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 "logging/log.hpp" 65 #include "logging/logStream.hpp" 66 #include "memory/allocation.hpp" 67 #include "memory/binaryTreeDictionary.inline.hpp" 68 #include "memory/iterator.inline.hpp" 69 #include "memory/padded.hpp" 70 #include "memory/resourceArea.hpp" 71 #include "oops/access.inline.hpp" 72 #include "oops/oop.inline.hpp" 73 #include "prims/jvmtiExport.hpp" 74 #include "runtime/atomic.hpp" 75 #include "runtime/flags/flagSetting.hpp" 76 #include "runtime/globals_extension.hpp" 77 #include "runtime/handles.inline.hpp" 78 #include "runtime/java.hpp" 79 #include "runtime/orderAccess.hpp" 80 #include "runtime/timer.hpp" 81 #include "runtime/vmThread.hpp" 82 #include "services/memoryService.hpp" 83 #include "services/runtimeService.hpp" 84 #include "utilities/align.hpp" 85 #include "utilities/stack.inline.hpp" 86 87 // statics 88 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; 89 bool CMSCollector::_full_gc_requested = false; 90 GCCause::Cause CMSCollector::_full_gc_cause = GCCause::_no_gc; 91 92 ////////////////////////////////////////////////////////////////// 93 // In support of CMS/VM thread synchronization 94 ////////////////////////////////////////////////////////////////// 95 // We split use of the CGC_lock into 2 "levels". 96 // The low-level locking is of the usual CGC_lock monitor. We introduce 97 // a higher level "token" (hereafter "CMS token") built on top of the 98 // low level monitor (hereafter "CGC lock"). 99 // The token-passing protocol gives priority to the VM thread. The 100 // CMS-lock doesn't provide any fairness guarantees, but clients 101 // should ensure that it is only held for very short, bounded 102 // durations. 103 // 104 // When either of the CMS thread or the VM thread is involved in 105 // collection operations during which it does not want the other 106 // thread to interfere, it obtains the CMS token. 107 // 108 // If either thread tries to get the token while the other has 109 // it, that thread waits. However, if the VM thread and CMS thread 110 // both want the token, then the VM thread gets priority while the 111 // CMS thread waits. This ensures, for instance, that the "concurrent" 112 // phases of the CMS thread's work do not block out the VM thread 113 // for long periods of time as the CMS thread continues to hog 114 // the token. (See bug 4616232). 115 // 116 // The baton-passing functions are, however, controlled by the 117 // flags _foregroundGCShouldWait and _foregroundGCIsActive, 118 // and here the low-level CMS lock, not the high level token, 119 // ensures mutual exclusion. 120 // 121 // Two important conditions that we have to satisfy: 122 // 1. if a thread does a low-level wait on the CMS lock, then it 123 // relinquishes the CMS token if it were holding that token 124 // when it acquired the low-level CMS lock. 125 // 2. any low-level notifications on the low-level lock 126 // should only be sent when a thread has relinquished the token. 127 // 128 // In the absence of either property, we'd have potential deadlock. 129 // 130 // We protect each of the CMS (concurrent and sequential) phases 131 // with the CMS _token_, not the CMS _lock_. 132 // 133 // The only code protected by CMS lock is the token acquisition code 134 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the 135 // baton-passing code. 136 // 137 // Unfortunately, i couldn't come up with a good abstraction to factor and 138 // hide the naked CGC_lock manipulation in the baton-passing code 139 // further below. That's something we should try to do. Also, the proof 140 // of correctness of this 2-level locking scheme is far from obvious, 141 // and potentially quite slippery. We have an uneasy suspicion, for instance, 142 // that there may be a theoretical possibility of delay/starvation in the 143 // low-level lock/wait/notify scheme used for the baton-passing because of 144 // potential interference with the priority scheme embodied in the 145 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() 146 // invocation further below and marked with "XXX 20011219YSR". 147 // Indeed, as we note elsewhere, this may become yet more slippery 148 // in the presence of multiple CMS and/or multiple VM threads. XXX 149 150 class CMSTokenSync: public StackObj { 151 private: 152 bool _is_cms_thread; 153 public: 154 CMSTokenSync(bool is_cms_thread): 155 _is_cms_thread(is_cms_thread) { 156 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), 157 "Incorrect argument to constructor"); 158 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); 159 } 160 161 ~CMSTokenSync() { 162 assert(_is_cms_thread ? 163 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : 164 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 165 "Incorrect state"); 166 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); 167 } 168 }; 169 170 // Convenience class that does a CMSTokenSync, and then acquires 171 // upto three locks. 172 class CMSTokenSyncWithLocks: public CMSTokenSync { 173 private: 174 // Note: locks are acquired in textual declaration order 175 // and released in the opposite order 176 MutexLockerEx _locker1, _locker2, _locker3; 177 public: 178 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, 179 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): 180 CMSTokenSync(is_cms_thread), 181 _locker1(mutex1, Mutex::_no_safepoint_check_flag), 182 _locker2(mutex2, Mutex::_no_safepoint_check_flag), 183 _locker3(mutex3, Mutex::_no_safepoint_check_flag) 184 { } 185 }; 186 187 188 ////////////////////////////////////////////////////////////////// 189 // Concurrent Mark-Sweep Generation ///////////////////////////// 190 ////////////////////////////////////////////////////////////////// 191 192 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) 193 194 // This struct contains per-thread things necessary to support parallel 195 // young-gen collection. 196 class CMSParGCThreadState: public CHeapObj<mtGC> { 197 public: 198 CompactibleFreeListSpaceLAB lab; 199 PromotionInfo promo; 200 201 // Constructor. 202 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { 203 promo.setSpace(cfls); 204 } 205 }; 206 207 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( 208 ReservedSpace rs, size_t initial_byte_size, CardTableRS* ct) : 209 CardGeneration(rs, initial_byte_size, ct), 210 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), 211 _did_compact(false) 212 { 213 HeapWord* bottom = (HeapWord*) _virtual_space.low(); 214 HeapWord* end = (HeapWord*) _virtual_space.high(); 215 216 _direct_allocated_words = 0; 217 NOT_PRODUCT( 218 _numObjectsPromoted = 0; 219 _numWordsPromoted = 0; 220 _numObjectsAllocated = 0; 221 _numWordsAllocated = 0; 222 ) 223 224 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end)); 225 NOT_PRODUCT(debug_cms_space = _cmsSpace;) 226 _cmsSpace->_old_gen = this; 227 228 _gc_stats = new CMSGCStats(); 229 230 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass 231 // offsets match. The ability to tell free chunks from objects 232 // depends on this property. 233 debug_only( 234 FreeChunk* junk = NULL; 235 assert(UseCompressedClassPointers || 236 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), 237 "Offset of FreeChunk::_prev within FreeChunk must match" 238 " that of OopDesc::_klass within OopDesc"); 239 ) 240 241 _par_gc_thread_states = NEW_C_HEAP_ARRAY(CMSParGCThreadState*, ParallelGCThreads, mtGC); 242 for (uint i = 0; i < ParallelGCThreads; i++) { 243 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); 244 } 245 246 _incremental_collection_failed = false; 247 // The "dilatation_factor" is the expansion that can occur on 248 // account of the fact that the minimum object size in the CMS 249 // generation may be larger than that in, say, a contiguous young 250 // generation. 251 // Ideally, in the calculation below, we'd compute the dilatation 252 // factor as: MinChunkSize/(promoting_gen's min object size) 253 // Since we do not have such a general query interface for the 254 // promoting generation, we'll instead just use the minimum 255 // object size (which today is a header's worth of space); 256 // note that all arithmetic is in units of HeapWords. 257 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); 258 assert(_dilatation_factor >= 1.0, "from previous assert"); 259 } 260 261 262 // The field "_initiating_occupancy" represents the occupancy percentage 263 // at which we trigger a new collection cycle. Unless explicitly specified 264 // via CMSInitiatingOccupancyFraction (argument "io" below), it 265 // is calculated by: 266 // 267 // Let "f" be MinHeapFreeRatio in 268 // 269 // _initiating_occupancy = 100-f + 270 // f * (CMSTriggerRatio/100) 271 // where CMSTriggerRatio is the argument "tr" below. 272 // 273 // That is, if we assume the heap is at its desired maximum occupancy at the 274 // end of a collection, we let CMSTriggerRatio of the (purported) free 275 // space be allocated before initiating a new collection cycle. 276 // 277 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, uintx tr) { 278 assert(io <= 100 && tr <= 100, "Check the arguments"); 279 if (io >= 0) { 280 _initiating_occupancy = (double)io / 100.0; 281 } else { 282 _initiating_occupancy = ((100 - MinHeapFreeRatio) + 283 (double)(tr * MinHeapFreeRatio) / 100.0) 284 / 100.0; 285 } 286 } 287 288 void ConcurrentMarkSweepGeneration::ref_processor_init() { 289 assert(collector() != NULL, "no collector"); 290 collector()->ref_processor_init(); 291 } 292 293 void CMSCollector::ref_processor_init() { 294 if (_ref_processor == NULL) { 295 // Allocate and initialize a reference processor 296 _ref_processor = 297 new ReferenceProcessor(&_span_based_discoverer, 298 (ParallelGCThreads > 1) && ParallelRefProcEnabled, // mt processing 299 ParallelGCThreads, // mt processing degree 300 _cmsGen->refs_discovery_is_mt(), // mt discovery 301 MAX2(ConcGCThreads, ParallelGCThreads), // mt discovery degree 302 _cmsGen->refs_discovery_is_atomic(), // discovery is not atomic 303 &_is_alive_closure, // closure for liveness info 304 false); // disable adjusting number of processing threads 305 // Initialize the _ref_processor field of CMSGen 306 _cmsGen->set_ref_processor(_ref_processor); 307 308 } 309 } 310 311 AdaptiveSizePolicy* CMSCollector::size_policy() { 312 return CMSHeap::heap()->size_policy(); 313 } 314 315 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { 316 317 const char* gen_name = "old"; 318 GenCollectorPolicy* gcp = CMSHeap::heap()->gen_policy(); 319 // Generation Counters - generation 1, 1 subspace 320 _gen_counters = new GenerationCounters(gen_name, 1, 1, 321 gcp->min_old_size(), gcp->max_old_size(), &_virtual_space); 322 323 _space_counters = new GSpaceCounters(gen_name, 0, 324 _virtual_space.reserved_size(), 325 this, _gen_counters); 326 } 327 328 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): 329 _cms_gen(cms_gen) 330 { 331 assert(alpha <= 100, "bad value"); 332 _saved_alpha = alpha; 333 334 // Initialize the alphas to the bootstrap value of 100. 335 _gc0_alpha = _cms_alpha = 100; 336 337 _cms_begin_time.update(); 338 _cms_end_time.update(); 339 340 _gc0_duration = 0.0; 341 _gc0_period = 0.0; 342 _gc0_promoted = 0; 343 344 _cms_duration = 0.0; 345 _cms_period = 0.0; 346 _cms_allocated = 0; 347 348 _cms_used_at_gc0_begin = 0; 349 _cms_used_at_gc0_end = 0; 350 _allow_duty_cycle_reduction = false; 351 _valid_bits = 0; 352 } 353 354 double CMSStats::cms_free_adjustment_factor(size_t free) const { 355 // TBD: CR 6909490 356 return 1.0; 357 } 358 359 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { 360 } 361 362 // If promotion failure handling is on use 363 // the padded average size of the promotion for each 364 // young generation collection. 365 double CMSStats::time_until_cms_gen_full() const { 366 size_t cms_free = _cms_gen->cmsSpace()->free(); 367 CMSHeap* heap = CMSHeap::heap(); 368 size_t expected_promotion = MIN2(heap->young_gen()->capacity(), 369 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average()); 370 if (cms_free > expected_promotion) { 371 // Start a cms collection if there isn't enough space to promote 372 // for the next young collection. Use the padded average as 373 // a safety factor. 374 cms_free -= expected_promotion; 375 376 // Adjust by the safety factor. 377 double cms_free_dbl = (double)cms_free; 378 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor) / 100.0; 379 // Apply a further correction factor which tries to adjust 380 // for recent occurance of concurrent mode failures. 381 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); 382 cms_free_dbl = cms_free_dbl * cms_adjustment; 383 384 log_trace(gc)("CMSStats::time_until_cms_gen_full: cms_free " SIZE_FORMAT " expected_promotion " SIZE_FORMAT, 385 cms_free, expected_promotion); 386 log_trace(gc)(" cms_free_dbl %f cms_consumption_rate %f", cms_free_dbl, cms_consumption_rate() + 1.0); 387 // Add 1 in case the consumption rate goes to zero. 388 return cms_free_dbl / (cms_consumption_rate() + 1.0); 389 } 390 return 0.0; 391 } 392 393 // Compare the duration of the cms collection to the 394 // time remaining before the cms generation is empty. 395 // Note that the time from the start of the cms collection 396 // to the start of the cms sweep (less than the total 397 // duration of the cms collection) can be used. This 398 // has been tried and some applications experienced 399 // promotion failures early in execution. This was 400 // possibly because the averages were not accurate 401 // enough at the beginning. 402 double CMSStats::time_until_cms_start() const { 403 // We add "gc0_period" to the "work" calculation 404 // below because this query is done (mostly) at the 405 // end of a scavenge, so we need to conservatively 406 // account for that much possible delay 407 // in the query so as to avoid concurrent mode failures 408 // due to starting the collection just a wee bit too 409 // late. 410 double work = cms_duration() + gc0_period(); 411 double deadline = time_until_cms_gen_full(); 412 // If a concurrent mode failure occurred recently, we want to be 413 // more conservative and halve our expected time_until_cms_gen_full() 414 if (work > deadline) { 415 log_develop_trace(gc)("CMSCollector: collect because of anticipated promotion before full %3.7f + %3.7f > %3.7f ", 416 cms_duration(), gc0_period(), time_until_cms_gen_full()); 417 return 0.0; 418 } 419 return work - deadline; 420 } 421 422 #ifndef PRODUCT 423 void CMSStats::print_on(outputStream *st) const { 424 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); 425 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, 426 gc0_duration(), gc0_period(), gc0_promoted()); 427 st->print(",cms_dur=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, 428 cms_duration(), cms_period(), cms_allocated()); 429 st->print(",cms_since_beg=%g,cms_since_end=%g", 430 cms_time_since_begin(), cms_time_since_end()); 431 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, 432 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); 433 434 if (valid()) { 435 st->print(",promo_rate=%g,cms_alloc_rate=%g", 436 promotion_rate(), cms_allocation_rate()); 437 st->print(",cms_consumption_rate=%g,time_until_full=%g", 438 cms_consumption_rate(), time_until_cms_gen_full()); 439 } 440 st->cr(); 441 } 442 #endif // #ifndef PRODUCT 443 444 CMSCollector::CollectorState CMSCollector::_collectorState = 445 CMSCollector::Idling; 446 bool CMSCollector::_foregroundGCIsActive = false; 447 bool CMSCollector::_foregroundGCShouldWait = false; 448 449 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, 450 CardTableRS* ct, 451 ConcurrentMarkSweepPolicy* cp): 452 _overflow_list(NULL), 453 _conc_workers(NULL), // may be set later 454 _completed_initialization(false), 455 _collection_count_start(0), 456 _should_unload_classes(CMSClassUnloadingEnabled), 457 _concurrent_cycles_since_last_unload(0), 458 _roots_scanning_options(GenCollectedHeap::SO_None), 459 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), 460 _verifying(false), 461 _collector_policy(cp), 462 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 463 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), 464 _gc_tracer_cm(new (ResourceObj::C_HEAP, mtGC) CMSTracer()), 465 _gc_timer_cm(new (ResourceObj::C_HEAP, mtGC) ConcurrentGCTimer()), 466 _cms_start_registered(false), 467 _cmsGen(cmsGen), 468 // Adjust span to cover old (cms) gen 469 _span(cmsGen->reserved()), 470 _ct(ct), 471 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), 472 _modUnionTable((CardTable::card_shift - LogHeapWordSize), 473 -1 /* lock-free */, "No_lock" /* dummy */), 474 _restart_addr(NULL), 475 _ser_pmc_preclean_ovflw(0), 476 _ser_pmc_remark_ovflw(0), 477 _par_pmc_remark_ovflw(0), 478 _ser_kac_preclean_ovflw(0), 479 _ser_kac_ovflw(0), 480 _par_kac_ovflw(0), 481 #ifndef PRODUCT 482 _num_par_pushes(0), 483 #endif 484 _span_based_discoverer(_span), 485 _ref_processor(NULL), // will be set later 486 // Construct the is_alive_closure with _span & markBitMap 487 _is_alive_closure(_span, &_markBitMap), 488 _modUnionClosurePar(&_modUnionTable), 489 _between_prologue_and_epilogue(false), 490 _abort_preclean(false), 491 _start_sampling(false), 492 _stats(cmsGen), 493 _eden_chunk_lock(new Mutex(Mutex::leaf + 1, "CMS_eden_chunk_lock", true, 494 //verify that this lock should be acquired with safepoint check. 