 // conditions hold: 1242 // . our current occupancy exceeds the configured initiating occupancy 1243 // for this generation, or 1244 // . we recently needed to expand this space and have not, since that 1245 // expansion, done a collection of this generation, or 1246 // . the underlying space believes that it may be a good idea to initiate 1247 // a concurrent collection (this may be based on criteria such as the 1248 // following: the space uses linear allocation and linear allocation is 1249 // going to fail, or there is believed to be excessive fragmentation in 1250 // the generation, etc... or ... 1251 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1252 // the case of the old generation; see CR 6543076): 1253 // we may be approaching a point at which allocation requests may fail because 1254 // we will be out of sufficient free space given allocation rate estimates.] 1255 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1256 1257 assert_lock_strong(freelistLock()); 1258 if (occupancy() > initiating_occupancy()) { 1259 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1260 short_name(), occupancy(), initiating_occupancy()); 1261 return true; 1262 } 1263 if (UseCMSInitiatingOccupancyOnly) { 1264 return false; 1265 } 1266 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1267 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1268 return true; 1269 } 1270 return false; 1271 } 1272 1273 void ConcurrentMarkSweepGeneration::collect(bool full, 1274 bool clear_all_soft_refs, 1275 size_t size, 1276 bool tlab) 1277 { 1278 collector()->collect(full, clear_all_soft_refs, size, tlab); 1279 } 1280 1281 void CMSCollector::collect(bool full, 1282 bool clear_all_soft_refs, 1283 size_t size, 1284 bool tlab) 1285 { 1286 // The following "if" branch is present for defensive reasons. 1287 // In the current uses of this interface, it can be replaced with: 1288 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1289 // But I am not placing that assert here to allow future 1290 // generality in invoking this interface. 1291 if (GCLocker::is_active()) { 1292 // A consistency test for GCLocker 1293 assert(GCLocker::needs_gc(), "Should have been set already"); 1294 // Skip this foreground collection, instead 1295 // expanding the heap if necessary. 1296 // Need the free list locks for the call to free() in compute_new_size() 1297 compute_new_size(); 1298 return; 1299 } 1300 acquire_control_and_collect(full, clear_all_soft_refs); 1301 } 1302 1303 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1304 CMSHeap* heap = CMSHeap::heap(); 1305 unsigned int gc_count = heap->total_full_collections(); 1306 if (gc_count == full_gc_count) { 1307 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1308 _full_gc_requested = true; 1309 _full_gc_cause = cause; 1310 CGC_lock->notify(); // nudge CMS thread 1311 } else { 1312 assert(gc_count > full_gc_count, "Error: causal loop"); 1313 } 1314 } 1315 1316 bool CMSCollector::is_external_interruption() { 1317 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1318 return GCCause::is_user_requested_gc(cause) || 1319 GCCause::is_serviceability_requested_gc(cause); 1320 } 1321 1322 void CMSCollector::report_concurrent_mode_interruption() { 1323 if (is_external_interruption()) { 1324 log_debug(gc)("Concurrent mode interrupted"); 1325 } else { 1326 log_debug(gc)("Concurrent mode failure"); 1327 _gc_tracer_cm->report_concurrent_mode_failure(); 1328 } 1329 } 1330 1331 1332 // The foreground and background collectors need to coordinate in order 1333 // to make sure that they do not mutually interfere with CMS collections. 1334 // When a background collection is active, 1335 // the foreground collector may need to take over (preempt) and 1336 // synchronously complete an ongoing collection. Depending on the 1337 // frequency of the background collections and the heap usage 1338 // of the application, this preemption can be seldom or frequent. 1339 // There are only certain 1340 // points in the background collection that the "collection-baton" 1341 // can be passed to the foreground collector. 1342 // 1343 // The foreground collector will wait for the baton before 1344 // starting any part of the collection. The foreground collector 1345 // will only wait at one location. 1346 // 1347 // The background collector will yield the baton before starting a new 1348 // phase of the collection (e.g., before initial marking, marking from roots, 1349 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1350 // of the loop which switches the phases. The background collector does some 1351 // of the phases (initial mark, final re-mark) with the world stopped. 1352 // Because of locking involved in stopping the world, 1353 // the foreground collector should not block waiting for the background 1354 // collector when it is doing a stop-the-world phase. The background 1355 // collector will yield the baton at an additional point just before 1356 // it enters a stop-the-world phase. Once the world is stopped, the 1357 // background collector checks the phase of the collection. If the 1358 // phase has not changed, it proceeds with the collection. If the 1359 // phase has changed, it skips that phase of the collection. See 1360 // the comments on the use of the Heap_lock in collect_in_background(). 1361 // 1362 // Variable used in baton passing. 1363 // _foregroundGCIsActive - Set to true by the foreground collector when 1364 // it wants the baton. The foreground clears it when it has finished 1365 // the collection. 1366 // _foregroundGCShouldWait - Set to true by the background collector 1367 // when it is running. The foreground collector waits while 1368 // _foregroundGCShouldWait is true. 1369 // CGC_lock - monitor used to protect access to the above variables 1370 // and to notify the foreground and background collectors. 1371 // _collectorState - current state of the CMS collection. 1372 // 1373 // The foreground collector 1374 // acquires the CGC_lock 1375 // sets _foregroundGCIsActive 1376 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1377 // various locks acquired in preparation for the collection 1378 // are released so as not to block the background collector 1379 // that is in the midst of a collection 1380 // proceeds with the collection 1381 // clears _foregroundGCIsActive 1382 // returns 1383 // 1384 // The background collector in a loop iterating on the phases of the 1385 // collection 1386 // acquires the CGC_lock 1387 // sets _foregroundGCShouldWait 1388 // if _foregroundGCIsActive is set 1389 // clears _foregroundGCShouldWait, notifies _CGC_lock 1390 // waits on _CGC_lock for _foregroundGCIsActive to become false 1391 // and exits the loop. 1392 // otherwise 1393 // proceed with that phase of the collection 1394 // if the phase is a stop-the-world phase, 1395 // yield the baton once more just before enqueueing 1396 // the stop-world CMS operation (executed by the VM thread). 1397 // returns after all phases of the collection are done 1398 // 1399 1400 void CMSCollector::acquire_control_and_collect(bool full, 1401 bool clear_all_soft_refs) { 1402 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1403 assert(!Thread::current()->is_ConcurrentGC_thread(), 1404 "shouldn't try to acquire control from self!"); 1405 1406 // Start the protocol for acquiring control of the 1407 // collection from the background collector (aka CMS thread). 1408 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1409 "VM thread should have CMS token"); 1410 // Remember the possibly interrupted state of an ongoing 1411 // concurrent collection 1412 CollectorState first_state = _collectorState; 1413 1414 // Signal to a possibly ongoing concurrent collection that 1415 // we want to do a foreground collection. 1416 _foregroundGCIsActive = true; 1417 1418 // release locks and wait for a notify from the background collector 1419 // releasing the locks in only necessary for phases which 1420 // do yields to improve the granularity of the collection. 1421 assert_lock_strong(bitMapLock()); 1422 // We need to lock the Free list lock for the space that we are 1423 // currently collecting. 1424 assert(haveFreelistLocks(), "Must be holding free list locks"); 1425 bitMapLock()->unlock(); 1426 releaseFreelistLocks(); 1427 { 1428 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1429 if (_foregroundGCShouldWait) { 1430 // We are going to be waiting for action for the CMS thread; 1431 // it had better not be gone (for instance at shutdown)! 1432 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1433 "CMS thread must be running"); 1434 // Wait here until the background collector gives us the go-ahead 1435 ConcurrentMarkSweepThread::clear_CMS_flag( 1436 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1437 // Get a possibly blocked CMS thread going: 1438 // Note that we set _foregroundGCIsActive true above, 1439 // without protection of the CGC_lock. 1440 CGC_lock->notify(); 1441 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1442 "Possible deadlock"); 1443 while (_foregroundGCShouldWait) { 1444 // wait for notification 1445 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1446 // Possibility of delay/starvation here, since CMS token does 1447 // not know to give priority to VM thread? Actually, i think 1448 // there wouldn't be any delay/starvation, but the proof of 1449 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1450 } 1451 ConcurrentMarkSweepThread::set_CMS_flag( 1452 ConcurrentMarkSweepThread::CMS_vm_has_token); 1453 } 1454 } 1455 // The CMS_token is already held. Get back the other locks. 1456 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1457 "VM thread should have CMS token"); 1458 getFreelistLocks(); 1459 bitMapLock()->lock_without_safepoint_check(); 1460 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1461 p2i(Thread::current()), first_state); 1462 log_debug(gc, state)(" gets control with state %d", _collectorState); 1463 1464 // Inform cms gen if this was due to partial collection failing. 1465 // The CMS gen may use this fact to determine its expansion policy. 1466 CMSHeap* heap = CMSHeap::heap(); 1467 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1468 assert(!_cmsGen->incremental_collection_failed(), 1469 "Should have been noticed, reacted to and cleared"); 1470 _cmsGen->set_incremental_collection_failed(); 1471 } 1472 1473 if (first_state > Idling) { 1474 report_concurrent_mode_interruption(); 1475 } 1476 1477 set_did_compact(true); 1478 1479 // If the collection is being acquired from the background 1480 // collector, there may be references on the discovered 1481 // references lists. Abandon those references, since some 1482 // of them may have become unreachable after concurrent 1483 // discovery; the STW compacting collector will redo discovery 1484 // more precisely, without being subject to floating garbage. 1485 // Leaving otherwise unreachable references in the discovered 1486 // lists would require special handling. 1487 ref_processor()->disable_discovery(); 1488 ref_processor()->abandon_partial_discovery(); 1489 ref_processor()->verify_no_references_recorded(); 1490 1491 if (first_state > Idling) { 1492 save_heap_summary(); 1493 } 1494 1495 do_compaction_work(clear_all_soft_refs); 1496 1497 // Has the GC time limit been exceeded? 1498 size_t max_eden_size = _young_gen->max_eden_size(); 1499 GCCause::Cause gc_cause = heap->gc_cause(); 1500 size_policy()->check_gc_overhead_limit(_young_gen->used(), 1501 _young_gen->eden()->used(), 1502 _cmsGen->max_capacity(), 1503 max_eden_size, 1504 full, 1505 gc_cause, 1506 heap->soft_ref_policy()); 1507 1508 // Reset the expansion cause, now that we just completed 1509 // a collection cycle. 1510 clear_expansion_cause(); 1511 _foregroundGCIsActive = false; 1512 return; 1513 } 1514 1515 // Resize the tenured generation 1516 // after obtaining the free list locks for the 1517 // two generations. 1518 void CMSCollector::compute_new_size() { 1519 assert_locked_or_safepoint(Heap_lock); 1520 FreelistLocker z(this); 1521 MetaspaceGC::compute_new_size(); 1522 _cmsGen->compute_new_size_free_list(); 1523 } 1524 1525 // A work method used by the foreground collector to do 1526 // a mark-sweep-compact. 1527 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1528 CMSHeap* heap = CMSHeap::heap(); 1529 1530 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1531 gc_timer->register_gc_start(); 1532 1533 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1534 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1535 1536 heap->pre_full_gc_dump(gc_timer); 1537 1538 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1539 1540 // Temporarily widen the span of the weak reference processing to 1541 // the entire heap. 1542 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1543 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1544 // Temporarily, clear the "is_alive_non_header" field of the 1545 // reference processor. 1546 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1547 // Temporarily make reference _processing_ single threaded (non-MT). 1548 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1549 // Temporarily make refs discovery atomic 1550 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1551 // Temporarily make reference _discovery_ single threaded (non-MT) 1552 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1553 1554 ref_processor()->set_enqueuing_is_done(false); 1555 ref_processor()->enable_discovery(); 1556 ref_processor()->setup_policy(clear_all_soft_refs); 1557 // If an asynchronous collection finishes, the _modUnionTable is 1558 // all clear. If we are assuming the collection from an asynchronous 1559 // collection, clear the _modUnionTable. 1560 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1561 "_modUnionTable should be clear if the baton was not passed"); 1562 _modUnionTable.clear_all(); 1563 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1564 "mod union for klasses should be clear if the baton was passed"); 1565 _ct->cld_rem_set()->clear_mod_union(); 1566 1567 1568 // We must adjust the allocation statistics being maintained 1569 // in the free list space. We do so by reading and clearing 1570 // the sweep timer and updating the block flux rate estimates below. 1571 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1572 if (_inter_sweep_timer.is_active()) { 1573 _inter_sweep_timer.stop(); 1574 // Note that we do not use this sample to update the _inter_sweep_estimate. 1575 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1576 _inter_sweep_estimate.padded_average(), 1577 _intra_sweep_estimate.padded_average()); 1578 } 1579 1580 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1581 #ifdef ASSERT 1582 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1583 size_t free_size = cms_space->free(); 1584 assert(free_size == 1585 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1586 * HeapWordSize, 1587 "All the free space should be compacted into one chunk at top"); 1588 assert(cms_space->dictionary()->total_chunk_size( 1589 debug_only(cms_space->freelistLock())) == 0 || 1590 cms_space->totalSizeInIndexedFreeLists() == 0, 1591 "All the free space should be in a single chunk"); 1592 size_t num = cms_space->totalCount(); 1593 assert((free_size == 0 && num == 0) || 1594 (free_size > 0 && (num == 1 || num == 2)), 1595 "There should be at most 2 free chunks after compaction"); 1596 #endif // ASSERT 1597 _collectorState = Resetting; 1598 assert(_restart_addr == NULL, 1599 "Should have been NULL'd before baton was passed"); 1600 reset_stw(); 1601 _cmsGen->reset_after_compaction(); 1602 _concurrent_cycles_since_last_unload = 0; 1603 1604 // Clear any data recorded in the PLAB chunk arrays. 1605 if (_survivor_plab_array != NULL) { 1606 reset_survivor_plab_arrays(); 1607 } 1608 1609 // Adjust the per-size allocation stats for the next epoch. 1610 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1611 // Restart the "inter sweep timer" for the next epoch. 1612 _inter_sweep_timer.reset(); 1613 _inter_sweep_timer.start(); 1614 1615 // No longer a need to do a concurrent collection for Metaspace. 1616 MetaspaceGC::set_should_concurrent_collect(false); 1617 1618 heap->post_full_gc_dump(gc_timer); 1619 1620 gc_timer->register_gc_end(); 1621 1622 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1623 1624 // For a mark-sweep-compact, compute_new_size() will be called 1625 // in the heap's do_collection() method. 1626 } 1627 1628 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1629 Log(gc, heap) log; 1630 if (!log.is_trace()) { 1631 return; 1632 } 1633 1634 ContiguousSpace* eden_space = _young_gen->eden(); 1635 ContiguousSpace* from_space = _young_gen->from(); 1636 ContiguousSpace* to_space = _young_gen->to(); 1637 // Eden 1638 if (_eden_chunk_array != NULL) { 1639 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1640 p2i(eden_space->bottom()), p2i(eden_space->top()), 1641 p2i(eden_space->end()), eden_space->capacity()); 1642 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1643 _eden_chunk_index, _eden_chunk_capacity); 1644 for (size_t i = 0; i < _eden_chunk_index; i++) { 1645 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1646 } 1647 } 1648 // Survivor 1649 if (_survivor_chunk_array != NULL) { 1650 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1651 p2i(from_space->bottom()), p2i(from_space->top()), 1652 p2i(from_space->end()), from_space->capacity()); 1653 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1654 _survivor_chunk_index, _survivor_chunk_capacity); 1655 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1656 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1657 } 1658 } 1659 } 1660 1661 void CMSCollector::getFreelistLocks() const { 1662 // Get locks for all free lists in all generations that this 1663 // collector is responsible for 1664 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1665 } 1666 1667 void CMSCollector::releaseFreelistLocks() const { 1668 // Release locks for all free lists in all generations that this 1669 // collector is responsible for 1670 _cmsGen->freelistLock()->unlock(); 1671 } 1672 1673 bool CMSCollector::haveFreelistLocks() const { 1674 // Check locks for all free lists in all generations that this 1675 // collector is responsible for 1676 assert_lock_strong(_cmsGen->freelistLock()); 1677 PRODUCT_ONLY(ShouldNotReachHere()); 1678 return true; 1679 } 1680 1681 // A utility class that is used by the CMS collector to 1682 // temporarily "release" the foreground collector from its 1683 // usual obligation to wait for the background collector to 1684 // complete an ongoing phase before proceeding. 1685 class ReleaseForegroundGC: public StackObj { 1686 private: 1687 CMSCollector* _c; 1688 public: 1689 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1690 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1691 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1692 // allow a potentially blocked foreground collector to proceed 1693 _c->_foregroundGCShouldWait = false; 1694 if (_c->_foregroundGCIsActive) { 1695 CGC_lock->notify(); 1696 } 1697 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1698 "Possible deadlock"); 1699 } 1700 1701 ~ReleaseForegroundGC() { 1702 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1703 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1704 _c->_foregroundGCShouldWait = true; 1705 } 1706 }; 1707 1708 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1709 assert(Thread::current()->is_ConcurrentGC_thread(), 1710 "A CMS asynchronous collection is only allowed on a CMS thread."); 1711 1712 CMSHeap* heap = CMSHeap::heap(); 1713 { 1714 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1715 MutexLockerEx hl(Heap_lock, safepoint_check); 1716 FreelistLocker fll(this); 1717 MutexLockerEx x(CGC_lock, safepoint_check); 1718 if (_foregroundGCIsActive) { 1719 // The foreground collector is. Skip this 1720 // background collection. 1721 assert(!_foregroundGCShouldWait, "Should be clear"); 1722 return; 1723 } else { 1724 assert(_collectorState == Idling, "Should be idling before start."); 1725 _collectorState = InitialMarking; 1726 register_gc_start(cause); 1727 // Reset the expansion cause, now that we are about to begin 1728 // a new cycle. 1729 clear_expansion_cause(); 1730 1731 // Clear the MetaspaceGC flag since a concurrent collection 1732 // is starting but also clear it after the collection. 1733 MetaspaceGC::set_should_concurrent_collect(false); 1734 } 1735 // Decide if we want to enable class unloading as part of the 1736 // ensuing concurrent GC cycle. 1737 update_should_unload_classes(); 1738 _full_gc_requested = false; // acks all outstanding full gc requests 1739 _full_gc_cause = GCCause::_no_gc; 1740 // Signal that we are about to start a collection 1741 heap->increment_total_full_collections(); // ... starting a collection cycle 1742 _collection_count_start = heap->total_full_collections(); 1743 } 1744 1745 size_t prev_used = _cmsGen->used(); 1746 1747 // The change of the collection state is normally done at this level; 1748 // the exceptions are phases that are executed while the world is 1749 // stopped. For those phases the change of state is done while the 1750 // world is stopped. For baton passing purposes this allows the 1751 // background collector to finish the phase and change state atomically. 1752 // The foreground collector cannot wait on a phase that is done 1753 // while the world is stopped because the foreground collector already 1754 // has the world stopped and would deadlock. 1755 while (_collectorState != Idling) { 1756 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1757 p2i(Thread::current()), _collectorState); 1758 // The foreground collector 1759 // holds the Heap_lock throughout its collection. 1760 // holds the CMS token (but not the lock) 1761 // except while it is waiting for the background collector to yield. 1762 // 1763 // The foreground collector should be blocked (not for long) 1764 // if the background collector is about to start a phase 1765 // executed with world stopped. If the background 1766 // collector has already started such a phase, the 1767 // foreground collector is blocked waiting for the 1768 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1769 // are executed in the VM thread. 1770 // 1771 // The locking order is 1772 // PendingListLock (PLL) -- if applicable (FinalMarking) 1773 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1774 // CMS token (claimed in 1775 // stop_world_and_do() --> 1776 // safepoint_synchronize() --> 1777 // CMSThread::synchronize()) 1778 1779 { 1780 // Check if the FG collector wants us to yield. 1781 CMSTokenSync x(true); // is cms thread 1782 if (waitForForegroundGC()) { 1783 // We yielded to a foreground GC, nothing more to be 1784 // done this round. 1785 assert(_foregroundGCShouldWait == false, "We set it to false in " 1786 "waitForForegroundGC()"); 1787 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1788 p2i(Thread::current()), _collectorState); 1789 return; 1790 } else { 1791 // The background collector can run but check to see if the 1792 // foreground collector has done a collection while the 1793 // background collector was waiting to get the CGC_lock 1794 // above. If yes, break so that _foregroundGCShouldWait 1795 // is cleared before returning. 1796 if (_collectorState == Idling) { 1797 break; 1798 } 1799 } 1800 } 1801 1802 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1803 "should be waiting"); 1804 1805 switch (_collectorState) { 1806 case InitialMarking: 1807 { 1808 ReleaseForegroundGC x(this); 1809 stats().record_cms_begin(); 1810 VM_CMS_Initial_Mark initial_mark_op(this); 1811 VMThread::execute(&initial_mark_op); 1812 } 1813 // The collector state may be any legal state at this point 1814 // since the background collector may have yielded to the 1815 // foreground collector. 1816 break; 1817 case Marking: 1818 // initial marking in checkpointRootsInitialWork has been completed 1819 if (markFromRoots()) { // we were successful 1820 assert(_collectorState == Precleaning, "Collector state should " 1821 "have changed"); 1822 } else { 1823 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1824 } 1825 break; 1826 case Precleaning: 1827 // marking from roots in markFromRoots has been completed 1828 preclean(); 1829 assert(_collectorState == AbortablePreclean || 1830 _collectorState == FinalMarking, 1831 "Collector state should have changed"); 1832 break; 1833 case AbortablePreclean: 1834 abortable_preclean(); 1835 assert(_collectorState == FinalMarking, "Collector state should " 1836 "have changed"); 1837 break; 1838 case FinalMarking: 1839 { 1840 ReleaseForegroundGC x(this); 1841 1842 VM_CMS_Final_Remark final_remark_op(this); 1843 VMThread::execute(&final_remark_op); 1844 } 1845 assert(_foregroundGCShouldWait, "block post-condition"); 1846 break; 1847 case Sweeping: 1848 // final marking in checkpointRootsFinal has been completed 1849 sweep(); 1850 assert(_collectorState == Resizing, "Collector state change " 1851 "to Resizing must be done under the free_list_lock"); 1852 1853 case Resizing: { 1854 // Sweeping has been completed... 1855 // At this point the background collection has completed. 1856 // Don't move the call to compute_new_size() down 1857 // into code that might be executed if the background 1858 // collection was preempted. 1859 { 1860 ReleaseForegroundGC x(this); // unblock FG collection 1861 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1862 CMSTokenSync z(true); // not strictly needed. 1863 if (_collectorState == Resizing) { 1864 compute_new_size(); 1865 save_heap_summary(); 1866 _collectorState = Resetting; 1867 } else { 1868 assert(_collectorState == Idling, "The state should only change" 1869 " because the foreground collector has finished the collection"); 1870 } 1871 } 1872 break; 1873 } 1874 case Resetting: 1875 // CMS heap resizing has been completed 1876 reset_concurrent(); 1877 assert(_collectorState == Idling, "Collector state should " 1878 "have changed"); 1879 1880 MetaspaceGC::set_should_concurrent_collect(false); 1881 1882 stats().record_cms_end(); 1883 // Don't move the concurrent_phases_end() and compute_new_size() 1884 // calls to here because a preempted background collection 1885 // has it's state set to "Resetting". 1886 break; 1887 case Idling: 1888 default: 1889 ShouldNotReachHere(); 1890 break; 1891 } 1892 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1893 p2i(Thread::current()), _collectorState); 1894 assert(_foregroundGCShouldWait, "block post-condition"); 1895 } 1896 1897 // Should this be in gc_epilogue? 1898 heap->counters()->update_counters(); 1899 1900 { 1901 // Clear _foregroundGCShouldWait and, in the event that the 1902 // foreground collector is waiting, notify it, before 1903 // returning. 1904 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1905 _foregroundGCShouldWait = false; 1906 if (_foregroundGCIsActive) { 1907 CGC_lock->notify(); 1908 } 1909 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1910 "Possible deadlock"); 1911 } 1912 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1913 p2i(Thread::current()), _collectorState); 1914 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1915 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1916 } 1917 1918 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1919 _cms_start_registered = true; 1920 _gc_timer_cm->register_gc_start(); 1921 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1922 } 1923 1924 void CMSCollector::register_gc_end() { 1925 if (_cms_start_registered) { 1926 report_heap_summary(GCWhen::AfterGC); 1927 1928 _gc_timer_cm->register_gc_end(); 1929 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1930 _cms_start_registered = false; 1931 } 1932 } 1933 1934 void CMSCollector::save_heap_summary() { 1935 CMSHeap* heap = CMSHeap::heap(); 1936 _last_heap_summary = heap->create_heap_summary(); 1937 _last_metaspace_summary = heap->create_metaspace_summary(); 1938 } 1939 1940 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1941 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1942 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1943 } 1944 1945 bool CMSCollector::waitForForegroundGC() { 1946 bool res = false; 1947 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1948 "CMS thread should have CMS token"); 1949 // Block the foreground collector until the 1950 // background collectors decides whether to 1951 // yield. 1952 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1953 _foregroundGCShouldWait = true; 1954 if (_foregroundGCIsActive) { 1955 // The background collector yields to the 1956 // foreground collector and returns a value 1957 // indicating that it has yielded. The foreground 1958 // collector can proceed. 1959 res = true; 1960 _foregroundGCShouldWait = false; 1961 ConcurrentMarkSweepThread::clear_CMS_flag( 1962 ConcurrentMarkSweepThread::CMS_cms_has_token); 1963 ConcurrentMarkSweepThread::set_CMS_flag( 1964 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1965 // Get a possibly blocked foreground thread going 1966 CGC_lock->notify(); 1967 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1968 p2i(Thread::current()), _collectorState); 1969 while (_foregroundGCIsActive) { 1970 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1971 } 1972 ConcurrentMarkSweepThread::set_CMS_flag( 1973 ConcurrentMarkSweepThread::CMS_cms_has_token); 1974 ConcurrentMarkSweepThread::clear_CMS_flag( 1975 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1976 } 1977 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1978 p2i(Thread::current()), _collectorState); 1979 return res; 1980 } 1981 1982 // Because of the need to lock the free lists and other structures in 1983 // the collector, common to all the generations that the collector is 1984 // collecting, we need the gc_prologues of individual CMS generations 1985 // delegate to their collector. It may have been simpler had the 1986 // current infrastructure allowed one to call a prologue on a 1987 // collector. In the absence of that we have the generation's 1988 // prologue delegate to the collector, which delegates back 1989 // some "local" work to a worker method in the individual generations 1990 // that it's responsible for collecting, while itself doing any 1991 // work common to all generations it's responsible for. A similar 1992 // comment applies to the gc_epilogue()'s. 1993 // The role of the variable _between_prologue_and_epilogue is to 1994 // enforce the invocation protocol. 1995 void CMSCollector::gc_prologue(bool full) { 1996 // Call gc_prologue_work() for the CMSGen 1997 // we are responsible for. 1998 1999 // The following locking discipline assumes that we are only called 2000 // when the world is stopped. 2001 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2002 2003 // The CMSCollector prologue must call the gc_prologues for the 2004 // "generations" that it's responsible 2005 // for. 2006 2007 assert( Thread::current()->is_VM_thread() 2008 || ( CMSScavengeBeforeRemark 2009 && Thread::current()->is_ConcurrentGC_thread()), 2010 "Incorrect thread type for prologue execution"); 2011 2012 if (_between_prologue_and_epilogue) { 2013 // We have already been invoked; this is a gc_prologue delegation 2014 // from yet another CMS generation that we are responsible for, just 2015 // ignore it since all relevant work has already been done. 2016 return; 2017 } 2018 2019 // set a bit saying prologue has been called; cleared in epilogue 2020 _between_prologue_and_epilogue = true; 2021 // Claim locks for common data structures, then call gc_prologue_work() 2022 // for each CMSGen. 2023 2024 getFreelistLocks(); // gets free list locks on constituent spaces 2025 bitMapLock()->lock_without_safepoint_check(); 2026 2027 // Should call gc_prologue_work() for all cms gens we are responsible for 2028 bool duringMarking = _collectorState >= Marking 2029 && _collectorState < Sweeping; 2030 2031 // The young collections clear the modified oops state, which tells if 2032 // there are any modified oops in the class. The remark phase also needs 2033 // that information. Tell the young collection to save the union of all 2034 // modified klasses. 2035 if (duringMarking) { 2036 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2037 } 2038 2039 bool registerClosure = duringMarking; 2040 2041 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2042 2043 if (!full) { 2044 stats().record_gc0_begin(); 2045 } 2046 } 2047 2048 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2049 2050 _capacity_at_prologue = capacity(); 2051 _used_at_prologue = used(); 2052 2053 // We enable promotion tracking so that card-scanning can recognize 2054 // which objects have been promoted during this GC and skip them. 2055 for (uint i = 0; i < ParallelGCThreads; i++) { 2056 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2057 } 2058 2059 // Delegate to CMScollector which knows how to coordinate between 2060 // this and any other CMS generations that it is responsible for 2061 // collecting. 2062 collector()->gc_prologue(full); 2063 } 2064 2065 // This is a "private" interface for use by this generation's CMSCollector. 2066 // Not to be called directly by any other entity (for instance, 2067 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2068 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2069 bool registerClosure, ModUnionClosure* modUnionClosure) { 2070 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2071 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2072 "Should be NULL"); 2073 if (registerClosure) { 2074 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2075 } 2076 cmsSpace()->gc_prologue(); 2077 // Clear stat counters 2078 NOT_PRODUCT( 2079 assert(_numObjectsPromoted == 0, "check"); 2080 assert(_numWordsPromoted == 0, "check"); 2081 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2082 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2083 _numObjectsAllocated = 0; 2084 _numWordsAllocated = 0; 2085 ) 2086 } 2087 2088 void CMSCollector::gc_epilogue(bool full) { 2089 // The following locking discipline assumes that we are only called 2090 // when the world is stopped. 2091 assert(SafepointSynchronize::is_at_safepoint(), 2092 "world is stopped assumption"); 2093 2094 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2095 // if linear allocation blocks need to be appropriately marked to allow the 2096 // the blocks to be parsable. We also check here whether we need to nudge the 2097 // CMS collector thread to start a new cycle (if it's not already active). 2098 assert( Thread::current()->is_VM_thread() 2099 || ( CMSScavengeBeforeRemark 2100 && Thread::current()->is_ConcurrentGC_thread()), 2101 "Incorrect thread type for epilogue execution"); 2102 2103 if (!_between_prologue_and_epilogue) { 2104 // We have already been invoked; this is a gc_epilogue delegation 2105 // from yet another CMS generation that we are responsible for, just 2106 // ignore it since all relevant work has already been done. 2107 return; 2108 } 2109 assert(haveFreelistLocks(), "must have freelist locks"); 2110 assert_lock_strong(bitMapLock()); 2111 2112 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2113 2114 _cmsGen->gc_epilogue_work(full); 2115 2116 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2117 // in case sampling was not already enabled, enable it 2118 _start_sampling = true; 2119 } 2120 // reset _eden_chunk_array so sampling starts afresh 2121 _eden_chunk_index = 0; 2122 2123 size_t cms_used = _cmsGen->cmsSpace()->used(); 2124 2125 // update performance counters - this uses a special version of 2126 // update_counters() that allows the utilization to be passed as a 2127 // parameter, avoiding multiple calls to used(). 2128 // 2129 _cmsGen->update_counters(cms_used); 2130 2131 bitMapLock()->unlock(); 2132 releaseFreelistLocks(); 2133 2134 if (!CleanChunkPoolAsync) { 2135 Chunk::clean_chunk_pool(); 2136 } 2137 2138 set_did_compact(false); 2139 _between_prologue_and_epilogue = false; // ready for next cycle 2140 } 2141 2142 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2143 collector()->gc_epilogue(full); 2144 2145 // When using ParNew, promotion tracking should have already been 2146 // disabled. However, the prologue (which enables promotion 2147 // tracking) and epilogue are called irrespective of the type of 2148 // GC. So they will also be called before and after Full GCs, during 2149 // which promotion tracking will not be explicitly disabled. So, 2150 // it's safer to also disable it here too (to be symmetric with 2151 // enabling it in the prologue). 2152 for (uint i = 0; i < ParallelGCThreads; i++) { 2153 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2154 } 2155 } 2156 2157 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2158 assert(!incremental_collection_failed(), "Should have been cleared"); 2159 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2160 cmsSpace()->gc_epilogue(); 2161 // Print stat counters 2162 NOT_PRODUCT( 2163 assert(_numObjectsAllocated == 0, "check"); 2164 assert(_numWordsAllocated == 0, "check"); 2165 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2166 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2167 _numObjectsPromoted = 0; 2168 _numWordsPromoted = 0; 2169 ) 2170 2171 // Call down the chain in contiguous_available needs the freelistLock 2172 // so print this out before releasing the freeListLock. 2173 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2174 } 2175 2176 #ifndef PRODUCT 2177 bool CMSCollector::have_cms_token() { 2178 Thread* thr = Thread::current(); 2179 if (thr->is_VM_thread()) { 2180 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2181 } else if (thr->is_ConcurrentGC_thread()) { 2182 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2183 } else if (thr->is_GC_task_thread()) { 2184 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2185 ParGCRareEvent_lock->owned_by_self(); 2186 } 2187 return false; 2188 } 2189 2190 // Check reachability of the given heap address in CMS generation, 2191 // treating all other generations as roots. 2192 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2193 // We could "guarantee" below, rather than assert, but I'll 2194 // leave these as "asserts" so that an adventurous debugger 2195 // could try this in the product build provided some subset of 2196 // the conditions were met, provided they were interested in the 2197 // results and knew that the computation below wouldn't interfere 2198 // with other concurrent computations mutating the structures 2199 // being read or written. 2200 assert(SafepointSynchronize::is_at_safepoint(), 2201 "Else mutations in object graph will make answer suspect"); 2202 assert(have_cms_token(), "Should hold cms token"); 2203 assert(haveFreelistLocks(), "must hold free list locks"); 2204 assert_lock_strong(bitMapLock()); 2205 2206 // Clear the marking bit map array before starting, but, just 2207 // for kicks, first report if the given address is already marked 2208 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2209 _markBitMap.isMarked(addr) ? "" : " not"); 2210 2211 if (verify_after_remark()) { 2212 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2213 bool result = verification_mark_bm()->isMarked(addr); 2214 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2215 result ? "IS" : "is NOT"); 2216 return result; 2217 } else { 2218 tty->print_cr("Could not compute result"); 2219 return false; 2220 } 2221 } 2222 #endif 2223 2224 void 2225 CMSCollector::print_on_error(outputStream* st) { 2226 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2227 if (collector != NULL) { 2228 CMSBitMap* bitmap = &collector->_markBitMap; 2229 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2230 bitmap->print_on_error(st, " Bits: "); 2231 2232 st->cr(); 2233 2234 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2235 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2236 mut_bitmap->print_on_error(st, " Bits: "); 2237 } 2238 } 2239 2240 //////////////////////////////////////////////////////// 2241 // CMS Verification Support 2242 //////////////////////////////////////////////////////// 2243 // Following the remark phase, the following invariant 2244 // should hold -- each object in the CMS heap which is 2245 // marked in markBitMap() should be marked in the verification_mark_bm(). 2246 2247 class VerifyMarkedClosure: public BitMapClosure { 2248 CMSBitMap* _marks; 2249 bool _failed; 2250 2251 public: 2252 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2253 2254 bool do_bit(size_t offset) { 2255 HeapWord* addr = _marks->offsetToHeapWord(offset); 2256 if (!_marks->isMarked(addr)) { 2257 Log(gc, verify) log; 2258 ResourceMark rm; 2259 LogStream ls(log.error()); 2260 oop(addr)->print_on(&ls); 2261 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2262 _failed = true; 2263 } 2264 return true; 2265 } 2266 2267 bool failed() { return _failed; } 2268 }; 2269 2270 bool CMSCollector::verify_after_remark() { 2271 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2272 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2273 static bool init = false; 2274 2275 assert(SafepointSynchronize::is_at_safepoint(), 2276 "Else mutations in object graph will make answer suspect"); 2277 assert(have_cms_token(), 2278 "Else there may be mutual interference in use of " 2279 " verification data structures"); 2280 assert(_collectorState > Marking && _collectorState <= Sweeping, 2281 "Else marking info checked here may be obsolete"); 2282 assert(haveFreelistLocks(), "must hold free list locks"); 2283 assert_lock_strong(bitMapLock()); 2284 2285 2286 // Allocate marking bit map if not already allocated 2287 if (!init) { // first time 2288 if (!verification_mark_bm()->allocate(_span)) { 2289 return false; 2290 } 2291 init = true; 2292 } 2293 2294 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2295 2296 // Turn off refs discovery -- so we will be tracing through refs. 2297 // This is as intended, because by this time 2298 // GC must already have cleared any refs that need to be cleared, 2299 // and traced those that need to be marked; moreover, 2300 // the marking done here is not going to interfere in any 2301 // way with the marking information used by GC. 2302 NoRefDiscovery no_discovery(ref_processor()); 2303 2304 #if COMPILER2_OR_JVMCI 2305 DerivedPointerTableDeactivate dpt_deact; 2306 #endif 2307 2308 // Clear any marks from a previous round 2309 verification_mark_bm()->clear_all(); 2310 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2311 verify_work_stacks_empty(); 2312 2313 CMSHeap* heap = CMSHeap::heap(); 2314 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2315 // Update the saved marks which may affect the root scans. 2316 heap->save_marks(); 2317 2318 if (CMSRemarkVerifyVariant == 1) { 2319 // In this first variant of verification, we complete 2320 // all marking, then check if the new marks-vector is 2321 // a subset of the CMS marks-vector. 2322 verify_after_remark_work_1(); 2323 } else { 2324 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2325 // In this second variant of verification, we flag an error 2326 // (i.e. an object reachable in the new marks-vector not reachable 2327 // in the CMS marks-vector) immediately, also indicating the 2328 // identify of an object (A) that references the unmarked object (B) -- 2329 // presumably, a mutation to A failed to be picked up by preclean/remark? 2330 verify_after_remark_work_2(); 2331 } 2332 2333 return true; 2334 } 2335 2336 void CMSCollector::verify_after_remark_work_1() { 2337 ResourceMark rm; 2338 HandleMark hm; 2339 CMSHeap* heap = CMSHeap::heap(); 2340 2341 // Get a clear set of claim bits for the roots processing to work with. 2342 ClassLoaderDataGraph::clear_claimed_marks(); 2343 2344 // Mark from roots one level into CMS 2345 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2346 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2347 2348 { 2349 StrongRootsScope srs(1); 2350 2351 heap->cms_process_roots(&srs, 2352 true, // young gen as roots 2353 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2354 should_unload_classes(), 2355 ¬Older, 2356 NULL); 2357 } 2358 2359 // Now mark from the roots 2360 MarkFromRootsClosure markFromRootsClosure(this, _span, 2361 verification_mark_bm(), verification_mark_stack(), 2362 false /* don't yield */, true /* verifying */); 2363 assert(_restart_addr == NULL, "Expected pre-condition"); 2364 verification_mark_bm()->iterate(&markFromRootsClosure); 2365 while (_restart_addr != NULL) { 2366 // Deal with stack overflow: by restarting at the indicated 2367 // address. 2368 HeapWord* ra = _restart_addr; 2369 markFromRootsClosure.reset(ra); 2370 _restart_addr = NULL; 2371 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2372 } 2373 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2374 verify_work_stacks_empty(); 2375 2376 // Marking completed -- now verify that each bit marked in 2377 // verification_mark_bm() is also marked in markBitMap(); flag all 2378 // errors by printing corresponding objects. 2379 VerifyMarkedClosure vcl(markBitMap()); 2380 verification_mark_bm()->iterate(&vcl); 2381 if (vcl.failed()) { 2382 Log(gc, verify) log; 2383 log.error("Failed marking verification after remark"); 2384 ResourceMark rm; 2385 LogStream ls(log.error()); 2386 heap->print_on(&ls); 2387 fatal("CMS: failed marking verification after remark"); 2388 } 2389 } 2390 2391 class VerifyCLDOopsCLDClosure : public CLDClosure { 2392 class VerifyCLDOopsClosure : public OopClosure { 2393 CMSBitMap* _bitmap; 2394 public: 2395 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2396 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2397 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2398 } _oop_closure; 2399 public: 2400 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2401 void do_cld(ClassLoaderData* cld) { 2402 cld->oops_do(&_oop_closure, ClassLoaderData::_claim_none, false); 2403 } 2404 }; 2405 2406 void CMSCollector::verify_after_remark_work_2() { 2407 ResourceMark rm; 2408 HandleMark hm; 2409 CMSHeap* heap = CMSHeap::heap(); 2410 2411 // Get a clear set of claim bits for the roots processing to work with. 2412 ClassLoaderDataGraph::clear_claimed_marks(); 2413 2414 // Mark from roots one level into CMS 2415 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2416 markBitMap()); 2417 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2418 2419 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2420 2421 { 2422 StrongRootsScope srs(1); 2423 2424 heap->cms_process_roots(&srs, 2425 true, // young gen as roots 2426 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2427 should_unload_classes(), 2428 ¬Older, 2429 &cld_closure); 2430 } 2431 2432 // Now mark from the roots 2433 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2434 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2435 assert(_restart_addr == NULL, "Expected pre-condition"); 2436 verification_mark_bm()->iterate(&markFromRootsClosure); 2437 while (_restart_addr != NULL) { 2438 // Deal with stack overflow: by restarting at the indicated 2439 // address. 2440 HeapWord* ra = _restart_addr; 2441 markFromRootsClosure.reset(ra); 2442 _restart_addr = NULL; 2443 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2444 } 2445 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2446 verify_work_stacks_empty(); 2447 2448 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2449 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2450 2451 // Marking completed -- now verify that each bit marked in 2452 // verification_mark_bm() is also marked in markBitMap(); flag all 2453 // errors by printing corresponding objects. 2454 VerifyMarkedClosure vcl(markBitMap()); 2455 verification_mark_bm()->iterate(&vcl); 2456 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2457 } 2458 2459 void ConcurrentMarkSweepGeneration::save_marks() { 2460 // delegate to CMS space 2461 cmsSpace()->save_marks(); 2462 } 2463 2464 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2465 return cmsSpace()->no_allocs_since_save_marks(); 2466 } 2467 2468 void 2469 ConcurrentMarkSweepGeneration::oop_iterate(OopIterateClosure* cl) { 2470 if (freelistLock()->owned_by_self()) { 2471 Generation::oop_iterate(cl); 2472 } else { 2473 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2474 Generation::oop_iterate(cl); 2475 } 2476 } 2477 2478 void 2479 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2480 if (freelistLock()->owned_by_self()) { 2481 Generation::object_iterate(cl); 2482 } else { 2483 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2484 Generation::object_iterate(cl); 2485 } 2486 } 2487 2488 void 2489 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2490 if (freelistLock()->owned_by_self()) { 2491 Generation::safe_object_iterate(cl); 2492 } else { 2493 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2494 Generation::safe_object_iterate(cl); 2495 } 2496 } 2497 2498 void 2499 ConcurrentMarkSweepGeneration::post_compact() { 2500 } 2501 2502 void 2503 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2504 // Fix the linear allocation blocks to look like free blocks. 2505 2506 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2507 // are not called when the heap is verified during universe initialization and 2508 // at vm shutdown. 2509 if (freelistLock()->owned_by_self()) { 2510 cmsSpace()->prepare_for_verify(); 2511 } else { 2512 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2513 cmsSpace()->prepare_for_verify(); 2514 } 2515 } 2516 2517 void 2518 ConcurrentMarkSweepGeneration::verify() { 2519 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2520 // are not called when the heap is verified during universe initialization and 2521 // at vm shutdown. 2522 if (freelistLock()->owned_by_self()) { 2523 cmsSpace()->verify(); 2524 } else { 2525 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2526 cmsSpace()->verify(); 2527 } 2528 } 2529 2530 void CMSCollector::verify() { 2531 _cmsGen->verify(); 2532 } 2533 2534 #ifndef PRODUCT 2535 bool CMSCollector::overflow_list_is_empty() const { 2536 assert(_num_par_pushes >= 0, "Inconsistency"); 2537 if (_overflow_list == NULL) { 2538 assert(_num_par_pushes == 0, "Inconsistency"); 2539 } 2540 return _overflow_list == NULL; 2541 } 2542 2543 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2544 // merely consolidate assertion checks that appear to occur together frequently. 2545 void CMSCollector::verify_work_stacks_empty() const { 2546 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2547 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2548 } 2549 2550 void CMSCollector::verify_overflow_empty() const { 2551 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2552 assert(no_preserved_marks(), "No preserved marks"); 2553 } 2554 #endif // PRODUCT 2555 2556 // Decide if we want to enable class unloading as part of the 2557 // ensuing concurrent GC cycle. We will collect and 2558 // unload classes if it's the case that: 2559 // (a) class unloading is enabled at the command line, and 2560 // (b) old gen is getting really full 2561 // NOTE: Provided there is no change in the state of the heap between 2562 // calls to this method, it should have idempotent results. Moreover, 2563 // its results should be monotonically increasing (i.e. going from 0 to 1, 2564 // but not 1 to 0) between successive calls between which the heap was 2565 // not collected. For the implementation below, it must thus rely on 2566 // the property that concurrent_cycles_since_last_unload() 2567 // will not decrease unless a collection cycle happened and that 2568 // _cmsGen->is_too_full() are 2569 // themselves also monotonic in that sense. See check_monotonicity() 2570 // below. 2571 void CMSCollector::update_should_unload_classes() { 2572 _should_unload_classes = false; 2573 if (CMSClassUnloadingEnabled) { 2574 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2575 CMSClassUnloadingMaxInterval) 2576 || _cmsGen->is_too_full(); 2577 } 2578 } 2579 2580 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2581 bool res = should_concurrent_collect(); 2582 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2583 return res; 2584 } 2585 2586 void CMSCollector::setup_cms_unloading_and_verification_state() { 2587 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2588 || VerifyBeforeExit; 2589 const int rso = GenCollectedHeap::SO_AllCodeCache; 2590 2591 // We set the proper root for this CMS cycle here. 2592 if (should_unload_classes()) { // Should unload classes this cycle 2593 remove_root_scanning_option(rso); // Shrink the root set appropriately 2594 set_verifying(should_verify); // Set verification state for this cycle 2595 return; // Nothing else needs to be done at this time 2596 } 2597 2598 // Not unloading classes this cycle 2599 assert(!should_unload_classes(), "Inconsistency!"); 2600 2601 // If we are not unloading classes then add SO_AllCodeCache to root 2602 // scanning options. 2603 add_root_scanning_option(rso); 2604 2605 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2606 set_verifying(true); 2607 } else if (verifying() && !should_verify) { 2608 // We were verifying, but some verification flags got disabled. 2609 set_verifying(false); 2610 // Exclude symbols, strings and code cache elements from root scanning to 2611 // reduce IM and RM pauses. 2612 remove_root_scanning_option(rso); 2613 } 2614 } 2615 2616 2617 #ifndef PRODUCT 2618 HeapWord* CMSCollector::block_start(const void* p) const { 2619 const HeapWord* addr = (HeapWord*)p; 2620 if (_span.contains(p)) { 2621 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2622 return _cmsGen->cmsSpace()->block_start(p); 2623 } 2624 } 2625 return NULL; 2626 } 2627 #endif 2628 2629 HeapWord* 2630 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2631 bool tlab, 2632 bool parallel) { 2633 CMSSynchronousYieldRequest yr; 2634 assert(!tlab, "Can't deal with TLAB allocation"); 2635 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2636 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2637 if (GCExpandToAllocateDelayMillis > 0) { 2638 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2639 } 2640 return have_lock_and_allocate(word_size, tlab); 2641 } 2642 2643 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2644 size_t bytes, 2645 size_t expand_bytes, 2646 CMSExpansionCause::Cause cause) 2647 { 2648 2649 bool success = expand(bytes, expand_bytes); 2650 2651 // remember why we expanded; this information is used 2652 // by shouldConcurrentCollect() when making decisions on whether to start 2653 // a new CMS cycle. 2654 if (success) { 2655 set_expansion_cause(cause); 2656 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2657 } 2658 } 2659 2660 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2661 HeapWord* res = NULL; 2662 MutexLocker x(ParGCRareEvent_lock); 2663 while (true) { 2664 // Expansion by some other thread might make alloc OK now: 2665 res = ps->lab.alloc(word_sz); 2666 if (res != NULL) return res; 2667 // If there's not enough expansion space available, give up. 2668 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2669 return NULL; 2670 } 2671 // Otherwise, we try expansion. 2672 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2673 // Now go around the loop and try alloc again; 2674 // A competing par_promote might beat us to the expansion space, 2675 // so we may go around the loop again if promotion fails again. 2676 if (GCExpandToAllocateDelayMillis > 0) { 2677 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2678 } 2679 } 2680 } 2681 2682 2683 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2684 PromotionInfo* promo) { 2685 MutexLocker x(ParGCRareEvent_lock); 2686 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2687 while (true) { 2688 // Expansion by some other thread might make alloc OK now: 2689 if (promo->ensure_spooling_space()) { 2690 assert(promo->has_spooling_space(), 2691 "Post-condition of successful ensure_spooling_space()"); 2692 return true; 2693 } 2694 // If there's not enough expansion space available, give up. 2695 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2696 return false; 2697 } 2698 // Otherwise, we try expansion. 2699 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2700 // Now go around the loop and try alloc again; 2701 // A competing allocation might beat us to the expansion space, 2702 // so we may go around the loop again if allocation fails again. 2703 if (GCExpandToAllocateDelayMillis > 0) { 2704 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2705 } 2706 } 2707 } 2708 2709 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2710 // Only shrink if a compaction was done so that all the free space 2711 // in the generation is in a contiguous block at the end. 2712 if (did_compact()) { 2713 CardGeneration::shrink(bytes); 2714 } 2715 } 2716 2717 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2718 assert_locked_or_safepoint(Heap_lock); 2719 } 2720 2721 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2722 assert_locked_or_safepoint(Heap_lock); 2723 assert_lock_strong(freelistLock()); 2724 log_trace(gc)("Shrinking of CMS not yet implemented"); 2725 return; 2726 } 2727 2728 2729 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2730 // phases. 2731 class CMSPhaseAccounting: public StackObj { 2732 public: 2733 CMSPhaseAccounting(CMSCollector *collector, 2734 const char *title); 2735 ~CMSPhaseAccounting(); 2736 2737 private: 2738 CMSCollector *_collector; 2739 const char *_title; 2740 GCTraceConcTime(Info, gc) _trace_time; 2741 2742 public: 2743 // Not MT-safe; so do not pass around these StackObj's 2744 // where they may be accessed by other threads. 2745 double wallclock_millis() { 2746 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2747 } 2748 }; 2749 2750 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2751 const char *title) : 2752 _collector(collector), _title(title), _trace_time(title) { 2753 2754 _collector->resetYields(); 2755 _collector->resetTimer(); 2756 _collector->startTimer(); 2757 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2758 } 2759 2760 CMSPhaseAccounting::~CMSPhaseAccounting() { 2761 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2762 _collector->stopTimer(); 2763 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2764 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2765 } 2766 2767 // CMS work 2768 2769 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2770 class CMSParMarkTask : public AbstractGangTask { 2771 protected: 2772 CMSCollector* _collector; 2773 uint _n_workers; 2774 OopStorage::ParState<false, false> _par_state_string; 2775 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2776 AbstractGangTask(name), 2777 _collector(collector), 2778 _n_workers(n_workers), 2779 _par_state_string(StringTable::weak_storage()) {} 2780 // Work method in support of parallel rescan ... of young gen spaces 2781 void do_young_space_rescan(OopsInGenClosure* cl, 2782 ContiguousSpace* space, 2783 HeapWord** chunk_array, size_t chunk_top); 2784 void work_on_young_gen_roots(OopsInGenClosure* cl); 2785 }; 2786 2787 // Parallel initial mark task 2788 class CMSParInitialMarkTask: public CMSParMarkTask { 2789 StrongRootsScope* _strong_roots_scope; 2790 public: 2791 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2792 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2793 _strong_roots_scope(strong_roots_scope) {} 2794 void work(uint worker_id); 2795 }; 2796 2797 // Checkpoint the roots into this generation from outside 2798 // this generation. [Note this initial checkpoint need only 2799 // be approximate -- we'll do a catch up phase subsequently.] 2800 void CMSCollector::checkpointRootsInitial() { 2801 assert(_collectorState == InitialMarking, "Wrong collector state"); 2802 check_correct_thread_executing(); 2803 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2804 2805 save_heap_summary(); 2806 report_heap_summary(GCWhen::BeforeGC); 2807 2808 ReferenceProcessor* rp = ref_processor(); 2809 assert(_restart_addr == NULL, "Control point invariant"); 2810 { 2811 // acquire locks for subsequent manipulations 2812 MutexLockerEx x(bitMapLock(), 2813 Mutex::_no_safepoint_check_flag); 2814 checkpointRootsInitialWork(); 2815 // enable ("weak") refs discovery 2816 rp->enable_discovery(); 2817 _collectorState = Marking; 2818 } 2819 } 2820 2821 void CMSCollector::checkpointRootsInitialWork() { 2822 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2823 assert(_collectorState == InitialMarking, "just checking"); 2824 2825 // Already have locks. 2826 assert_lock_strong(bitMapLock()); 2827 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2828 2829 // Setup the verification and class unloading state for this 2830 // CMS collection cycle. 2831 setup_cms_unloading_and_verification_state(); 2832 2833 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2834 2835 // Reset all the PLAB chunk arrays if necessary. 2836 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2837 reset_survivor_plab_arrays(); 2838 } 2839 2840 ResourceMark rm; 2841 HandleMark hm; 2842 2843 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2844 CMSHeap* heap = CMSHeap::heap(); 2845 2846 verify_work_stacks_empty(); 2847 verify_overflow_empty(); 2848 2849 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2850 // Update the saved marks which may affect the root scans. 2851 heap->save_marks(); 2852 2853 // weak reference processing has not started yet. 2854 ref_processor()->set_enqueuing_is_done(false); 2855 2856 // Need to remember all newly created CLDs, 2857 // so that we can guarantee that the remark finds them. 2858 ClassLoaderDataGraph::remember_new_clds(true); 2859 2860 // Whenever a CLD is found, it will be claimed before proceeding to mark 2861 // the klasses. The claimed marks need to be cleared before marking starts. 2862 ClassLoaderDataGraph::clear_claimed_marks(); 2863 2864 print_eden_and_survivor_chunk_arrays(); 2865 2866 { 2867 #if COMPILER2_OR_JVMCI 2868 DerivedPointerTableDeactivate dpt_deact; 2869 #endif 2870 if (CMSParallelInitialMarkEnabled) { 2871 // The parallel version. 2872 WorkGang* workers = heap->workers(); 2873 assert(workers != NULL, "Need parallel worker threads."); 2874 uint n_workers = workers->active_workers(); 2875 2876 StrongRootsScope srs(n_workers); 2877 2878 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2879 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2880 // If the total workers is greater than 1, then multiple workers 2881 // may be used at some time and the initialization has been set 2882 // such that the single threaded path cannot be used. 2883 if (workers->total_workers() > 1) { 2884 workers->run_task(&tsk); 2885 } else { 2886 tsk.work(0); 2887 } 2888 } else { 2889 // The serial version. 2890 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2891 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2892 2893 StrongRootsScope srs(1); 2894 2895 heap->cms_process_roots(&srs, 2896 true, // young gen as roots 2897 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2898 should_unload_classes(), 2899 ¬Older, 2900 &cld_closure); 2901 } 2902 } 2903 2904 // Clear mod-union table; it will be dirtied in the prologue of 2905 // CMS generation per each young generation collection. 