495 Monitor::_safepoint_check_sometimes)), 496 _eden_chunk_array(NULL), // may be set in ctor body 497 _eden_chunk_index(0), // -- ditto -- 498 _eden_chunk_capacity(0), // -- ditto -- 499 _survivor_chunk_array(NULL), // -- ditto -- 500 _survivor_chunk_index(0), // -- ditto -- 501 _survivor_chunk_capacity(0), // -- ditto -- 502 _survivor_plab_array(NULL) // -- ditto -- 503 { 504 // Now expand the span and allocate the collection support structures 505 // (MUT, marking bit map etc.) to cover both generations subject to 506 // collection. 507 508 // For use by dirty card to oop closures. 509 _cmsGen->cmsSpace()->set_collector(this); 510 511 // Allocate MUT and marking bit map 512 { 513 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); 514 if (!_markBitMap.allocate(_span)) { 515 log_warning(gc)("Failed to allocate CMS Bit Map"); 516 return; 517 } 518 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); 519 } 520 { 521 _modUnionTable.allocate(_span); 522 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); 523 } 524 525 if (!_markStack.allocate(MarkStackSize)) { 526 log_warning(gc)("Failed to allocate CMS Marking Stack"); 527 return; 528 } 529 530 // Support for multi-threaded concurrent phases 531 if (CMSConcurrentMTEnabled) { 532 if (FLAG_IS_DEFAULT(ConcGCThreads)) { 533 // just for now 534 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3) / 4); 535 } 536 if (ConcGCThreads > 1) { 537 _conc_workers = new YieldingFlexibleWorkGang("CMS Thread", 538 ConcGCThreads, true); 539 if (_conc_workers == NULL) { 540 log_warning(gc)("GC/CMS: _conc_workers allocation failure: forcing -CMSConcurrentMTEnabled"); 541 CMSConcurrentMTEnabled = false; 542 } else { 543 _conc_workers->initialize_workers(); 544 } 545 } else { 546 CMSConcurrentMTEnabled = false; 547 } 548 } 549 if (!CMSConcurrentMTEnabled) { 550 ConcGCThreads = 0; 551 } else { 552 // Turn off CMSCleanOnEnter optimization temporarily for 553 // the MT case where it's not fixed yet; see 6178663. 554 CMSCleanOnEnter = false; 555 } 556 assert((_conc_workers != NULL) == (ConcGCThreads > 1), 557 "Inconsistency"); 558 log_debug(gc)("ConcGCThreads: %u", ConcGCThreads); 559 log_debug(gc)("ParallelGCThreads: %u", ParallelGCThreads); 560 561 // Parallel task queues; these are shared for the 562 // concurrent and stop-world phases of CMS, but 563 // are not shared with parallel scavenge (ParNew). 564 { 565 uint i; 566 uint num_queues = MAX2(ParallelGCThreads, ConcGCThreads); 567 568 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled 569 || ParallelRefProcEnabled) 570 && num_queues > 0) { 571 _task_queues = new OopTaskQueueSet(num_queues); 572 if (_task_queues == NULL) { 573 log_warning(gc)("task_queues allocation failure."); 574 return; 575 } 576 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; 577 for (i = 0; i < num_queues; i++) { 578 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); 579 if (q == NULL) { 580 log_warning(gc)("work_queue allocation failure."); 581 return; 582 } 583 _task_queues->register_queue(i, q); 584 } 585 for (i = 0; i < num_queues; i++) { 586 _task_queues->queue(i)->initialize(); 587 } 588 } 589 } 590 591 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); 592 593 // Clip CMSBootstrapOccupancy between 0 and 100. 594 _bootstrap_occupancy = CMSBootstrapOccupancy / 100.0; 595 596 // Now tell CMS generations the identity of their collector 597 ConcurrentMarkSweepGeneration::set_collector(this); 598 599 // Create & start a CMS thread for this CMS collector 600 _cmsThread = ConcurrentMarkSweepThread::start(this); 601 assert(cmsThread() != NULL, "CMS Thread should have been created"); 602 assert(cmsThread()->collector() == this, 603 "CMS Thread should refer to this gen"); 604 assert(CGC_lock != NULL, "Where's the CGC_lock?"); 605 606 // Support for parallelizing young gen rescan 607 CMSHeap* heap = CMSHeap::heap(); 608 _young_gen = heap->young_gen(); 609 if (heap->supports_inline_contig_alloc()) { 610 _top_addr = heap->top_addr(); 611 _end_addr = heap->end_addr(); 612 assert(_young_gen != NULL, "no _young_gen"); 613 _eden_chunk_index = 0; 614 _eden_chunk_capacity = (_young_gen->max_capacity() + CMSSamplingGrain) / CMSSamplingGrain; 615 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity, mtGC); 616 } 617 618 // Support for parallelizing survivor space rescan 619 if ((CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) || CMSParallelInitialMarkEnabled) { 620 const size_t max_plab_samples = 621 _young_gen->max_survivor_size() / (PLAB::min_size() * HeapWordSize); 622 623 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads, mtGC); 624 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 625 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads, mtGC); 626 _survivor_chunk_capacity = max_plab_samples; 627 for (uint i = 0; i < ParallelGCThreads; i++) { 628 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples, mtGC); 629 ChunkArray* cur = ::new (&_survivor_plab_array[i]) ChunkArray(vec, max_plab_samples); 630 assert(cur->end() == 0, "Should be 0"); 631 assert(cur->array() == vec, "Should be vec"); 632 assert(cur->capacity() == max_plab_samples, "Error"); 633 } 634 } 635 636 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) 637 _gc_counters = new CollectorCounters("CMS", 1); 638 _cgc_counters = new CollectorCounters("CMS stop-the-world phases", 2); 639 _completed_initialization = true; 640 _inter_sweep_timer.start(); // start of time 641 } 642 643 const char* ConcurrentMarkSweepGeneration::name() const { 644 return "concurrent mark-sweep generation"; 645 } 646 void ConcurrentMarkSweepGeneration::update_counters() { 647 if (UsePerfData) { 648 _space_counters->update_all(); 649 _gen_counters->update_all(); 650 } 651 } 652 653 // this is an optimized version of update_counters(). it takes the 654 // used value as a parameter rather than computing it. 655 // 656 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { 657 if (UsePerfData) { 658 _space_counters->update_used(used); 659 _space_counters->update_capacity(); 660 _gen_counters->update_all(); 661 } 662 } 663 664 void ConcurrentMarkSweepGeneration::print() const { 665 Generation::print(); 666 cmsSpace()->print(); 667 } 668 669 #ifndef PRODUCT 670 void ConcurrentMarkSweepGeneration::print_statistics() { 671 cmsSpace()->printFLCensus(0); 672 } 673 #endif 674 675 size_t 676 ConcurrentMarkSweepGeneration::contiguous_available() const { 677 // dld proposes an improvement in precision here. If the committed 678 // part of the space ends in a free block we should add that to 679 // uncommitted size in the calculation below. Will make this 680 // change later, staying with the approximation below for the 681 // time being. -- ysr. 682 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); 683 } 684 685 size_t 686 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { 687 return _cmsSpace->max_alloc_in_words() * HeapWordSize; 688 } 689 690 size_t ConcurrentMarkSweepGeneration::max_available() const { 691 return free() + _virtual_space.uncommitted_size(); 692 } 693 694 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe(size_t max_promotion_in_bytes) const { 695 size_t available = max_available(); 696 size_t av_promo = (size_t)gc_stats()->avg_promoted()->padded_average(); 697 bool res = (available >= av_promo) || (available >= max_promotion_in_bytes); 698 log_trace(gc, promotion)("CMS: promo attempt is%s safe: available(" SIZE_FORMAT ") %s av_promo(" SIZE_FORMAT "), max_promo(" SIZE_FORMAT ")", 699 res? "":" not", available, res? ">=":"<", av_promo, max_promotion_in_bytes); 700 return res; 701 } 702 703 // At a promotion failure dump information on block layout in heap 704 // (cms old generation). 705 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { 706 Log(gc, promotion) log; 707 if (log.is_trace()) { 708 LogStream ls(log.trace()); 709 cmsSpace()->dump_at_safepoint_with_locks(collector(), &ls); 710 } 711 } 712 713 void ConcurrentMarkSweepGeneration::reset_after_compaction() { 714 // Clear the promotion information. These pointers can be adjusted 715 // along with all the other pointers into the heap but 716 // compaction is expected to be a rare event with 717 // a heap using cms so don't do it without seeing the need. 718 for (uint i = 0; i < ParallelGCThreads; i++) { 719 _par_gc_thread_states[i]->promo.reset(); 720 } 721 } 722 723 void ConcurrentMarkSweepGeneration::compute_new_size() { 724 assert_locked_or_safepoint(Heap_lock); 725 726 // If incremental collection failed, we just want to expand 727 // to the limit. 728 if (incremental_collection_failed()) { 729 clear_incremental_collection_failed(); 730 grow_to_reserved(); 731 return; 732 } 733 734 // The heap has been compacted but not reset yet. 735 // Any metric such as free() or used() will be incorrect. 736 737 CardGeneration::compute_new_size(); 738 739 // Reset again after a possible resizing 740 if (did_compact()) { 741 cmsSpace()->reset_after_compaction(); 742 } 743 } 744 745 void ConcurrentMarkSweepGeneration::compute_new_size_free_list() { 746 assert_locked_or_safepoint(Heap_lock); 747 748 // If incremental collection failed, we just want to expand 749 // to the limit. 750 if (incremental_collection_failed()) { 751 clear_incremental_collection_failed(); 752 grow_to_reserved(); 753 return; 754 } 755 756 double free_percentage = ((double) free()) / capacity(); 757 double desired_free_percentage = (double) MinHeapFreeRatio / 100; 758 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; 759 760 // compute expansion delta needed for reaching desired free percentage 761 if (free_percentage < desired_free_percentage) { 762 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 763 assert(desired_capacity >= capacity(), "invalid expansion size"); 764 size_t expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); 765 Log(gc) log; 766 if (log.is_trace()) { 767 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 768 log.trace("From compute_new_size: "); 769 log.trace(" Free fraction %f", free_percentage); 770 log.trace(" Desired free fraction %f", desired_free_percentage); 771 log.trace(" Maximum free fraction %f", maximum_free_percentage); 772 log.trace(" Capacity " SIZE_FORMAT, capacity() / 1000); 773 log.trace(" Desired capacity " SIZE_FORMAT, desired_capacity / 1000); 774 CMSHeap* heap = CMSHeap::heap(); 775 size_t young_size = heap->young_gen()->capacity(); 776 log.trace(" Young gen size " SIZE_FORMAT, young_size / 1000); 777 log.trace(" unsafe_max_alloc_nogc " SIZE_FORMAT, unsafe_max_alloc_nogc() / 1000); 778 log.trace(" contiguous available " SIZE_FORMAT, contiguous_available() / 1000); 779 log.trace(" Expand by " SIZE_FORMAT " (bytes)", expand_bytes); 780 } 781 // safe if expansion fails 782 expand_for_gc_cause(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); 783 log.trace(" Expanded free fraction %f", ((double) free()) / capacity()); 784 } else { 785 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); 786 assert(desired_capacity <= capacity(), "invalid expansion size"); 787 size_t shrink_bytes = capacity() - desired_capacity; 788 // Don't shrink unless the delta is greater than the minimum shrink we want 789 if (shrink_bytes >= MinHeapDeltaBytes) { 790 shrink_free_list_by(shrink_bytes); 791 } 792 } 793 } 794 795 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { 796 return cmsSpace()->freelistLock(); 797 } 798 799 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, bool tlab) { 800 CMSSynchronousYieldRequest yr; 801 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 802 return have_lock_and_allocate(size, tlab); 803 } 804 805 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, 806 bool tlab /* ignored */) { 807 assert_lock_strong(freelistLock()); 808 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); 809 HeapWord* res = cmsSpace()->allocate(adjustedSize); 810 // Allocate the object live (grey) if the background collector has 811 // started marking. This is necessary because the marker may 812 // have passed this address and consequently this object will 813 // not otherwise be greyed and would be incorrectly swept up. 814 // Note that if this object contains references, the writing 815 // of those references will dirty the card containing this object 816 // allowing the object to be blackened (and its references scanned) 817 // either during a preclean phase or at the final checkpoint. 818 if (res != NULL) { 819 // We may block here with an uninitialized object with 820 // its mark-bit or P-bits not yet set. Such objects need 821 // to be safely navigable by block_start(). 822 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); 823 assert(!((FreeChunk*)res)->is_free(), "Error, block will look free but show wrong size"); 824 collector()->direct_allocated(res, adjustedSize); 825 _direct_allocated_words += adjustedSize; 826 // allocation counters 827 NOT_PRODUCT( 828 _numObjectsAllocated++; 829 _numWordsAllocated += (int)adjustedSize; 830 ) 831 } 832 return res; 833 } 834 835 // In the case of direct allocation by mutators in a generation that 836 // is being concurrently collected, the object must be allocated 837 // live (grey) if the background collector has started marking. 838 // This is necessary because the marker may 839 // have passed this address and consequently this object will 840 // not otherwise be greyed and would be incorrectly swept up. 841 // Note that if this object contains references, the writing 842 // of those references will dirty the card containing this object 843 // allowing the object to be blackened (and its references scanned) 844 // either during a preclean phase or at the final checkpoint. 845 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { 846 assert(_markBitMap.covers(start, size), "Out of bounds"); 847 if (_collectorState >= Marking) { 848 MutexLockerEx y(_markBitMap.lock(), 849 Mutex::_no_safepoint_check_flag); 850 // [see comments preceding SweepClosure::do_blk() below for details] 851 // 852 // Can the P-bits be deleted now? JJJ 853 // 854 // 1. need to mark the object as live so it isn't collected 855 // 2. need to mark the 2nd bit to indicate the object may be uninitialized 856 // 3. need to mark the end of the object so marking, precleaning or sweeping 857 // can skip over uninitialized or unparsable objects. An allocated 858 // object is considered uninitialized for our purposes as long as 859 // its klass word is NULL. All old gen objects are parsable 860 // as soon as they are initialized.) 