2906 2907 assert(_modUnionTable.isAllClear(), 2908 "Was cleared in most recent final checkpoint phase" 2909 " or no bits are set in the gc_prologue before the start of the next " 2910 "subsequent marking phase."); 2911 2912 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2913 2914 // Save the end of the used_region of the constituent generations 2915 // to be used to limit the extent of sweep in each generation. 2916 save_sweep_limits(); 2917 verify_overflow_empty(); 2918 } 2919 2920 bool CMSCollector::markFromRoots() { 2921 // we might be tempted to assert that: 2922 // assert(!SafepointSynchronize::is_at_safepoint(), 2923 // "inconsistent argument?"); 2924 // However that wouldn't be right, because it's possible that 2925 // a safepoint is indeed in progress as a young generation 2926 // stop-the-world GC happens even as we mark in this generation. 2927 assert(_collectorState == Marking, "inconsistent state?"); 2928 check_correct_thread_executing(); 2929 verify_overflow_empty(); 2930 2931 // Weak ref discovery note: We may be discovering weak 2932 // refs in this generation concurrent (but interleaved) with 2933 // weak ref discovery by the young generation collector. 2934 2935 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2936 GCTraceCPUTime tcpu; 2937 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2938 bool res = markFromRootsWork(); 2939 if (res) { 2940 _collectorState = Precleaning; 2941 } else { // We failed and a foreground collection wants to take over 2942 assert(_foregroundGCIsActive, "internal state inconsistency"); 2943 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2944 log_debug(gc)("bailing out to foreground collection"); 2945 } 2946 verify_overflow_empty(); 2947 return res; 2948 } 2949 2950 bool CMSCollector::markFromRootsWork() { 2951 // iterate over marked bits in bit map, doing a full scan and mark 2952 // from these roots using the following algorithm: 2953 // . if oop is to the right of the current scan pointer, 2954 // mark corresponding bit (we'll process it later) 2955 // . else (oop is to left of current scan pointer) 2956 // push oop on marking stack 2957 // . drain the marking stack 2958 2959 // Note that when we do a marking step we need to hold the 2960 // bit map lock -- recall that direct allocation (by mutators) 2961 // and promotion (by the young generation collector) is also 2962 // marking the bit map. [the so-called allocate live policy.] 2963 // Because the implementation of bit map marking is not 2964 // robust wrt simultaneous marking of bits in the same word, 2965 // we need to make sure that there is no such interference 2966 // between concurrent such updates. 2967 2968 // already have locks 2969 assert_lock_strong(bitMapLock()); 2970 2971 verify_work_stacks_empty(); 2972 verify_overflow_empty(); 2973 bool result = false; 2974 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2975 result = do_marking_mt(); 2976 } else { 2977 result = do_marking_st(); 2978 } 2979 return result; 2980 } 2981 2982 // Forward decl 2983 class CMSConcMarkingTask; 2984 2985 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2986 CMSCollector* _collector; 2987 CMSConcMarkingTask* _task; 2988 public: 2989 virtual void yield(); 2990 2991 // "n_threads" is the number of threads to be terminated. 2992 // "queue_set" is a set of work queues of other threads. 2993 // "collector" is the CMS collector associated with this task terminator. 2994 // "yield" indicates whether we need the gang as a whole to yield. 2995 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2996 ParallelTaskTerminator(n_threads, queue_set), 2997 _collector(collector) { } 2998 2999 void set_task(CMSConcMarkingTask* task) { 3000 _task = task; 3001 } 3002 }; 3003 3004 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3005 CMSConcMarkingTask* _task; 3006 public: 3007 bool should_exit_termination(); 3008 void set_task(CMSConcMarkingTask* task) { 3009 _task = task; 3010 } 3011 }; 3012 3013 // MT Concurrent Marking Task 3014 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3015 CMSCollector* _collector; 3016 uint _n_workers; // requested/desired # workers 3017 bool _result; 3018 CompactibleFreeListSpace* _cms_space; 3019 char _pad_front[64]; // padding to ... 3020 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3021 char _pad_back[64]; 3022 HeapWord* _restart_addr; 3023 3024 // Exposed here for yielding support 3025 Mutex* const _bit_map_lock; 3026 3027 // The per thread work queues, available here for stealing 3028 OopTaskQueueSet* _task_queues; 3029 3030 // Termination (and yielding) support 3031 CMSConcMarkingTerminator _term; 3032 CMSConcMarkingTerminatorTerminator _term_term; 3033 3034 public: 3035 CMSConcMarkingTask(CMSCollector* collector, 3036 CompactibleFreeListSpace* cms_space, 3037 YieldingFlexibleWorkGang* workers, 3038 OopTaskQueueSet* task_queues): 3039 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3040 _collector(collector), 3041 _n_workers(0), 3042 _result(true), 3043 _cms_space(cms_space), 3044 _bit_map_lock(collector->bitMapLock()), 3045 _task_queues(task_queues), 3046 _term(_n_workers, task_queues, _collector) 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->try_claim_task(/* 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 MetadataVisitingOopIterateClosure { 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 MetadataVisitingOopIterateClosure(collector->ref_processor()), 3322 _collector(collector), 3323 _task(task), 3324 _span(collector->_span), 3325 _bit_map(bit_map), 3326 _overflow_stack(overflow_stack), 3327 _work_queue(work_queue) 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::trim_queue(size_t max) { 3386 while (_work_queue->size() > max) { 3387 oop new_oop; 3388 if (_work_queue->pop_local(new_oop)) { 3389 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3390 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3391 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3392 new_oop->oop_iterate(this); // do_oop() above 3393 do_yield_check(); 3394 } 3395 } 3396 } 3397 3398 // Upon stack overflow, we discard (part of) the stack, 3399 // remembering the least address amongst those discarded 3400 // in CMSCollector's _restart_address. 3401 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3402 // We need to do this under a mutex to prevent other 3403 // workers from interfering with the work done below. 3404 MutexLockerEx ml(_overflow_stack->par_lock(), 3405 Mutex::_no_safepoint_check_flag); 3406 // Remember the least grey address discarded 3407 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3408 _collector->lower_restart_addr(ra); 3409 _overflow_stack->reset(); // discard stack contents 3410 _overflow_stack->expand(); // expand the stack if possible 3411 } 3412 3413 3414 void CMSConcMarkingTask::do_work_steal(int i) { 3415 OopTaskQueue* work_q = work_queue(i); 3416 oop obj_to_scan; 3417 CMSBitMap* bm = &(_collector->_markBitMap); 3418 CMSMarkStack* ovflw = &(_collector->_markStack); 3419 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3420 while (true) { 3421 cl.trim_queue(0); 3422 assert(work_q->size() == 0, "Should have been emptied above"); 3423 if (get_work_from_overflow_stack(ovflw, work_q)) { 3424 // Can't assert below because the work obtained from the 3425 // overflow stack may already have been stolen from us. 3426 // assert(work_q->size() > 0, "Work from overflow stack"); 3427 continue; 3428 } else if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 3429 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3430 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3431 obj_to_scan->oop_iterate(&cl); 3432 } else if (terminator()->offer_termination(&_term_term)) { 3433 assert(work_q->size() == 0, "Impossible!"); 3434 break; 3435 } else if (yielding() || should_yield()) { 3436 yield(); 3437 } 3438 } 3439 } 3440 3441 // This is run by the CMS (coordinator) thread. 3442 void CMSConcMarkingTask::coordinator_yield() { 3443 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3444 "CMS thread should hold CMS token"); 3445 // First give up the locks, then yield, then re-lock 3446 // We should probably use a constructor/destructor idiom to 3447 // do this unlock/lock or modify the MutexUnlocker class to 3448 // serve our purpose. XXX 3449 assert_lock_strong(_bit_map_lock); 3450 _bit_map_lock->unlock(); 3451 ConcurrentMarkSweepThread::desynchronize(true); 3452 _collector->stopTimer(); 3453 _collector->incrementYields(); 3454 3455 // It is possible for whichever thread initiated the yield request 3456 // not to get a chance to wake up and take the bitmap lock between 3457 // this thread releasing it and reacquiring it. So, while the 3458 // should_yield() flag is on, let's sleep for a bit to give the 3459 // other thread a chance to wake up. The limit imposed on the number 3460 // of iterations is defensive, to avoid any unforseen circumstances 3461 // putting us into an infinite loop. Since it's always been this 3462 // (coordinator_yield()) method that was observed to cause the 3463 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3464 // which is by default non-zero. For the other seven methods that 3465 // also perform the yield operation, as are using a different 3466 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3467 // can enable the sleeping for those methods too, if necessary. 3468 // See 6442774. 3469 // 3470 // We really need to reconsider the synchronization between the GC 3471 // thread and the yield-requesting threads in the future and we 3472 // should really use wait/notify, which is the recommended 3473 // way of doing this type of interaction. Additionally, we should 3474 // consolidate the eight methods that do the yield operation and they 3475 // are almost identical into one for better maintainability and 3476 // readability. See 6445193. 3477 // 3478 // Tony 2006.06.29 3479 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3480 ConcurrentMarkSweepThread::should_yield() && 3481 !CMSCollector::foregroundGCIsActive(); ++i) { 3482 os::sleep(Thread::current(), 1, false); 3483 } 3484 3485 ConcurrentMarkSweepThread::synchronize(true); 3486 _bit_map_lock->lock_without_safepoint_check(); 3487 _collector->startTimer(); 3488 } 3489 3490 bool CMSCollector::do_marking_mt() { 3491 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3492 uint num_workers = WorkerPolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3493 conc_workers()->active_workers(), 3494 Threads::number_of_non_daemon_threads()); 3495 num_workers = conc_workers()->update_active_workers(num_workers); 3496 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3497 3498 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3499 3500 CMSConcMarkingTask tsk(this, 3501 cms_space, 3502 conc_workers(), 3503 task_queues()); 3504 3505 // Since the actual number of workers we get may be different 3506 // from the number we requested above, do we need to do anything different 3507 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3508 // class?? XXX 3509 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3510 3511 // Refs discovery is already non-atomic. 3512 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3513 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3514 conc_workers()->start_task(&tsk); 3515 while (tsk.yielded()) { 3516 tsk.coordinator_yield(); 3517 conc_workers()->continue_task(&tsk); 3518 } 3519 // If the task was aborted, _restart_addr will be non-NULL 3520 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3521 while (_restart_addr != NULL) { 3522 // XXX For now we do not make use of ABORTED state and have not 3523 // yet implemented the right abort semantics (even in the original 3524 // single-threaded CMS case). That needs some more investigation 3525 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3526 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3527 // If _restart_addr is non-NULL, a marking stack overflow 3528 // occurred; we need to do a fresh marking iteration from the 3529 // indicated restart address. 3530 if (_foregroundGCIsActive) { 3531 // We may be running into repeated stack overflows, having 3532 // reached the limit of the stack size, while making very 3533 // slow forward progress. It may be best to bail out and 3534 // let the foreground collector do its job. 3535 // Clear _restart_addr, so that foreground GC 3536 // works from scratch. This avoids the headache of 3537 // a "rescan" which would otherwise be needed because 3538 // of the dirty mod union table & card table. 3539 _restart_addr = NULL; 3540 return false; 3541 } 3542 // Adjust the task to restart from _restart_addr 3543 tsk.reset(_restart_addr); 3544 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3545 _restart_addr); 3546 _restart_addr = NULL; 3547 // Get the workers going again 3548 conc_workers()->start_task(&tsk); 3549 while (tsk.yielded()) { 3550 tsk.coordinator_yield(); 3551 conc_workers()->continue_task(&tsk); 3552 } 3553 } 3554 assert(tsk.completed(), "Inconsistency"); 3555 assert(tsk.result() == true, "Inconsistency"); 3556 return true; 3557 } 3558 3559 bool CMSCollector::do_marking_st() { 3560 ResourceMark rm; 3561 HandleMark hm; 3562 3563 // Temporarily make refs discovery single threaded (non-MT) 3564 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3565 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3566 &_markStack, CMSYield); 3567 // the last argument to iterate indicates whether the iteration 3568 // should be incremental with periodic yields. 3569 _markBitMap.iterate(&markFromRootsClosure); 3570 // If _restart_addr is non-NULL, a marking stack overflow 3571 // occurred; we need to do a fresh iteration from the 3572 // indicated restart address. 3573 while (_restart_addr != NULL) { 3574 if (_foregroundGCIsActive) { 3575 // We may be running into repeated stack overflows, having 3576 // reached the limit of the stack size, while making very 3577 // slow forward progress. It may be best to bail out and 3578 // let the foreground collector do its job. 3579 // Clear _restart_addr, so that foreground GC 3580 // works from scratch. This avoids the headache of 3581 // a "rescan" which would otherwise be needed because 3582 // of the dirty mod union table & card table. 3583 _restart_addr = NULL; 3584 return false; // indicating failure to complete marking 3585 } 3586 // Deal with stack overflow: 3587 // we restart marking from _restart_addr 3588 HeapWord* ra = _restart_addr; 3589 markFromRootsClosure.reset(ra); 3590 _restart_addr = NULL; 3591 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3592 } 3593 return true; 3594 } 3595 3596 void CMSCollector::preclean() { 3597 check_correct_thread_executing(); 3598 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3599 verify_work_stacks_empty(); 3600 verify_overflow_empty(); 3601 _abort_preclean = false; 3602 if (CMSPrecleaningEnabled) { 3603 if (!CMSEdenChunksRecordAlways) { 3604 _eden_chunk_index = 0; 3605 } 3606 size_t used = get_eden_used(); 3607 size_t capacity = get_eden_capacity(); 3608 // Don't start sampling unless we will get sufficiently 3609 // many samples. 3610 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3611 * CMSScheduleRemarkEdenPenetration)) { 3612 _start_sampling = true; 3613 } else { 3614 _start_sampling = false; 3615 } 3616 GCTraceCPUTime tcpu; 3617 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3618 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3619 } 3620 CMSTokenSync x(true); // is cms thread 3621 if (CMSPrecleaningEnabled) { 3622 sample_eden(); 3623 _collectorState = AbortablePreclean; 3624 } else { 3625 _collectorState = FinalMarking; 3626 } 3627 verify_work_stacks_empty(); 3628 verify_overflow_empty(); 3629 } 3630 3631 // Try and schedule the remark such that young gen 3632 // occupancy is CMSScheduleRemarkEdenPenetration %. 3633 void CMSCollector::abortable_preclean() { 3634 check_correct_thread_executing(); 3635 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3636 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3637 3638 // If Eden's current occupancy is below this threshold, 3639 // immediately schedule the remark; else preclean 3640 // past the next scavenge in an effort to 3641 // schedule the pause as described above. By choosing 3642 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3643 // we will never do an actual abortable preclean cycle. 3644 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3645 GCTraceCPUTime tcpu; 3646 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3647 // We need more smarts in the abortable preclean 3648 // loop below to deal with cases where allocation 3649 // in young gen is very very slow, and our precleaning 3650 // is running a losing race against a horde of 3651 // mutators intent on flooding us with CMS updates 3652 // (dirty cards). 3653 // One, admittedly dumb, strategy is to give up 3654 // after a certain number of abortable precleaning loops 3655 // or after a certain maximum time. We want to make 3656 // this smarter in the next iteration. 3657 // XXX FIX ME!!! YSR 3658 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3659 while (!(should_abort_preclean() || 3660 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3661 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3662 cumworkdone += workdone; 3663 loops++; 3664 // Voluntarily terminate abortable preclean phase if we have 3665 // been at it for too long. 3666 if ((CMSMaxAbortablePrecleanLoops != 0) && 3667 loops >= CMSMaxAbortablePrecleanLoops) { 3668 log_debug(gc)(" CMS: abort preclean due to loops "); 3669 break; 3670 } 3671 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3672 log_debug(gc)(" CMS: abort preclean due to time "); 3673 break; 3674 } 3675 // If we are doing little work each iteration, we should 3676 // take a short break. 3677 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3678 // Sleep for some time, waiting for work to accumulate 3679 stopTimer(); 3680 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3681 startTimer(); 3682 waited++; 3683 } 3684 } 3685 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3686 loops, waited, cumworkdone); 3687 } 3688 CMSTokenSync x(true); // is cms thread 3689 if (_collectorState != Idling) { 3690 assert(_collectorState == AbortablePreclean, 3691 "Spontaneous state transition?"); 3692 _collectorState = FinalMarking; 3693 } // Else, a foreground collection completed this CMS cycle. 3694 return; 3695 } 3696 3697 // Respond to an Eden sampling opportunity 3698 void CMSCollector::sample_eden() { 3699 // Make sure a young gc cannot sneak in between our 3700 // reading and recording of a sample. 3701 assert(Thread::current()->is_ConcurrentGC_thread(), 3702 "Only the cms thread may collect Eden samples"); 3703 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3704 "Should collect samples while holding CMS token"); 3705 if (!_start_sampling) { 3706 return; 3707 } 3708 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3709 // is populated by the young generation. 3710 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3711 if (_eden_chunk_index < _eden_chunk_capacity) { 3712 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3713 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3714 "Unexpected state of Eden"); 3715 // We'd like to check that what we just sampled is an oop-start address; 3716 // however, we cannot do that here since the object may not yet have been 3717 // initialized. So we'll instead do the check when we _use_ this sample 3718 // later. 3719 if (_eden_chunk_index == 0 || 3720 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3721 _eden_chunk_array[_eden_chunk_index-1]) 3722 >= CMSSamplingGrain)) { 3723 _eden_chunk_index++; // commit sample 3724 } 3725 } 3726 } 3727 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3728 size_t used = get_eden_used(); 3729 size_t capacity = get_eden_capacity(); 3730 assert(used <= capacity, "Unexpected state of Eden"); 3731 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3732 _abort_preclean = true; 3733 } 3734 } 3735 } 3736 3737 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3738 assert(_collectorState == Precleaning || 3739 _collectorState == AbortablePreclean, "incorrect state"); 3740 ResourceMark rm; 3741 HandleMark hm; 3742 3743 // Precleaning is currently not MT but the reference processor 3744 // may be set for MT. Disable it temporarily here. 3745 ReferenceProcessor* rp = ref_processor(); 3746 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3747 3748 // Do one pass of scrubbing the discovered reference lists 3749 // to remove any reference objects with strongly-reachable 3750 // referents. 3751 if (clean_refs) { 3752 CMSPrecleanRefsYieldClosure yield_cl(this); 3753 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3754 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3755 &_markStack, true /* preclean */); 3756 CMSDrainMarkingStackClosure complete_trace(this, 3757 _span, &_markBitMap, &_markStack, 3758 &keep_alive, true /* preclean */); 3759 3760 // We don't want this step to interfere with a young 3761 // collection because we don't want to take CPU 3762 // or memory bandwidth away from the young GC threads 3763 // (which may be as many as there are CPUs). 3764 // Note that we don't need to protect ourselves from 3765 // interference with mutators because they can't 3766 // manipulate the discovered reference lists nor affect 3767 // the computed reachability of the referents, the 3768 // only properties manipulated by the precleaning 3769 // of these reference lists. 3770 stopTimer(); 3771 CMSTokenSyncWithLocks x(true /* is cms thread */, 3772 bitMapLock()); 3773 startTimer(); 3774 sample_eden(); 3775 3776 // The following will yield to allow foreground 3777 // collection to proceed promptly. XXX YSR: 3778 // The code in this method may need further 3779 // tweaking for better performance and some restructuring 3780 // for cleaner interfaces. 3781 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3782 rp->preclean_discovered_references( 3783 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3784 gc_timer); 3785 } 3786 3787 if (clean_survivor) { // preclean the active survivor space(s) 3788 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3789 &_markBitMap, &_modUnionTable, 3790 &_markStack, true /* precleaning phase */); 3791 stopTimer(); 3792 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3793 bitMapLock()); 3794 startTimer(); 3795 unsigned int before_count = 3796 CMSHeap::heap()->total_collections(); 3797 SurvivorSpacePrecleanClosure 3798 sss_cl(this, _span, &_markBitMap, &_markStack, 3799 &pam_cl, before_count, CMSYield); 3800 _young_gen->from()->object_iterate_careful(&sss_cl); 3801 _young_gen->to()->object_iterate_careful(&sss_cl); 3802 } 3803 MarkRefsIntoAndScanClosure 3804 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3805 &_markStack, this, CMSYield, 3806 true /* precleaning phase */); 3807 // CAUTION: The following closure has persistent state that may need to 3808 // be reset upon a decrease in the sequence of addresses it 3809 // processes. 3810 ScanMarkedObjectsAgainCarefullyClosure 3811 smoac_cl(this, _span, 3812 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3813 3814 // Preclean dirty cards in ModUnionTable and CardTable using 3815 // appropriate convergence criterion; 3816 // repeat CMSPrecleanIter times unless we find that 3817 // we are losing. 3818 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3819 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3820 "Bad convergence multiplier"); 3821 assert(CMSPrecleanThreshold >= 100, 3822 "Unreasonably low CMSPrecleanThreshold"); 3823 3824 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3825 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3826 numIter < CMSPrecleanIter; 3827 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3828 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3829 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3830 // Either there are very few dirty cards, so re-mark 3831 // pause will be small anyway, or our pre-cleaning isn't 3832 // that much faster than the rate at which cards are being 3833 // dirtied, so we might as well stop and re-mark since 3834 // precleaning won't improve our re-mark time by much. 3835 if (curNumCards <= CMSPrecleanThreshold || 3836 (numIter > 0 && 3837 (curNumCards * CMSPrecleanDenominator > 3838 lastNumCards * CMSPrecleanNumerator))) { 3839 numIter++; 3840 cumNumCards += curNumCards; 3841 break; 3842 } 3843 } 3844 3845 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3846 3847 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3848 cumNumCards += curNumCards; 3849 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3850 curNumCards, cumNumCards, numIter); 3851 return cumNumCards; // as a measure of useful work done 3852 } 3853 3854 // PRECLEANING NOTES: 3855 // Precleaning involves: 3856 // . reading the bits of the modUnionTable and clearing the set bits. 3857 // . For the cards corresponding to the set bits, we scan the 3858 // objects on those cards. This means we need the free_list_lock 3859 // so that we can safely iterate over the CMS space when scanning 3860 // for oops. 3861 // . When we scan the objects, we'll be both reading and setting 3862 // marks in the marking bit map, so we'll need the marking bit map. 3863 // . For protecting _collector_state transitions, we take the CGC_lock. 3864 // Note that any races in the reading of of card table entries by the 3865 // CMS thread on the one hand and the clearing of those entries by the 3866 // VM thread or the setting of those entries by the mutator threads on the 3867 // other are quite benign. However, for efficiency it makes sense to keep 3868 // the VM thread from racing with the CMS thread while the latter is 3869 // dirty card info to the modUnionTable. We therefore also use the 3870 // CGC_lock to protect the reading of the card table and the mod union 3871 // table by the CM thread. 3872 // . We run concurrently with mutator updates, so scanning 3873 // needs to be done carefully -- we should not try to scan 3874 // potentially uninitialized objects. 3875 // 3876 // Locking strategy: While holding the CGC_lock, we scan over and 3877 // reset a maximal dirty range of the mod union / card tables, then lock 3878 // the free_list_lock and bitmap lock to do a full marking, then 3879 // release these locks; and repeat the cycle. This allows for a 3880 // certain amount of fairness in the sharing of these locks between 3881 // the CMS collector on the one hand, and the VM thread and the 3882 // mutators on the other. 3883 3884 // NOTE: preclean_mod_union_table() and preclean_card_table() 3885 // further below are largely identical; if you need to modify 3886 // one of these methods, please check the other method too. 3887 3888 size_t CMSCollector::preclean_mod_union_table( 3889 ConcurrentMarkSweepGeneration* old_gen, 3890 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3891 verify_work_stacks_empty(); 3892 verify_overflow_empty(); 3893 3894 // strategy: starting with the first card, accumulate contiguous 3895 // ranges of dirty cards; clear these cards, then scan the region 3896 // covered by these cards. 3897 3898 // Since all of the MUT is committed ahead, we can just use 3899 // that, in case the generations expand while we are precleaning. 3900 // It might also be fine to just use the committed part of the 3901 // generation, but we might potentially miss cards when the 3902 // generation is rapidly expanding while we are in the midst 3903 // of precleaning. 3904 HeapWord* startAddr = old_gen->reserved().start(); 3905 HeapWord* endAddr = old_gen->reserved().end(); 3906 3907 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3908 3909 size_t numDirtyCards, cumNumDirtyCards; 3910 HeapWord *nextAddr, *lastAddr; 3911 for (cumNumDirtyCards = numDirtyCards = 0, 3912 nextAddr = lastAddr = startAddr; 3913 nextAddr < endAddr; 3914 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3915 3916 ResourceMark rm; 3917 HandleMark hm; 3918 3919 MemRegion dirtyRegion; 3920 { 3921 stopTimer(); 3922 // Potential yield point 3923 CMSTokenSync ts(true); 3924 startTimer(); 3925 sample_eden(); 3926 // Get dirty region starting at nextOffset (inclusive), 3927 // simultaneously clearing it. 3928 dirtyRegion = 3929 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3930 assert(dirtyRegion.start() >= nextAddr, 3931 "returned region inconsistent?"); 3932 } 3933 // Remember where the next search should begin. 3934 // The returned region (if non-empty) is a right open interval, 3935 // so lastOffset is obtained from the right end of that 3936 // interval. 3937 lastAddr = dirtyRegion.end(); 3938 // Should do something more transparent and less hacky XXX 3939 numDirtyCards = 3940 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3941 3942 // We'll scan the cards in the dirty region (with periodic 3943 // yields for foreground GC as needed). 3944 if (!dirtyRegion.is_empty()) { 3945 assert(numDirtyCards > 0, "consistency check"); 3946 HeapWord* stop_point = NULL; 3947 stopTimer(); 3948 // Potential yield point 3949 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3950 bitMapLock()); 3951 startTimer(); 3952 { 3953 verify_work_stacks_empty(); 3954 verify_overflow_empty(); 3955 sample_eden(); 3956 stop_point = 3957 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3958 } 3959 if (stop_point != NULL) { 3960 // The careful iteration stopped early either because it found an 3961 // uninitialized object, or because we were in the midst of an 3962 // "abortable preclean", which should now be aborted. Redirty 3963 // the bits corresponding to the partially-scanned or unscanned 3964 // cards. We'll either restart at the next block boundary or 3965 // abort the preclean. 3966 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3967 "Should only be AbortablePreclean."); 3968 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3969 if (should_abort_preclean()) { 3970 break; // out of preclean loop 3971 } else { 3972 // Compute the next address at which preclean should pick up; 3973 // might need bitMapLock in order to read P-bits. 3974 lastAddr = next_card_start_after_block(stop_point); 3975 } 3976 } 3977 } else { 3978 assert(lastAddr == endAddr, "consistency check"); 3979 assert(numDirtyCards == 0, "consistency check"); 3980 break; 3981 } 3982 } 3983 verify_work_stacks_empty(); 3984 verify_overflow_empty(); 3985 return cumNumDirtyCards; 3986 } 3987 3988 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3989 // below are largely identical; if you need to modify 3990 // one of these methods, please check the other method too. 3991 3992 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3993 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3994 // strategy: it's similar to precleamModUnionTable above, in that 3995 // we accumulate contiguous ranges of dirty cards, mark these cards 3996 // precleaned, then scan the region covered by these cards. 3997 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3998 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3999 4000 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 4001 4002 size_t numDirtyCards, cumNumDirtyCards; 4003 HeapWord *lastAddr, *nextAddr; 4004 4005 for (cumNumDirtyCards = numDirtyCards = 0, 4006 nextAddr = lastAddr = startAddr; 4007 nextAddr < endAddr; 4008 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4009 4010 ResourceMark rm; 4011 HandleMark hm; 4012 4013 MemRegion dirtyRegion; 4014 { 4015 // See comments in "Precleaning notes" above on why we 4016 // do this locking. XXX Could the locking overheads be 4017 // too high when dirty cards are sparse? [I don't think so.] 4018 stopTimer(); 4019 CMSTokenSync x(true); // is cms thread 4020 startTimer(); 4021 sample_eden(); 4022 // Get and clear dirty region from card table 4023 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4024 true, 4025 CardTable::precleaned_card_val()); 4026 4027 assert(dirtyRegion.start() >= nextAddr, 4028 "returned region inconsistent?"); 4029 } 4030 lastAddr = dirtyRegion.end(); 4031 numDirtyCards = 4032 dirtyRegion.word_size()/CardTable::card_size_in_words; 4033 4034 if (!dirtyRegion.is_empty()) { 4035 stopTimer(); 4036 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4037 startTimer(); 4038 sample_eden(); 4039 verify_work_stacks_empty(); 4040 verify_overflow_empty(); 4041 HeapWord* stop_point = 4042 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4043 if (stop_point != NULL) { 4044 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4045 "Should only be AbortablePreclean."); 4046 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4047 if (should_abort_preclean()) { 4048 break; // out of preclean loop 4049 } else { 4050 // Compute the next address at which preclean should pick up. 4051 lastAddr = next_card_start_after_block(stop_point); 4052 } 4053 } 4054 } else { 4055 break; 4056 } 4057 } 4058 verify_work_stacks_empty(); 4059 verify_overflow_empty(); 4060 return cumNumDirtyCards; 4061 } 4062 4063 class PrecleanCLDClosure : public CLDClosure { 4064 MetadataVisitingOopsInGenClosure* _cm_closure; 4065 public: 4066 PrecleanCLDClosure(MetadataVisitingOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4067 void do_cld(ClassLoaderData* cld) { 4068 if (cld->has_accumulated_modified_oops()) { 4069 cld->clear_accumulated_modified_oops(); 4070 4071 _cm_closure->do_cld(cld); 4072 } 4073 } 4074 }; 4075 4076 // The freelist lock is needed to prevent asserts, is it really needed? 4077 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4078 // Needed to walk CLDG 4079 MutexLocker ml(ClassLoaderDataGraph_lock); 4080 4081 cl->set_freelistLock(freelistLock); 4082 4083 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4084 4085 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4086 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4087 PrecleanCLDClosure preclean_closure(cl); 4088 ClassLoaderDataGraph::cld_do(&preclean_closure); 4089 4090 verify_work_stacks_empty(); 4091 verify_overflow_empty(); 4092 } 4093 4094 void CMSCollector::checkpointRootsFinal() { 4095 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4096 check_correct_thread_executing(); 4097 // world is stopped at this checkpoint 4098 assert(SafepointSynchronize::is_at_safepoint(), 4099 "world should be stopped"); 4100 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4101 4102 verify_work_stacks_empty(); 4103 verify_overflow_empty(); 4104 4105 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4106 _young_gen->used() / K, _young_gen->capacity() / K); 4107 { 4108 if (CMSScavengeBeforeRemark) { 4109 CMSHeap* heap = CMSHeap::heap(); 4110 // Temporarily set flag to false, GCH->do_collection will 4111 // expect it to be false and set to true 4112 FlagSetting fl(heap->_is_gc_active, false); 4113 4114 heap->do_collection(true, // full (i.e. force, see below) 4115 false, // !clear_all_soft_refs 4116 0, // size 4117 false, // is_tlab 4118 GenCollectedHeap::YoungGen // type 4119 ); 4120 } 4121 FreelistLocker x(this); 4122 MutexLockerEx y(bitMapLock(), 4123 Mutex::_no_safepoint_check_flag); 4124 checkpointRootsFinalWork(); 4125 } 4126 verify_work_stacks_empty(); 4127 verify_overflow_empty(); 4128 } 4129 4130 void CMSCollector::checkpointRootsFinalWork() { 4131 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4132 4133 assert(haveFreelistLocks(), "must have free list locks"); 4134 assert_lock_strong(bitMapLock()); 4135 4136 ResourceMark rm; 4137 HandleMark hm; 4138 4139 CMSHeap* heap = CMSHeap::heap(); 4140 4141 if (should_unload_classes()) { 4142 CodeCache::gc_prologue(); 4143 } 4144 assert(haveFreelistLocks(), "must have free list locks"); 4145 assert_lock_strong(bitMapLock()); 4146 4147 // We might assume that we need not fill TLAB's when 4148 // CMSScavengeBeforeRemark is set, because we may have just done 4149 // a scavenge which would have filled all TLAB's -- and besides 4150 // Eden would be empty. This however may not always be the case -- 4151 // for instance although we asked for a scavenge, it may not have 4152 // happened because of a JNI critical section. We probably need 4153 // a policy for deciding whether we can in that case wait until 4154 // the critical section releases and then do the remark following 4155 // the scavenge, and skip it here. In the absence of that policy, 4156 // or of an indication of whether the scavenge did indeed occur, 4157 // we cannot rely on TLAB's having been filled and must do 4158 // so here just in case a scavenge did not happen. 4159 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4160 // Update the saved marks which may affect the root scans. 4161 heap->save_marks(); 4162 4163 print_eden_and_survivor_chunk_arrays(); 4164 4165 { 4166 #if COMPILER2_OR_JVMCI 4167 DerivedPointerTableDeactivate dpt_deact; 4168 #endif 4169 4170 // Note on the role of the mod union table: 4171 // Since the marker in "markFromRoots" marks concurrently with 4172 // mutators, it is possible for some reachable objects not to have been 4173 // scanned. For instance, an only reference to an object A was 4174 // placed in object B after the marker scanned B. Unless B is rescanned, 4175 // A would be collected. Such updates to references in marked objects 4176 // are detected via the mod union table which is the set of all cards 4177 // dirtied since the first checkpoint in this GC cycle and prior to 4178 // the most recent young generation GC, minus those cleaned up by the 4179 // concurrent precleaning. 4180 if (CMSParallelRemarkEnabled) { 4181 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4182 do_remark_parallel(); 4183 } else { 4184 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4185 do_remark_non_parallel(); 4186 } 4187 } 4188 verify_work_stacks_empty(); 4189 verify_overflow_empty(); 4190 4191 { 4192 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4193 refProcessingWork(); 4194 } 4195 verify_work_stacks_empty(); 4196 verify_overflow_empty(); 4197 4198 if (should_unload_classes()) { 4199 CodeCache::gc_epilogue(); 4200 } 4201 JvmtiExport::gc_epilogue(); 4202 4203 // If we encountered any (marking stack / work queue) overflow 4204 // events during the current CMS cycle, take appropriate 4205 // remedial measures, where possible, so as to try and avoid 4206 // recurrence of that condition. 4207 assert(_markStack.isEmpty(), "No grey objects"); 4208 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4209 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4210 if (ser_ovflw > 0) { 4211 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4212 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4213 _markStack.expand(); 4214 _ser_pmc_remark_ovflw = 0; 4215 _ser_pmc_preclean_ovflw = 0; 4216 _ser_kac_preclean_ovflw = 0; 4217 _ser_kac_ovflw = 0; 4218 } 4219 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4220 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4221 _par_pmc_remark_ovflw, _par_kac_ovflw); 4222 _par_pmc_remark_ovflw = 0; 4223 _par_kac_ovflw = 0; 4224 } 4225 if (_markStack._hit_limit > 0) { 4226 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4227 _markStack._hit_limit); 4228 } 4229 if (_markStack._failed_double > 0) { 4230 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4231 _markStack._failed_double, _markStack.capacity()); 4232 } 4233 _markStack._hit_limit = 0; 4234 _markStack._failed_double = 0; 4235 4236 if ((VerifyAfterGC || VerifyDuringGC) && 4237 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4238 verify_after_remark(); 4239 } 4240 4241 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4242 4243 // Change under the freelistLocks. 4244 _collectorState = Sweeping; 4245 // Call isAllClear() under bitMapLock 4246 assert(_modUnionTable.isAllClear(), 4247 "Should be clear by end of the final marking"); 4248 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4249 "Should be clear by end of the final marking"); 4250 } 4251 4252 void CMSParInitialMarkTask::work(uint worker_id) { 4253 elapsedTimer _timer; 4254 ResourceMark rm; 4255 HandleMark hm; 4256 4257 // ---------- scan from roots -------------- 4258 _timer.start(); 4259 CMSHeap* heap = CMSHeap::heap(); 4260 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4261 4262 // ---------- young gen roots -------------- 4263 { 4264 work_on_young_gen_roots(&par_mri_cl); 4265 _timer.stop(); 4266 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4267 } 4268 4269 // ---------- remaining roots -------------- 4270 _timer.reset(); 4271 _timer.start(); 4272 4273 CLDToOopClosure cld_closure(&par_mri_cl, ClassLoaderData::_claim_strong); 4274 4275 heap->cms_process_roots(_strong_roots_scope, 4276 false, // yg was scanned above 4277 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4278 _collector->should_unload_classes(), 4279 &par_mri_cl, 4280 &cld_closure, 4281 &_par_state_string); 4282 4283 assert(_collector->should_unload_classes() 4284 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4285 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4286 _timer.stop(); 4287 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4288 } 4289 4290 // Parallel remark task 4291 class CMSParRemarkTask: public CMSParMarkTask { 4292 CompactibleFreeListSpace* _cms_space; 4293 4294 // The per-thread work queues, available here for stealing. 4295 OopTaskQueueSet* _task_queues; 4296 ParallelTaskTerminator _term; 4297 StrongRootsScope* _strong_roots_scope; 4298 4299 public: 4300 // A value of 0 passed to n_workers will cause the number of 4301 // workers to be taken from the active workers in the work gang. 4302 CMSParRemarkTask(CMSCollector* collector, 4303 CompactibleFreeListSpace* cms_space, 4304 uint n_workers, WorkGang* workers, 4305 OopTaskQueueSet* task_queues, 4306 StrongRootsScope* strong_roots_scope): 4307 CMSParMarkTask("Rescan roots and grey objects in parallel", 4308 collector, n_workers), 4309 _cms_space(cms_space), 4310 _task_queues(task_queues), 4311 _term(n_workers, task_queues), 4312 _strong_roots_scope(strong_roots_scope) { } 4313 4314 OopTaskQueueSet* task_queues() { return _task_queues; } 4315 4316 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4317 4318 ParallelTaskTerminator* terminator() { return &_term; } 4319 uint n_workers() { return _n_workers; } 4320 4321 void work(uint worker_id); 4322 4323 private: 4324 // ... of dirty cards in old space 4325 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4326 ParMarkRefsIntoAndScanClosure* cl); 4327 4328 // ... work stealing for the above 4329 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl); 4330 }; 4331 4332 class RemarkCLDClosure : public CLDClosure { 4333 CLDToOopClosure _cm_closure; 4334 public: 4335 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure, ClassLoaderData::_claim_strong) {} 4336 void do_cld(ClassLoaderData* cld) { 4337 // Check if we have modified any oops in the CLD during the concurrent marking. 4338 if (cld->has_accumulated_modified_oops()) { 4339 cld->clear_accumulated_modified_oops(); 4340 4341 // We could have transfered the current modified marks to the accumulated marks, 4342 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4343 } else if (cld->has_modified_oops()) { 4344 // Don't clear anything, this info is needed by the next young collection. 4345 } else { 4346 // No modified oops in the ClassLoaderData. 4347 return; 4348 } 4349 4350 // The klass has modified fields, need to scan the klass. 4351 _cm_closure.do_cld(cld); 4352 } 4353 }; 4354 4355 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4356 ParNewGeneration* young_gen = _collector->_young_gen; 4357 ContiguousSpace* eden_space = young_gen->eden(); 4358 ContiguousSpace* from_space = young_gen->from(); 4359 ContiguousSpace* to_space = young_gen->to(); 4360 4361 HeapWord** eca = _collector->_eden_chunk_array; 4362 size_t ect = _collector->_eden_chunk_index; 4363 HeapWord** sca = _collector->_survivor_chunk_array; 4364 size_t sct = _collector->_survivor_chunk_index; 4365 4366 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4367 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4368 4369 do_young_space_rescan(cl, to_space, NULL, 0); 4370 do_young_space_rescan(cl, from_space, sca, sct); 4371 do_young_space_rescan(cl, eden_space, eca, ect); 4372 } 4373 4374 // work_queue(i) is passed to the closure 4375 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4376 // also is passed to do_dirty_card_rescan_tasks() and to 4377 // do_work_steal() to select the i-th task_queue. 4378 4379 void CMSParRemarkTask::work(uint worker_id) { 4380 elapsedTimer _timer; 4381 ResourceMark rm; 4382 HandleMark hm; 4383 4384 // ---------- rescan from roots -------------- 4385 _timer.start(); 4386 CMSHeap* heap = CMSHeap::heap(); 4387 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4388 _collector->_span, _collector->ref_processor(), 4389 &(_collector->_markBitMap), 4390 work_queue(worker_id)); 4391 4392 // Rescan young gen roots first since these are likely 4393 // coarsely partitioned and may, on that account, constitute 4394 // the critical path; thus, it's best to start off that 4395 // work first. 4396 // ---------- young gen roots -------------- 4397 { 4398 work_on_young_gen_roots(&par_mrias_cl); 4399 _timer.stop(); 4400 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4401 } 4402 4403 // ---------- remaining roots -------------- 4404 _timer.reset(); 4405 _timer.start(); 4406 heap->cms_process_roots(_strong_roots_scope, 4407 false, // yg was scanned above 4408 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4409 _collector->should_unload_classes(), 4410 &par_mrias_cl, 4411 NULL, // The dirty klasses will be handled below 4412 &_par_state_string); 4413 4414 assert(_collector->should_unload_classes() 4415 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4416 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4417 _timer.stop(); 4418 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4419 4420 // ---------- unhandled CLD scanning ---------- 4421 if (worker_id == 0) { // Single threaded at the moment. 4422 _timer.reset(); 4423 _timer.start(); 4424 4425 // Scan all new class loader data objects and new dependencies that were 4426 // introduced during concurrent marking. 4427 ResourceMark rm; 4428 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4429 for (int i = 0; i < array->length(); i++) { 4430 Devirtualizer::do_cld(&par_mrias_cl, array->at(i)); 4431 } 4432 4433 // We don't need to keep track of new CLDs anymore. 4434 ClassLoaderDataGraph::remember_new_clds(false); 4435 4436 _timer.stop(); 4437 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4438 } 4439 4440 // We might have added oops to ClassLoaderData::_handles during the 4441 // concurrent marking phase. These oops do not always point to newly allocated objects 4442 // that are guaranteed to be kept alive. Hence, 4443 // we do have to revisit the _handles block during the remark phase. 4444 4445 // ---------- dirty CLD scanning ---------- 4446 if (worker_id == 0) { // Single threaded at the moment. 4447 _timer.reset(); 4448 _timer.start(); 4449 4450 // Scan all classes that was dirtied during the concurrent marking phase. 4451 RemarkCLDClosure remark_closure(&par_mrias_cl); 4452 ClassLoaderDataGraph::cld_do(&remark_closure); 4453 4454 _timer.stop(); 4455 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4456 } 4457 4458 4459 // ---------- rescan dirty cards ------------ 4460 _timer.reset(); 4461 _timer.start(); 4462 4463 // Do the rescan tasks for each of the two spaces 4464 // (cms_space) in turn. 4465 // "worker_id" is passed to select the task_queue for "worker_id" 4466 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4467 _timer.stop(); 4468 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4469 4470 // ---------- steal work from other threads ... 4471 // ---------- ... and drain overflow list. 4472 _timer.reset(); 4473 _timer.start(); 4474 do_work_steal(worker_id, &par_mrias_cl); 4475 _timer.stop(); 4476 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4477 } 4478 4479 void 4480 CMSParMarkTask::do_young_space_rescan( 4481 OopsInGenClosure* cl, ContiguousSpace* space, 4482 HeapWord** chunk_array, size_t chunk_top) { 4483 // Until all tasks completed: 4484 // . claim an unclaimed task 4485 // . compute region boundaries corresponding to task claimed 4486 // using chunk_array 4487 // . par_oop_iterate(cl) over that region 4488 4489 ResourceMark rm; 4490 HandleMark hm; 4491 4492 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4493 4494 uint nth_task = 0; 4495 uint n_tasks = pst->n_tasks(); 4496 4497 if (n_tasks > 0) { 4498 assert(pst->valid(), "Uninitialized use?"); 4499 HeapWord *start, *end; 4500 while (pst->try_claim_task(/* reference */ nth_task)) { 4501 // We claimed task # nth_task; compute its boundaries. 4502 if (chunk_top == 0) { // no samples were taken 4503 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4504 start = space->bottom(); 4505 end = space->top(); 4506 } else if (nth_task == 0) { 4507 start = space->bottom(); 4508 end = chunk_array[nth_task]; 4509 } else if (nth_task < (uint)chunk_top) { 4510 assert(nth_task >= 1, "Control point invariant"); 4511 start = chunk_array[nth_task - 1]; 4512 end = chunk_array[nth_task]; 4513 } else { 4514 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4515 start = chunk_array[chunk_top - 1]; 4516 end = space->top(); 4517 } 4518 MemRegion mr(start, end); 4519 // Verify that mr is in space 4520 assert(mr.is_empty() || space->used_region().contains(mr), 4521 "Should be in space"); 4522 // Verify that "start" is an object boundary 4523 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4524 "Should be an oop"); 4525 space->par_oop_iterate(mr, cl); 4526 } 4527 pst->all_tasks_completed(); 4528 } 4529 } 4530 4531 void 4532 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4533 CompactibleFreeListSpace* sp, int i, 4534 ParMarkRefsIntoAndScanClosure* cl) { 4535 // Until all tasks completed: 4536 // . claim an unclaimed task 4537 // . compute region boundaries corresponding to task claimed 4538 // . transfer dirty bits ct->mut for that region 4539 // . apply rescanclosure to dirty mut bits for that region 4540 4541 ResourceMark rm; 4542 HandleMark hm; 4543 4544 OopTaskQueue* work_q = work_queue(i); 4545 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4546 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4547 // CAUTION: This closure has state that persists across calls to 4548 // the work method dirty_range_iterate_clear() in that it has 4549 // embedded in it a (subtype of) UpwardsObjectClosure. The 4550 // use of that state in the embedded UpwardsObjectClosure instance 4551 // assumes that the cards are always iterated (even if in parallel 4552 // by several threads) in monotonically increasing order per each 4553 // thread. This is true of the implementation below which picks 4554 // card ranges (chunks) in monotonically increasing order globally 4555 // and, a-fortiori, in monotonically increasing order per thread 4556 // (the latter order being a subsequence of the former). 4557 // If the work code below is ever reorganized into a more chaotic 4558 // work-partitioning form than the current "sequential tasks" 4559 // paradigm, the use of that persistent state will have to be 4560 // revisited and modified appropriately. See also related 4561 // bug 4756801 work on which should examine this code to make 4562 // sure that the changes there do not run counter to the 4563 // assumptions made here and necessary for correctness and 4564 // efficiency. Note also that this code might yield inefficient 4565 // behavior in the case of very large objects that span one or 4566 // more work chunks. Such objects would potentially be scanned 4567 // several times redundantly. Work on 4756801 should try and 4568 // address that performance anomaly if at all possible. XXX 4569 MemRegion full_span = _collector->_span; 4570 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4571 MarkFromDirtyCardsClosure 4572 greyRescanClosure(_collector, full_span, // entire span of interest 4573 sp, bm, work_q, cl); 4574 4575 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4576 assert(pst->valid(), "Uninitialized use?"); 4577 uint nth_task = 0; 4578 const int alignment = CardTable::card_size * BitsPerWord; 4579 MemRegion span = sp->used_region(); 4580 HeapWord* start_addr = span.start(); 4581 HeapWord* end_addr = align_up(span.end(), alignment); 4582 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4583 assert(is_aligned(start_addr, alignment), "Check alignment"); 4584 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4585 4586 while (pst->try_claim_task(/* reference */ nth_task)) { 4587 // Having claimed the nth_task, compute corresponding mem-region, 4588 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4589 // The alignment restriction ensures that we do not need any 4590 // synchronization with other gang-workers while setting or 4591 // clearing bits in thus chunk of the MUT. 4592 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4593 start_addr + (nth_task+1)*chunk_size); 4594 // The last chunk's end might be way beyond end of the 4595 // used region. In that case pull back appropriately. 4596 if (this_span.end() > end_addr) { 4597 this_span.set_end(end_addr); 4598 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4599 } 4600 // Iterate over the dirty cards covering this chunk, marking them 4601 // precleaned, and setting the corresponding bits in the mod union 4602 // table. Since we have been careful to partition at Card and MUT-word 4603 // boundaries no synchronization is needed between parallel threads. 4604 _collector->_ct->dirty_card_iterate(this_span, 4605 &modUnionClosure); 4606 4607 // Having transferred these marks into the modUnionTable, 4608 // rescan the marked objects on the dirty cards in the modUnionTable. 4609 // Even if this is at a synchronous collection, the initial marking 4610 // may have been done during an asynchronous collection so there 4611 // may be dirty bits in the mod-union table. 4612 _collector->_modUnionTable.dirty_range_iterate_clear( 4613 this_span, &greyRescanClosure); 4614 _collector->_modUnionTable.verifyNoOneBitsInRange( 4615 this_span.start(), 4616 this_span.end()); 4617 } 4618 pst->all_tasks_completed(); // declare that i am done 4619 } 4620 4621 // . see if we can share work_queues with ParNew? XXX 4622 void 4623 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl) { 4624 OopTaskQueue* work_q = work_queue(i); 4625 NOT_PRODUCT(int num_steals = 0;) 4626 oop obj_to_scan; 4627 CMSBitMap* bm = &(_collector->_markBitMap); 4628 4629 while (true) { 4630 // Completely finish any left over work from (an) earlier round(s) 4631 cl->trim_queue(0); 4632 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4633 (size_t)ParGCDesiredObjsFromOverflowList); 4634 // Now check if there's any work in the overflow list 4635 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4636 // only affects the number of attempts made to get work from the 4637 // overflow list and does not affect the number of workers. Just 4638 // pass ParallelGCThreads so this behavior is unchanged. 4639 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4640 work_q, 4641 ParallelGCThreads)) { 4642 // found something in global overflow list; 4643 // not yet ready to go stealing work from others. 4644 // We'd like to assert(work_q->size() != 0, ...) 4645 // because we just took work from the overflow list, 4646 // but of course we can't since all of that could have 4647 // been already stolen from us. 4648 // "He giveth and He taketh away." 4649 continue; 4650 } 4651 // Verify that we have no work before we resort to stealing 4652 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4653 // Try to steal from other queues that have work 4654 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 4655 NOT_PRODUCT(num_steals++;) 4656 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4657 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4658 // Do scanning work 4659 obj_to_scan->oop_iterate(cl); 4660 // Loop around, finish this work, and try to steal some more 4661 } else if (terminator()->offer_termination()) { 4662 break; // nirvana from the infinite cycle 4663 } 4664 } 4665 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4666 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4667 "Else our work is not yet done"); 4668 } 4669 4670 // Record object boundaries in _eden_chunk_array by sampling the eden 4671 // top in the slow-path eden object allocation code path and record 4672 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4673 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4674 // sampling in sample_eden() that activates during the part of the 4675 // preclean phase. 4676 void CMSCollector::sample_eden_chunk() { 4677 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4678 if (_eden_chunk_lock->try_lock()) { 4679 // Record a sample. This is the critical section. The contents 4680 // of the _eden_chunk_array have to be non-decreasing in the 4681 // address order. 4682 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4683 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4684 "Unexpected state of Eden"); 4685 if (_eden_chunk_index == 0 || 4686 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4687 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4688 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4689 _eden_chunk_index++; // commit sample 4690 } 4691 _eden_chunk_lock->unlock(); 4692 } 4693 } 4694 } 4695 4696 // Return a thread-local PLAB recording array, as appropriate. 4697 void* CMSCollector::get_data_recorder(int thr_num) { 4698 if (_survivor_plab_array != NULL && 4699 (CMSPLABRecordAlways || 4700 (_collectorState > Marking && _collectorState < FinalMarking))) { 4701 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4702 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4703 ca->reset(); // clear it so that fresh data is recorded 4704 return (void*) ca; 4705 } else { 4706 return NULL; 4707 } 4708 } 4709 4710 // Reset all the thread-local PLAB recording arrays 4711 void CMSCollector::reset_survivor_plab_arrays() { 4712 for (uint i = 0; i < ParallelGCThreads; i++) { 4713 _survivor_plab_array[i].reset(); 4714 } 4715 } 4716 4717 // Merge the per-thread plab arrays into the global survivor chunk 4718 // array which will provide the partitioning of the survivor space 4719 // for CMS initial scan and rescan. 4720 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4721 int no_of_gc_threads) { 4722 assert(_survivor_plab_array != NULL, "Error"); 4723 assert(_survivor_chunk_array != NULL, "Error"); 4724 assert(_collectorState == FinalMarking || 4725 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4726 for (int j = 0; j < no_of_gc_threads; j++) { 4727 _cursor[j] = 0; 4728 } 4729 HeapWord* top = surv->top(); 4730 size_t i; 4731 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4732 HeapWord* min_val = top; // Higher than any PLAB address 4733 uint min_tid = 0; // position of min_val this round 4734 for (int j = 0; j < no_of_gc_threads; j++) { 4735 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4736 if (_cursor[j] == cur_sca->end()) { 4737 continue; 4738 } 4739 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4740 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4741 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4742 if (cur_val < min_val) { 4743 min_tid = j; 4744 min_val = cur_val; 4745 } else { 4746 assert(cur_val < top, "All recorded addresses should be less"); 4747 } 4748 } 4749 // At this point min_val and min_tid are respectively 4750 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4751 // and the thread (j) that witnesses that address. 4752 // We record this address in the _survivor_chunk_array[i] 4753 // and increment _cursor[min_tid] prior to the next round i. 4754 if (min_val == top) { 4755 break; 4756 } 4757 _survivor_chunk_array[i] = min_val; 4758 _cursor[min_tid]++; 4759 } 4760 // We are all done; record the size of the _survivor_chunk_array 4761 _survivor_chunk_index = i; // exclusive: [0, i) 4762 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4763 // Verify that we used up all the recorded entries 4764 #ifdef ASSERT 4765 size_t total = 0; 4766 for (int j = 0; j < no_of_gc_threads; j++) { 4767 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4768 total += _cursor[j]; 4769 } 4770 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4771 // Check that the merged array is in sorted order 4772 if (total > 0) { 4773 for (size_t i = 0; i < total - 1; i++) { 4774 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4775 i, p2i(_survivor_chunk_array[i])); 4776 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4777 "Not sorted"); 4778 } 4779 } 4780 #endif // ASSERT 4781 } 4782 4783 // Set up the space's par_seq_tasks structure for work claiming 4784 // for parallel initial scan and rescan of young gen. 4785 // See ParRescanTask where this is currently used. 4786 void 4787 CMSCollector:: 4788 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4789 assert(n_threads > 0, "Unexpected n_threads argument"); 4790 4791 // Eden space 4792 if (!_young_gen->eden()->is_empty()) { 4793 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4794 assert(!pst->valid(), "Clobbering existing data?"); 4795 // Each valid entry in [0, _eden_chunk_index) represents a task. 4796 size_t n_tasks = _eden_chunk_index + 1; 4797 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4798 // Sets the condition for completion of the subtask (how many threads 4799 // need to finish in order to be done). 4800 pst->set_n_threads(n_threads); 4801 pst->set_n_tasks((int)n_tasks); 4802 } 4803 4804 // Merge the survivor plab arrays into _survivor_chunk_array 4805 if (_survivor_plab_array != NULL) { 4806 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4807 } else { 4808 assert(_survivor_chunk_index == 0, "Error"); 4809 } 4810 4811 // To space 4812 { 4813 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4814 assert(!pst->valid(), "Clobbering existing data?"); 4815 // Sets the condition for completion of the subtask (how many threads 4816 // need to finish in order to be done). 4817 pst->set_n_threads(n_threads); 4818 pst->set_n_tasks(1); 4819 assert(pst->valid(), "Error"); 4820 } 4821 4822 // From space 4823 { 4824 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4825 assert(!pst->valid(), "Clobbering existing data?"); 4826 size_t n_tasks = _survivor_chunk_index + 1; 4827 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4828 // Sets the condition for completion of the subtask (how many threads 4829 // need to finish in order to be done). 4830 pst->set_n_threads(n_threads); 4831 pst->set_n_tasks((int)n_tasks); 4832 assert(pst->valid(), "Error"); 4833 } 4834 } 4835 4836 // Parallel version of remark 4837 void CMSCollector::do_remark_parallel() { 4838 CMSHeap* heap = CMSHeap::heap(); 4839 WorkGang* workers = heap->workers(); 4840 assert(workers != NULL, "Need parallel worker threads."); 4841 // Choose to use the number of GC workers most recently set 4842 // into "active_workers". 