861 _markBitMap.mark(start); // object is live 862 _markBitMap.mark(start + 1); // object is potentially uninitialized? 863 _markBitMap.mark(start + size - 1); 864 // mark end of object 865 } 866 // check that oop looks uninitialized 867 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); 868 } 869 870 void CMSCollector::promoted(bool par, HeapWord* start, 871 bool is_obj_array, size_t obj_size) { 872 assert(_markBitMap.covers(start), "Out of bounds"); 873 // See comment in direct_allocated() about when objects should 874 // be allocated live. 875 if (_collectorState >= Marking) { 876 // we already hold the marking bit map lock, taken in 877 // the prologue 878 if (par) { 879 _markBitMap.par_mark(start); 880 } else { 881 _markBitMap.mark(start); 882 } 883 // We don't need to mark the object as uninitialized (as 884 // in direct_allocated above) because this is being done with the 885 // world stopped and the object will be initialized by the 886 // time the marking, precleaning or sweeping get to look at it. 887 // But see the code for copying objects into the CMS generation, 888 // where we need to ensure that concurrent readers of the 889 // block offset table are able to safely navigate a block that 890 // is in flux from being free to being allocated (and in 891 // transition while being copied into) and subsequently 892 // becoming a bona-fide object when the copy/promotion is complete. 893 assert(SafepointSynchronize::is_at_safepoint(), 894 "expect promotion only at safepoints"); 895 896 if (_collectorState < Sweeping) { 897 // Mark the appropriate cards in the modUnionTable, so that 898 // this object gets scanned before the sweep. If this is 899 // not done, CMS generation references in the object might 900 // not get marked. 901 // For the case of arrays, which are otherwise precisely 902 // marked, we need to dirty the entire array, not just its head. 903 if (is_obj_array) { 904 // The [par_]mark_range() method expects mr.end() below to 905 // be aligned to the granularity of a bit's representation 906 // in the heap. In the case of the MUT below, that's a 907 // card size. 908 MemRegion mr(start, 909 align_up(start + obj_size, 910 CardTable::card_size /* bytes */)); 911 if (par) { 912 _modUnionTable.par_mark_range(mr); 913 } else { 914 _modUnionTable.mark_range(mr); 915 } 916 } else { // not an obj array; we can just mark the head 917 if (par) { 918 _modUnionTable.par_mark(start); 919 } else { 920 _modUnionTable.mark(start); 921 } 922 } 923 } 924 } 925 } 926 927 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { 928 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); 929 // allocate, copy and if necessary update promoinfo -- 930 // delegate to underlying space. 931 assert_lock_strong(freelistLock()); 932 933 #ifndef PRODUCT 934 if (CMSHeap::heap()->promotion_should_fail()) { 935 return NULL; 936 } 937 #endif // #ifndef PRODUCT 938 939 oop res = _cmsSpace->promote(obj, obj_size); 940 if (res == NULL) { 941 // expand and retry 942 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords 943 expand_for_gc_cause(s*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_promotion); 944 // Since this is the old generation, we don't try to promote 945 // into a more senior generation. 946 res = _cmsSpace->promote(obj, obj_size); 947 } 948 if (res != NULL) { 949 // See comment in allocate() about when objects should 950 // be allocated live. 951 assert(oopDesc::is_oop(obj), "Will dereference klass pointer below"); 952 collector()->promoted(false, // Not parallel 953 (HeapWord*)res, obj->is_objArray(), obj_size); 954 // promotion counters 955 NOT_PRODUCT( 956 _numObjectsPromoted++; 957 _numWordsPromoted += 958 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); 959 ) 960 } 961 return res; 962 } 963 964 965 // IMPORTANT: Notes on object size recognition in CMS. 966 // --------------------------------------------------- 967 // A block of storage in the CMS generation is always in 968 // one of three states. A free block (FREE), an allocated 969 // object (OBJECT) whose size() method reports the correct size, 970 // and an intermediate state (TRANSIENT) in which its size cannot 971 // be accurately determined. 972 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) 973 // ----------------------------------------------------- 974 // FREE: klass_word & 1 == 1; mark_word holds block size 975 // 976 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; 977 // obj->size() computes correct size 978 // 979 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 980 // 981 // STATE IDENTIFICATION: (64 bit+COOPS) 982 // ------------------------------------ 983 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size 984 // 985 // OBJECT: klass_word installed; klass_word != 0; 986 // obj->size() computes correct size 987 // 988 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT 989 // 990 // 991 // STATE TRANSITION DIAGRAM 992 // 993 // mut / parnew mut / parnew 994 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| 995 // ^ | 996 // |------------------------ DEAD <------------------------------------| 997 // sweep mut 998 // 999 // While a block is in TRANSIENT state its size cannot be determined 1000 // so readers will either need to come back later or stall until 1001 // the size can be determined. Note that for the case of direct 1002 // allocation, P-bits, when available, may be used to determine the 1003 // size of an object that may not yet have been initialized. 1004 1005 // Things to support parallel young-gen collection. 1006 oop 1007 ConcurrentMarkSweepGeneration::par_promote(int thread_num, 1008 oop old, markOop m, 1009 size_t word_sz) { 1010 #ifndef PRODUCT 1011 if (CMSHeap::heap()->promotion_should_fail()) { 1012 return NULL; 1013 } 1014 #endif // #ifndef PRODUCT 1015 1016 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1017 PromotionInfo* promoInfo = &ps->promo; 1018 // if we are tracking promotions, then first ensure space for 1019 // promotion (including spooling space for saving header if necessary). 1020 // then allocate and copy, then track promoted info if needed. 1021 // When tracking (see PromotionInfo::track()), the mark word may 1022 // be displaced and in this case restoration of the mark word 1023 // occurs in the (oop_since_save_marks_)iterate phase. 1024 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { 1025 // Out of space for allocating spooling buffers; 1026 // try expanding and allocating spooling buffers. 1027 if (!expand_and_ensure_spooling_space(promoInfo)) { 1028 return NULL; 1029 } 1030 } 1031 assert(!promoInfo->tracking() || promoInfo->has_spooling_space(), "Control point invariant"); 1032 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); 1033 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); 1034 if (obj_ptr == NULL) { 1035 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); 1036 if (obj_ptr == NULL) { 1037 return NULL; 1038 } 1039 } 1040 oop obj = oop(obj_ptr); 1041 OrderAccess::storestore(); 1042 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1043 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1044 // IMPORTANT: See note on object initialization for CMS above. 1045 // Otherwise, copy the object. Here we must be careful to insert the 1046 // klass pointer last, since this marks the block as an allocated object. 1047 // Except with compressed oops it's the mark word. 1048 HeapWord* old_ptr = (HeapWord*)old; 1049 // Restore the mark word copied above. 1050 obj->set_mark_raw(m); 1051 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1052 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1053 OrderAccess::storestore(); 1054 1055 if (UseCompressedClassPointers) { 1056 // Copy gap missed by (aligned) header size calculation below 1057 obj->set_klass_gap(old->klass_gap()); 1058 } 1059 if (word_sz > (size_t)oopDesc::header_size()) { 1060 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), 1061 obj_ptr + oopDesc::header_size(), 1062 word_sz - oopDesc::header_size()); 1063 } 1064 1065 // Now we can track the promoted object, if necessary. We take care 1066 // to delay the transition from uninitialized to full object 1067 // (i.e., insertion of klass pointer) until after, so that it 1068 // atomically becomes a promoted object. 1069 if (promoInfo->tracking()) { 1070 promoInfo->track((PromotedObject*)obj, old->klass()); 1071 } 1072 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); 1073 assert(!((FreeChunk*)obj_ptr)->is_free(), "Error, block will look free but show wrong size"); 1074 assert(oopDesc::is_oop(old), "Will use and dereference old klass ptr below"); 1075 1076 // Finally, install the klass pointer (this should be volatile). 1077 OrderAccess::storestore(); 1078 obj->set_klass(old->klass()); 1079 // We should now be able to calculate the right size for this object 1080 assert(oopDesc::is_oop(obj) && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); 1081 1082 collector()->promoted(true, // parallel 1083 obj_ptr, old->is_objArray(), word_sz); 1084 1085 NOT_PRODUCT( 1086 Atomic::inc(&_numObjectsPromoted); 1087 Atomic::add(alloc_sz, &_numWordsPromoted); 1088 ) 1089 1090 return obj; 1091 } 1092 1093 void 1094 ConcurrentMarkSweepGeneration:: 1095 par_promote_alloc_done(int thread_num) { 1096 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1097 ps->lab.retire(thread_num); 1098 } 1099 1100 void 1101 ConcurrentMarkSweepGeneration:: 1102 par_oop_since_save_marks_iterate_done(int thread_num) { 1103 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; 1104 ParScanWithoutBarrierClosure* dummy_cl = NULL; 1105 ps->promo.promoted_oops_iterate(dummy_cl); 1106 1107 // Because card-scanning has been completed, subsequent phases 1108 // (e.g., reference processing) will not need to recognize which 1109 // objects have been promoted during this GC. So, we can now disable 1110 // promotion tracking. 1111 ps->promo.stopTrackingPromotions(); 1112 } 1113 1114 bool ConcurrentMarkSweepGeneration::should_collect(bool full, 1115 size_t size, 1116 bool tlab) 1117 { 1118 // We allow a STW collection only if a full 1119 // collection was requested. 1120 return full || should_allocate(size, tlab); // FIX ME !!! 1121 // This and promotion failure handling are connected at the 1122 // hip and should be fixed by untying them. 1123 } 1124 1125 bool CMSCollector::shouldConcurrentCollect() { 1126 LogTarget(Trace, gc) log; 1127 1128 if (_full_gc_requested) { 1129 log.print("CMSCollector: collect because of explicit gc request (or GCLocker)"); 1130 return true; 1131 } 1132 1133 FreelistLocker x(this); 1134 // ------------------------------------------------------------------ 1135 // Print out lots of information which affects the initiation of 1136 // a collection. 1137 if (log.is_enabled() && stats().valid()) { 1138 log.print("CMSCollector shouldConcurrentCollect: "); 1139 1140 LogStream out(log); 1141 stats().print_on(&out); 1142 1143 log.print("time_until_cms_gen_full %3.7f", stats().time_until_cms_gen_full()); 1144 log.print("free=" SIZE_FORMAT, _cmsGen->free()); 1145 log.print("contiguous_available=" SIZE_FORMAT, _cmsGen->contiguous_available()); 1146 log.print("promotion_rate=%g", stats().promotion_rate()); 1147 log.print("cms_allocation_rate=%g", stats().cms_allocation_rate()); 1148 log.print("occupancy=%3.7f", _cmsGen->occupancy()); 1149 log.print("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); 1150 log.print("cms_time_since_begin=%3.7f", stats().cms_time_since_begin()); 1151 log.print("cms_time_since_end=%3.7f", stats().cms_time_since_end()); 1152 log.print("metadata initialized %d", MetaspaceGC::should_concurrent_collect()); 1153 } 1154 // ------------------------------------------------------------------ 1155 1156 // If the estimated time to complete a cms collection (cms_duration()) 1157 // is less than the estimated time remaining until the cms generation 1158 // is full, start a collection. 1159 if (!UseCMSInitiatingOccupancyOnly) { 1160 if (stats().valid()) { 1161 if (stats().time_until_cms_start() == 0.0) { 1162 return true; 1163 } 1164 } else { 1165 // We want to conservatively collect somewhat early in order 1166 // to try and "bootstrap" our CMS/promotion statistics; 1167 // this branch will not fire after the first successful CMS 1168 // collection because the stats should then be valid. 1169 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { 1170 log.print(" CMSCollector: collect for bootstrapping statistics: occupancy = %f, boot occupancy = %f", 1171 _cmsGen->occupancy(), _bootstrap_occupancy); 1172 return true; 1173 } 1174 } 1175 } 1176 1177 // Otherwise, we start a collection cycle if 1178 // old gen want a collection cycle started. Each may use 1179 // an appropriate criterion for making this decision. 1180 // XXX We need to make sure that the gen expansion 1181 // criterion dovetails well with this. XXX NEED TO FIX THIS 1182 if (_cmsGen->should_concurrent_collect()) { 1183 log.print("CMS old gen initiated"); 1184 return true; 1185 } 1186 1187 // We start a collection if we believe an incremental collection may fail; 1188 // this is not likely to be productive in practice because it's probably too 1189 // late anyway. 1190 CMSHeap* heap = CMSHeap::heap(); 1191 if (heap->incremental_collection_will_fail(true /* consult_young */)) { 1192 log.print("CMSCollector: collect because incremental collection will fail "); 1193 return true; 1194 } 1195 1196 if (MetaspaceGC::should_concurrent_collect()) { 1197 log.print("CMSCollector: collect for metadata allocation "); 1198 return true; 1199 } 1200 1201 // CMSTriggerInterval starts a CMS cycle if enough time has passed. 1202 if (CMSTriggerInterval >= 0) { 1203 if (CMSTriggerInterval == 0) { 1204 // Trigger always 1205 return true; 1206 } 1207 1208 // Check the CMS time since begin (we do not check the stats validity 1209 // as we want to be able to trigger the first CMS cycle as well) 1210 if (stats().cms_time_since_begin() >= (CMSTriggerInterval / ((double) MILLIUNITS))) { 1211 if (stats().valid()) { 1212 log.print("CMSCollector: collect because of trigger interval (time since last begin %3.7f secs)", 1213 stats().cms_time_since_begin()); 1214 } else { 1215 log.print("CMSCollector: collect because of trigger interval (first collection)"); 1216 } 1217 return true; 1218 } 1219 } 1220 1221 return false; 1222 } 1223 1224 void CMSCollector::set_did_compact(bool v) { _cmsGen->set_did_compact(v); } 1225 1226 // Clear _expansion_cause fields of constituent generations 1227 void CMSCollector::clear_expansion_cause() { 1228 _cmsGen->clear_expansion_cause(); 1229 } 1230 1231 // We should be conservative in starting a collection cycle. To 1232 // start too eagerly runs the risk of collecting too often in the 1233 // extreme. To collect too rarely falls back on full collections, 1234 // which works, even if not optimum in terms of concurrent work. 1235 // As a work around for too eagerly collecting, use the flag 1236 // UseCMSInitiatingOccupancyOnly. This also has the advantage of 1237 // giving the user an easily understandable way of controlling the 1238 // collections. 1239 // We want to start a new collection cycle if any of the following 1240 // conditions hold: 1241 // . our current occupancy exceeds the configured initiating occupancy 1242 // for this generation, or 1243 // . we recently needed to expand this space and have not, since that 1244 // expansion, done a collection of this generation, or 1245 // . the underlying space believes that it may be a good idea to initiate 1246 // a concurrent collection (this may be based on criteria such as the 1247 // following: the space uses linear allocation and linear allocation is 1248 // going to fail, or there is believed to be excessive fragmentation in 1249 // the generation, etc... or ... 