4843 uint n_workers = workers->active_workers(); 4844 4845 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4846 4847 StrongRootsScope srs(n_workers); 4848 4849 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4850 4851 // We won't be iterating over the cards in the card table updating 4852 // the younger_gen cards, so we shouldn't call the following else 4853 // the verification code as well as subsequent younger_refs_iterate 4854 // code would get confused. XXX 4855 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4856 4857 // The young gen rescan work will not be done as part of 4858 // process_roots (which currently doesn't know how to 4859 // parallelize such a scan), but rather will be broken up into 4860 // a set of parallel tasks (via the sampling that the [abortable] 4861 // preclean phase did of eden, plus the [two] tasks of 4862 // scanning the [two] survivor spaces. Further fine-grain 4863 // parallelization of the scanning of the survivor spaces 4864 // themselves, and of precleaning of the young gen itself 4865 // is deferred to the future. 4866 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4867 4868 // The dirty card rescan work is broken up into a "sequence" 4869 // of parallel tasks (per constituent space) that are dynamically 4870 // claimed by the parallel threads. 4871 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4872 4873 // It turns out that even when we're using 1 thread, doing the work in a 4874 // separate thread causes wide variance in run times. We can't help this 4875 // in the multi-threaded case, but we special-case n=1 here to get 4876 // repeatable measurements of the 1-thread overhead of the parallel code. 4877 if (n_workers > 1) { 4878 // Make refs discovery MT-safe, if it isn't already: it may not 4879 // necessarily be so, since it's possible that we are doing 4880 // ST marking. 4881 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4882 workers->run_task(&tsk); 4883 } else { 4884 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4885 tsk.work(0); 4886 } 4887 4888 // restore, single-threaded for now, any preserved marks 4889 // as a result of work_q overflow 4890 restore_preserved_marks_if_any(); 4891 } 4892 4893 // Non-parallel version of remark 4894 void CMSCollector::do_remark_non_parallel() { 4895 ResourceMark rm; 4896 HandleMark hm; 4897 CMSHeap* heap = CMSHeap::heap(); 4898 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4899 4900 MarkRefsIntoAndScanClosure 4901 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4902 &_markStack, this, 4903 false /* should_yield */, false /* not precleaning */); 4904 MarkFromDirtyCardsClosure 4905 markFromDirtyCardsClosure(this, _span, 4906 NULL, // space is set further below 4907 &_markBitMap, &_markStack, &mrias_cl); 4908 { 4909 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4910 // Iterate over the dirty cards, setting the corresponding bits in the 4911 // mod union table. 4912 { 4913 ModUnionClosure modUnionClosure(&_modUnionTable); 4914 _ct->dirty_card_iterate(_cmsGen->used_region(), 4915 &modUnionClosure); 4916 } 4917 // Having transferred these marks into the modUnionTable, we just need 4918 // to rescan the marked objects on the dirty cards in the modUnionTable. 4919 // The initial marking may have been done during an asynchronous 4920 // collection so there may be dirty bits in the mod-union table. 4921 const int alignment = CardTable::card_size * BitsPerWord; 4922 { 4923 // ... First handle dirty cards in CMS gen 4924 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4925 MemRegion ur = _cmsGen->used_region(); 4926 HeapWord* lb = ur.start(); 4927 HeapWord* ub = align_up(ur.end(), alignment); 4928 MemRegion cms_span(lb, ub); 4929 _modUnionTable.dirty_range_iterate_clear(cms_span, 4930 &markFromDirtyCardsClosure); 4931 verify_work_stacks_empty(); 4932 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4933 } 4934 } 4935 if (VerifyDuringGC && 4936 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4937 HandleMark hm; // Discard invalid handles created during verification 4938 Universe::verify(); 4939 } 4940 { 4941 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4942 4943 verify_work_stacks_empty(); 4944 4945 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4946 StrongRootsScope srs(1); 4947 4948 heap->cms_process_roots(&srs, 4949 true, // young gen as roots 4950 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4951 should_unload_classes(), 4952 &mrias_cl, 4953 NULL); // The dirty klasses will be handled below 4954 4955 assert(should_unload_classes() 4956 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4957 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4958 } 4959 4960 { 4961 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4962 4963 verify_work_stacks_empty(); 4964 4965 // Scan all class loader data objects that might have been introduced 4966 // during concurrent marking. 4967 ResourceMark rm; 4968 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4969 for (int i = 0; i < array->length(); i++) { 4970 Devirtualizer::do_cld(&mrias_cl, array->at(i)); 4971 } 4972 4973 // We don't need to keep track of new CLDs anymore. 4974 ClassLoaderDataGraph::remember_new_clds(false); 4975 4976 verify_work_stacks_empty(); 4977 } 4978 4979 // We might have added oops to ClassLoaderData::_handles during the 4980 // concurrent marking phase. These oops do not point to newly allocated objects 4981 // that are guaranteed to be kept alive. Hence, 4982 // we do have to revisit the _handles block during the remark phase. 4983 { 4984 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4985 4986 verify_work_stacks_empty(); 4987 4988 RemarkCLDClosure remark_closure(&mrias_cl); 4989 ClassLoaderDataGraph::cld_do(&remark_closure); 4990 4991 verify_work_stacks_empty(); 4992 } 4993 4994 verify_work_stacks_empty(); 4995 // Restore evacuated mark words, if any, used for overflow list links 4996 restore_preserved_marks_if_any(); 4997 4998 verify_overflow_empty(); 4999 } 5000 5001 //////////////////////////////////////////////////////// 5002 // Parallel Reference Processing Task Proxy Class 5003 //////////////////////////////////////////////////////// 5004 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5005 OopTaskQueueSet* _queues; 5006 ParallelTaskTerminator _terminator; 5007 public: 5008 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5009 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5010 ParallelTaskTerminator* terminator() { return &_terminator; } 5011 OopTaskQueueSet* queues() { return _queues; } 5012 }; 5013 5014 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5015 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5016 CMSCollector* _collector; 5017 CMSBitMap* _mark_bit_map; 5018 const MemRegion _span; 5019 ProcessTask& _task; 5020 5021 public: 5022 CMSRefProcTaskProxy(ProcessTask& task, 5023 CMSCollector* collector, 5024 const MemRegion& span, 5025 CMSBitMap* mark_bit_map, 5026 AbstractWorkGang* workers, 5027 OopTaskQueueSet* task_queues): 5028 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5029 task_queues, 5030 workers->active_workers()), 5031 _collector(collector), 5032 _mark_bit_map(mark_bit_map), 5033 _span(span), 5034 _task(task) 5035 { 5036 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5037 "Inconsistency in _span"); 5038 } 5039 5040 OopTaskQueueSet* task_queues() { return queues(); } 5041 5042 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5043 5044 void do_work_steal(int i, 5045 CMSParDrainMarkingStackClosure* drain, 5046 CMSParKeepAliveClosure* keep_alive); 5047 5048 virtual void work(uint worker_id); 5049 }; 5050 5051 void CMSRefProcTaskProxy::work(uint worker_id) { 5052 ResourceMark rm; 5053 HandleMark hm; 5054 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5055 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5056 _mark_bit_map, 5057 work_queue(worker_id)); 5058 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5059 _mark_bit_map, 5060 work_queue(worker_id)); 5061 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5062 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5063 if (_task.marks_oops_alive()) { 5064 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive); 5065 } 5066 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5067 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5068 } 5069 5070 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5071 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5072 _span(span), 5073 _work_queue(work_queue), 5074 _bit_map(bit_map), 5075 _mark_and_push(collector, span, bit_map, work_queue), 5076 _low_water_mark(MIN2((work_queue->max_elems()/4), 5077 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5078 { } 5079 5080 // . see if we can share work_queues with ParNew? XXX 5081 void CMSRefProcTaskProxy::do_work_steal(int i, 5082 CMSParDrainMarkingStackClosure* drain, 5083 CMSParKeepAliveClosure* keep_alive) { 5084 OopTaskQueue* work_q = work_queue(i); 5085 NOT_PRODUCT(int num_steals = 0;) 5086 oop obj_to_scan; 5087 5088 while (true) { 5089 // Completely finish any left over work from (an) earlier round(s) 5090 drain->trim_queue(0); 5091 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5092 (size_t)ParGCDesiredObjsFromOverflowList); 5093 // Now check if there's any work in the overflow list 5094 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5095 // only affects the number of attempts made to get work from the 5096 // overflow list and does not affect the number of workers. Just 5097 // pass ParallelGCThreads so this behavior is unchanged. 5098 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5099 work_q, 5100 ParallelGCThreads)) { 5101 // Found something in global overflow list; 5102 // not yet ready to go stealing work from others. 5103 // We'd like to assert(work_q->size() != 0, ...) 5104 // because we just took work from the overflow list, 5105 // but of course we can't, since all of that might have 5106 // been already stolen from us. 5107 continue; 5108 } 5109 // Verify that we have no work before we resort to stealing 5110 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5111 // Try to steal from other queues that have work 5112 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 5113 NOT_PRODUCT(num_steals++;) 5114 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5115 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5116 // Do scanning work 5117 obj_to_scan->oop_iterate(keep_alive); 5118 // Loop around, finish this work, and try to steal some more 5119 } else if (terminator()->offer_termination()) { 5120 break; // nirvana from the infinite cycle 5121 } 5122 } 5123 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5124 } 5125 5126 void CMSRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 5127 CMSHeap* heap = CMSHeap::heap(); 5128 WorkGang* workers = heap->workers(); 5129 assert(workers != NULL, "Need parallel worker threads."); 5130 assert(workers->active_workers() == ergo_workers, 5131 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 5132 ergo_workers, workers->active_workers()); 5133 CMSRefProcTaskProxy rp_task(task, &_collector, 5134 _collector.ref_processor_span(), 5135 _collector.markBitMap(), 5136 workers, _collector.task_queues()); 5137 workers->run_task(&rp_task, workers->active_workers()); 5138 } 5139 5140 void CMSCollector::refProcessingWork() { 5141 ResourceMark rm; 5142 HandleMark hm; 5143 5144 ReferenceProcessor* rp = ref_processor(); 5145 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5146 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5147 // Process weak references. 5148 rp->setup_policy(false); 5149 verify_work_stacks_empty(); 5150 5151 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 5152 { 5153 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5154 5155 // Setup keep_alive and complete closures. 5156 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5157 &_markStack, false /* !preclean */); 5158 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5159 _span, &_markBitMap, &_markStack, 5160 &cmsKeepAliveClosure, false /* !preclean */); 5161 5162 ReferenceProcessorStats stats; 5163 if (rp->processing_is_mt()) { 5164 // Set the degree of MT here. If the discovery is done MT, there 5165 // may have been a different number of threads doing the discovery 5166 // and a different number of discovered lists may have Ref objects. 5167 // That is OK as long as the Reference lists are balanced (see 5168 // balance_all_queues() and balance_queues()). 5169 CMSHeap* heap = CMSHeap::heap(); 5170 uint active_workers = ParallelGCThreads; 5171 WorkGang* workers = heap->workers(); 5172 if (workers != NULL) { 5173 active_workers = workers->active_workers(); 5174 // The expectation is that active_workers will have already 5175 // been set to a reasonable value. If it has not been set, 5176 // investigate. 5177 assert(active_workers > 0, "Should have been set during scavenge"); 5178 } 5179 rp->set_active_mt_degree(active_workers); 5180 CMSRefProcTaskExecutor task_executor(*this); 5181 stats = rp->process_discovered_references(&_is_alive_closure, 5182 &cmsKeepAliveClosure, 5183 &cmsDrainMarkingStackClosure, 5184 &task_executor, 5185 &pt); 5186 } else { 5187 stats = rp->process_discovered_references(&_is_alive_closure, 5188 &cmsKeepAliveClosure, 5189 &cmsDrainMarkingStackClosure, 5190 NULL, 5191 &pt); 5192 } 5193 _gc_tracer_cm->report_gc_reference_stats(stats); 5194 pt.print_all_references(); 5195 } 5196 5197 // This is the point where the entire marking should have completed. 5198 verify_work_stacks_empty(); 5199 5200 { 5201 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5202 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5203 } 5204 5205 if (should_unload_classes()) { 5206 { 5207 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5208 5209 // Unload classes and purge the SystemDictionary. 5210 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm); 5211 5212 // Unload nmethods. 5213 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5214 5215 // Prune dead klasses from subklass/sibling/implementor lists. 5216 Klass::clean_weak_klass_links(purged_class); 5217 } 5218 5219 { 5220 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5221 // Clean up unreferenced symbols in symbol table. 5222 SymbolTable::unlink(); 5223 } 5224 5225 { 5226 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5227 // Delete entries for dead interned strings. 5228 StringTable::unlink(&_is_alive_closure); 5229 } 5230 } 5231 5232 // Restore any preserved marks as a result of mark stack or 5233 // work queue overflow 5234 restore_preserved_marks_if_any(); // done single-threaded for now 5235 5236 rp->set_enqueuing_is_done(true); 5237 rp->verify_no_references_recorded(); 5238 } 5239 5240 #ifndef PRODUCT 5241 void CMSCollector::check_correct_thread_executing() { 5242 Thread* t = Thread::current(); 5243 // Only the VM thread or the CMS thread should be here. 5244 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5245 "Unexpected thread type"); 5246 // If this is the vm thread, the foreground process 5247 // should not be waiting. Note that _foregroundGCIsActive is 5248 // true while the foreground collector is waiting. 5249 if (_foregroundGCShouldWait) { 5250 // We cannot be the VM thread 5251 assert(t->is_ConcurrentGC_thread(), 5252 "Should be CMS thread"); 5253 } else { 5254 // We can be the CMS thread only if we are in a stop-world 5255 // phase of CMS collection. 5256 if (t->is_ConcurrentGC_thread()) { 5257 assert(_collectorState == InitialMarking || 5258 _collectorState == FinalMarking, 5259 "Should be a stop-world phase"); 5260 // The CMS thread should be holding the CMS_token. 5261 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5262 "Potential interference with concurrently " 5263 "executing VM thread"); 5264 } 5265 } 5266 } 5267 #endif 5268 5269 void CMSCollector::sweep() { 5270 assert(_collectorState == Sweeping, "just checking"); 5271 check_correct_thread_executing(); 5272 verify_work_stacks_empty(); 5273 verify_overflow_empty(); 5274 increment_sweep_count(); 5275 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5276 5277 _inter_sweep_timer.stop(); 5278 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5279 5280 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5281 _intra_sweep_timer.reset(); 5282 _intra_sweep_timer.start(); 5283 { 5284 GCTraceCPUTime tcpu; 5285 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5286 // First sweep the old gen 5287 { 5288 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5289 bitMapLock()); 5290 sweepWork(_cmsGen); 5291 } 5292 5293 // Update Universe::_heap_*_at_gc figures. 5294 // We need all the free list locks to make the abstract state 5295 // transition from Sweeping to Resetting. See detailed note 5296 // further below. 5297 { 5298 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5299 // Update heap occupancy information which is used as 5300 // input to soft ref clearing policy at the next gc. 5301 Universe::update_heap_info_at_gc(); 5302 _collectorState = Resizing; 5303 } 5304 } 5305 verify_work_stacks_empty(); 5306 verify_overflow_empty(); 5307 5308 if (should_unload_classes()) { 5309 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5310 // requires that the virtual spaces are stable and not deleted. 5311 ClassLoaderDataGraph::set_should_purge(true); 5312 } 5313 5314 _intra_sweep_timer.stop(); 5315 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5316 5317 _inter_sweep_timer.reset(); 5318 _inter_sweep_timer.start(); 5319 5320 // We need to use a monotonically non-decreasing time in ms 5321 // or we will see time-warp warnings and os::javaTimeMillis() 5322 // does not guarantee monotonicity. 5323 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5324 update_time_of_last_gc(now); 5325 5326 // NOTE on abstract state transitions: 5327 // Mutators allocate-live and/or mark the mod-union table dirty 5328 // based on the state of the collection. The former is done in 5329 // the interval [Marking, Sweeping] and the latter in the interval 5330 // [Marking, Sweeping). Thus the transitions into the Marking state 5331 // and out of the Sweeping state must be synchronously visible 5332 // globally to the mutators. 5333 // The transition into the Marking state happens with the world 5334 // stopped so the mutators will globally see it. Sweeping is 5335 // done asynchronously by the background collector so the transition 5336 // from the Sweeping state to the Resizing state must be done 5337 // under the freelistLock (as is the check for whether to 5338 // allocate-live and whether to dirty the mod-union table). 5339 assert(_collectorState == Resizing, "Change of collector state to" 5340 " Resizing must be done under the freelistLocks (plural)"); 5341 5342 // Now that sweeping has been completed, we clear 5343 // the incremental_collection_failed flag, 5344 // thus inviting a younger gen collection to promote into 5345 // this generation. If such a promotion may still fail, 5346 // the flag will be set again when a young collection is 5347 // attempted. 5348 CMSHeap* heap = CMSHeap::heap(); 5349 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5350 heap->update_full_collections_completed(_collection_count_start); 5351 } 5352 5353 // FIX ME!!! Looks like this belongs in CFLSpace, with 5354 // CMSGen merely delegating to it. 5355 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5356 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5357 HeapWord* minAddr = _cmsSpace->bottom(); 5358 HeapWord* largestAddr = 5359 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5360 if (largestAddr == NULL) { 5361 // The dictionary appears to be empty. In this case 5362 // try to coalesce at the end of the heap. 5363 largestAddr = _cmsSpace->end(); 5364 } 5365 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5366 size_t nearLargestOffset = 5367 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5368 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5369 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5370 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5371 } 5372 5373 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5374 return addr >= _cmsSpace->nearLargestChunk(); 5375 } 5376 5377 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5378 return _cmsSpace->find_chunk_at_end(); 5379 } 5380 5381 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5382 bool full) { 5383 // If the young generation has been collected, gather any statistics 5384 // that are of interest at this point. 5385 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5386 if (!full && current_is_young) { 5387 // Gather statistics on the young generation collection. 5388 collector()->stats().record_gc0_end(used()); 5389 } 5390 } 5391 5392 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5393 // We iterate over the space(s) underlying this generation, 5394 // checking the mark bit map to see if the bits corresponding 5395 // to specific blocks are marked or not. Blocks that are 5396 // marked are live and are not swept up. All remaining blocks 5397 // are swept up, with coalescing on-the-fly as we sweep up 5398 // contiguous free and/or garbage blocks: 5399 // We need to ensure that the sweeper synchronizes with allocators 5400 // and stop-the-world collectors. In particular, the following 5401 // locks are used: 5402 // . CMS token: if this is held, a stop the world collection cannot occur 5403 // . freelistLock: if this is held no allocation can occur from this 5404 // generation by another thread 5405 // . bitMapLock: if this is held, no other thread can access or update 5406 // 5407 5408 // Note that we need to hold the freelistLock if we use 5409 // block iterate below; else the iterator might go awry if 5410 // a mutator (or promotion) causes block contents to change 5411 // (for instance if the allocator divvies up a block). 5412 // If we hold the free list lock, for all practical purposes 5413 // young generation GC's can't occur (they'll usually need to 5414 // promote), so we might as well prevent all young generation 5415 // GC's while we do a sweeping step. For the same reason, we might 5416 // as well take the bit map lock for the entire duration 5417 5418 // check that we hold the requisite locks 5419 assert(have_cms_token(), "Should hold cms token"); 5420 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5421 assert_lock_strong(old_gen->freelistLock()); 5422 assert_lock_strong(bitMapLock()); 5423 5424 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5425 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5426 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5427 _inter_sweep_estimate.padded_average(), 5428 _intra_sweep_estimate.padded_average()); 5429 old_gen->setNearLargestChunk(); 5430 5431 { 5432 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5433 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5434 // We need to free-up/coalesce garbage/blocks from a 5435 // co-terminal free run. This is done in the SweepClosure 5436 // destructor; so, do not remove this scope, else the 5437 // end-of-sweep-census below will be off by a little bit. 5438 } 5439 old_gen->cmsSpace()->sweep_completed(); 5440 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5441 if (should_unload_classes()) { // unloaded classes this cycle, 5442 _concurrent_cycles_since_last_unload = 0; // ... reset count 5443 } else { // did not unload classes, 5444 _concurrent_cycles_since_last_unload++; // ... increment count 5445 } 5446 } 5447 5448 // Reset CMS data structures (for now just the marking bit map) 5449 // preparatory for the next cycle. 5450 void CMSCollector::reset_concurrent() { 5451 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5452 5453 // If the state is not "Resetting", the foreground thread 5454 // has done a collection and the resetting. 5455 if (_collectorState != Resetting) { 5456 assert(_collectorState == Idling, "The state should only change" 5457 " because the foreground collector has finished the collection"); 5458 return; 5459 } 5460 5461 { 5462 // Clear the mark bitmap (no grey objects to start with) 5463 // for the next cycle. 5464 GCTraceCPUTime tcpu; 5465 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5466 5467 HeapWord* curAddr = _markBitMap.startWord(); 5468 while (curAddr < _markBitMap.endWord()) { 5469 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5470 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5471 _markBitMap.clear_large_range(chunk); 5472 if (ConcurrentMarkSweepThread::should_yield() && 5473 !foregroundGCIsActive() && 5474 CMSYield) { 5475 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5476 "CMS thread should hold CMS token"); 5477 assert_lock_strong(bitMapLock()); 5478 bitMapLock()->unlock(); 5479 ConcurrentMarkSweepThread::desynchronize(true); 5480 stopTimer(); 5481 incrementYields(); 5482 5483 // See the comment in coordinator_yield() 5484 for (unsigned i = 0; i < CMSYieldSleepCount && 5485 ConcurrentMarkSweepThread::should_yield() && 5486 !CMSCollector::foregroundGCIsActive(); ++i) { 5487 os::sleep(Thread::current(), 1, false); 5488 } 5489 5490 ConcurrentMarkSweepThread::synchronize(true); 5491 bitMapLock()->lock_without_safepoint_check(); 5492 startTimer(); 5493 } 5494 curAddr = chunk.end(); 5495 } 5496 // A successful mostly concurrent collection has been done. 5497 // Because only the full (i.e., concurrent mode failure) collections 5498 // are being measured for gc overhead limits, clean the "near" flag 5499 // and count. 5500 size_policy()->reset_gc_overhead_limit_count(); 5501 _collectorState = Idling; 5502 } 5503 5504 register_gc_end(); 5505 } 5506 5507 // Same as above but for STW paths 5508 void CMSCollector::reset_stw() { 5509 // already have the lock 5510 assert(_collectorState == Resetting, "just checking"); 5511 assert_lock_strong(bitMapLock()); 5512 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5513 _markBitMap.clear_all(); 5514 _collectorState = Idling; 5515 register_gc_end(); 5516 } 5517 5518 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5519 GCTraceCPUTime tcpu; 5520 TraceCollectorStats tcs_cgc(cgc_counters()); 5521 5522 switch (op) { 5523 case CMS_op_checkpointRootsInitial: { 5524 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5525 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5526 checkpointRootsInitial(); 5527 break; 5528 } 5529 case CMS_op_checkpointRootsFinal: { 5530 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5531 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5532 checkpointRootsFinal(); 5533 break; 5534 } 5535 default: 5536 fatal("No such CMS_op"); 5537 } 5538 } 5539 5540 #ifndef PRODUCT 5541 size_t const CMSCollector::skip_header_HeapWords() { 5542 return FreeChunk::header_size(); 5543 } 5544 5545 // Try and collect here conditions that should hold when 5546 // CMS thread is exiting. The idea is that the foreground GC 5547 // thread should not be blocked if it wants to terminate 5548 // the CMS thread and yet continue to run the VM for a while 5549 // after that. 5550 void CMSCollector::verify_ok_to_terminate() const { 5551 assert(Thread::current()->is_ConcurrentGC_thread(), 5552 "should be called by CMS thread"); 5553 assert(!_foregroundGCShouldWait, "should be false"); 5554 // We could check here that all the various low-level locks 5555 // are not held by the CMS thread, but that is overkill; see 5556 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5557 // is checked. 5558 } 5559 #endif 5560 5561 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5562 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5563 "missing Printezis mark?"); 5564 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5565 size_t size = pointer_delta(nextOneAddr + 1, addr); 5566 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5567 "alignment problem"); 5568 assert(size >= 3, "Necessary for Printezis marks to work"); 5569 return size; 5570 } 5571 5572 // A variant of the above (block_size_using_printezis_bits()) except 5573 // that we return 0 if the P-bits are not yet set. 5574 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5575 if (_markBitMap.isMarked(addr + 1)) { 5576 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5577 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5578 size_t size = pointer_delta(nextOneAddr + 1, addr); 5579 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5580 "alignment problem"); 5581 assert(size >= 3, "Necessary for Printezis marks to work"); 5582 return size; 5583 } 5584 return 0; 5585 } 5586 5587 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5588 size_t sz = 0; 5589 oop p = (oop)addr; 5590 if (p->klass_or_null_acquire() != NULL) { 5591 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5592 } else { 5593 sz = block_size_using_printezis_bits(addr); 5594 } 5595 assert(sz > 0, "size must be nonzero"); 5596 HeapWord* next_block = addr + sz; 5597 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5598 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5599 align_down((uintptr_t)next_card, CardTable::card_size), 5600 "must be different cards"); 5601 return next_card; 5602 } 5603 5604 5605 // CMS Bit Map Wrapper ///////////////////////////////////////// 5606 5607 // Construct a CMS bit map infrastructure, but don't create the 5608 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5609 // further below. 5610 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5611 _shifter(shifter), 5612 _bm(), 5613 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5614 Monitor::_safepoint_check_sometimes) : NULL) 5615 { 5616 _bmStartWord = 0; 5617 _bmWordSize = 0; 5618 } 5619 5620 bool CMSBitMap::allocate(MemRegion mr) { 5621 _bmStartWord = mr.start(); 5622 _bmWordSize = mr.word_size(); 5623 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5624 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5625 if (!brs.is_reserved()) { 5626 log_warning(gc)("CMS bit map allocation failure"); 5627 return false; 5628 } 5629 // For now we'll just commit all of the bit map up front. 5630 // Later on we'll try to be more parsimonious with swap. 5631 if (!_virtual_space.initialize(brs, brs.size())) { 5632 log_warning(gc)("CMS bit map backing store failure"); 5633 return false; 5634 } 5635 assert(_virtual_space.committed_size() == brs.size(), 5636 "didn't reserve backing store for all of CMS bit map?"); 5637 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5638 _bmWordSize, "inconsistency in bit map sizing"); 5639 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5640 5641 // bm.clear(); // can we rely on getting zero'd memory? verify below 5642 assert(isAllClear(), 5643 "Expected zero'd memory from ReservedSpace constructor"); 5644 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5645 "consistency check"); 5646 return true; 5647 } 5648 5649 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5650 HeapWord *next_addr, *end_addr, *last_addr; 5651 assert_locked(); 5652 assert(covers(mr), "out-of-range error"); 5653 // XXX assert that start and end are appropriately aligned 5654 for (next_addr = mr.start(), end_addr = mr.end(); 5655 next_addr < end_addr; next_addr = last_addr) { 5656 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5657 last_addr = dirty_region.end(); 5658 if (!dirty_region.is_empty()) { 5659 cl->do_MemRegion(dirty_region); 5660 } else { 5661 assert(last_addr == end_addr, "program logic"); 5662 return; 5663 } 5664 } 5665 } 5666 5667 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5668 _bm.print_on_error(st, prefix); 5669 } 5670 5671 #ifndef PRODUCT 5672 void CMSBitMap::assert_locked() const { 5673 CMSLockVerifier::assert_locked(lock()); 5674 } 5675 5676 bool CMSBitMap::covers(MemRegion mr) const { 5677 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5678 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5679 "size inconsistency"); 5680 return (mr.start() >= _bmStartWord) && 5681 (mr.end() <= endWord()); 5682 } 5683 5684 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5685 return (start >= _bmStartWord && (start + size) <= endWord()); 5686 } 5687 5688 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5689 // verify that there are no 1 bits in the interval [left, right) 5690 FalseBitMapClosure falseBitMapClosure; 5691 iterate(&falseBitMapClosure, left, right); 5692 } 5693 5694 void CMSBitMap::region_invariant(MemRegion mr) 5695 { 5696 assert_locked(); 5697 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5698 assert(!mr.is_empty(), "unexpected empty region"); 5699 assert(covers(mr), "mr should be covered by bit map"); 5700 // convert address range into offset range 5701 size_t start_ofs = heapWordToOffset(mr.start()); 5702 // Make sure that end() is appropriately aligned 5703 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5704 "Misaligned mr.end()"); 5705 size_t end_ofs = heapWordToOffset(mr.end()); 5706 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5707 } 5708 5709 #endif 5710 5711 bool CMSMarkStack::allocate(size_t size) { 5712 // allocate a stack of the requisite depth 5713 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5714 size * sizeof(oop))); 5715 if (!rs.is_reserved()) { 5716 log_warning(gc)("CMSMarkStack allocation failure"); 5717 return false; 5718 } 5719 if (!_virtual_space.initialize(rs, rs.size())) { 5720 log_warning(gc)("CMSMarkStack backing store failure"); 5721 return false; 5722 } 5723 assert(_virtual_space.committed_size() == rs.size(), 5724 "didn't reserve backing store for all of CMS stack?"); 5725 _base = (oop*)(_virtual_space.low()); 5726 _index = 0; 5727 _capacity = size; 5728 NOT_PRODUCT(_max_depth = 0); 5729 return true; 5730 } 5731 5732 // XXX FIX ME !!! In the MT case we come in here holding a 5733 // leaf lock. For printing we need to take a further lock 5734 // which has lower rank. We need to recalibrate the two 5735 // lock-ranks involved in order to be able to print the 5736 // messages below. (Or defer the printing to the caller. 5737 // For now we take the expedient path of just disabling the 5738 // messages for the problematic case.) 5739 void CMSMarkStack::expand() { 5740 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5741 if (_capacity == MarkStackSizeMax) { 5742 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5743 // We print a warning message only once per CMS cycle. 5744 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5745 } 5746 return; 5747 } 5748 // Double capacity if possible 5749 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5750 // Do not give up existing stack until we have managed to 5751 // get the double capacity that we desired. 5752 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5753 new_capacity * sizeof(oop))); 5754 if (rs.is_reserved()) { 5755 // Release the backing store associated with old stack 5756 _virtual_space.release(); 5757 // Reinitialize virtual space for new stack 5758 if (!_virtual_space.initialize(rs, rs.size())) { 5759 fatal("Not enough swap for expanded marking stack"); 5760 } 5761 _base = (oop*)(_virtual_space.low()); 5762 _index = 0; 5763 _capacity = new_capacity; 5764 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5765 // Failed to double capacity, continue; 5766 // we print a detail message only once per CMS cycle. 5767 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5768 _capacity / K, new_capacity / K); 5769 } 5770 } 5771 5772 5773 // Closures 5774 // XXX: there seems to be a lot of code duplication here; 5775 // should refactor and consolidate common code. 5776 5777 // This closure is used to mark refs into the CMS generation in 5778 // the CMS bit map. Called at the first checkpoint. This closure 5779 // assumes that we do not need to re-mark dirty cards; if the CMS 5780 // generation on which this is used is not an oldest 5781 // generation then this will lose younger_gen cards! 5782 5783 MarkRefsIntoClosure::MarkRefsIntoClosure( 5784 MemRegion span, CMSBitMap* bitMap): 5785 _span(span), 5786 _bitMap(bitMap) 5787 { 5788 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5789 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5790 } 5791 5792 void MarkRefsIntoClosure::do_oop(oop obj) { 5793 // if p points into _span, then mark corresponding bit in _markBitMap 5794 assert(oopDesc::is_oop(obj), "expected an oop"); 5795 HeapWord* addr = (HeapWord*)obj; 5796 if (_span.contains(addr)) { 5797 // this should be made more efficient 5798 _bitMap->mark(addr); 5799 } 5800 } 5801 5802 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5803 MemRegion span, CMSBitMap* bitMap): 5804 _span(span), 5805 _bitMap(bitMap) 5806 { 5807 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5808 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5809 } 5810 5811 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5812 // if p points into _span, then mark corresponding bit in _markBitMap 5813 assert(oopDesc::is_oop(obj), "expected an oop"); 5814 HeapWord* addr = (HeapWord*)obj; 5815 if (_span.contains(addr)) { 5816 // this should be made more efficient 5817 _bitMap->par_mark(addr); 5818 } 5819 } 5820 5821 // A variant of the above, used for CMS marking verification. 5822 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5823 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5824 _span(span), 5825 _verification_bm(verification_bm), 5826 _cms_bm(cms_bm) 5827 { 5828 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5829 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5830 } 5831 5832 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5833 // if p points into _span, then mark corresponding bit in _markBitMap 5834 assert(oopDesc::is_oop(obj), "expected an oop"); 5835 HeapWord* addr = (HeapWord*)obj; 5836 if (_span.contains(addr)) { 5837 _verification_bm->mark(addr); 5838 if (!_cms_bm->isMarked(addr)) { 5839 Log(gc, verify) log; 5840 ResourceMark rm; 5841 LogStream ls(log.error()); 5842 oop(addr)->print_on(&ls); 5843 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5844 fatal("... aborting"); 5845 } 5846 } 5847 } 5848 5849 ////////////////////////////////////////////////// 5850 // MarkRefsIntoAndScanClosure 5851 ////////////////////////////////////////////////// 5852 5853 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5854 ReferenceDiscoverer* rd, 5855 CMSBitMap* bit_map, 5856 CMSBitMap* mod_union_table, 5857 CMSMarkStack* mark_stack, 5858 CMSCollector* collector, 5859 bool should_yield, 5860 bool concurrent_precleaning): 5861 _span(span), 5862 _bit_map(bit_map), 5863 _mark_stack(mark_stack), 5864 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5865 mark_stack, concurrent_precleaning), 5866 _collector(collector), 5867 _freelistLock(NULL), 5868 _yield(should_yield), 5869 _concurrent_precleaning(concurrent_precleaning) 5870 { 5871 // FIXME: Should initialize in base class constructor. 5872 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5873 set_ref_discoverer_internal(rd); 5874 } 5875 5876 // This closure is used to mark refs into the CMS generation at the 5877 // second (final) checkpoint, and to scan and transitively follow 5878 // the unmarked oops. It is also used during the concurrent precleaning 5879 // phase while scanning objects on dirty cards in the CMS generation. 5880 // The marks are made in the marking bit map and the marking stack is 5881 // used for keeping the (newly) grey objects during the scan. 5882 // The parallel version (Par_...) appears further below. 5883 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5884 if (obj != NULL) { 5885 assert(oopDesc::is_oop(obj), "expected an oop"); 5886 HeapWord* addr = (HeapWord*)obj; 5887 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5888 assert(_collector->overflow_list_is_empty(), 5889 "overflow list should be empty"); 5890 if (_span.contains(addr) && 5891 !_bit_map->isMarked(addr)) { 5892 // mark bit map (object is now grey) 5893 _bit_map->mark(addr); 5894 // push on marking stack (stack should be empty), and drain the 5895 // stack by applying this closure to the oops in the oops popped 5896 // from the stack (i.e. blacken the grey objects) 5897 bool res = _mark_stack->push(obj); 5898 assert(res, "Should have space to push on empty stack"); 5899 do { 5900 oop new_oop = _mark_stack->pop(); 5901 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5902 assert(_bit_map->isMarked((HeapWord*)new_oop), 5903 "only grey objects on this stack"); 5904 // iterate over the oops in this oop, marking and pushing 5905 // the ones in CMS heap (i.e. in _span). 5906 new_oop->oop_iterate(&_pushAndMarkClosure); 5907 // check if it's time to yield 5908 do_yield_check(); 5909 } while (!_mark_stack->isEmpty() || 5910 (!_concurrent_precleaning && take_from_overflow_list())); 5911 // if marking stack is empty, and we are not doing this 5912 // during precleaning, then check the overflow list 5913 } 5914 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5915 assert(_collector->overflow_list_is_empty(), 5916 "overflow list was drained above"); 5917 5918 assert(_collector->no_preserved_marks(), 5919 "All preserved marks should have been restored above"); 5920 } 5921 } 5922 5923 void MarkRefsIntoAndScanClosure::do_yield_work() { 5924 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5925 "CMS thread should hold CMS token"); 5926 assert_lock_strong(_freelistLock); 5927 assert_lock_strong(_bit_map->lock()); 5928 // relinquish the free_list_lock and bitMaplock() 5929 _bit_map->lock()->unlock(); 5930 _freelistLock->unlock(); 5931 ConcurrentMarkSweepThread::desynchronize(true); 5932 _collector->stopTimer(); 5933 _collector->incrementYields(); 5934 5935 // See the comment in coordinator_yield() 5936 for (unsigned i = 0; 5937 i < CMSYieldSleepCount && 5938 ConcurrentMarkSweepThread::should_yield() && 5939 !CMSCollector::foregroundGCIsActive(); 5940 ++i) { 5941 os::sleep(Thread::current(), 1, false); 5942 } 5943 5944 ConcurrentMarkSweepThread::synchronize(true); 5945 _freelistLock->lock_without_safepoint_check(); 5946 _bit_map->lock()->lock_without_safepoint_check(); 5947 _collector->startTimer(); 5948 } 5949 5950 /////////////////////////////////////////////////////////// 5951 // ParMarkRefsIntoAndScanClosure: a parallel version of 5952 // MarkRefsIntoAndScanClosure 5953 /////////////////////////////////////////////////////////// 5954 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5955 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 5956 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5957 _span(span), 5958 _bit_map(bit_map), 5959 _work_queue(work_queue), 5960 _low_water_mark(MIN2((work_queue->max_elems()/4), 5961 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 5962 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 5963 { 5964 // FIXME: Should initialize in base class constructor. 5965 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5966 set_ref_discoverer_internal(rd); 5967 } 5968 5969 // This closure is used to mark refs into the CMS generation at the 5970 // second (final) checkpoint, and to scan and transitively follow 5971 // the unmarked oops. The marks are made in the marking bit map and 5972 // the work_queue is used for keeping the (newly) grey objects during 5973 // the scan phase whence they are also available for stealing by parallel 5974 // threads. Since the marking bit map is shared, updates are 5975 // synchronized (via CAS). 5976 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 5977 if (obj != NULL) { 5978 // Ignore mark word because this could be an already marked oop 5979 // that may be chained at the end of the overflow list. 5980 assert(oopDesc::is_oop(obj, true), "expected an oop"); 5981 HeapWord* addr = (HeapWord*)obj; 5982 if (_span.contains(addr) && 5983 !_bit_map->isMarked(addr)) { 5984 // mark bit map (object will become grey): 5985 // It is possible for several threads to be 5986 // trying to "claim" this object concurrently; 5987 // the unique thread that succeeds in marking the 5988 // object first will do the subsequent push on 5989 // to the work queue (or overflow list). 5990 if (_bit_map->par_mark(addr)) { 5991 // push on work_queue (which may not be empty), and trim the 5992 // queue to an appropriate length by applying this closure to 5993 // the oops in the oops popped from the stack (i.e. blacken the 5994 // grey objects) 5995 bool res = _work_queue->push(obj); 5996 assert(res, "Low water mark should be less than capacity?"); 5997 trim_queue(_low_water_mark); 5998 } // Else, another thread claimed the object 5999 } 6000 } 6001 } 6002 6003 // This closure is used to rescan the marked objects on the dirty cards 6004 // in the mod union table and the card table proper. 6005 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6006 oop p, MemRegion mr) { 6007 6008 size_t size = 0; 6009 HeapWord* addr = (HeapWord*)p; 6010 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6011 assert(_span.contains(addr), "we are scanning the CMS generation"); 6012 // check if it's time to yield 6013 if (do_yield_check()) { 6014 // We yielded for some foreground stop-world work, 6015 // and we have been asked to abort this ongoing preclean cycle. 6016 return 0; 6017 } 6018 if (_bitMap->isMarked(addr)) { 6019 // it's marked; is it potentially uninitialized? 6020 if (p->klass_or_null_acquire() != NULL) { 6021 // an initialized object; ignore mark word in verification below 6022 // since we are running concurrent with mutators 6023 assert(oopDesc::is_oop(p, true), "should be an oop"); 6024 if (p->is_objArray()) { 6025 // objArrays are precisely marked; restrict scanning 6026 // to dirty cards only. 6027 size = CompactibleFreeListSpace::adjustObjectSize( 6028 p->oop_iterate_size(_scanningClosure, mr)); 6029 } else { 6030 // A non-array may have been imprecisely marked; we need 6031 // to scan object in its entirety. 6032 size = CompactibleFreeListSpace::adjustObjectSize( 6033 p->oop_iterate_size(_scanningClosure)); 6034 } 6035 #ifdef ASSERT 6036 size_t direct_size = 6037 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6038 assert(size == direct_size, "Inconsistency in size"); 6039 assert(size >= 3, "Necessary for Printezis marks to work"); 6040 HeapWord* start_pbit = addr + 1; 6041 HeapWord* end_pbit = addr + size - 1; 6042 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6043 "inconsistent Printezis mark"); 6044 // Verify inner mark bits (between Printezis bits) are clear, 6045 // but don't repeat if there are multiple dirty regions for 6046 // the same object, to avoid potential O(N^2) performance. 6047 if (addr != _last_scanned_object) { 6048 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6049 _last_scanned_object = addr; 6050 } 6051 #endif // ASSERT 6052 } else { 6053 // An uninitialized object. 6054 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6055 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6056 size = pointer_delta(nextOneAddr + 1, addr); 6057 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6058 "alignment problem"); 6059 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6060 // will dirty the card when the klass pointer is installed in the 6061 // object (signaling the completion of initialization). 6062 } 6063 } else { 6064 // Either a not yet marked object or an uninitialized object 6065 if (p->klass_or_null_acquire() == NULL) { 6066 // An uninitialized object, skip to the next card, since 6067 // we may not be able to read its P-bits yet. 6068 assert(size == 0, "Initial value"); 6069 } else { 6070 // An object not (yet) reached by marking: we merely need to 6071 // compute its size so as to go look at the next block. 6072 assert(oopDesc::is_oop(p, true), "should be an oop"); 6073 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6074 } 6075 } 6076 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6077 return size; 6078 } 6079 6080 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6081 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6082 "CMS thread should hold CMS token"); 6083 assert_lock_strong(_freelistLock); 6084 assert_lock_strong(_bitMap->lock()); 6085 // relinquish the free_list_lock and bitMaplock() 6086 _bitMap->lock()->unlock(); 6087 _freelistLock->unlock(); 6088 ConcurrentMarkSweepThread::desynchronize(true); 6089 _collector->stopTimer(); 6090 _collector->incrementYields(); 6091 6092 // See the comment in coordinator_yield() 6093 for (unsigned i = 0; i < CMSYieldSleepCount && 6094 ConcurrentMarkSweepThread::should_yield() && 6095 !CMSCollector::foregroundGCIsActive(); ++i) { 6096 os::sleep(Thread::current(), 1, false); 6097 } 6098 6099 ConcurrentMarkSweepThread::synchronize(true); 6100 _freelistLock->lock_without_safepoint_check(); 6101 _bitMap->lock()->lock_without_safepoint_check(); 6102 _collector->startTimer(); 6103 } 6104 6105 6106 ////////////////////////////////////////////////////////////////// 6107 // SurvivorSpacePrecleanClosure 6108 ////////////////////////////////////////////////////////////////// 6109 // This (single-threaded) closure is used to preclean the oops in 6110 // the survivor spaces. 6111 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6112 6113 HeapWord* addr = (HeapWord*)p; 6114 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6115 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6116 assert(p->klass_or_null() != NULL, "object should be initialized"); 6117 // an initialized object; ignore mark word in verification below 6118 // since we are running concurrent with mutators 6119 assert(oopDesc::is_oop(p, true), "should be an oop"); 6120 // Note that we do not yield while we iterate over 6121 // the interior oops of p, pushing the relevant ones 6122 // on our marking stack. 6123 size_t size = p->oop_iterate_size(_scanning_closure); 6124 do_yield_check(); 6125 // Observe that below, we do not abandon the preclean 6126 // phase as soon as we should; rather we empty the 6127 // marking stack before returning. This is to satisfy 6128 // some existing assertions. In general, it may be a 6129 // good idea to abort immediately and complete the marking 6130 // from the grey objects at a later time. 6131 while (!_mark_stack->isEmpty()) { 6132 oop new_oop = _mark_stack->pop(); 6133 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6134 assert(_bit_map->isMarked((HeapWord*)new_oop), 6135 "only grey objects on this stack"); 6136 // iterate over the oops in this oop, marking and pushing 6137 // the ones in CMS heap (i.e. in _span). 6138 new_oop->oop_iterate(_scanning_closure); 6139 // check if it's time to yield 6140 do_yield_check(); 6141 } 6142 unsigned int after_count = 6143 CMSHeap::heap()->total_collections(); 6144 bool abort = (_before_count != after_count) || 6145 _collector->should_abort_preclean(); 6146 return abort ? 0 : size; 6147 } 6148 6149 void SurvivorSpacePrecleanClosure::do_yield_work() { 6150 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6151 "CMS thread should hold CMS token"); 6152 assert_lock_strong(_bit_map->lock()); 6153 // Relinquish the bit map lock 6154 _bit_map->lock()->unlock(); 6155 ConcurrentMarkSweepThread::desynchronize(true); 6156 _collector->stopTimer(); 6157 _collector->incrementYields(); 6158 6159 // See the comment in coordinator_yield() 6160 for (unsigned i = 0; i < CMSYieldSleepCount && 6161 ConcurrentMarkSweepThread::should_yield() && 6162 !CMSCollector::foregroundGCIsActive(); ++i) { 6163 os::sleep(Thread::current(), 1, false); 6164 } 6165 6166 ConcurrentMarkSweepThread::synchronize(true); 6167 _bit_map->lock()->lock_without_safepoint_check(); 6168 _collector->startTimer(); 6169 } 6170 6171 // This closure is used to rescan the marked objects on the dirty cards 6172 // in the mod union table and the card table proper. In the parallel 6173 // case, although the bitMap is shared, we do a single read so the 6174 // isMarked() query is "safe". 6175 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6176 // Ignore mark word because we are running concurrent with mutators 6177 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6178 HeapWord* addr = (HeapWord*)p; 6179 assert(_span.contains(addr), "we are scanning the CMS generation"); 6180 bool is_obj_array = false; 6181 #ifdef ASSERT 6182 if (!_parallel) { 6183 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6184 assert(_collector->overflow_list_is_empty(), 6185 "overflow list should be empty"); 6186 6187 } 6188 #endif // ASSERT 6189 if (_bit_map->isMarked(addr)) { 6190 // Obj arrays are precisely marked, non-arrays are not; 6191 // so we scan objArrays precisely and non-arrays in their 6192 // entirety. 6193 if (p->is_objArray()) { 6194 is_obj_array = true; 6195 if (_parallel) { 6196 p->oop_iterate(_par_scan_closure, mr); 6197 } else { 6198 p->oop_iterate(_scan_closure, mr); 6199 } 6200 } else { 6201 if (_parallel) { 6202 p->oop_iterate(_par_scan_closure); 6203 } else { 6204 p->oop_iterate(_scan_closure); 6205 } 6206 } 6207 } 6208 #ifdef ASSERT 6209 if (!_parallel) { 6210 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6211 assert(_collector->overflow_list_is_empty(), 6212 "overflow list should be empty"); 6213 6214 } 6215 #endif // ASSERT 6216 return is_obj_array; 6217 } 6218 6219 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6220 MemRegion span, 6221 CMSBitMap* bitMap, CMSMarkStack* markStack, 6222 bool should_yield, bool verifying): 6223 _collector(collector), 6224 _span(span), 6225 _bitMap(bitMap), 6226 _mut(&collector->_modUnionTable), 6227 _markStack(markStack), 6228 _yield(should_yield), 6229 _skipBits(0) 6230 { 6231 assert(_markStack->isEmpty(), "stack should be empty"); 6232 _finger = _bitMap->startWord(); 6233 _threshold = _finger; 6234 assert(_collector->_restart_addr == NULL, "Sanity check"); 6235 assert(_span.contains(_finger), "Out of bounds _finger?"); 6236 DEBUG_ONLY(_verifying = verifying;) 6237 } 6238 6239 void MarkFromRootsClosure::reset(HeapWord* addr) { 6240 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6241 assert(_span.contains(addr), "Out of bounds _finger?"); 6242 _finger = addr; 6243 _threshold = align_up(_finger, CardTable::card_size); 6244 } 6245 6246 // Should revisit to see if this should be restructured for 6247 // greater efficiency. 6248 bool MarkFromRootsClosure::do_bit(size_t offset) { 6249 if (_skipBits > 0) { 6250 _skipBits--; 6251 return true; 6252 } 6253 // convert offset into a HeapWord* 6254 HeapWord* addr = _bitMap->startWord() + offset; 6255 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6256 "address out of range"); 6257 assert(_bitMap->isMarked(addr), "tautology"); 6258 if (_bitMap->isMarked(addr+1)) { 6259 // this is an allocated but not yet initialized object 6260 assert(_skipBits == 0, "tautology"); 6261 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6262 oop p = oop(addr); 6263 if (p->klass_or_null_acquire() == NULL) { 6264 DEBUG_ONLY(if (!_verifying) {) 6265 // We re-dirty the cards on which this object lies and increase 6266 // the _threshold so that we'll come back to scan this object 6267 // during the preclean or remark phase. (CMSCleanOnEnter) 6268 if (CMSCleanOnEnter) { 6269 size_t sz = _collector->block_size_using_printezis_bits(addr); 6270 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6271 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6272 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6273 // Bump _threshold to end_card_addr; note that 6274 // _threshold cannot possibly exceed end_card_addr, anyhow. 6275 // This prevents future clearing of the card as the scan proceeds 6276 // to the right. 6277 assert(_threshold <= end_card_addr, 6278 "Because we are just scanning into this object"); 6279 if (_threshold < end_card_addr) { 6280 _threshold = end_card_addr; 6281 } 6282 if (p->klass_or_null_acquire() != NULL) { 6283 // Redirty the range of cards... 6284 _mut->mark_range(redirty_range); 6285 } // ...else the setting of klass will dirty the card anyway. 6286 } 6287 DEBUG_ONLY(}) 6288 return true; 6289 } 6290 } 6291 scanOopsInOop(addr); 6292 return true; 6293 } 6294 6295 // We take a break if we've been at this for a while, 6296 // so as to avoid monopolizing the locks involved. 6297 void MarkFromRootsClosure::do_yield_work() { 6298 // First give up the locks, then yield, then re-lock 6299 // We should probably use a constructor/destructor idiom to 6300 // do this unlock/lock or modify the MutexUnlocker class to 6301 // serve our purpose. XXX 6302 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6303 "CMS thread should hold CMS token"); 6304 assert_lock_strong(_bitMap->lock()); 6305 _bitMap->lock()->unlock(); 6306 ConcurrentMarkSweepThread::desynchronize(true); 6307 _collector->stopTimer(); 6308 _collector->incrementYields(); 6309 6310 // See the comment in coordinator_yield() 6311 for (unsigned i = 0; i < CMSYieldSleepCount && 6312 ConcurrentMarkSweepThread::should_yield() && 6313 !CMSCollector::foregroundGCIsActive(); ++i) { 6314 os::sleep(Thread::current(), 1, false); 6315 } 6316 6317 ConcurrentMarkSweepThread::synchronize(true); 6318 _bitMap->lock()->lock_without_safepoint_check(); 6319 _collector->startTimer(); 6320 } 6321 6322 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6323 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6324 assert(_markStack->isEmpty(), 6325 "should drain stack to limit stack usage"); 6326 // convert ptr to an oop preparatory to scanning 6327 oop obj = oop(ptr); 6328 // Ignore mark word in verification below, since we 6329 // may be running concurrent with mutators. 6330 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6331 assert(_finger <= ptr, "_finger runneth ahead"); 6332 // advance the finger to right end of this object 6333 _finger = ptr + obj->size(); 6334 assert(_finger > ptr, "we just incremented it above"); 6335 // On large heaps, it may take us some time to get through 6336 // the marking phase. During 6337 // this time it's possible that a lot of mutations have 6338 // accumulated in the card table and the mod union table -- 6339 // these mutation records are redundant until we have 6340 // actually traced into the corresponding card. 6341 // Here, we check whether advancing the finger would make 6342 // us cross into a new card, and if so clear corresponding 6343 // cards in the MUT (preclean them in the card-table in the 6344 // future). 6345 6346 DEBUG_ONLY(if (!_verifying) {) 6347 // The clean-on-enter optimization is disabled by default, 6348 // until we fix 6178663. 6349 if (CMSCleanOnEnter && (_finger > _threshold)) { 6350 // [_threshold, _finger) represents the interval 6351 // of cards to be cleared in MUT (or precleaned in card table). 6352 // The set of cards to be cleared is all those that overlap 6353 // with the interval [_threshold, _finger); note that 6354 // _threshold is always kept card-aligned but _finger isn't 6355 // always card-aligned. 6356 HeapWord* old_threshold = _threshold; 6357 assert(is_aligned(old_threshold, CardTable::card_size), 6358 "_threshold should always be card-aligned"); 6359 _threshold = align_up(_finger, CardTable::card_size); 6360 MemRegion mr(old_threshold, _threshold); 6361 assert(!mr.is_empty(), "Control point invariant"); 6362 assert(_span.contains(mr), "Should clear within span"); 6363 _mut->clear_range(mr); 6364 } 6365 DEBUG_ONLY(}) 6366 // Note: the finger doesn't advance while we drain 6367 // the stack below. 6368 PushOrMarkClosure pushOrMarkClosure(_collector, 6369 _span, _bitMap, _markStack, 6370 _finger, this); 6371 bool res = _markStack->push(obj); 6372 assert(res, "Empty non-zero size stack should have space for single push"); 6373 while (!_markStack->isEmpty()) { 6374 oop new_oop = _markStack->pop(); 6375 // Skip verifying header mark word below because we are 6376 // running concurrent with mutators. 6377 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6378 // now scan this oop's oops 6379 new_oop->oop_iterate(&pushOrMarkClosure); 6380 do_yield_check(); 6381 } 6382 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6383 } 6384 6385 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6386 CMSCollector* collector, MemRegion span, 6387 CMSBitMap* bit_map, 6388 OopTaskQueue* work_queue, 6389 CMSMarkStack* overflow_stack): 6390 _collector(collector), 6391 _whole_span(collector->_span), 6392 _span(span), 6393 _bit_map(bit_map), 6394 _mut(&collector->_modUnionTable), 6395 _work_queue(work_queue), 6396 _overflow_stack(overflow_stack), 6397 _skip_bits(0), 6398 _task(task) 6399 { 6400 assert(_work_queue->size() == 0, "work_queue should be empty"); 6401 _finger = span.start(); 6402 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6403 assert(_span.contains(_finger), "Out of bounds _finger?"); 6404 } 6405 6406 // Should revisit to see if this should be restructured for 6407 // greater efficiency. 6408 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6409 if (_skip_bits > 0) { 6410 _skip_bits--; 6411 return true; 6412 } 6413 // convert offset into a HeapWord* 6414 HeapWord* addr = _bit_map->startWord() + offset; 6415 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6416 "address out of range"); 6417 assert(_bit_map->isMarked(addr), "tautology"); 6418 if (_bit_map->isMarked(addr+1)) { 6419 // this is an allocated object that might not yet be initialized 6420 assert(_skip_bits == 0, "tautology"); 6421 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6422 oop p = oop(addr); 6423 if (p->klass_or_null_acquire() == NULL) { 6424 // in the case of Clean-on-Enter optimization, redirty card 6425 // and avoid clearing card by increasing the threshold. 6426 return true; 6427 } 6428 } 6429 scan_oops_in_oop(addr); 6430 return true; 6431 } 6432 6433 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6434 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6435 // Should we assert that our work queue is empty or 6436 // below some drain limit? 6437 assert(_work_queue->size() == 0, 6438 "should drain stack to limit stack usage"); 6439 // convert ptr to an oop preparatory to scanning 6440 oop obj = oop(ptr); 6441 // Ignore mark word in verification below, since we 6442 // may be running concurrent with mutators. 6443 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6444 assert(_finger <= ptr, "_finger runneth ahead"); 6445 // advance the finger to right end of this object 6446 _finger = ptr + obj->size(); 6447 assert(_finger > ptr, "we just incremented it above"); 6448 // On large heaps, it may take us some time to get through 6449 // the marking phase. During 6450 // this time it's possible that a lot of mutations have 6451 // accumulated in the card table and the mod union table -- 6452 // these mutation records are redundant until we have 6453 // actually traced into the corresponding card. 6454 // Here, we check whether advancing the finger would make 6455 // us cross into a new card, and if so clear corresponding 6456 // cards in the MUT (preclean them in the card-table in the 6457 // future). 6458 6459 // The clean-on-enter optimization is disabled by default, 6460 // until we fix 6178663. 6461 if (CMSCleanOnEnter && (_finger > _threshold)) { 6462 // [_threshold, _finger) represents the interval 6463 // of cards to be cleared in MUT (or precleaned in card table). 6464 // The set of cards to be cleared is all those that overlap 6465 // with the interval [_threshold, _finger); note that 6466 // _threshold is always kept card-aligned but _finger isn't 6467 // always card-aligned. 6468 HeapWord* old_threshold = _threshold; 6469 assert(is_aligned(old_threshold, CardTable::card_size), 6470 "_threshold should always be card-aligned"); 6471 _threshold = align_up(_finger, CardTable::card_size); 6472 MemRegion mr(old_threshold, _threshold); 6473 assert(!mr.is_empty(), "Control point invariant"); 6474 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6475 _mut->clear_range(mr); 6476 } 6477 6478 // Note: the local finger doesn't advance while we drain 6479 // the stack below, but the global finger sure can and will. 6480 HeapWord* volatile* gfa = _task->global_finger_addr(); 6481 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6482 _span, _bit_map, 6483 _work_queue, 6484 _overflow_stack, 6485 _finger, 6486 gfa, this); 6487 bool res = _work_queue->push(obj); // overflow could occur here 6488 assert(res, "Will hold once we use workqueues"); 6489 while (true) { 6490 oop new_oop; 6491 if (!