1250 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for 1251 // the case of the old generation; see CR 6543076): 1252 // we may be approaching a point at which allocation requests may fail because 1253 // we will be out of sufficient free space given allocation rate estimates.] 1254 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { 1255 1256 assert_lock_strong(freelistLock()); 1257 if (occupancy() > initiating_occupancy()) { 1258 log_trace(gc)(" %s: collect because of occupancy %f / %f ", 1259 short_name(), occupancy(), initiating_occupancy()); 1260 return true; 1261 } 1262 if (UseCMSInitiatingOccupancyOnly) { 1263 return false; 1264 } 1265 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { 1266 log_trace(gc)(" %s: collect because expanded for allocation ", short_name()); 1267 return true; 1268 } 1269 return false; 1270 } 1271 1272 void ConcurrentMarkSweepGeneration::collect(bool full, 1273 bool clear_all_soft_refs, 1274 size_t size, 1275 bool tlab) 1276 { 1277 collector()->collect(full, clear_all_soft_refs, size, tlab); 1278 } 1279 1280 void CMSCollector::collect(bool full, 1281 bool clear_all_soft_refs, 1282 size_t size, 1283 bool tlab) 1284 { 1285 // The following "if" branch is present for defensive reasons. 1286 // In the current uses of this interface, it can be replaced with: 1287 // assert(!GCLocker.is_active(), "Can't be called otherwise"); 1288 // But I am not placing that assert here to allow future 1289 // generality in invoking this interface. 1290 if (GCLocker::is_active()) { 1291 // A consistency test for GCLocker 1292 assert(GCLocker::needs_gc(), "Should have been set already"); 1293 // Skip this foreground collection, instead 1294 // expanding the heap if necessary. 1295 // Need the free list locks for the call to free() in compute_new_size() 1296 compute_new_size(); 1297 return; 1298 } 1299 acquire_control_and_collect(full, clear_all_soft_refs); 1300 } 1301 1302 void CMSCollector::request_full_gc(unsigned int full_gc_count, GCCause::Cause cause) { 1303 CMSHeap* heap = CMSHeap::heap(); 1304 unsigned int gc_count = heap->total_full_collections(); 1305 if (gc_count == full_gc_count) { 1306 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); 1307 _full_gc_requested = true; 1308 _full_gc_cause = cause; 1309 CGC_lock->notify(); // nudge CMS thread 1310 } else { 1311 assert(gc_count > full_gc_count, "Error: causal loop"); 1312 } 1313 } 1314 1315 bool CMSCollector::is_external_interruption() { 1316 GCCause::Cause cause = CMSHeap::heap()->gc_cause(); 1317 return GCCause::is_user_requested_gc(cause) || 1318 GCCause::is_serviceability_requested_gc(cause); 1319 } 1320 1321 void CMSCollector::report_concurrent_mode_interruption() { 1322 if (is_external_interruption()) { 1323 log_debug(gc)("Concurrent mode interrupted"); 1324 } else { 1325 log_debug(gc)("Concurrent mode failure"); 1326 _gc_tracer_cm->report_concurrent_mode_failure(); 1327 } 1328 } 1329 1330 1331 // The foreground and background collectors need to coordinate in order 1332 // to make sure that they do not mutually interfere with CMS collections. 1333 // When a background collection is active, 1334 // the foreground collector may need to take over (preempt) and 1335 // synchronously complete an ongoing collection. Depending on the 1336 // frequency of the background collections and the heap usage 1337 // of the application, this preemption can be seldom or frequent. 1338 // There are only certain 1339 // points in the background collection that the "collection-baton" 1340 // can be passed to the foreground collector. 1341 // 1342 // The foreground collector will wait for the baton before 1343 // starting any part of the collection. The foreground collector 1344 // will only wait at one location. 1345 // 1346 // The background collector will yield the baton before starting a new 1347 // phase of the collection (e.g., before initial marking, marking from roots, 1348 // precleaning, final re-mark, sweep etc.) This is normally done at the head 1349 // of the loop which switches the phases. The background collector does some 1350 // of the phases (initial mark, final re-mark) with the world stopped. 1351 // Because of locking involved in stopping the world, 1352 // the foreground collector should not block waiting for the background 1353 // collector when it is doing a stop-the-world phase. The background 1354 // collector will yield the baton at an additional point just before 1355 // it enters a stop-the-world phase. Once the world is stopped, the 1356 // background collector checks the phase of the collection. If the 1357 // phase has not changed, it proceeds with the collection. If the 1358 // phase has changed, it skips that phase of the collection. See 1359 // the comments on the use of the Heap_lock in collect_in_background(). 1360 // 1361 // Variable used in baton passing. 1362 // _foregroundGCIsActive - Set to true by the foreground collector when 1363 // it wants the baton. The foreground clears it when it has finished 1364 // the collection. 1365 // _foregroundGCShouldWait - Set to true by the background collector 1366 // when it is running. The foreground collector waits while 1367 // _foregroundGCShouldWait is true. 1368 // CGC_lock - monitor used to protect access to the above variables 1369 // and to notify the foreground and background collectors. 1370 // _collectorState - current state of the CMS collection. 1371 // 1372 // The foreground collector 1373 // acquires the CGC_lock 1374 // sets _foregroundGCIsActive 1375 // waits on the CGC_lock for _foregroundGCShouldWait to be false 1376 // various locks acquired in preparation for the collection 1377 // are released so as not to block the background collector 1378 // that is in the midst of a collection 1379 // proceeds with the collection 1380 // clears _foregroundGCIsActive 1381 // returns 1382 // 1383 // The background collector in a loop iterating on the phases of the 1384 // collection 1385 // acquires the CGC_lock 1386 // sets _foregroundGCShouldWait 1387 // if _foregroundGCIsActive is set 1388 // clears _foregroundGCShouldWait, notifies _CGC_lock 1389 // waits on _CGC_lock for _foregroundGCIsActive to become false 1390 // and exits the loop. 1391 // otherwise 1392 // proceed with that phase of the collection 1393 // if the phase is a stop-the-world phase, 1394 // yield the baton once more just before enqueueing 1395 // the stop-world CMS operation (executed by the VM thread). 1396 // returns after all phases of the collection are done 1397 // 1398 1399 void CMSCollector::acquire_control_and_collect(bool full, 1400 bool clear_all_soft_refs) { 1401 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); 1402 assert(!Thread::current()->is_ConcurrentGC_thread(), 1403 "shouldn't try to acquire control from self!"); 1404 1405 // Start the protocol for acquiring control of the 1406 // collection from the background collector (aka CMS thread). 1407 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1408 "VM thread should have CMS token"); 1409 // Remember the possibly interrupted state of an ongoing 1410 // concurrent collection 1411 CollectorState first_state = _collectorState; 1412 1413 // Signal to a possibly ongoing concurrent collection that 1414 // we want to do a foreground collection. 1415 _foregroundGCIsActive = true; 1416 1417 // release locks and wait for a notify from the background collector 1418 // releasing the locks in only necessary for phases which 1419 // do yields to improve the granularity of the collection. 1420 assert_lock_strong(bitMapLock()); 1421 // We need to lock the Free list lock for the space that we are 1422 // currently collecting. 1423 assert(haveFreelistLocks(), "Must be holding free list locks"); 1424 bitMapLock()->unlock(); 1425 releaseFreelistLocks(); 1426 { 1427 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1428 if (_foregroundGCShouldWait) { 1429 // We are going to be waiting for action for the CMS thread; 1430 // it had better not be gone (for instance at shutdown)! 1431 assert(ConcurrentMarkSweepThread::cmst() != NULL && !ConcurrentMarkSweepThread::cmst()->has_terminated(), 1432 "CMS thread must be running"); 1433 // Wait here until the background collector gives us the go-ahead 1434 ConcurrentMarkSweepThread::clear_CMS_flag( 1435 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token 1436 // Get a possibly blocked CMS thread going: 1437 // Note that we set _foregroundGCIsActive true above, 1438 // without protection of the CGC_lock. 1439 CGC_lock->notify(); 1440 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), 1441 "Possible deadlock"); 1442 while (_foregroundGCShouldWait) { 1443 // wait for notification 1444 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1445 // Possibility of delay/starvation here, since CMS token does 1446 // not know to give priority to VM thread? Actually, i think 1447 // there wouldn't be any delay/starvation, but the proof of 1448 // that "fact" (?) appears non-trivial. XXX 20011219YSR 1449 } 1450 ConcurrentMarkSweepThread::set_CMS_flag( 1451 ConcurrentMarkSweepThread::CMS_vm_has_token); 1452 } 1453 } 1454 // The CMS_token is already held. Get back the other locks. 1455 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), 1456 "VM thread should have CMS token"); 1457 getFreelistLocks(); 1458 bitMapLock()->lock_without_safepoint_check(); 1459 log_debug(gc, state)("CMS foreground collector has asked for control " INTPTR_FORMAT " with first state %d", 1460 p2i(Thread::current()), first_state); 1461 log_debug(gc, state)(" gets control with state %d", _collectorState); 1462 1463 // Inform cms gen if this was due to partial collection failing. 1464 // The CMS gen may use this fact to determine its expansion policy. 1465 CMSHeap* heap = CMSHeap::heap(); 1466 if (heap->incremental_collection_will_fail(false /* don't consult_young */)) { 1467 assert(!_cmsGen->incremental_collection_failed(), 1468 "Should have been noticed, reacted to and cleared"); 1469 _cmsGen->set_incremental_collection_failed(); 1470 } 1471 1472 if (first_state > Idling) { 1473 report_concurrent_mode_interruption(); 1474 } 1475 1476 set_did_compact(true); 1477 1478 // If the collection is being acquired from the background 1479 // collector, there may be references on the discovered 1480 // references lists. Abandon those references, since some 1481 // of them may have become unreachable after concurrent 1482 // discovery; the STW compacting collector will redo discovery 1483 // more precisely, without being subject to floating garbage. 1484 // Leaving otherwise unreachable references in the discovered 1485 // lists would require special handling. 1486 ref_processor()->disable_discovery(); 1487 ref_processor()->abandon_partial_discovery(); 1488 ref_processor()->verify_no_references_recorded(); 1489 1490 if (first_state > Idling) { 1491 save_heap_summary(); 1492 } 1493 1494 do_compaction_work(clear_all_soft_refs); 1495 1496 // Has the GC time limit been exceeded? 1497 size_t max_eden_size = _young_gen->max_eden_size(); 1498 GCCause::Cause gc_cause = heap->gc_cause(); 1499 size_policy()->check_gc_overhead_limit(_young_gen->eden()->used(), 1500 _cmsGen->max_capacity(), 1501 max_eden_size, 1502 full, 1503 gc_cause, 1504 heap->soft_ref_policy()); 1505 1506 // Reset the expansion cause, now that we just completed 1507 // a collection cycle. 1508 clear_expansion_cause(); 1509 _foregroundGCIsActive = false; 1510 return; 1511 } 1512 1513 // Resize the tenured generation 1514 // after obtaining the free list locks for the 1515 // two generations. 1516 void CMSCollector::compute_new_size() { 1517 assert_locked_or_safepoint(Heap_lock); 1518 FreelistLocker z(this); 1519 MetaspaceGC::compute_new_size(); 1520 _cmsGen->compute_new_size_free_list(); 1521 } 1522 1523 // A work method used by the foreground collector to do 1524 // a mark-sweep-compact. 1525 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { 1526 CMSHeap* heap = CMSHeap::heap(); 1527 1528 STWGCTimer* gc_timer = GenMarkSweep::gc_timer(); 1529 gc_timer->register_gc_start(); 1530 1531 SerialOldTracer* gc_tracer = GenMarkSweep::gc_tracer(); 1532 gc_tracer->report_gc_start(heap->gc_cause(), gc_timer->gc_start()); 1533 1534 heap->pre_full_gc_dump(gc_timer); 1535 1536 GCTraceTime(Trace, gc, phases) t("CMS:MSC"); 1537 1538 // Temporarily widen the span of the weak reference processing to 1539 // the entire heap. 1540 MemRegion new_span(CMSHeap::heap()->reserved_region()); 1541 ReferenceProcessorSpanMutator rp_mut_span(ref_processor(), new_span); 1542 // Temporarily, clear the "is_alive_non_header" field of the 1543 // reference processor. 1544 ReferenceProcessorIsAliveMutator rp_mut_closure(ref_processor(), NULL); 1545 // Temporarily make reference _processing_ single threaded (non-MT). 1546 ReferenceProcessorMTProcMutator rp_mut_mt_processing(ref_processor(), false); 1547 // Temporarily make refs discovery atomic 1548 ReferenceProcessorAtomicMutator rp_mut_atomic(ref_processor(), true); 1549 // Temporarily make reference _discovery_ single threaded (non-MT) 1550 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 1551 1552 ref_processor()->set_enqueuing_is_done(false); 1553 ref_processor()->enable_discovery(); 1554 ref_processor()->setup_policy(clear_all_soft_refs); 1555 // If an asynchronous collection finishes, the _modUnionTable is 1556 // all clear. If we are assuming the collection from an asynchronous 1557 // collection, clear the _modUnionTable. 1558 assert(_collectorState != Idling || _modUnionTable.isAllClear(), 1559 "_modUnionTable should be clear if the baton was not passed"); 1560 _modUnionTable.clear_all(); 1561 assert(_collectorState != Idling || _ct->cld_rem_set()->mod_union_is_clear(), 1562 "mod union for klasses should be clear if the baton was passed"); 1563 _ct->cld_rem_set()->clear_mod_union(); 1564 1565 1566 // We must adjust the allocation statistics being maintained 1567 // in the free list space. We do so by reading and clearing 1568 // the sweep timer and updating the block flux rate estimates below. 1569 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); 1570 if (_inter_sweep_timer.is_active()) { 1571 _inter_sweep_timer.stop(); 1572 // Note that we do not use this sample to update the _inter_sweep_estimate. 1573 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 1574 _inter_sweep_estimate.padded_average(), 1575 _intra_sweep_estimate.padded_average()); 1576 } 1577 1578 GenMarkSweep::invoke_at_safepoint(ref_processor(), clear_all_soft_refs); 1579 #ifdef ASSERT 1580 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 1581 size_t free_size = cms_space->free(); 1582 assert(free_size == 1583 pointer_delta(cms_space->end(), cms_space->compaction_top()) 1584 * HeapWordSize, 1585 "All the free space should be compacted into one chunk at top"); 1586 assert(cms_space->dictionary()->total_chunk_size( 1587 debug_only(cms_space->freelistLock())) == 0 || 1588 cms_space->totalSizeInIndexedFreeLists() == 0, 1589 "All the free space should be in a single chunk"); 1590 size_t num = cms_space->totalCount(); 1591 assert((free_size == 0 && num == 0) || 1592 (free_size > 0 && (num == 1 || num == 2)), 1593 "There should be at most 2 free chunks after compaction"); 1594 #endif // ASSERT 1595 _collectorState = Resetting; 1596 assert(_restart_addr == NULL, 1597 "Should have been NULL'd before baton was passed"); 1598 reset_stw(); 1599 _cmsGen->reset_after_compaction(); 1600 _concurrent_cycles_since_last_unload = 0; 1601 1602 // Clear any data recorded in the PLAB chunk arrays. 1603 if (_survivor_plab_array != NULL) { 1604 reset_survivor_plab_arrays(); 1605 } 1606 1607 // Adjust the per-size allocation stats for the next epoch. 1608 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); 1609 // Restart the "inter sweep timer" for the next epoch. 1610 _inter_sweep_timer.reset(); 1611 _inter_sweep_timer.start(); 1612 1613 // No longer a need to do a concurrent collection for Metaspace. 1614 MetaspaceGC::set_should_concurrent_collect(false); 1615 1616 heap->post_full_gc_dump(gc_timer); 1617 1618 gc_timer->register_gc_end(); 1619 1620 gc_tracer->report_gc_end(gc_timer->gc_end(), gc_timer->time_partitions()); 1621 1622 // For a mark-sweep-compact, compute_new_size() will be called 1623 // in the heap's do_collection() method. 1624 } 1625 1626 void CMSCollector::print_eden_and_survivor_chunk_arrays() { 1627 Log(gc, heap) log; 1628 if (!log.is_trace()) { 1629 return; 1630 } 1631 1632 ContiguousSpace* eden_space = _young_gen->eden(); 1633 ContiguousSpace* from_space = _young_gen->from(); 1634 ContiguousSpace* to_space = _young_gen->to(); 1635 // Eden 1636 if (_eden_chunk_array != NULL) { 1637 log.trace("eden " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1638 p2i(eden_space->bottom()), p2i(eden_space->top()), 1639 p2i(eden_space->end()), eden_space->capacity()); 1640 log.trace("_eden_chunk_index=" SIZE_FORMAT ", _eden_chunk_capacity=" SIZE_FORMAT, 1641 _eden_chunk_index, _eden_chunk_capacity); 1642 for (size_t i = 0; i < _eden_chunk_index; i++) { 1643 log.trace("_eden_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_eden_chunk_array[i])); 1644 } 1645 } 1646 // Survivor 1647 if (_survivor_chunk_array != NULL) { 1648 log.trace("survivor " PTR_FORMAT "-" PTR_FORMAT "-" PTR_FORMAT "(" SIZE_FORMAT ")", 1649 p2i(from_space->bottom()), p2i(from_space->top()), 1650 p2i(from_space->end()), from_space->capacity()); 1651 log.trace("_survivor_chunk_index=" SIZE_FORMAT ", _survivor_chunk_capacity=" SIZE_FORMAT, 1652 _survivor_chunk_index, _survivor_chunk_capacity); 1653 for (size_t i = 0; i < _survivor_chunk_index; i++) { 1654 log.trace("_survivor_chunk_array[" SIZE_FORMAT "]=" PTR_FORMAT, i, p2i(_survivor_chunk_array[i])); 1655 } 1656 } 1657 } 1658 1659 void CMSCollector::getFreelistLocks() const { 1660 // Get locks for all free lists in all generations that this 1661 // collector is responsible for 1662 _cmsGen->freelistLock()->lock_without_safepoint_check(); 1663 } 1664 1665 void CMSCollector::releaseFreelistLocks() const { 1666 // Release locks for all free lists in all generations that this 1667 // collector is responsible for 1668 _cmsGen->freelistLock()->unlock(); 1669 } 1670 1671 bool CMSCollector::haveFreelistLocks() const { 1672 // Check locks for all free lists in all generations that this 1673 // collector is responsible for 1674 assert_lock_strong(_cmsGen->freelistLock()); 1675 PRODUCT_ONLY(ShouldNotReachHere()); 1676 return true; 1677 } 1678 1679 // A utility class that is used by the CMS collector to 1680 // temporarily "release" the foreground collector from its 1681 // usual obligation to wait for the background collector to 1682 // complete an ongoing phase before proceeding. 1683 class ReleaseForegroundGC: public StackObj { 1684 private: 1685 CMSCollector* _c; 1686 public: 1687 ReleaseForegroundGC(CMSCollector* c) : _c(c) { 1688 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); 1689 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1690 // allow a potentially blocked foreground collector to proceed 1691 _c->_foregroundGCShouldWait = false; 1692 if (_c->_foregroundGCIsActive) { 1693 CGC_lock->notify(); 1694 } 1695 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1696 "Possible deadlock"); 1697 } 1698 1699 ~ReleaseForegroundGC() { 1700 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); 1701 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1702 _c->_foregroundGCShouldWait = true; 1703 } 1704 }; 1705 1706 void CMSCollector::collect_in_background(GCCause::Cause cause) { 1707 assert(Thread::current()->is_ConcurrentGC_thread(), 1708 "A CMS asynchronous collection is only allowed on a CMS thread."); 1709 1710 CMSHeap* heap = CMSHeap::heap(); 1711 { 1712 bool safepoint_check = Mutex::_no_safepoint_check_flag; 1713 MutexLockerEx hl(Heap_lock, safepoint_check); 1714 FreelistLocker fll(this); 1715 MutexLockerEx x(CGC_lock, safepoint_check); 1716 if (_foregroundGCIsActive) { 1717 // The foreground collector is. Skip this 1718 // background collection. 1719 assert(!_foregroundGCShouldWait, "Should be clear"); 1720 return; 1721 } else { 1722 assert(_collectorState == Idling, "Should be idling before start."); 1723 _collectorState = InitialMarking; 1724 register_gc_start(cause); 1725 // Reset the expansion cause, now that we are about to begin 1726 // a new cycle. 1727 clear_expansion_cause(); 1728 1729 // Clear the MetaspaceGC flag since a concurrent collection 1730 // is starting but also clear it after the collection. 1731 MetaspaceGC::set_should_concurrent_collect(false); 1732 } 1733 // Decide if we want to enable class unloading as part of the 1734 // ensuing concurrent GC cycle. 1735 update_should_unload_classes(); 1736 _full_gc_requested = false; // acks all outstanding full gc requests 1737 _full_gc_cause = GCCause::_no_gc; 1738 // Signal that we are about to start a collection 1739 heap->increment_total_full_collections(); // ... starting a collection cycle 1740 _collection_count_start = heap->total_full_collections(); 1741 } 1742 1743 size_t prev_used = _cmsGen->used(); 1744 1745 // The change of the collection state is normally done at this level; 1746 // the exceptions are phases that are executed while the world is 1747 // stopped. For those phases the change of state is done while the 1748 // world is stopped. For baton passing purposes this allows the 1749 // background collector to finish the phase and change state atomically. 1750 // The foreground collector cannot wait on a phase that is done 1751 // while the world is stopped because the foreground collector already 1752 // has the world stopped and would deadlock. 1753 while (_collectorState != Idling) { 1754 log_debug(gc, state)("Thread " INTPTR_FORMAT " in CMS state %d", 1755 p2i(Thread::current()), _collectorState); 1756 // The foreground collector 1757 // holds the Heap_lock throughout its collection. 1758 // holds the CMS token (but not the lock) 1759 // except while it is waiting for the background collector to yield. 1760 // 1761 // The foreground collector should be blocked (not for long) 1762 // if the background collector is about to start a phase 1763 // executed with world stopped. If the background 1764 // collector has already started such a phase, the 1765 // foreground collector is blocked waiting for the 1766 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) 1767 // are executed in the VM thread. 1768 // 1769 // The locking order is 1770 // PendingListLock (PLL) -- if applicable (FinalMarking) 1771 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) 1772 // CMS token (claimed in 1773 // stop_world_and_do() --> 1774 // safepoint_synchronize() --> 1775 // CMSThread::synchronize()) 1776 1777 { 1778 // Check if the FG collector wants us to yield. 1779 CMSTokenSync x(true); // is cms thread 1780 if (waitForForegroundGC()) { 1781 // We yielded to a foreground GC, nothing more to be 1782 // done this round. 1783 assert(_foregroundGCShouldWait == false, "We set it to false in " 1784 "waitForForegroundGC()"); 1785 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1786 p2i(Thread::current()), _collectorState); 1787 return; 1788 } else { 1789 // The background collector can run but check to see if the 1790 // foreground collector has done a collection while the 1791 // background collector was waiting to get the CGC_lock 1792 // above. If yes, break so that _foregroundGCShouldWait 1793 // is cleared before returning. 1794 if (_collectorState == Idling) { 1795 break; 1796 } 1797 } 1798 } 1799 1800 assert(_foregroundGCShouldWait, "Foreground collector, if active, " 1801 "should be waiting"); 1802 1803 switch (_collectorState) { 1804 case InitialMarking: 1805 { 1806 ReleaseForegroundGC x(this); 1807 stats().record_cms_begin(); 1808 VM_CMS_Initial_Mark initial_mark_op(this); 1809 VMThread::execute(&initial_mark_op); 1810 } 1811 // The collector state may be any legal state at this point 1812 // since the background collector may have yielded to the 1813 // foreground collector. 1814 break; 1815 case Marking: 1816 // initial marking in checkpointRootsInitialWork has been completed 1817 if (markFromRoots()) { // we were successful 1818 assert(_collectorState == Precleaning, "Collector state should " 1819 "have changed"); 1820 } else { 1821 assert(_foregroundGCIsActive, "Internal state inconsistency"); 1822 } 1823 break; 1824 case Precleaning: 1825 // marking from roots in markFromRoots has been completed 1826 preclean(); 1827 assert(_collectorState == AbortablePreclean || 1828 _collectorState == FinalMarking, 1829 "Collector state should have changed"); 1830 break; 1831 case AbortablePreclean: 1832 abortable_preclean(); 1833 assert(_collectorState == FinalMarking, "Collector state should " 1834 "have changed"); 1835 break; 1836 case FinalMarking: 1837 { 1838 ReleaseForegroundGC x(this); 1839 1840 VM_CMS_Final_Remark final_remark_op(this); 1841 VMThread::execute(&final_remark_op); 1842 } 1843 assert(_foregroundGCShouldWait, "block post-condition"); 1844 break; 1845 case Sweeping: 1846 // final marking in checkpointRootsFinal has been completed 1847 sweep(); 1848 assert(_collectorState == Resizing, "Collector state change " 1849 "to Resizing must be done under the free_list_lock"); 1850 1851 case Resizing: { 1852 // Sweeping has been completed... 1853 // At this point the background collection has completed. 1854 // Don't move the call to compute_new_size() down 1855 // into code that might be executed if the background 1856 // collection was preempted. 1857 { 1858 ReleaseForegroundGC x(this); // unblock FG collection 1859 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); 1860 CMSTokenSync z(true); // not strictly needed. 1861 if (_collectorState == Resizing) { 1862 compute_new_size(); 1863 save_heap_summary(); 1864 _collectorState = Resetting; 1865 } else { 1866 assert(_collectorState == Idling, "The state should only change" 1867 " because the foreground collector has finished the collection"); 1868 } 1869 } 1870 break; 1871 } 1872 case Resetting: 1873 // CMS heap resizing has been completed 1874 reset_concurrent(); 1875 assert(_collectorState == Idling, "Collector state should " 1876 "have changed"); 1877 1878 MetaspaceGC::set_should_concurrent_collect(false); 1879 1880 stats().record_cms_end(); 1881 // Don't move the concurrent_phases_end() and compute_new_size() 1882 // calls to here because a preempted background collection 1883 // has it's state set to "Resetting". 1884 break; 1885 case Idling: 1886 default: 1887 ShouldNotReachHere(); 1888 break; 1889 } 1890 log_debug(gc, state)(" Thread " INTPTR_FORMAT " done - next CMS state %d", 1891 p2i(Thread::current()), _collectorState); 1892 assert(_foregroundGCShouldWait, "block post-condition"); 1893 } 1894 1895 // Should this be in gc_epilogue? 1896 heap->counters()->update_counters(); 1897 1898 { 1899 // Clear _foregroundGCShouldWait and, in the event that the 1900 // foreground collector is waiting, notify it, before 1901 // returning. 1902 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1903 _foregroundGCShouldWait = false; 1904 if (_foregroundGCIsActive) { 1905 CGC_lock->notify(); 1906 } 1907 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1908 "Possible deadlock"); 1909 } 1910 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " exiting collection CMS state %d", 1911 p2i(Thread::current()), _collectorState); 1912 log_info(gc, heap)("Old: " SIZE_FORMAT "K->" SIZE_FORMAT "K(" SIZE_FORMAT "K)", 1913 prev_used / K, _cmsGen->used()/K, _cmsGen->capacity() /K); 1914 } 1915 1916 void CMSCollector::register_gc_start(GCCause::Cause cause) { 1917 _cms_start_registered = true; 1918 _gc_timer_cm->register_gc_start(); 1919 _gc_tracer_cm->report_gc_start(cause, _gc_timer_cm->gc_start()); 1920 } 1921 1922 void CMSCollector::register_gc_end() { 1923 if (_cms_start_registered) { 1924 report_heap_summary(GCWhen::AfterGC); 1925 1926 _gc_timer_cm->register_gc_end(); 1927 _gc_tracer_cm->report_gc_end(_gc_timer_cm->gc_end(), _gc_timer_cm->time_partitions()); 1928 _cms_start_registered = false; 1929 } 1930 } 1931 1932 void CMSCollector::save_heap_summary() { 1933 CMSHeap* heap = CMSHeap::heap(); 1934 _last_heap_summary = heap->create_heap_summary(); 1935 _last_metaspace_summary = heap->create_metaspace_summary(); 1936 } 1937 1938 void CMSCollector::report_heap_summary(GCWhen::Type when) { 1939 _gc_tracer_cm->report_gc_heap_summary(when, _last_heap_summary); 1940 _gc_tracer_cm->report_metaspace_summary(when, _last_metaspace_summary); 1941 } 1942 1943 bool CMSCollector::waitForForegroundGC() { 1944 bool res = false; 1945 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 1946 "CMS thread should have CMS token"); 1947 // Block the foreground collector until the 1948 // background collectors decides whether to 1949 // yield. 1950 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 1951 _foregroundGCShouldWait = true; 1952 if (_foregroundGCIsActive) { 1953 // The background collector yields to the 1954 // foreground collector and returns a value 1955 // indicating that it has yielded. The foreground 1956 // collector can proceed. 1957 res = true; 1958 _foregroundGCShouldWait = false; 1959 ConcurrentMarkSweepThread::clear_CMS_flag( 1960 ConcurrentMarkSweepThread::CMS_cms_has_token); 1961 ConcurrentMarkSweepThread::set_CMS_flag( 1962 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1963 // Get a possibly blocked foreground thread going 1964 CGC_lock->notify(); 1965 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", 1966 p2i(Thread::current()), _collectorState); 1967 while (_foregroundGCIsActive) { 1968 CGC_lock->wait(Mutex::_no_safepoint_check_flag); 1969 } 1970 ConcurrentMarkSweepThread::set_CMS_flag( 1971 ConcurrentMarkSweepThread::CMS_cms_has_token); 1972 ConcurrentMarkSweepThread::clear_CMS_flag( 1973 ConcurrentMarkSweepThread::CMS_cms_wants_token); 1974 } 1975 log_debug(gc, state)("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", 1976 p2i(Thread::current()), _collectorState); 1977 return res; 1978 } 1979 1980 // Because of the need to lock the free lists and other structures in 1981 // the collector, common to all the generations that the collector is 1982 // collecting, we need the gc_prologues of individual CMS generations 1983 // delegate to their collector. It may have been simpler had the 1984 // current infrastructure allowed one to call a prologue on a 1985 // collector. In the absence of that we have the generation's 1986 // prologue delegate to the collector, which delegates back 1987 // some "local" work to a worker method in the individual generations 1988 // that it's responsible for collecting, while itself doing any 1989 // work common to all generations it's responsible for. A similar 1990 // comment applies to the gc_epilogue()'s. 1991 // The role of the variable _between_prologue_and_epilogue is to 1992 // enforce the invocation protocol. 1993 void CMSCollector::gc_prologue(bool full) { 1994 // Call gc_prologue_work() for the CMSGen 1995 // we are responsible for. 1996 1997 // The following locking discipline assumes that we are only called 1998 // when the world is stopped. 1999 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); 2000 2001 // The CMSCollector prologue must call the gc_prologues for the 2002 // "generations" that it's responsible 2003 // for. 2004 2005 assert( Thread::current()->is_VM_thread() 2006 || ( CMSScavengeBeforeRemark 2007 && Thread::current()->is_ConcurrentGC_thread()), 2008 "Incorrect thread type for prologue execution"); 2009 2010 if (_between_prologue_and_epilogue) { 2011 // We have already been invoked; this is a gc_prologue delegation 2012 // from yet another CMS generation that we are responsible for, just 2013 // ignore it since all relevant work has already been done. 2014 return; 2015 } 2016 2017 // set a bit saying prologue has been called; cleared in epilogue 2018 _between_prologue_and_epilogue = true; 2019 // Claim locks for common data structures, then call gc_prologue_work() 2020 // for each CMSGen. 