_work_queue->pop_local(new_oop)) { 6492 // We emptied our work_queue; check if there's stuff that can 6493 // be gotten from the overflow stack. 6494 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6495 _overflow_stack, _work_queue)) { 6496 do_yield_check(); 6497 continue; 6498 } else { // done 6499 break; 6500 } 6501 } 6502 // Skip verifying header mark word below because we are 6503 // running concurrent with mutators. 6504 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6505 // now scan this oop's oops 6506 new_oop->oop_iterate(&pushOrMarkClosure); 6507 do_yield_check(); 6508 } 6509 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6510 } 6511 6512 // Yield in response to a request from VM Thread or 6513 // from mutators. 6514 void ParMarkFromRootsClosure::do_yield_work() { 6515 assert(_task != NULL, "sanity"); 6516 _task->yield(); 6517 } 6518 6519 // A variant of the above used for verifying CMS marking work. 6520 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6521 MemRegion span, 6522 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6523 CMSMarkStack* mark_stack): 6524 _collector(collector), 6525 _span(span), 6526 _verification_bm(verification_bm), 6527 _cms_bm(cms_bm), 6528 _mark_stack(mark_stack), 6529 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6530 mark_stack) 6531 { 6532 assert(_mark_stack->isEmpty(), "stack should be empty"); 6533 _finger = _verification_bm->startWord(); 6534 assert(_collector->_restart_addr == NULL, "Sanity check"); 6535 assert(_span.contains(_finger), "Out of bounds _finger?"); 6536 } 6537 6538 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6539 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6540 assert(_span.contains(addr), "Out of bounds _finger?"); 6541 _finger = addr; 6542 } 6543 6544 // Should revisit to see if this should be restructured for 6545 // greater efficiency. 6546 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6547 // convert offset into a HeapWord* 6548 HeapWord* addr = _verification_bm->startWord() + offset; 6549 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6550 "address out of range"); 6551 assert(_verification_bm->isMarked(addr), "tautology"); 6552 assert(_cms_bm->isMarked(addr), "tautology"); 6553 6554 assert(_mark_stack->isEmpty(), 6555 "should drain stack to limit stack usage"); 6556 // convert addr to an oop preparatory to scanning 6557 oop obj = oop(addr); 6558 assert(oopDesc::is_oop(obj), "should be an oop"); 6559 assert(_finger <= addr, "_finger runneth ahead"); 6560 // advance the finger to right end of this object 6561 _finger = addr + obj->size(); 6562 assert(_finger > addr, "we just incremented it above"); 6563 // Note: the finger doesn't advance while we drain 6564 // the stack below. 6565 bool res = _mark_stack->push(obj); 6566 assert(res, "Empty non-zero size stack should have space for single push"); 6567 while (!_mark_stack->isEmpty()) { 6568 oop new_oop = _mark_stack->pop(); 6569 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6570 // now scan this oop's oops 6571 new_oop->oop_iterate(&_pam_verify_closure); 6572 } 6573 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6574 return true; 6575 } 6576 6577 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6578 CMSCollector* collector, MemRegion span, 6579 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6580 CMSMarkStack* mark_stack): 6581 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6582 _collector(collector), 6583 _span(span), 6584 _verification_bm(verification_bm), 6585 _cms_bm(cms_bm), 6586 _mark_stack(mark_stack) 6587 { } 6588 6589 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6590 oop obj = RawAccess<>::oop_load(p); 6591 do_oop(obj); 6592 } 6593 6594 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6595 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6596 6597 // Upon stack overflow, we discard (part of) the stack, 6598 // remembering the least address amongst those discarded 6599 // in CMSCollector's _restart_address. 6600 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6601 // Remember the least grey address discarded 6602 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6603 _collector->lower_restart_addr(ra); 6604 _mark_stack->reset(); // discard stack contents 6605 _mark_stack->expand(); // expand the stack if possible 6606 } 6607 6608 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6609 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6610 HeapWord* addr = (HeapWord*)obj; 6611 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6612 // Oop lies in _span and isn't yet grey or black 6613 _verification_bm->mark(addr); // now grey 6614 if (!_cms_bm->isMarked(addr)) { 6615 Log(gc, verify) log; 6616 ResourceMark rm; 6617 LogStream ls(log.error()); 6618 oop(addr)->print_on(&ls); 6619 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6620 fatal("... aborting"); 6621 } 6622 6623 if (!_mark_stack->push(obj)) { // stack overflow 6624 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6625 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6626 handle_stack_overflow(addr); 6627 } 6628 // anything including and to the right of _finger 6629 // will be scanned as we iterate over the remainder of the 6630 // bit map 6631 } 6632 } 6633 6634 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6635 MemRegion span, 6636 CMSBitMap* bitMap, CMSMarkStack* markStack, 6637 HeapWord* finger, MarkFromRootsClosure* parent) : 6638 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6639 _collector(collector), 6640 _span(span), 6641 _bitMap(bitMap), 6642 _markStack(markStack), 6643 _finger(finger), 6644 _parent(parent) 6645 { } 6646 6647 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6648 MemRegion span, 6649 CMSBitMap* bit_map, 6650 OopTaskQueue* work_queue, 6651 CMSMarkStack* overflow_stack, 6652 HeapWord* finger, 6653 HeapWord* volatile* global_finger_addr, 6654 ParMarkFromRootsClosure* parent) : 6655 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6656 _collector(collector), 6657 _whole_span(collector->_span), 6658 _span(span), 6659 _bit_map(bit_map), 6660 _work_queue(work_queue), 6661 _overflow_stack(overflow_stack), 6662 _finger(finger), 6663 _global_finger_addr(global_finger_addr), 6664 _parent(parent) 6665 { } 6666 6667 // Assumes thread-safe access by callers, who are 6668 // responsible for mutual exclusion. 6669 void CMSCollector::lower_restart_addr(HeapWord* low) { 6670 assert(_span.contains(low), "Out of bounds addr"); 6671 if (_restart_addr == NULL) { 6672 _restart_addr = low; 6673 } else { 6674 _restart_addr = MIN2(_restart_addr, low); 6675 } 6676 } 6677 6678 // Upon stack overflow, we discard (part of) the stack, 6679 // remembering the least address amongst those discarded 6680 // in CMSCollector's _restart_address. 6681 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6682 // Remember the least grey address discarded 6683 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6684 _collector->lower_restart_addr(ra); 6685 _markStack->reset(); // discard stack contents 6686 _markStack->expand(); // expand the stack if possible 6687 } 6688 6689 // Upon stack overflow, we discard (part of) the stack, 6690 // remembering the least address amongst those discarded 6691 // in CMSCollector's _restart_address. 6692 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6693 // We need to do this under a mutex to prevent other 6694 // workers from interfering with the work done below. 6695 MutexLockerEx ml(_overflow_stack->par_lock(), 6696 Mutex::_no_safepoint_check_flag); 6697 // Remember the least grey address discarded 6698 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6699 _collector->lower_restart_addr(ra); 6700 _overflow_stack->reset(); // discard stack contents 6701 _overflow_stack->expand(); // expand the stack if possible 6702 } 6703 6704 void PushOrMarkClosure::do_oop(oop obj) { 6705 // Ignore mark word because we are running concurrent with mutators. 6706 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6707 HeapWord* addr = (HeapWord*)obj; 6708 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6709 // Oop lies in _span and isn't yet grey or black 6710 _bitMap->mark(addr); // now grey 6711 if (addr < _finger) { 6712 // the bit map iteration has already either passed, or 6713 // sampled, this bit in the bit map; we'll need to 6714 // use the marking stack to scan this oop's oops. 6715 bool simulate_overflow = false; 6716 NOT_PRODUCT( 6717 if (CMSMarkStackOverflowALot && 6718 _collector->simulate_overflow()) { 6719 // simulate a stack overflow 6720 simulate_overflow = true; 6721 } 6722 ) 6723 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6724 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6725 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6726 handle_stack_overflow(addr); 6727 } 6728 } 6729 // anything including and to the right of _finger 6730 // will be scanned as we iterate over the remainder of the 6731 // bit map 6732 do_yield_check(); 6733 } 6734 } 6735 6736 void ParPushOrMarkClosure::do_oop(oop obj) { 6737 // Ignore mark word because we are running concurrent with mutators. 6738 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6739 HeapWord* addr = (HeapWord*)obj; 6740 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6741 // Oop lies in _span and isn't yet grey or black 6742 // We read the global_finger (volatile read) strictly after marking oop 6743 bool res = _bit_map->par_mark(addr); // now grey 6744 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6745 // Should we push this marked oop on our stack? 6746 // -- if someone else marked it, nothing to do 6747 // -- if target oop is above global finger nothing to do 6748 // -- if target oop is in chunk and above local finger 6749 // then nothing to do 6750 // -- else push on work queue 6751 if ( !res // someone else marked it, they will deal with it 6752 || (addr >= *gfa) // will be scanned in a later task 6753 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6754 return; 6755 } 6756 // the bit map iteration has already either passed, or 6757 // sampled, this bit in the bit map; we'll need to 6758 // use the marking stack to scan this oop's oops. 6759 bool simulate_overflow = false; 6760 NOT_PRODUCT( 6761 if (CMSMarkStackOverflowALot && 6762 _collector->simulate_overflow()) { 6763 // simulate a stack overflow 6764 simulate_overflow = true; 6765 } 6766 ) 6767 if (simulate_overflow || 6768 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6769 // stack overflow 6770 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6771 // We cannot assert that the overflow stack is full because 6772 // it may have been emptied since. 6773 assert(simulate_overflow || 6774 _work_queue->size() == _work_queue->max_elems(), 6775 "Else push should have succeeded"); 6776 handle_stack_overflow(addr); 6777 } 6778 do_yield_check(); 6779 } 6780 } 6781 6782 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6783 MemRegion span, 6784 ReferenceDiscoverer* rd, 6785 CMSBitMap* bit_map, 6786 CMSBitMap* mod_union_table, 6787 CMSMarkStack* mark_stack, 6788 bool concurrent_precleaning): 6789 MetadataVisitingOopIterateClosure(rd), 6790 _collector(collector), 6791 _span(span), 6792 _bit_map(bit_map), 6793 _mod_union_table(mod_union_table), 6794 _mark_stack(mark_stack), 6795 _concurrent_precleaning(concurrent_precleaning) 6796 { 6797 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6798 } 6799 6800 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6801 // the non-parallel version (the parallel version appears further below.) 6802 void PushAndMarkClosure::do_oop(oop obj) { 6803 // Ignore mark word verification. If during concurrent precleaning, 6804 // the object monitor may be locked. If during the checkpoint 6805 // phases, the object may already have been reached by a different 6806 // path and may be at the end of the global overflow list (so 6807 // the mark word may be NULL). 6808 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6809 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6810 HeapWord* addr = (HeapWord*)obj; 6811 // Check if oop points into the CMS generation 6812 // and is not marked 6813 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6814 // a white object ... 6815 _bit_map->mark(addr); // ... now grey 6816 // push on the marking stack (grey set) 6817 bool simulate_overflow = false; 6818 NOT_PRODUCT( 6819 if (CMSMarkStackOverflowALot && 6820 _collector->simulate_overflow()) { 6821 // simulate a stack overflow 6822 simulate_overflow = true; 6823 } 6824 ) 6825 if (simulate_overflow || !_mark_stack->push(obj)) { 6826 if (_concurrent_precleaning) { 6827 // During precleaning we can just dirty the appropriate card(s) 6828 // in the mod union table, thus ensuring that the object remains 6829 // in the grey set and continue. In the case of object arrays 6830 // we need to dirty all of the cards that the object spans, 6831 // since the rescan of object arrays will be limited to the 6832 // dirty cards. 6833 // Note that no one can be interfering with us in this action 6834 // of dirtying the mod union table, so no locking or atomics 6835 // are required. 6836 if (obj->is_objArray()) { 6837 size_t sz = obj->size(); 6838 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6839 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6840 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6841 _mod_union_table->mark_range(redirty_range); 6842 } else { 6843 _mod_union_table->mark(addr); 6844 } 6845 _collector->_ser_pmc_preclean_ovflw++; 6846 } else { 6847 // During the remark phase, we need to remember this oop 6848 // in the overflow list. 6849 _collector->push_on_overflow_list(obj); 6850 _collector->_ser_pmc_remark_ovflw++; 6851 } 6852 } 6853 } 6854 } 6855 6856 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6857 MemRegion span, 6858 ReferenceDiscoverer* rd, 6859 CMSBitMap* bit_map, 6860 OopTaskQueue* work_queue): 6861 MetadataVisitingOopIterateClosure(rd), 6862 _collector(collector), 6863 _span(span), 6864 _bit_map(bit_map), 6865 _work_queue(work_queue) 6866 { 6867 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6868 } 6869 6870 // Grey object rescan during second checkpoint phase -- 6871 // the parallel version. 6872 void ParPushAndMarkClosure::do_oop(oop obj) { 6873 // In the assert below, we ignore the mark word because 6874 // this oop may point to an already visited object that is 6875 // on the overflow stack (in which case the mark word has 6876 // been hijacked for chaining into the overflow stack -- 6877 // if this is the last object in the overflow stack then 6878 // its mark word will be NULL). Because this object may 6879 // have been subsequently popped off the global overflow 6880 // stack, and the mark word possibly restored to the prototypical 6881 // value, by the time we get to examined this failing assert in 6882 // the debugger, is_oop_or_null(false) may subsequently start 6883 // to hold. 6884 assert(oopDesc::is_oop_or_null(obj, true), 6885 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6886 HeapWord* addr = (HeapWord*)obj; 6887 // Check if oop points into the CMS generation 6888 // and is not marked 6889 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6890 // a white object ... 6891 // If we manage to "claim" the object, by being the 6892 // first thread to mark it, then we push it on our 6893 // marking stack 6894 if (_bit_map->par_mark(addr)) { // ... now grey 6895 // push on work queue (grey set) 6896 bool simulate_overflow = false; 6897 NOT_PRODUCT( 6898 if (CMSMarkStackOverflowALot && 6899 _collector->par_simulate_overflow()) { 6900 // simulate a stack overflow 6901 simulate_overflow = true; 6902 } 6903 ) 6904 if (simulate_overflow || !_work_queue->push(obj)) { 6905 _collector->par_push_on_overflow_list(obj); 6906 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6907 } 6908 } // Else, some other thread got there first 6909 } 6910 } 6911 6912 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6913 Mutex* bml = _collector->bitMapLock(); 6914 assert_lock_strong(bml); 6915 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6916 "CMS thread should hold CMS token"); 6917 6918 bml->unlock(); 6919 ConcurrentMarkSweepThread::desynchronize(true); 6920 6921 _collector->stopTimer(); 6922 _collector->incrementYields(); 6923 6924 // See the comment in coordinator_yield() 6925 for (unsigned i = 0; i < CMSYieldSleepCount && 6926 ConcurrentMarkSweepThread::should_yield() && 6927 !CMSCollector::foregroundGCIsActive(); ++i) { 6928 os::sleep(Thread::current(), 1, false); 6929 } 6930 6931 ConcurrentMarkSweepThread::synchronize(true); 6932 bml->lock(); 6933 6934 _collector->startTimer(); 6935 } 6936 6937 bool CMSPrecleanRefsYieldClosure::should_return() { 6938 if (ConcurrentMarkSweepThread::should_yield()) { 6939 do_yield_work(); 6940 } 6941 return _collector->foregroundGCIsActive(); 6942 } 6943 6944 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6945 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 6946 "mr should be aligned to start at a card boundary"); 6947 // We'd like to assert: 6948 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 6949 // "mr should be a range of cards"); 6950 // However, that would be too strong in one case -- the last 6951 // partition ends at _unallocated_block which, in general, can be 6952 // an arbitrary boundary, not necessarily card aligned. 6953 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 6954 _space->object_iterate_mem(mr, &_scan_cl); 6955 } 6956 6957 SweepClosure::SweepClosure(CMSCollector* collector, 6958 ConcurrentMarkSweepGeneration* g, 6959 CMSBitMap* bitMap, bool should_yield) : 6960 _collector(collector), 6961 _g(g), 6962 _sp(g->cmsSpace()), 6963 _limit(_sp->sweep_limit()), 6964 _freelistLock(_sp->freelistLock()), 6965 _bitMap(bitMap), 6966 _inFreeRange(false), // No free range at beginning of sweep 6967 _freeRangeInFreeLists(false), // No free range at beginning of sweep 6968 _lastFreeRangeCoalesced(false), 6969 _yield(should_yield), 6970 _freeFinger(g->used_region().start()) 6971 { 6972 NOT_PRODUCT( 6973 _numObjectsFreed = 0; 6974 _numWordsFreed = 0; 6975 _numObjectsLive = 0; 6976 _numWordsLive = 0; 6977 _numObjectsAlreadyFree = 0; 6978 _numWordsAlreadyFree = 0; 6979 _last_fc = NULL; 6980 6981 _sp->initializeIndexedFreeListArrayReturnedBytes(); 6982 _sp->dictionary()->initialize_dict_returned_bytes(); 6983 ) 6984 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 6985 "sweep _limit out of bounds"); 6986 log_develop_trace(gc, sweep)("===================="); 6987 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 6988 } 6989 6990 void SweepClosure::print_on(outputStream* st) const { 6991 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 6992 p2i(_sp->bottom()), p2i(_sp->end())); 6993 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 6994 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 6995 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 6996 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 6997 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 6998 } 6999 7000 #ifndef PRODUCT 7001 // Assertion checking only: no useful work in product mode -- 7002 // however, if any of the flags below become product flags, 7003 // you may need to review this code to see if it needs to be 7004 // enabled in product mode. 7005 SweepClosure::~SweepClosure() { 7006 assert_lock_strong(_freelistLock); 7007 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7008 "sweep _limit out of bounds"); 7009 if (inFreeRange()) { 7010 Log(gc, sweep) log; 7011 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7012 ResourceMark rm; 7013 LogStream ls(log.error()); 7014 print_on(&ls); 7015 ShouldNotReachHere(); 7016 } 7017 7018 if (log_is_enabled(Debug, gc, sweep)) { 7019 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7020 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7021 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7022 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7023 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7024 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7025 } 7026 7027 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7028 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7029 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7030 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7031 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7032 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7033 } 7034 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7035 log_develop_trace(gc, sweep)("================"); 7036 } 7037 #endif // PRODUCT 7038 7039 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7040 bool freeRangeInFreeLists) { 7041 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7042 p2i(freeFinger), freeRangeInFreeLists); 7043 assert(!inFreeRange(), "Trampling existing free range"); 7044 set_inFreeRange(true); 7045 set_lastFreeRangeCoalesced(false); 7046 7047 set_freeFinger(freeFinger); 7048 set_freeRangeInFreeLists(freeRangeInFreeLists); 7049 if (CMSTestInFreeList) { 7050 if (freeRangeInFreeLists) { 7051 FreeChunk* fc = (FreeChunk*) freeFinger; 7052 assert(fc->is_free(), "A chunk on the free list should be free."); 7053 assert(fc->size() > 0, "Free range should have a size"); 7054 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7055 } 7056 } 7057 } 7058 7059 // Note that the sweeper runs concurrently with mutators. Thus, 7060 // it is possible for direct allocation in this generation to happen 7061 // in the middle of the sweep. Note that the sweeper also coalesces 7062 // contiguous free blocks. Thus, unless the sweeper and the allocator 7063 // synchronize appropriately freshly allocated blocks may get swept up. 7064 // This is accomplished by the sweeper locking the free lists while 7065 // it is sweeping. Thus blocks that are determined to be free are 7066 // indeed free. There is however one additional complication: 7067 // blocks that have been allocated since the final checkpoint and 7068 // mark, will not have been marked and so would be treated as 7069 // unreachable and swept up. To prevent this, the allocator marks 7070 // the bit map when allocating during the sweep phase. This leads, 7071 // however, to a further complication -- objects may have been allocated 7072 // but not yet initialized -- in the sense that the header isn't yet 7073 // installed. The sweeper can not then determine the size of the block 7074 // in order to skip over it. To deal with this case, we use a technique 7075 // (due to Printezis) to encode such uninitialized block sizes in the 7076 // bit map. Since the bit map uses a bit per every HeapWord, but the 7077 // CMS generation has a minimum object size of 3 HeapWords, it follows 7078 // that "normal marks" won't be adjacent in the bit map (there will 7079 // always be at least two 0 bits between successive 1 bits). We make use 7080 // of these "unused" bits to represent uninitialized blocks -- the bit 7081 // corresponding to the start of the uninitialized object and the next 7082 // bit are both set. Finally, a 1 bit marks the end of the object that 7083 // started with the two consecutive 1 bits to indicate its potentially 7084 // uninitialized state. 7085 7086 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7087 FreeChunk* fc = (FreeChunk*)addr; 7088 size_t res; 7089 7090 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7091 // than "addr == _limit" because although _limit was a block boundary when 7092 // we started the sweep, it may no longer be one because heap expansion 7093 // may have caused us to coalesce the block ending at the address _limit 7094 // with a newly expanded chunk (this happens when _limit was set to the 7095 // previous _end of the space), so we may have stepped past _limit: 7096 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7097 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7098 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7099 "sweep _limit out of bounds"); 7100 assert(addr < _sp->end(), "addr out of bounds"); 7101 // Flush any free range we might be holding as a single 7102 // coalesced chunk to the appropriate free list. 7103 if (inFreeRange()) { 7104 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7105 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7106 flush_cur_free_chunk(freeFinger(), 7107 pointer_delta(addr, freeFinger())); 7108 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7109 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7110 lastFreeRangeCoalesced() ? 1 : 0); 7111 } 7112 7113 // help the iterator loop finish 7114 return pointer_delta(_sp->end(), addr); 7115 } 7116 7117 assert(addr < _limit, "sweep invariant"); 7118 // check if we should yield 7119 do_yield_check(addr); 7120 if (fc->is_free()) { 7121 // Chunk that is already free 7122 res = fc->size(); 7123 do_already_free_chunk(fc); 7124 debug_only(_sp->verifyFreeLists()); 7125 // If we flush the chunk at hand in lookahead_and_flush() 7126 // and it's coalesced with a preceding chunk, then the 7127 // process of "mangling" the payload of the coalesced block 7128 // will cause erasure of the size information from the 7129 // (erstwhile) header of all the coalesced blocks but the 7130 // first, so the first disjunct in the assert will not hold 7131 // in that specific case (in which case the second disjunct 7132 // will hold). 7133 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7134 "Otherwise the size info doesn't change at this step"); 7135 NOT_PRODUCT( 7136 _numObjectsAlreadyFree++; 7137 _numWordsAlreadyFree += res; 7138 ) 7139 NOT_PRODUCT(_last_fc = fc;) 7140 } else if (!_bitMap->isMarked(addr)) { 7141 // Chunk is fresh garbage 7142 res = do_garbage_chunk(fc); 7143 debug_only(_sp->verifyFreeLists()); 7144 NOT_PRODUCT( 7145 _numObjectsFreed++; 7146 _numWordsFreed += res; 7147 ) 7148 } else { 7149 // Chunk that is alive. 7150 res = do_live_chunk(fc); 7151 debug_only(_sp->verifyFreeLists()); 7152 NOT_PRODUCT( 7153 _numObjectsLive++; 7154 _numWordsLive += res; 7155 ) 7156 } 7157 return res; 7158 } 7159 7160 // For the smart allocation, record following 7161 // split deaths - a free chunk is removed from its free list because 7162 // it is being split into two or more chunks. 7163 // split birth - a free chunk is being added to its free list because 7164 // a larger free chunk has been split and resulted in this free chunk. 7165 // coal death - a free chunk is being removed from its free list because 7166 // it is being coalesced into a large free chunk. 7167 // coal birth - a free chunk is being added to its free list because 7168 // it was created when two or more free chunks where coalesced into 7169 // this free chunk. 7170 // 7171 // These statistics are used to determine the desired number of free 7172 // chunks of a given size. The desired number is chosen to be relative 7173 // to the end of a CMS sweep. The desired number at the end of a sweep 7174 // is the 7175 // count-at-end-of-previous-sweep (an amount that was enough) 7176 // - count-at-beginning-of-current-sweep (the excess) 7177 // + split-births (gains in this size during interval) 7178 // - split-deaths (demands on this size during interval) 7179 // where the interval is from the end of one sweep to the end of the 7180 // next. 7181 // 7182 // When sweeping the sweeper maintains an accumulated chunk which is 7183 // the chunk that is made up of chunks that have been coalesced. That 7184 // will be termed the left-hand chunk. A new chunk of garbage that 7185 // is being considered for coalescing will be referred to as the 7186 // right-hand chunk. 7187 // 7188 // When making a decision on whether to coalesce a right-hand chunk with 7189 // the current left-hand chunk, the current count vs. the desired count 7190 // of the left-hand chunk is considered. Also if the right-hand chunk 7191 // is near the large chunk at the end of the heap (see 7192 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7193 // left-hand chunk is coalesced. 7194 // 7195 // When making a decision about whether to split a chunk, the desired count 7196 // vs. the current count of the candidate to be split is also considered. 7197 // If the candidate is underpopulated (currently fewer chunks than desired) 7198 // a chunk of an overpopulated (currently more chunks than desired) size may 7199 // be chosen. The "hint" associated with a free list, if non-null, points 7200 // to a free list which may be overpopulated. 7201 // 7202 7203 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7204 const size_t size = fc->size(); 7205 // Chunks that cannot be coalesced are not in the 7206 // free lists. 7207 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7208 assert(_sp->verify_chunk_in_free_list(fc), 7209 "free chunk should be in free lists"); 7210 } 7211 // a chunk that is already free, should not have been 7212 // marked in the bit map 7213 HeapWord* const addr = (HeapWord*) fc; 7214 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7215 // Verify that the bit map has no bits marked between 7216 // addr and purported end of this block. 7217 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7218 7219 // Some chunks cannot be coalesced under any circumstances. 7220 // See the definition of cantCoalesce(). 7221 if (!fc->cantCoalesce()) { 7222 // This chunk can potentially be coalesced. 7223 // All the work is done in 7224 do_post_free_or_garbage_chunk(fc, size); 7225 // Note that if the chunk is not coalescable (the else arm 7226 // below), we unconditionally flush, without needing to do 7227 // a "lookahead," as we do below. 7228 if (inFreeRange()) lookahead_and_flush(fc, size); 7229 } else { 7230 // Code path common to both original and adaptive free lists. 7231 7232 // cant coalesce with previous block; this should be treated 7233 // as the end of a free run if any 7234 if (inFreeRange()) { 7235 // we kicked some butt; time to pick up the garbage 7236 assert(freeFinger() < addr, "freeFinger points too high"); 7237 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7238 } 7239 // else, nothing to do, just continue 7240 } 7241 } 7242 7243 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7244 // This is a chunk of garbage. It is not in any free list. 7245 // Add it to a free list or let it possibly be coalesced into 7246 // a larger chunk. 7247 HeapWord* const addr = (HeapWord*) fc; 7248 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7249 7250 // Verify that the bit map has no bits marked between 7251 // addr and purported end of just dead object. 7252 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7253 do_post_free_or_garbage_chunk(fc, size); 7254 7255 assert(_limit >= addr + size, 7256 "A freshly garbage chunk can't possibly straddle over _limit"); 7257 if (inFreeRange()) lookahead_and_flush(fc, size); 7258 return size; 7259 } 7260 7261 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7262 HeapWord* addr = (HeapWord*) fc; 7263 // The sweeper has just found a live object. Return any accumulated 7264 // left hand chunk to the free lists. 7265 if (inFreeRange()) { 7266 assert(freeFinger() < addr, "freeFinger points too high"); 7267 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7268 } 7269 7270 // This object is live: we'd normally expect this to be 7271 // an oop, and like to assert the following: 7272 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7273 // However, as we commented above, this may be an object whose 7274 // header hasn't yet been initialized. 