2021 2022 getFreelistLocks(); // gets free list locks on constituent spaces 2023 bitMapLock()->lock_without_safepoint_check(); 2024 2025 // Should call gc_prologue_work() for all cms gens we are responsible for 2026 bool duringMarking = _collectorState >= Marking 2027 && _collectorState < Sweeping; 2028 2029 // The young collections clear the modified oops state, which tells if 2030 // there are any modified oops in the class. The remark phase also needs 2031 // that information. Tell the young collection to save the union of all 2032 // modified klasses. 2033 if (duringMarking) { 2034 _ct->cld_rem_set()->set_accumulate_modified_oops(true); 2035 } 2036 2037 bool registerClosure = duringMarking; 2038 2039 _cmsGen->gc_prologue_work(full, registerClosure, &_modUnionClosurePar); 2040 2041 if (!full) { 2042 stats().record_gc0_begin(); 2043 } 2044 } 2045 2046 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { 2047 2048 _capacity_at_prologue = capacity(); 2049 _used_at_prologue = used(); 2050 2051 // We enable promotion tracking so that card-scanning can recognize 2052 // which objects have been promoted during this GC and skip them. 2053 for (uint i = 0; i < ParallelGCThreads; i++) { 2054 _par_gc_thread_states[i]->promo.startTrackingPromotions(); 2055 } 2056 2057 // Delegate to CMScollector which knows how to coordinate between 2058 // this and any other CMS generations that it is responsible for 2059 // collecting. 2060 collector()->gc_prologue(full); 2061 } 2062 2063 // This is a "private" interface for use by this generation's CMSCollector. 2064 // Not to be called directly by any other entity (for instance, 2065 // GenCollectedHeap, which calls the "public" gc_prologue method above). 2066 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, 2067 bool registerClosure, ModUnionClosure* modUnionClosure) { 2068 assert(!incremental_collection_failed(), "Shouldn't be set yet"); 2069 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, 2070 "Should be NULL"); 2071 if (registerClosure) { 2072 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); 2073 } 2074 cmsSpace()->gc_prologue(); 2075 // Clear stat counters 2076 NOT_PRODUCT( 2077 assert(_numObjectsPromoted == 0, "check"); 2078 assert(_numWordsPromoted == 0, "check"); 2079 log_develop_trace(gc, alloc)("Allocated " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes concurrently", 2080 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); 2081 _numObjectsAllocated = 0; 2082 _numWordsAllocated = 0; 2083 ) 2084 } 2085 2086 void CMSCollector::gc_epilogue(bool full) { 2087 // The following locking discipline assumes that we are only called 2088 // when the world is stopped. 2089 assert(SafepointSynchronize::is_at_safepoint(), 2090 "world is stopped assumption"); 2091 2092 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks 2093 // if linear allocation blocks need to be appropriately marked to allow the 2094 // the blocks to be parsable. We also check here whether we need to nudge the 2095 // CMS collector thread to start a new cycle (if it's not already active). 2096 assert( Thread::current()->is_VM_thread() 2097 || ( CMSScavengeBeforeRemark 2098 && Thread::current()->is_ConcurrentGC_thread()), 2099 "Incorrect thread type for epilogue execution"); 2100 2101 if (!_between_prologue_and_epilogue) { 2102 // We have already been invoked; this is a gc_epilogue delegation 2103 // from yet another CMS generation that we are responsible for, just 2104 // ignore it since all relevant work has already been done. 2105 return; 2106 } 2107 assert(haveFreelistLocks(), "must have freelist locks"); 2108 assert_lock_strong(bitMapLock()); 2109 2110 _ct->cld_rem_set()->set_accumulate_modified_oops(false); 2111 2112 _cmsGen->gc_epilogue_work(full); 2113 2114 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { 2115 // in case sampling was not already enabled, enable it 2116 _start_sampling = true; 2117 } 2118 // reset _eden_chunk_array so sampling starts afresh 2119 _eden_chunk_index = 0; 2120 2121 size_t cms_used = _cmsGen->cmsSpace()->used(); 2122 2123 // update performance counters - this uses a special version of 2124 // update_counters() that allows the utilization to be passed as a 2125 // parameter, avoiding multiple calls to used(). 2126 // 2127 _cmsGen->update_counters(cms_used); 2128 2129 bitMapLock()->unlock(); 2130 releaseFreelistLocks(); 2131 2132 if (!CleanChunkPoolAsync) { 2133 Chunk::clean_chunk_pool(); 2134 } 2135 2136 set_did_compact(false); 2137 _between_prologue_and_epilogue = false; // ready for next cycle 2138 } 2139 2140 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { 2141 collector()->gc_epilogue(full); 2142 2143 // When using ParNew, promotion tracking should have already been 2144 // disabled. However, the prologue (which enables promotion 2145 // tracking) and epilogue are called irrespective of the type of 2146 // GC. So they will also be called before and after Full GCs, during 2147 // which promotion tracking will not be explicitly disabled. So, 2148 // it's safer to also disable it here too (to be symmetric with 2149 // enabling it in the prologue). 2150 for (uint i = 0; i < ParallelGCThreads; i++) { 2151 _par_gc_thread_states[i]->promo.stopTrackingPromotions(); 2152 } 2153 } 2154 2155 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { 2156 assert(!incremental_collection_failed(), "Should have been cleared"); 2157 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); 2158 cmsSpace()->gc_epilogue(); 2159 // Print stat counters 2160 NOT_PRODUCT( 2161 assert(_numObjectsAllocated == 0, "check"); 2162 assert(_numWordsAllocated == 0, "check"); 2163 log_develop_trace(gc, promotion)("Promoted " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 2164 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); 2165 _numObjectsPromoted = 0; 2166 _numWordsPromoted = 0; 2167 ) 2168 2169 // Call down the chain in contiguous_available needs the freelistLock 2170 // so print this out before releasing the freeListLock. 2171 log_develop_trace(gc)(" Contiguous available " SIZE_FORMAT " bytes ", contiguous_available()); 2172 } 2173 2174 #ifndef PRODUCT 2175 bool CMSCollector::have_cms_token() { 2176 Thread* thr = Thread::current(); 2177 if (thr->is_VM_thread()) { 2178 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); 2179 } else if (thr->is_ConcurrentGC_thread()) { 2180 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); 2181 } else if (thr->is_GC_task_thread()) { 2182 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && 2183 ParGCRareEvent_lock->owned_by_self(); 2184 } 2185 return false; 2186 } 2187 2188 // Check reachability of the given heap address in CMS generation, 2189 // treating all other generations as roots. 2190 bool CMSCollector::is_cms_reachable(HeapWord* addr) { 2191 // We could "guarantee" below, rather than assert, but I'll 2192 // leave these as "asserts" so that an adventurous debugger 2193 // could try this in the product build provided some subset of 2194 // the conditions were met, provided they were interested in the 2195 // results and knew that the computation below wouldn't interfere 2196 // with other concurrent computations mutating the structures 2197 // being read or written. 2198 assert(SafepointSynchronize::is_at_safepoint(), 2199 "Else mutations in object graph will make answer suspect"); 2200 assert(have_cms_token(), "Should hold cms token"); 2201 assert(haveFreelistLocks(), "must hold free list locks"); 2202 assert_lock_strong(bitMapLock()); 2203 2204 // Clear the marking bit map array before starting, but, just 2205 // for kicks, first report if the given address is already marked 2206 tty->print_cr("Start: Address " PTR_FORMAT " is%s marked", p2i(addr), 2207 _markBitMap.isMarked(addr) ? "" : " not"); 2208 2209 if (verify_after_remark()) { 2210 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2211 bool result = verification_mark_bm()->isMarked(addr); 2212 tty->print_cr("TransitiveMark: Address " PTR_FORMAT " %s marked", p2i(addr), 2213 result ? "IS" : "is NOT"); 2214 return result; 2215 } else { 2216 tty->print_cr("Could not compute result"); 2217 return false; 2218 } 2219 } 2220 #endif 2221 2222 void 2223 CMSCollector::print_on_error(outputStream* st) { 2224 CMSCollector* collector = ConcurrentMarkSweepGeneration::_collector; 2225 if (collector != NULL) { 2226 CMSBitMap* bitmap = &collector->_markBitMap; 2227 st->print_cr("Marking Bits: (CMSBitMap*) " PTR_FORMAT, p2i(bitmap)); 2228 bitmap->print_on_error(st, " Bits: "); 2229 2230 st->cr(); 2231 2232 CMSBitMap* mut_bitmap = &collector->_modUnionTable; 2233 st->print_cr("Mod Union Table: (CMSBitMap*) " PTR_FORMAT, p2i(mut_bitmap)); 2234 mut_bitmap->print_on_error(st, " Bits: "); 2235 } 2236 } 2237 2238 //////////////////////////////////////////////////////// 2239 // CMS Verification Support 2240 //////////////////////////////////////////////////////// 2241 // Following the remark phase, the following invariant 2242 // should hold -- each object in the CMS heap which is 2243 // marked in markBitMap() should be marked in the verification_mark_bm(). 2244 2245 class VerifyMarkedClosure: public BitMapClosure { 2246 CMSBitMap* _marks; 2247 bool _failed; 2248 2249 public: 2250 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} 2251 2252 bool do_bit(size_t offset) { 2253 HeapWord* addr = _marks->offsetToHeapWord(offset); 2254 if (!_marks->isMarked(addr)) { 2255 Log(gc, verify) log; 2256 ResourceMark rm; 2257 LogStream ls(log.error()); 2258 oop(addr)->print_on(&ls); 2259 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 2260 _failed = true; 2261 } 2262 return true; 2263 } 2264 2265 bool failed() { return _failed; } 2266 }; 2267 2268 bool CMSCollector::verify_after_remark() { 2269 GCTraceTime(Info, gc, phases, verify) tm("Verifying CMS Marking."); 2270 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); 2271 static bool init = false; 2272 2273 assert(SafepointSynchronize::is_at_safepoint(), 2274 "Else mutations in object graph will make answer suspect"); 2275 assert(have_cms_token(), 2276 "Else there may be mutual interference in use of " 2277 " verification data structures"); 2278 assert(_collectorState > Marking && _collectorState <= Sweeping, 2279 "Else marking info checked here may be obsolete"); 2280 assert(haveFreelistLocks(), "must hold free list locks"); 2281 assert_lock_strong(bitMapLock()); 2282 2283 2284 // Allocate marking bit map if not already allocated 2285 if (!init) { // first time 2286 if (!verification_mark_bm()->allocate(_span)) { 2287 return false; 2288 } 2289 init = true; 2290 } 2291 2292 assert(verification_mark_stack()->isEmpty(), "Should be empty"); 2293 2294 // Turn off refs discovery -- so we will be tracing through refs. 2295 // This is as intended, because by this time 2296 // GC must already have cleared any refs that need to be cleared, 2297 // and traced those that need to be marked; moreover, 2298 // the marking done here is not going to interfere in any 2299 // way with the marking information used by GC. 2300 NoRefDiscovery no_discovery(ref_processor()); 2301 2302 #if COMPILER2_OR_JVMCI 2303 DerivedPointerTableDeactivate dpt_deact; 2304 #endif 2305 2306 // Clear any marks from a previous round 2307 verification_mark_bm()->clear_all(); 2308 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); 2309 verify_work_stacks_empty(); 2310 2311 CMSHeap* heap = CMSHeap::heap(); 2312 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2313 // Update the saved marks which may affect the root scans. 2314 heap->save_marks(); 2315 2316 if (CMSRemarkVerifyVariant == 1) { 2317 // In this first variant of verification, we complete 2318 // all marking, then check if the new marks-vector is 2319 // a subset of the CMS marks-vector. 2320 verify_after_remark_work_1(); 2321 } else { 2322 guarantee(CMSRemarkVerifyVariant == 2, "Range checking for CMSRemarkVerifyVariant should guarantee 1 or 2"); 2323 // In this second variant of verification, we flag an error 2324 // (i.e. an object reachable in the new marks-vector not reachable 2325 // in the CMS marks-vector) immediately, also indicating the 2326 // identify of an object (A) that references the unmarked object (B) -- 2327 // presumably, a mutation to A failed to be picked up by preclean/remark? 2328 verify_after_remark_work_2(); 2329 } 2330 2331 return true; 2332 } 2333 2334 void CMSCollector::verify_after_remark_work_1() { 2335 ResourceMark rm; 2336 HandleMark hm; 2337 CMSHeap* heap = CMSHeap::heap(); 2338 2339 // Get a clear set of claim bits for the roots processing to work with. 2340 ClassLoaderDataGraph::clear_claimed_marks(); 2341 2342 // Mark from roots one level into CMS 2343 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); 2344 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2345 2346 { 2347 StrongRootsScope srs(1); 2348 2349 heap->cms_process_roots(&srs, 2350 true, // young gen as roots 2351 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2352 should_unload_classes(), 2353 ¬Older, 2354 NULL); 2355 } 2356 2357 // Now mark from the roots 2358 MarkFromRootsClosure markFromRootsClosure(this, _span, 2359 verification_mark_bm(), verification_mark_stack(), 2360 false /* don't yield */, true /* verifying */); 2361 assert(_restart_addr == NULL, "Expected pre-condition"); 2362 verification_mark_bm()->iterate(&markFromRootsClosure); 2363 while (_restart_addr != NULL) { 2364 // Deal with stack overflow: by restarting at the indicated 2365 // address. 2366 HeapWord* ra = _restart_addr; 2367 markFromRootsClosure.reset(ra); 2368 _restart_addr = NULL; 2369 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2370 } 2371 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2372 verify_work_stacks_empty(); 2373 2374 // Marking completed -- now verify that each bit marked in 2375 // verification_mark_bm() is also marked in markBitMap(); flag all 2376 // errors by printing corresponding objects. 2377 VerifyMarkedClosure vcl(markBitMap()); 2378 verification_mark_bm()->iterate(&vcl); 2379 if (vcl.failed()) { 2380 Log(gc, verify) log; 2381 log.error("Failed marking verification after remark"); 2382 ResourceMark rm; 2383 LogStream ls(log.error()); 2384 heap->print_on(&ls); 2385 fatal("CMS: failed marking verification after remark"); 2386 } 2387 } 2388 2389 class VerifyCLDOopsCLDClosure : public CLDClosure { 2390 class VerifyCLDOopsClosure : public OopClosure { 2391 CMSBitMap* _bitmap; 2392 public: 2393 VerifyCLDOopsClosure(CMSBitMap* bitmap) : _bitmap(bitmap) { } 2394 void do_oop(oop* p) { guarantee(*p == NULL || _bitmap->isMarked((HeapWord*) *p), "Should be marked"); } 2395 void do_oop(narrowOop* p) { ShouldNotReachHere(); } 2396 } _oop_closure; 2397 public: 2398 VerifyCLDOopsCLDClosure(CMSBitMap* bitmap) : _oop_closure(bitmap) {} 2399 void do_cld(ClassLoaderData* cld) { 2400 cld->oops_do(&_oop_closure, ClassLoaderData::_claim_none, false); 2401 } 2402 }; 2403 2404 void CMSCollector::verify_after_remark_work_2() { 2405 ResourceMark rm; 2406 HandleMark hm; 2407 CMSHeap* heap = CMSHeap::heap(); 2408 2409 // Get a clear set of claim bits for the roots processing to work with. 2410 ClassLoaderDataGraph::clear_claimed_marks(); 2411 2412 // Mark from roots one level into CMS 2413 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), 2414 markBitMap()); 2415 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2416 2417 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2418 2419 { 2420 StrongRootsScope srs(1); 2421 2422 heap->cms_process_roots(&srs, 2423 true, // young gen as roots 2424 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2425 should_unload_classes(), 2426 ¬Older, 2427 &cld_closure); 2428 } 2429 2430 // Now mark from the roots 2431 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, 2432 verification_mark_bm(), markBitMap(), verification_mark_stack()); 2433 assert(_restart_addr == NULL, "Expected pre-condition"); 2434 verification_mark_bm()->iterate(&markFromRootsClosure); 2435 while (_restart_addr != NULL) { 2436 // Deal with stack overflow: by restarting at the indicated 2437 // address. 2438 HeapWord* ra = _restart_addr; 2439 markFromRootsClosure.reset(ra); 2440 _restart_addr = NULL; 2441 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); 2442 } 2443 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); 2444 verify_work_stacks_empty(); 2445 2446 VerifyCLDOopsCLDClosure verify_cld_oops(verification_mark_bm()); 2447 ClassLoaderDataGraph::cld_do(&verify_cld_oops); 2448 2449 // Marking completed -- now verify that each bit marked in 2450 // verification_mark_bm() is also marked in markBitMap(); flag all 2451 // errors by printing corresponding objects. 2452 VerifyMarkedClosure vcl(markBitMap()); 2453 verification_mark_bm()->iterate(&vcl); 2454 assert(!vcl.failed(), "Else verification above should not have succeeded"); 2455 } 2456 2457 void ConcurrentMarkSweepGeneration::save_marks() { 2458 // delegate to CMS space 2459 cmsSpace()->save_marks(); 2460 } 2461 2462 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { 2463 return cmsSpace()->no_allocs_since_save_marks(); 2464 } 2465 2466 void 2467 ConcurrentMarkSweepGeneration::oop_iterate(OopIterateClosure* cl) { 2468 if (freelistLock()->owned_by_self()) { 2469 Generation::oop_iterate(cl); 2470 } else { 2471 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2472 Generation::oop_iterate(cl); 2473 } 2474 } 2475 2476 void 2477 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { 2478 if (freelistLock()->owned_by_self()) { 2479 Generation::object_iterate(cl); 2480 } else { 2481 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2482 Generation::object_iterate(cl); 2483 } 2484 } 2485 2486 void 2487 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { 2488 if (freelistLock()->owned_by_self()) { 2489 Generation::safe_object_iterate(cl); 2490 } else { 2491 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2492 Generation::safe_object_iterate(cl); 2493 } 2494 } 2495 2496 void 2497 ConcurrentMarkSweepGeneration::post_compact() { 2498 } 2499 2500 void 2501 ConcurrentMarkSweepGeneration::prepare_for_verify() { 2502 // Fix the linear allocation blocks to look like free blocks. 2503 2504 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2505 // are not called when the heap is verified during universe initialization and 2506 // at vm shutdown. 2507 if (freelistLock()->owned_by_self()) { 2508 cmsSpace()->prepare_for_verify(); 2509 } else { 2510 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2511 cmsSpace()->prepare_for_verify(); 2512 } 2513 } 2514 2515 void 2516 ConcurrentMarkSweepGeneration::verify() { 2517 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those 2518 // are not called when the heap is verified during universe initialization and 2519 // at vm shutdown. 2520 if (freelistLock()->owned_by_self()) { 2521 cmsSpace()->verify(); 2522 } else { 2523 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); 2524 cmsSpace()->verify(); 2525 } 2526 } 2527 2528 void CMSCollector::verify() { 2529 _cmsGen->verify(); 2530 } 2531 2532 #ifndef PRODUCT 2533 bool CMSCollector::overflow_list_is_empty() const { 2534 assert(_num_par_pushes >= 0, "Inconsistency"); 2535 if (_overflow_list == NULL) { 2536 assert(_num_par_pushes == 0, "Inconsistency"); 2537 } 2538 return _overflow_list == NULL; 2539 } 2540 2541 // The methods verify_work_stacks_empty() and verify_overflow_empty() 2542 // merely consolidate assertion checks that appear to occur together frequently. 2543 void CMSCollector::verify_work_stacks_empty() const { 2544 assert(_markStack.isEmpty(), "Marking stack should be empty"); 2545 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2546 } 2547 2548 void CMSCollector::verify_overflow_empty() const { 2549 assert(overflow_list_is_empty(), "Overflow list should be empty"); 2550 assert(no_preserved_marks(), "No preserved marks"); 2551 } 2552 #endif // PRODUCT 2553 2554 // Decide if we want to enable class unloading as part of the 2555 // ensuing concurrent GC cycle. We will collect and 2556 // unload classes if it's the case that: 2557 // (a) class unloading is enabled at the command line, and 2558 // (b) old gen is getting really full 2559 // NOTE: Provided there is no change in the state of the heap between 2560 // calls to this method, it should have idempotent results. Moreover, 2561 // its results should be monotonically increasing (i.e. going from 0 to 1, 2562 // but not 1 to 0) between successive calls between which the heap was 2563 // not collected. For the implementation below, it must thus rely on 2564 // the property that concurrent_cycles_since_last_unload() 2565 // will not decrease unless a collection cycle happened and that 2566 // _cmsGen->is_too_full() are 2567 // themselves also monotonic in that sense. See check_monotonicity() 2568 // below. 2569 void CMSCollector::update_should_unload_classes() { 2570 _should_unload_classes = false; 2571 if (CMSClassUnloadingEnabled) { 2572 _should_unload_classes = (concurrent_cycles_since_last_unload() >= 2573 CMSClassUnloadingMaxInterval) 2574 || _cmsGen->is_too_full(); 2575 } 2576 } 2577 2578 bool ConcurrentMarkSweepGeneration::is_too_full() const { 2579 bool res = should_concurrent_collect(); 2580 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); 2581 return res; 2582 } 2583 2584 void CMSCollector::setup_cms_unloading_and_verification_state() { 2585 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC 2586 || VerifyBeforeExit; 2587 const int rso = GenCollectedHeap::SO_AllCodeCache; 2588 2589 // We set the proper root for this CMS cycle here. 2590 if (should_unload_classes()) { // Should unload classes this cycle 2591 remove_root_scanning_option(rso); // Shrink the root set appropriately 2592 set_verifying(should_verify); // Set verification state for this cycle 2593 return; // Nothing else needs to be done at this time 2594 } 2595 2596 // Not unloading classes this cycle 2597 assert(!should_unload_classes(), "Inconsistency!"); 2598 2599 // If we are not unloading classes then add SO_AllCodeCache to root 2600 // scanning options. 2601 add_root_scanning_option(rso); 2602 2603 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { 2604 set_verifying(true); 2605 } else if (verifying() && !should_verify) { 2606 // We were verifying, but some verification flags got disabled. 2607 set_verifying(false); 2608 // Exclude symbols, strings and code cache elements from root scanning to 2609 // reduce IM and RM pauses. 2610 remove_root_scanning_option(rso); 2611 } 2612 } 2613 2614 2615 #ifndef PRODUCT 2616 HeapWord* CMSCollector::block_start(const void* p) const { 2617 const HeapWord* addr = (HeapWord*)p; 2618 if (_span.contains(p)) { 2619 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { 2620 return _cmsGen->cmsSpace()->block_start(p); 2621 } 2622 } 2623 return NULL; 2624 } 2625 #endif 2626 2627 HeapWord* 2628 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, 2629 bool tlab, 2630 bool parallel) { 2631 CMSSynchronousYieldRequest yr; 2632 assert(!tlab, "Can't deal with TLAB allocation"); 2633 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); 2634 expand_for_gc_cause(word_size*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_satisfy_allocation); 2635 if (GCExpandToAllocateDelayMillis > 0) { 2636 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2637 } 2638 return have_lock_and_allocate(word_size, tlab); 2639 } 2640 2641 void ConcurrentMarkSweepGeneration::expand_for_gc_cause( 2642 size_t bytes, 2643 size_t expand_bytes, 2644 CMSExpansionCause::Cause cause) 2645 { 2646 2647 bool success = expand(bytes, expand_bytes); 2648 2649 // remember why we expanded; this information is used 2650 // by shouldConcurrentCollect() when making decisions on whether to start 2651 // a new CMS cycle. 2652 if (success) { 2653 set_expansion_cause(cause); 2654 log_trace(gc)("Expanded CMS gen for %s", CMSExpansionCause::to_string(cause)); 2655 } 2656 } 2657 2658 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { 2659 HeapWord* res = NULL; 2660 MutexLocker x(ParGCRareEvent_lock); 2661 while (true) { 2662 // Expansion by some other thread might make alloc OK now: 2663 res = ps->lab.alloc(word_sz); 2664 if (res != NULL) return res; 2665 // If there's not enough expansion space available, give up. 2666 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { 2667 return NULL; 2668 } 2669 // Otherwise, we try expansion. 2670 expand_for_gc_cause(word_sz*HeapWordSize, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_lab); 2671 // Now go around the loop and try alloc again; 2672 // A competing par_promote might beat us to the expansion space, 2673 // so we may go around the loop again if promotion fails again. 2674 if (GCExpandToAllocateDelayMillis > 0) { 2675 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2676 } 2677 } 2678 } 2679 2680 2681 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( 2682 PromotionInfo* promo) { 2683 MutexLocker x(ParGCRareEvent_lock); 2684 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; 2685 while (true) { 2686 // Expansion by some other thread might make alloc OK now: 2687 if (promo->ensure_spooling_space()) { 2688 assert(promo->has_spooling_space(), 2689 "Post-condition of successful ensure_spooling_space()"); 2690 return true; 2691 } 2692 // If there's not enough expansion space available, give up. 2693 if (_virtual_space.uncommitted_size() < refill_size_bytes) { 2694 return false; 2695 } 2696 // Otherwise, we try expansion. 2697 expand_for_gc_cause(refill_size_bytes, MinHeapDeltaBytes, CMSExpansionCause::_allocate_par_spooling_space); 2698 // Now go around the loop and try alloc again; 2699 // A competing allocation might beat us to the expansion space, 2700 // so we may go around the loop again if allocation fails again. 2701 if (GCExpandToAllocateDelayMillis > 0) { 2702 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); 2703 } 2704 } 2705 } 2706 2707 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { 2708 // Only shrink if a compaction was done so that all the free space 2709 // in the generation is in a contiguous block at the end. 2710 if (did_compact()) { 2711 CardGeneration::shrink(bytes); 2712 } 2713 } 2714 2715 void ConcurrentMarkSweepGeneration::assert_correct_size_change_locking() { 2716 assert_locked_or_safepoint(Heap_lock); 2717 } 2718 2719 void ConcurrentMarkSweepGeneration::shrink_free_list_by(size_t bytes) { 2720 assert_locked_or_safepoint(Heap_lock); 2721 assert_lock_strong(freelistLock()); 2722 log_trace(gc)("Shrinking of CMS not yet implemented"); 2723 return; 2724 } 2725 2726 2727 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent 2728 // phases. 2729 class CMSPhaseAccounting: public StackObj { 2730 public: 2731 CMSPhaseAccounting(CMSCollector *collector, 2732 const char *title); 2733 ~CMSPhaseAccounting(); 2734 2735 private: 2736 CMSCollector *_collector; 2737 const char *_title; 2738 GCTraceConcTime(Info, gc) _trace_time; 2739 2740 public: 2741 // Not MT-safe; so do not pass around these StackObj's 2742 // where they may be accessed by other threads. 2743 double wallclock_millis() { 2744 return TimeHelper::counter_to_millis(os::elapsed_counter() - _trace_time.start_time()); 2745 } 2746 }; 2747 2748 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, 2749 const char *title) : 2750 _collector(collector), _title(title), _trace_time(title) { 2751 2752 _collector->resetYields(); 2753 _collector->resetTimer(); 2754 _collector->startTimer(); 2755 _collector->gc_timer_cm()->register_gc_concurrent_start(title); 2756 } 2757 2758 CMSPhaseAccounting::~CMSPhaseAccounting() { 2759 _collector->gc_timer_cm()->register_gc_concurrent_end(); 2760 _collector->stopTimer(); 2761 log_debug(gc)("Concurrent active time: %.3fms", TimeHelper::counter_to_seconds(_collector->timerTicks())); 2762 log_trace(gc)(" (CMS %s yielded %d times)", _title, _collector->yields()); 2763 } 2764 2765 // CMS work 2766 2767 // The common parts of CMSParInitialMarkTask and CMSParRemarkTask. 2768 class CMSParMarkTask : public AbstractGangTask { 2769 protected: 2770 CMSCollector* _collector; 2771 uint _n_workers; 2772 OopStorage::ParState<false, false> _par_state_string; 2773 CMSParMarkTask(const char* name, CMSCollector* collector, uint n_workers) : 2774 AbstractGangTask(name), 2775 _collector(collector), 2776 _n_workers(n_workers), 2777 _par_state_string(StringTable::weak_storage()) {} 2778 // Work method in support of parallel rescan ... of young gen spaces 2779 void do_young_space_rescan(OopsInGenClosure* cl, 2780 ContiguousSpace* space, 2781 HeapWord** chunk_array, size_t chunk_top); 2782 void work_on_young_gen_roots(OopsInGenClosure* cl); 2783 }; 2784 2785 // Parallel initial mark task 2786 class CMSParInitialMarkTask: public CMSParMarkTask { 2787 StrongRootsScope* _strong_roots_scope; 2788 public: 2789 CMSParInitialMarkTask(CMSCollector* collector, StrongRootsScope* strong_roots_scope, uint n_workers) : 2790 CMSParMarkTask("Scan roots and young gen for initial mark in parallel", collector, n_workers), 2791 _strong_roots_scope(strong_roots_scope) {} 2792 void work(uint worker_id); 2793 }; 2794 2795 // Checkpoint the roots into this generation from outside 2796 // this generation. [Note this initial checkpoint need only 2797 // be approximate -- we'll do a catch up phase subsequently.] 2798 void CMSCollector::checkpointRootsInitial() { 2799 assert(_collectorState == InitialMarking, "Wrong collector state"); 2800 check_correct_thread_executing(); 2801 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 2802 2803 save_heap_summary(); 2804 report_heap_summary(GCWhen::BeforeGC); 2805 2806 ReferenceProcessor* rp = ref_processor(); 2807 assert(_restart_addr == NULL, "Control point invariant"); 2808 { 2809 // acquire locks for subsequent manipulations 2810 MutexLockerEx x(bitMapLock(), 2811 Mutex::_no_safepoint_check_flag); 2812 checkpointRootsInitialWork(); 2813 // enable ("weak") refs discovery 2814 rp->enable_discovery(); 2815 _collectorState = Marking; 2816 } 2817 } 2818 2819 void CMSCollector::checkpointRootsInitialWork() { 2820 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); 2821 assert(_collectorState == InitialMarking, "just checking"); 2822 2823 // Already have locks. 2824 assert_lock_strong(bitMapLock()); 2825 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); 2826 2827 // Setup the verification and class unloading state for this 2828 // CMS collection cycle. 2829 setup_cms_unloading_and_verification_state(); 2830 2831 GCTraceTime(Trace, gc, phases) ts("checkpointRootsInitialWork", _gc_timer_cm); 2832 2833 // Reset all the PLAB chunk arrays if necessary. 2834 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { 2835 reset_survivor_plab_arrays(); 2836 } 2837 2838 ResourceMark rm; 2839 HandleMark hm; 2840 2841 MarkRefsIntoClosure notOlder(_span, &_markBitMap); 2842 CMSHeap* heap = CMSHeap::heap(); 2843 2844 verify_work_stacks_empty(); 2845 verify_overflow_empty(); 2846 2847 heap->ensure_parsability(false); // fill TLABs, but no need to retire them 2848 // Update the saved marks which may affect the root scans. 2849 heap->save_marks(); 2850 2851 // weak reference processing has not started yet. 2852 ref_processor()->set_enqueuing_is_done(false); 2853 2854 // Need to remember all newly created CLDs, 2855 // so that we can guarantee that the remark finds them. 2856 ClassLoaderDataGraph::remember_new_clds(true); 2857 2858 // Whenever a CLD is found, it will be claimed before proceeding to mark 2859 // the klasses. The claimed marks need to be cleared before marking starts. 2860 ClassLoaderDataGraph::clear_claimed_marks(); 2861 2862 print_eden_and_survivor_chunk_arrays(); 2863 2864 { 2865 #if COMPILER2_OR_JVMCI 2866 DerivedPointerTableDeactivate dpt_deact; 2867 #endif 2868 if (CMSParallelInitialMarkEnabled) { 2869 // The parallel version. 