7275 size_t size; 7276 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7277 if (_bitMap->isMarked(addr + 1)) { 7278 // Determine the size from the bit map, rather than trying to 7279 // compute it from the object header. 7280 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7281 size = pointer_delta(nextOneAddr + 1, addr); 7282 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7283 "alignment problem"); 7284 7285 #ifdef ASSERT 7286 if (oop(addr)->klass_or_null_acquire() != NULL) { 7287 // Ignore mark word because we are running concurrent with mutators 7288 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7289 assert(size == 7290 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7291 "P-mark and computed size do not agree"); 7292 } 7293 #endif 7294 7295 } else { 7296 // This should be an initialized object that's alive. 7297 assert(oop(addr)->klass_or_null_acquire() != NULL, 7298 "Should be an initialized object"); 7299 // Ignore mark word because we are running concurrent with mutators 7300 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7301 // Verify that the bit map has no bits marked between 7302 // addr and purported end of this block. 7303 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7304 assert(size >= 3, "Necessary for Printezis marks to work"); 7305 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7306 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7307 } 7308 return size; 7309 } 7310 7311 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7312 size_t chunkSize) { 7313 // do_post_free_or_garbage_chunk() should only be called in the case 7314 // of the adaptive free list allocator. 7315 const bool fcInFreeLists = fc->is_free(); 7316 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7317 if (CMSTestInFreeList && fcInFreeLists) { 7318 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7319 } 7320 7321 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7322 7323 HeapWord* const fc_addr = (HeapWord*) fc; 7324 7325 bool coalesce = false; 7326 const size_t left = pointer_delta(fc_addr, freeFinger()); 7327 const size_t right = chunkSize; 7328 switch (FLSCoalescePolicy) { 7329 // numeric value forms a coalition aggressiveness metric 7330 case 0: { // never coalesce 7331 coalesce = false; 7332 break; 7333 } 7334 case 1: { // coalesce if left & right chunks on overpopulated lists 7335 coalesce = _sp->coalOverPopulated(left) && 7336 _sp->coalOverPopulated(right); 7337 break; 7338 } 7339 case 2: { // coalesce if left chunk on overpopulated list (default) 7340 coalesce = _sp->coalOverPopulated(left); 7341 break; 7342 } 7343 case 3: { // coalesce if left OR right chunk on overpopulated list 7344 coalesce = _sp->coalOverPopulated(left) || 7345 _sp->coalOverPopulated(right); 7346 break; 7347 } 7348 case 4: { // always coalesce 7349 coalesce = true; 7350 break; 7351 } 7352 default: 7353 ShouldNotReachHere(); 7354 } 7355 7356 // Should the current free range be coalesced? 7357 // If the chunk is in a free range and either we decided to coalesce above 7358 // or the chunk is near the large block at the end of the heap 7359 // (isNearLargestChunk() returns true), then coalesce this chunk. 7360 const bool doCoalesce = inFreeRange() 7361 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7362 if (doCoalesce) { 7363 // Coalesce the current free range on the left with the new 7364 // chunk on the right. If either is on a free list, 7365 // it must be removed from the list and stashed in the closure. 7366 if (freeRangeInFreeLists()) { 7367 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7368 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7369 "Size of free range is inconsistent with chunk size."); 7370 if (CMSTestInFreeList) { 7371 assert(_sp->verify_chunk_in_free_list(ffc), 7372 "Chunk is not in free lists"); 7373 } 7374 _sp->coalDeath(ffc->size()); 7375 _sp->removeFreeChunkFromFreeLists(ffc); 7376 set_freeRangeInFreeLists(false); 7377 } 7378 if (fcInFreeLists) { 7379 _sp->coalDeath(chunkSize); 7380 assert(fc->size() == chunkSize, 7381 "The chunk has the wrong size or is not in the free lists"); 7382 _sp->removeFreeChunkFromFreeLists(fc); 7383 } 7384 set_lastFreeRangeCoalesced(true); 7385 print_free_block_coalesced(fc); 7386 } else { // not in a free range and/or should not coalesce 7387 // Return the current free range and start a new one. 7388 if (inFreeRange()) { 7389 // In a free range but cannot coalesce with the right hand chunk. 7390 // Put the current free range into the free lists. 7391 flush_cur_free_chunk(freeFinger(), 7392 pointer_delta(fc_addr, freeFinger())); 7393 } 7394 // Set up for new free range. Pass along whether the right hand 7395 // chunk is in the free lists. 7396 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7397 } 7398 } 7399 7400 // Lookahead flush: 7401 // If we are tracking a free range, and this is the last chunk that 7402 // we'll look at because its end crosses past _limit, we'll preemptively 7403 // flush it along with any free range we may be holding on to. Note that 7404 // this can be the case only for an already free or freshly garbage 7405 // chunk. If this block is an object, it can never straddle 7406 // over _limit. The "straddling" occurs when _limit is set at 7407 // the previous end of the space when this cycle started, and 7408 // a subsequent heap expansion caused the previously co-terminal 7409 // free block to be coalesced with the newly expanded portion, 7410 // thus rendering _limit a non-block-boundary making it dangerous 7411 // for the sweeper to step over and examine. 7412 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7413 assert(inFreeRange(), "Should only be called if currently in a free range."); 7414 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7415 assert(_sp->used_region().contains(eob - 1), 7416 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7417 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7418 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7419 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7420 if (eob >= _limit) { 7421 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7422 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7423 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7424 "[" PTR_FORMAT "," PTR_FORMAT ")", 7425 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7426 // Return the storage we are tracking back into the free lists. 7427 log_develop_trace(gc, sweep)("Flushing ... "); 7428 assert(freeFinger() < eob, "Error"); 7429 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7430 } 7431 } 7432 7433 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7434 assert(inFreeRange(), "Should only be called if currently in a free range."); 7435 assert(size > 0, 7436 "A zero sized chunk cannot be added to the free lists."); 7437 if (!freeRangeInFreeLists()) { 7438 if (CMSTestInFreeList) { 7439 FreeChunk* fc = (FreeChunk*) chunk; 7440 fc->set_size(size); 7441 assert(!_sp->verify_chunk_in_free_list(fc), 7442 "chunk should not be in free lists yet"); 7443 } 7444 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7445 // A new free range is going to be starting. The current 7446 // free range has not been added to the free lists yet or 7447 // was removed so add it back. 7448 // If the current free range was coalesced, then the death 7449 // of the free range was recorded. Record a birth now. 7450 if (lastFreeRangeCoalesced()) { 7451 _sp->coalBirth(size); 7452 } 7453 _sp->addChunkAndRepairOffsetTable(chunk, size, 7454 lastFreeRangeCoalesced()); 7455 } else { 7456 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7457 } 7458 set_inFreeRange(false); 7459 set_freeRangeInFreeLists(false); 7460 } 7461 7462 // We take a break if we've been at this for a while, 7463 // so as to avoid monopolizing the locks involved. 7464 void SweepClosure::do_yield_work(HeapWord* addr) { 7465 // Return current free chunk being used for coalescing (if any) 7466 // to the appropriate freelist. After yielding, the next 7467 // free block encountered will start a coalescing range of 7468 // free blocks. If the next free block is adjacent to the 7469 // chunk just flushed, they will need to wait for the next 7470 // sweep to be coalesced. 7471 if (inFreeRange()) { 7472 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7473 } 7474 7475 // First give up the locks, then yield, then re-lock. 7476 // We should probably use a constructor/destructor idiom to 7477 // do this unlock/lock or modify the MutexUnlocker class to 7478 // serve our purpose. XXX 7479 assert_lock_strong(_bitMap->lock()); 7480 assert_lock_strong(_freelistLock); 7481 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7482 "CMS thread should hold CMS token"); 7483 _bitMap->lock()->unlock(); 7484 _freelistLock->unlock(); 7485 ConcurrentMarkSweepThread::desynchronize(true); 7486 _collector->stopTimer(); 7487 _collector->incrementYields(); 7488 7489 // See the comment in coordinator_yield() 7490 for (unsigned i = 0; i < CMSYieldSleepCount && 7491 ConcurrentMarkSweepThread::should_yield() && 7492 !CMSCollector::foregroundGCIsActive(); ++i) { 7493 os::sleep(Thread::current(), 1, false); 7494 } 7495 7496 ConcurrentMarkSweepThread::synchronize(true); 7497 _freelistLock->lock(); 7498 _bitMap->lock()->lock_without_safepoint_check(); 7499 _collector->startTimer(); 7500 } 7501 7502 #ifndef PRODUCT 7503 // This is actually very useful in a product build if it can 7504 // be called from the debugger. Compile it into the product 7505 // as needed. 7506 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7507 return debug_cms_space->verify_chunk_in_free_list(fc); 7508 } 7509 #endif 7510 7511 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7512 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7513 p2i(fc), fc->size()); 7514 } 7515 7516 // CMSIsAliveClosure 7517 bool CMSIsAliveClosure::do_object_b(oop obj) { 7518 HeapWord* addr = (HeapWord*)obj; 7519 return addr != NULL && 7520 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7521 } 7522 7523 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7524 MemRegion span, 7525 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7526 bool cpc): 7527 _collector(collector), 7528 _span(span), 7529 _mark_stack(mark_stack), 7530 _bit_map(bit_map), 7531 _concurrent_precleaning(cpc) { 7532 assert(!_span.is_empty(), "Empty span could spell trouble"); 7533 } 7534 7535 7536 // CMSKeepAliveClosure: the serial version 7537 void CMSKeepAliveClosure::do_oop(oop obj) { 7538 HeapWord* addr = (HeapWord*)obj; 7539 if (_span.contains(addr) && 7540 !_bit_map->isMarked(addr)) { 7541 _bit_map->mark(addr); 7542 bool simulate_overflow = false; 7543 NOT_PRODUCT( 7544 if (CMSMarkStackOverflowALot && 7545 _collector->simulate_overflow()) { 7546 // simulate a stack overflow 7547 simulate_overflow = true; 7548 } 7549 ) 7550 if (simulate_overflow || !_mark_stack->push(obj)) { 7551 if (_concurrent_precleaning) { 7552 // We dirty the overflown object and let the remark 7553 // phase deal with it. 7554 assert(_collector->overflow_list_is_empty(), "Error"); 7555 // In the case of object arrays, we need to dirty all of 7556 // the cards that the object spans. No locking or atomics 7557 // are needed since no one else can be mutating the mod union 7558 // table. 7559 if (obj->is_objArray()) { 7560 size_t sz = obj->size(); 7561 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7562 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7563 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7564 _collector->_modUnionTable.mark_range(redirty_range); 7565 } else { 7566 _collector->_modUnionTable.mark(addr); 7567 } 7568 _collector->_ser_kac_preclean_ovflw++; 7569 } else { 7570 _collector->push_on_overflow_list(obj); 7571 _collector->_ser_kac_ovflw++; 7572 } 7573 } 7574 } 7575 } 7576 7577 // CMSParKeepAliveClosure: a parallel version of the above. 7578 // The work queues are private to each closure (thread), 7579 // but (may be) available for stealing by other threads. 7580 void CMSParKeepAliveClosure::do_oop(oop obj) { 7581 HeapWord* addr = (HeapWord*)obj; 7582 if (_span.contains(addr) && 7583 !_bit_map->isMarked(addr)) { 7584 // In general, during recursive tracing, several threads 7585 // may be concurrently getting here; the first one to 7586 // "tag" it, claims it. 7587 if (_bit_map->par_mark(addr)) { 7588 bool res = _work_queue->push(obj); 7589 assert(res, "Low water mark should be much less than capacity"); 7590 // Do a recursive trim in the hope that this will keep 7591 // stack usage lower, but leave some oops for potential stealers 7592 trim_queue(_low_water_mark); 7593 } // Else, another thread got there first 7594 } 7595 } 7596 7597 void CMSParKeepAliveClosure::trim_queue(uint max) { 7598 while (_work_queue->size() > max) { 7599 oop new_oop; 7600 if (_work_queue->pop_local(new_oop)) { 7601 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7602 assert(_bit_map->isMarked((HeapWord*)new_oop), 7603 "no white objects on this stack!"); 7604 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7605 // iterate over the oops in this oop, marking and pushing 7606 // the ones in CMS heap (i.e. in _span). 7607 new_oop->oop_iterate(&_mark_and_push); 7608 } 7609 } 7610 } 7611 7612 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7613 CMSCollector* collector, 7614 MemRegion span, CMSBitMap* bit_map, 7615 OopTaskQueue* work_queue): 7616 _collector(collector), 7617 _span(span), 7618 _work_queue(work_queue), 7619 _bit_map(bit_map) { } 7620 7621 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7622 HeapWord* addr = (HeapWord*)obj; 7623 if (_span.contains(addr) && 7624 !_bit_map->isMarked(addr)) { 7625 if (_bit_map->par_mark(addr)) { 7626 bool simulate_overflow = false; 7627 NOT_PRODUCT( 7628 if (CMSMarkStackOverflowALot && 7629 _collector->par_simulate_overflow()) { 7630 // simulate a stack overflow 7631 simulate_overflow = true; 7632 } 7633 ) 7634 if (simulate_overflow || !_work_queue->push(obj)) { 7635 _collector->par_push_on_overflow_list(obj); 7636 _collector->_par_kac_ovflw++; 7637 } 7638 } // Else another thread got there already 7639 } 7640 } 7641 7642 ////////////////////////////////////////////////////////////////// 7643 // CMSExpansionCause ///////////////////////////// 7644 ////////////////////////////////////////////////////////////////// 7645 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7646 switch (cause) { 7647 case _no_expansion: 7648 return "No expansion"; 7649 case _satisfy_free_ratio: 7650 return "Free ratio"; 7651 case _satisfy_promotion: 7652 return "Satisfy promotion"; 7653 case _satisfy_allocation: 7654 return "allocation"; 7655 case _allocate_par_lab: 7656 return "Par LAB"; 7657 case _allocate_par_spooling_space: 7658 return "Par Spooling Space"; 7659 case _adaptive_size_policy: 7660 return "Ergonomics"; 7661 default: 7662 return "unknown"; 7663 } 7664 } 7665 7666 void CMSDrainMarkingStackClosure::do_void() { 7667 // the max number to take from overflow list at a time 7668 const size_t num = _mark_stack->capacity()/4; 7669 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7670 "Overflow list should be NULL during concurrent phases"); 7671 while (!_mark_stack->isEmpty() || 7672 // if stack is empty, check the overflow list 7673 _collector->take_from_overflow_list(num, _mark_stack)) { 7674 oop obj = _mark_stack->pop(); 7675 HeapWord* addr = (HeapWord*)obj; 7676 assert(_span.contains(addr), "Should be within span"); 7677 assert(_bit_map->isMarked(addr), "Should be marked"); 7678 assert(oopDesc::is_oop(obj), "Should be an oop"); 7679 obj->oop_iterate(_keep_alive); 7680 } 7681 } 7682 7683 void CMSParDrainMarkingStackClosure::do_void() { 7684 // drain queue 7685 trim_queue(0); 7686 } 7687 7688 // Trim our work_queue so its length is below max at return 7689 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7690 while (_work_queue->size() > max) { 7691 oop new_oop; 7692 if (_work_queue->pop_local(new_oop)) { 7693 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7694 assert(_bit_map->isMarked((HeapWord*)new_oop), 7695 "no white objects on this stack!"); 7696 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7697 // iterate over the oops in this oop, marking and pushing 7698 // the ones in CMS heap (i.e. in _span). 7699 new_oop->oop_iterate(&_mark_and_push); 7700 } 7701 } 7702 } 7703 7704 //////////////////////////////////////////////////////////////////// 7705 // Support for Marking Stack Overflow list handling and related code 7706 //////////////////////////////////////////////////////////////////// 7707 // Much of the following code is similar in shape and spirit to the 7708 // code used in ParNewGC. We should try and share that code 7709 // as much as possible in the future. 7710 7711 #ifndef PRODUCT 7712 // Debugging support for CMSStackOverflowALot 7713 7714 // It's OK to call this multi-threaded; the worst thing 7715 // that can happen is that we'll get a bunch of closely 7716 // spaced simulated overflows, but that's OK, in fact 7717 // probably good as it would exercise the overflow code 7718 // under contention. 7719 bool CMSCollector::simulate_overflow() { 7720 if (_overflow_counter-- <= 0) { // just being defensive 7721 _overflow_counter = CMSMarkStackOverflowInterval; 7722 return true; 7723 } else { 7724 return false; 7725 } 7726 } 7727 7728 bool CMSCollector::par_simulate_overflow() { 7729 return simulate_overflow(); 7730 } 7731 #endif 7732 7733 // Single-threaded 7734 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7735 assert(stack->isEmpty(), "Expected precondition"); 7736 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7737 size_t i = num; 7738 oop cur = _overflow_list; 7739 const markOop proto = markOopDesc::prototype(); 7740 NOT_PRODUCT(ssize_t n = 0;) 7741 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7742 next = oop(cur->mark_raw()); 7743 cur->set_mark_raw(proto); // until proven otherwise 7744 assert(oopDesc::is_oop(cur), "Should be an oop"); 7745 bool res = stack->push(cur); 7746 assert(res, "Bit off more than can chew?"); 7747 NOT_PRODUCT(n++;) 7748 } 7749 _overflow_list = cur; 7750 #ifndef PRODUCT 7751 assert(_num_par_pushes >= n, "Too many pops?"); 7752 _num_par_pushes -=n; 7753 #endif 7754 return !stack->isEmpty(); 7755 } 7756 7757 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7758 // (MT-safe) Get a prefix of at most "num" from the list. 7759 // The overflow list is chained through the mark word of 7760 // each object in the list. We fetch the entire list, 7761 // break off a prefix of the right size and return the 7762 // remainder. If other threads try to take objects from 7763 // the overflow list at that time, they will wait for 7764 // some time to see if data becomes available. If (and 7765 // only if) another thread places one or more object(s) 7766 // on the global list before we have returned the suffix 7767 // to the global list, we will walk down our local list 7768 // to find its end and append the global list to 7769 // our suffix before returning it. This suffix walk can 7770 // prove to be expensive (quadratic in the amount of traffic) 7771 // when there are many objects in the overflow list and 7772 // there is much producer-consumer contention on the list. 7773 // *NOTE*: The overflow list manipulation code here and 7774 // in ParNewGeneration:: are very similar in shape, 7775 // except that in the ParNew case we use the old (from/eden) 7776 // copy of the object to thread the list via its klass word. 7777 // Because of the common code, if you make any changes in 7778 // the code below, please check the ParNew version to see if 7779 // similar changes might be needed. 7780 // CR 6797058 has been filed to consolidate the common code. 7781 bool CMSCollector::par_take_from_overflow_list(size_t num, 7782 OopTaskQueue* work_q, 7783 int no_of_gc_threads) { 7784 assert(work_q->size() == 0, "First empty local work queue"); 7785 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7786 if (_overflow_list == NULL) { 7787 return false; 7788 } 7789 // Grab the entire list; we'll put back a suffix 7790 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7791 Thread* tid = Thread::current(); 7792 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7793 // set to ParallelGCThreads. 7794 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7795 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7796 // If the list is busy, we spin for a short while, 7797 // sleeping between attempts to get the list. 7798 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7799 os::sleep(tid, sleep_time_millis, false); 7800 if (_overflow_list == NULL) { 7801 // Nothing left to take 7802 return false; 7803 } else if (_overflow_list != BUSY) { 7804 // Try and grab the prefix 7805 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7806 } 7807 } 7808 // If the list was found to be empty, or we spun long 7809 // enough, we give up and return empty-handed. If we leave 7810 // the list in the BUSY state below, it must be the case that 7811 // some other thread holds the overflow list and will set it 7812 // to a non-BUSY state in the future. 7813 if (prefix == NULL || prefix == BUSY) { 7814 // Nothing to take or waited long enough 7815 if (prefix == NULL) { 7816 // Write back the NULL in case we overwrote it with BUSY above 7817 // and it is still the same value. 7818 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7819 } 7820 return false; 7821 } 7822 assert(prefix != NULL && prefix != BUSY, "Error"); 7823 size_t i = num; 7824 oop cur = prefix; 7825 // Walk down the first "num" objects, unless we reach the end. 7826 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7827 if (cur->mark_raw() == NULL) { 7828 // We have "num" or fewer elements in the list, so there 7829 // is nothing to return to the global list. 7830 // Write back the NULL in lieu of the BUSY we wrote 7831 // above, if it is still the same value. 7832 if (_overflow_list == BUSY) { 7833 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7834 } 7835 } else { 7836 // Chop off the suffix and return it to the global list. 7837 assert(cur->mark_raw() != BUSY, "Error"); 7838 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7839 cur->set_mark_raw(NULL); // break off suffix 7840 // It's possible that the list is still in the empty(busy) state 7841 // we left it in a short while ago; in that case we may be 7842 // able to place back the suffix without incurring the cost 7843 // of a walk down the list. 7844 oop observed_overflow_list = _overflow_list; 7845 oop cur_overflow_list = observed_overflow_list; 7846 bool attached = false; 7847 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7848 observed_overflow_list = 7849 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7850 if (cur_overflow_list == observed_overflow_list) { 7851 attached = true; 7852 break; 7853 } else cur_overflow_list = observed_overflow_list; 7854 } 7855 if (!attached) { 7856 // Too bad, someone else sneaked in (at least) an element; we'll need 7857 // to do a splice. Find tail of suffix so we can prepend suffix to global 7858 // list. 7859 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7860 oop suffix_tail = cur; 7861 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7862 "Tautology"); 7863 observed_overflow_list = _overflow_list; 7864 do { 7865 cur_overflow_list = observed_overflow_list; 7866 if (cur_overflow_list != BUSY) { 7867 // Do the splice ... 7868 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7869 } else { // cur_overflow_list == BUSY 7870 suffix_tail->set_mark_raw(NULL); 7871 } 7872 // ... and try to place spliced list back on overflow_list ... 7873 observed_overflow_list = 7874 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7875 } while (cur_overflow_list != observed_overflow_list); 7876 // ... until we have succeeded in doing so. 7877 } 7878 } 7879 7880 // Push the prefix elements on work_q 7881 assert(prefix != NULL, "control point invariant"); 7882 const markOop proto = markOopDesc::prototype(); 7883 oop next; 7884 NOT_PRODUCT(ssize_t n = 0;) 7885 for (cur = prefix; cur != NULL; cur = next) { 7886 next = oop(cur->mark_raw()); 7887 cur->set_mark_raw(proto); // until proven otherwise 7888 assert(oopDesc::is_oop(cur), "Should be an oop"); 7889 bool res = work_q->push(cur); 7890 assert(res, "Bit off more than we can chew?"); 7891 NOT_PRODUCT(n++;) 7892 } 7893 #ifndef PRODUCT 7894 assert(_num_par_pushes >= n, "Too many pops?"); 7895 Atomic::sub(n, &_num_par_pushes); 7896 #endif 7897 return true; 7898 } 7899 7900 // Single-threaded 7901 void CMSCollector::push_on_overflow_list(oop p) { 7902 NOT_PRODUCT(_num_par_pushes++;) 7903 assert(oopDesc::is_oop(p), "Not an oop"); 7904 preserve_mark_if_necessary(p); 7905 p->set_mark_raw((markOop)_overflow_list); 7906 _overflow_list = p; 7907 } 7908 7909 // Multi-threaded; use CAS to prepend to overflow list 7910 void CMSCollector::par_push_on_overflow_list(oop p) { 7911 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7912 assert(oopDesc::is_oop(p), "Not an oop"); 7913 par_preserve_mark_if_necessary(p); 7914 oop observed_overflow_list = _overflow_list; 7915 oop cur_overflow_list; 7916 do { 7917 cur_overflow_list = observed_overflow_list; 7918 if (cur_overflow_list != BUSY) { 7919 p->set_mark_raw(markOop(cur_overflow_list)); 7920 } else { 7921 p->set_mark_raw(NULL); 7922 } 7923 observed_overflow_list = 7924 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7925 } while (cur_overflow_list != observed_overflow_list); 7926 } 7927 #undef BUSY 7928 7929 // Single threaded 7930 // General Note on GrowableArray: pushes may silently fail 7931 // because we are (temporarily) out of C-heap for expanding 7932 // the stack. The problem is quite ubiquitous and affects 7933 // a lot of code in the JVM. The prudent thing for GrowableArray 7934 // to do (for now) is to exit with an error. However, that may 7935 // be too draconian in some cases because the caller may be 7936 // able to recover without much harm. For such cases, we 7937 // should probably introduce a "soft_push" method which returns 7938 // an indication of success or failure with the assumption that 7939 // the caller may be able to recover from a failure; code in 7940 // the VM can then be changed, incrementally, to deal with such 7941 // failures where possible, thus, incrementally hardening the VM 7942 // in such low resource situations. 7943 void CMSCollector::preserve_mark_work(oop p, markOop m) { 7944 _preserved_oop_stack.push(p); 7945 _preserved_mark_stack.push(m); 7946 assert(m == p->mark_raw(), "Mark word changed"); 7947 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7948 "bijection"); 7949 } 7950 7951 // Single threaded 7952 void CMSCollector::preserve_mark_if_necessary(oop p) { 7953 markOop m = p->mark_raw(); 7954 if (m->must_be_preserved(p)) { 7955 preserve_mark_work(p, m); 7956 } 7957 } 7958 7959 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 7960 markOop m = p->mark_raw(); 7961 if (m->must_be_preserved(p)) { 7962 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 7963 // Even though we read the mark word without holding 7964 // the lock, we are assured that it will not change 7965 // because we "own" this oop, so no other thread can 7966 // be trying to push it on the overflow list; see 7967 // the assertion in preserve_mark_work() that checks 7968 // that m == p->mark_raw(). 7969 preserve_mark_work(p, m); 7970 } 7971 } 7972 7973 // We should be able to do this multi-threaded, 7974 // a chunk of stack being a task (this is 7975 // correct because each oop only ever appears 7976 // once in the overflow list. However, it's 7977 // not very easy to completely overlap this with 7978 // other operations, so will generally not be done 7979 // until all work's been completed. Because we 7980 // expect the preserved oop stack (set) to be small, 7981 // it's probably fine to do this single-threaded. 7982 // We can explore cleverer concurrent/overlapped/parallel 7983 // processing of preserved marks if we feel the 7984 // need for this in the future. Stack overflow should 7985 // be so rare in practice and, when it happens, its 7986 // effect on performance so great that this will 7987 // likely just be in the noise anyway. 7988 void CMSCollector::restore_preserved_marks_if_any() { 7989 assert(SafepointSynchronize::is_at_safepoint(), 7990 "world should be stopped"); 7991 assert(Thread::current()->is_ConcurrentGC_thread() || 7992 Thread::current()->is_VM_thread(), 7993 "should be single-threaded"); 7994 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7995 "bijection"); 7996 7997 while (!_preserved_oop_stack.is_empty()) { 7998 oop p = _preserved_oop_stack.pop(); 7999 assert(oopDesc::is_oop(p), "Should be an oop"); 8000 assert(_span.contains(p), "oop should be in _span"); 8001 assert(p->mark_raw() == markOopDesc::prototype(), 8002 "Set when taken from overflow list"); 8003 markOop m = _preserved_mark_stack.pop(); 8004 p->set_mark_raw(m); 8005 } 8006 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8007 "stacks were cleared above"); 8008 } 8009 8010 #ifndef PRODUCT 8011 bool CMSCollector::no_preserved_marks() const { 8012 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8013 } 8014 #endif 8015 8016 // Transfer some number of overflown objects to usual marking 8017 // stack. Return true if some objects were transferred. 8018 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8019 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8020 (size_t)ParGCDesiredObjsFromOverflowList); 8021 8022 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8023 assert(_collector->overflow_list_is_empty() || res, 8024 "If list is not empty, we should have taken something"); 8025 assert(!res || !_mark_stack->isEmpty(), 8026 "If we took something, it should now be on our stack"); 8027 return res; 8028 } 8029 8030 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8031 size_t res = _sp->block_size_no_stall(addr, _collector); 8032 if (_sp->block_is_obj(addr)) { 8033 if (_live_bit_map->isMarked(addr)) { 8034 // It can't have been dead in a previous cycle 8035 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8036 } else { 8037 _dead_bit_map->mark(addr); // mark the dead object 8038 } 8039 } 8040 // Could be 0, if the block size could not be computed without stalling. 8041 return res; 8042 } 8043 8044 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8045 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8046 switch (phase) { 8047 case CMSCollector::InitialMarking: 8048 initialize(manager /* GC manager */ , 8049 cause /* cause of the GC */, 8050 true /* allMemoryPoolsAffected */, 8051 true /* recordGCBeginTime */, 8052 true /* recordPreGCUsage */, 8053 false /* recordPeakUsage */, 8054 false /* recordPostGCusage */, 8055 true /* recordAccumulatedGCTime */, 8056 false /* recordGCEndTime */, 8057 false /* countCollection */ ); 8058 break; 8059 8060 case CMSCollector::FinalMarking: 8061 initialize(manager /* GC manager */ , 8062 cause /* cause of the GC */, 8063 true /* allMemoryPoolsAffected */, 8064 false /* recordGCBeginTime */, 8065 false /* recordPreGCUsage */, 8066 false /* recordPeakUsage */, 8067 false /* recordPostGCusage */, 8068 true /* recordAccumulatedGCTime */, 8069 false /* recordGCEndTime */, 8070 false /* countCollection */ ); 8071 break; 8072 8073 case CMSCollector::Sweeping: 8074 initialize(manager /* GC manager */ , 8075 cause /* cause of the GC */, 8076 true /* allMemoryPoolsAffected */, 8077 false /* recordGCBeginTime */, 8078 false /* recordPreGCUsage */, 8079 true /* recordPeakUsage */, 8080 true /* recordPostGCusage */, 8081 false /* recordAccumulatedGCTime */, 8082 true /* recordGCEndTime */, 8083 true /* countCollection */ ); 8084 break; 8085 8086 default: 8087 ShouldNotReachHere(); 8088 } 8089 }