2870 WorkGang* workers = heap->workers(); 2871 assert(workers != NULL, "Need parallel worker threads."); 2872 uint n_workers = workers->active_workers(); 2873 2874 StrongRootsScope srs(n_workers); 2875 2876 CMSParInitialMarkTask tsk(this, &srs, n_workers); 2877 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 2878 // If the total workers is greater than 1, then multiple workers 2879 // may be used at some time and the initialization has been set 2880 // such that the single threaded path cannot be used. 2881 if (workers->total_workers() > 1) { 2882 workers->run_task(&tsk); 2883 } else { 2884 tsk.work(0); 2885 } 2886 } else { 2887 // The serial version. 2888 CLDToOopClosure cld_closure(¬Older, ClassLoaderData::_claim_strong); 2889 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 2890 2891 StrongRootsScope srs(1); 2892 2893 heap->cms_process_roots(&srs, 2894 true, // young gen as roots 2895 GenCollectedHeap::ScanningOption(roots_scanning_options()), 2896 should_unload_classes(), 2897 ¬Older, 2898 &cld_closure); 2899 } 2900 } 2901 2902 // Clear mod-union table; it will be dirtied in the prologue of 2903 // CMS generation per each young generation collection. 2904 2905 assert(_modUnionTable.isAllClear(), 2906 "Was cleared in most recent final checkpoint phase" 2907 " or no bits are set in the gc_prologue before the start of the next " 2908 "subsequent marking phase."); 2909 2910 assert(_ct->cld_rem_set()->mod_union_is_clear(), "Must be"); 2911 2912 // Save the end of the used_region of the constituent generations 2913 // to be used to limit the extent of sweep in each generation. 2914 save_sweep_limits(); 2915 verify_overflow_empty(); 2916 } 2917 2918 bool CMSCollector::markFromRoots() { 2919 // we might be tempted to assert that: 2920 // assert(!SafepointSynchronize::is_at_safepoint(), 2921 // "inconsistent argument?"); 2922 // However that wouldn't be right, because it's possible that 2923 // a safepoint is indeed in progress as a young generation 2924 // stop-the-world GC happens even as we mark in this generation. 2925 assert(_collectorState == Marking, "inconsistent state?"); 2926 check_correct_thread_executing(); 2927 verify_overflow_empty(); 2928 2929 // Weak ref discovery note: We may be discovering weak 2930 // refs in this generation concurrent (but interleaved) with 2931 // weak ref discovery by the young generation collector. 2932 2933 CMSTokenSyncWithLocks ts(true, bitMapLock()); 2934 GCTraceCPUTime tcpu; 2935 CMSPhaseAccounting pa(this, "Concurrent Mark"); 2936 bool res = markFromRootsWork(); 2937 if (res) { 2938 _collectorState = Precleaning; 2939 } else { // We failed and a foreground collection wants to take over 2940 assert(_foregroundGCIsActive, "internal state inconsistency"); 2941 assert(_restart_addr == NULL, "foreground will restart from scratch"); 2942 log_debug(gc)("bailing out to foreground collection"); 2943 } 2944 verify_overflow_empty(); 2945 return res; 2946 } 2947 2948 bool CMSCollector::markFromRootsWork() { 2949 // iterate over marked bits in bit map, doing a full scan and mark 2950 // from these roots using the following algorithm: 2951 // . if oop is to the right of the current scan pointer, 2952 // mark corresponding bit (we'll process it later) 2953 // . else (oop is to left of current scan pointer) 2954 // push oop on marking stack 2955 // . drain the marking stack 2956 2957 // Note that when we do a marking step we need to hold the 2958 // bit map lock -- recall that direct allocation (by mutators) 2959 // and promotion (by the young generation collector) is also 2960 // marking the bit map. [the so-called allocate live policy.] 2961 // Because the implementation of bit map marking is not 2962 // robust wrt simultaneous marking of bits in the same word, 2963 // we need to make sure that there is no such interference 2964 // between concurrent such updates. 2965 2966 // already have locks 2967 assert_lock_strong(bitMapLock()); 2968 2969 verify_work_stacks_empty(); 2970 verify_overflow_empty(); 2971 bool result = false; 2972 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { 2973 result = do_marking_mt(); 2974 } else { 2975 result = do_marking_st(); 2976 } 2977 return result; 2978 } 2979 2980 // Forward decl 2981 class CMSConcMarkingTask; 2982 2983 class CMSConcMarkingTerminator: public ParallelTaskTerminator { 2984 CMSCollector* _collector; 2985 CMSConcMarkingTask* _task; 2986 public: 2987 virtual void yield(); 2988 2989 // "n_threads" is the number of threads to be terminated. 2990 // "queue_set" is a set of work queues of other threads. 2991 // "collector" is the CMS collector associated with this task terminator. 2992 // "yield" indicates whether we need the gang as a whole to yield. 2993 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : 2994 ParallelTaskTerminator(n_threads, queue_set), 2995 _collector(collector) { } 2996 2997 void set_task(CMSConcMarkingTask* task) { 2998 _task = task; 2999 } 3000 }; 3001 3002 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { 3003 CMSConcMarkingTask* _task; 3004 public: 3005 bool should_exit_termination(); 3006 void set_task(CMSConcMarkingTask* task) { 3007 _task = task; 3008 } 3009 }; 3010 3011 // MT Concurrent Marking Task 3012 class CMSConcMarkingTask: public YieldingFlexibleGangTask { 3013 CMSCollector* _collector; 3014 uint _n_workers; // requested/desired # workers 3015 bool _result; 3016 CompactibleFreeListSpace* _cms_space; 3017 char _pad_front[64]; // padding to ... 3018 HeapWord* volatile _global_finger; // ... avoid sharing cache line 3019 char _pad_back[64]; 3020 HeapWord* _restart_addr; 3021 3022 // Exposed here for yielding support 3023 Mutex* const _bit_map_lock; 3024 3025 // The per thread work queues, available here for stealing 3026 OopTaskQueueSet* _task_queues; 3027 3028 // Termination (and yielding) support 3029 CMSConcMarkingTerminator _term; 3030 CMSConcMarkingTerminatorTerminator _term_term; 3031 3032 public: 3033 CMSConcMarkingTask(CMSCollector* collector, 3034 CompactibleFreeListSpace* cms_space, 3035 YieldingFlexibleWorkGang* workers, 3036 OopTaskQueueSet* task_queues): 3037 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), 3038 _collector(collector), 3039 _n_workers(0), 3040 _result(true), 3041 _cms_space(cms_space), 3042 _bit_map_lock(collector->bitMapLock()), 3043 _task_queues(task_queues), 3044 _term(_n_workers, task_queues, _collector) 3045 { 3046 _requested_size = _n_workers; 3047 _term.set_task(this); 3048 _term_term.set_task(this); 3049 _restart_addr = _global_finger = _cms_space->bottom(); 3050 } 3051 3052 3053 OopTaskQueueSet* task_queues() { return _task_queues; } 3054 3055 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 3056 3057 HeapWord* volatile* global_finger_addr() { return &_global_finger; } 3058 3059 CMSConcMarkingTerminator* terminator() { return &_term; } 3060 3061 virtual void set_for_termination(uint active_workers) { 3062 terminator()->reset_for_reuse(active_workers); 3063 } 3064 3065 void work(uint worker_id); 3066 bool should_yield() { 3067 return ConcurrentMarkSweepThread::should_yield() 3068 && !_collector->foregroundGCIsActive(); 3069 } 3070 3071 virtual void coordinator_yield(); // stuff done by coordinator 3072 bool result() { return _result; } 3073 3074 void reset(HeapWord* ra) { 3075 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); 3076 _restart_addr = _global_finger = ra; 3077 _term.reset_for_reuse(); 3078 } 3079 3080 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3081 OopTaskQueue* work_q); 3082 3083 private: 3084 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); 3085 void do_work_steal(int i); 3086 void bump_global_finger(HeapWord* f); 3087 }; 3088 3089 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { 3090 assert(_task != NULL, "Error"); 3091 return _task->yielding(); 3092 // Note that we do not need the disjunct || _task->should_yield() above 3093 // because we want terminating threads to yield only if the task 3094 // is already in the midst of yielding, which happens only after at least one 3095 // thread has yielded. 3096 } 3097 3098 void CMSConcMarkingTerminator::yield() { 3099 if (_task->should_yield()) { 3100 _task->yield(); 3101 } else { 3102 ParallelTaskTerminator::yield(); 3103 } 3104 } 3105 3106 //////////////////////////////////////////////////////////////// 3107 // Concurrent Marking Algorithm Sketch 3108 //////////////////////////////////////////////////////////////// 3109 // Until all tasks exhausted (both spaces): 3110 // -- claim next available chunk 3111 // -- bump global finger via CAS 3112 // -- find first object that starts in this chunk 3113 // and start scanning bitmap from that position 3114 // -- scan marked objects for oops 3115 // -- CAS-mark target, and if successful: 3116 // . if target oop is above global finger (volatile read) 3117 // nothing to do 3118 // . if target oop is in chunk and above local finger 3119 // then nothing to do 3120 // . else push on work-queue 3121 // -- Deal with possible overflow issues: 3122 // . local work-queue overflow causes stuff to be pushed on 3123 // global (common) overflow queue 3124 // . always first empty local work queue 3125 // . then get a batch of oops from global work queue if any 3126 // . then do work stealing 3127 // -- When all tasks claimed (both spaces) 3128 // and local work queue empty, 3129 // then in a loop do: 3130 // . check global overflow stack; steal a batch of oops and trace 3131 // . try to steal from other threads oif GOS is empty 3132 // . if neither is available, offer termination 3133 // -- Terminate and return result 3134 // 3135 void CMSConcMarkingTask::work(uint worker_id) { 3136 elapsedTimer _timer; 3137 ResourceMark rm; 3138 HandleMark hm; 3139 3140 DEBUG_ONLY(_collector->verify_overflow_empty();) 3141 3142 // Before we begin work, our work queue should be empty 3143 assert(work_queue(worker_id)->size() == 0, "Expected to be empty"); 3144 // Scan the bitmap covering _cms_space, tracing through grey objects. 3145 _timer.start(); 3146 do_scan_and_mark(worker_id, _cms_space); 3147 _timer.stop(); 3148 log_trace(gc, task)("Finished cms space scanning in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3149 3150 // ... do work stealing 3151 _timer.reset(); 3152 _timer.start(); 3153 do_work_steal(worker_id); 3154 _timer.stop(); 3155 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 3156 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); 3157 assert(work_queue(worker_id)->size() == 0, "Should have been emptied"); 3158 // Note that under the current task protocol, the 3159 // following assertion is true even of the spaces 3160 // expanded since the completion of the concurrent 3161 // marking. XXX This will likely change under a strict 3162 // ABORT semantics. 3163 // After perm removal the comparison was changed to 3164 // greater than or equal to from strictly greater than. 3165 // Before perm removal the highest address sweep would 3166 // have been at the end of perm gen but now is at the 3167 // end of the tenured gen. 3168 assert(_global_finger >= _cms_space->end(), 3169 "All tasks have been completed"); 3170 DEBUG_ONLY(_collector->verify_overflow_empty();) 3171 } 3172 3173 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { 3174 HeapWord* read = _global_finger; 3175 HeapWord* cur = read; 3176 while (f > read) { 3177 cur = read; 3178 read = Atomic::cmpxchg(f, &_global_finger, cur); 3179 if (cur == read) { 3180 // our cas succeeded 3181 assert(_global_finger >= f, "protocol consistency"); 3182 break; 3183 } 3184 } 3185 } 3186 3187 // This is really inefficient, and should be redone by 3188 // using (not yet available) block-read and -write interfaces to the 3189 // stack and the work_queue. XXX FIX ME !!! 3190 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, 3191 OopTaskQueue* work_q) { 3192 // Fast lock-free check 3193 if (ovflw_stk->length() == 0) { 3194 return false; 3195 } 3196 assert(work_q->size() == 0, "Shouldn't steal"); 3197 MutexLockerEx ml(ovflw_stk->par_lock(), 3198 Mutex::_no_safepoint_check_flag); 3199 // Grab up to 1/4 the size of the work queue 3200 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 3201 (size_t)ParGCDesiredObjsFromOverflowList); 3202 num = MIN2(num, ovflw_stk->length()); 3203 for (int i = (int) num; i > 0; i--) { 3204 oop cur = ovflw_stk->pop(); 3205 assert(cur != NULL, "Counted wrong?"); 3206 work_q->push(cur); 3207 } 3208 return num > 0; 3209 } 3210 3211 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { 3212 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 3213 int n_tasks = pst->n_tasks(); 3214 // We allow that there may be no tasks to do here because 3215 // we are restarting after a stack overflow. 3216 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); 3217 uint nth_task = 0; 3218 3219 HeapWord* aligned_start = sp->bottom(); 3220 if (sp->used_region().contains(_restart_addr)) { 3221 // Align down to a card boundary for the start of 0th task 3222 // for this space. 3223 aligned_start = align_down(_restart_addr, CardTable::card_size); 3224 } 3225 3226 size_t chunk_size = sp->marking_task_size(); 3227 while (pst->try_claim_task(/* reference */ nth_task)) { 3228 // Having claimed the nth task in this space, 3229 // compute the chunk that it corresponds to: 3230 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, 3231 aligned_start + (nth_task+1)*chunk_size); 3232 // Try and bump the global finger via a CAS; 3233 // note that we need to do the global finger bump 3234 // _before_ taking the intersection below, because 3235 // the task corresponding to that region will be 3236 // deemed done even if the used_region() expands 3237 // because of allocation -- as it almost certainly will 3238 // during start-up while the threads yield in the 3239 // closure below. 3240 HeapWord* finger = span.end(); 3241 bump_global_finger(finger); // atomically 3242 // There are null tasks here corresponding to chunks 3243 // beyond the "top" address of the space. 3244 span = span.intersection(sp->used_region()); 3245 if (!span.is_empty()) { // Non-null task 3246 HeapWord* prev_obj; 3247 assert(!span.contains(_restart_addr) || nth_task == 0, 3248 "Inconsistency"); 3249 if (nth_task == 0) { 3250 // For the 0th task, we'll not need to compute a block_start. 3251 if (span.contains(_restart_addr)) { 3252 // In the case of a restart because of stack overflow, 3253 // we might additionally skip a chunk prefix. 3254 prev_obj = _restart_addr; 3255 } else { 3256 prev_obj = span.start(); 3257 } 3258 } else { 3259 // We want to skip the first object because 3260 // the protocol is to scan any object in its entirety 3261 // that _starts_ in this span; a fortiori, any 3262 // object starting in an earlier span is scanned 3263 // as part of an earlier claimed task. 3264 // Below we use the "careful" version of block_start 3265 // so we do not try to navigate uninitialized objects. 3266 prev_obj = sp->block_start_careful(span.start()); 3267 // Below we use a variant of block_size that uses the 3268 // Printezis bits to avoid waiting for allocated 3269 // objects to become initialized/parsable. 3270 while (prev_obj < span.start()) { 3271 size_t sz = sp->block_size_no_stall(prev_obj, _collector); 3272 if (sz > 0) { 3273 prev_obj += sz; 3274 } else { 3275 // In this case we may end up doing a bit of redundant 3276 // scanning, but that appears unavoidable, short of 3277 // locking the free list locks; see bug 6324141. 3278 break; 3279 } 3280 } 3281 } 3282 if (prev_obj < span.end()) { 3283 MemRegion my_span = MemRegion(prev_obj, span.end()); 3284 // Do the marking work within a non-empty span -- 3285 // the last argument to the constructor indicates whether the 3286 // iteration should be incremental with periodic yields. 3287 ParMarkFromRootsClosure cl(this, _collector, my_span, 3288 &_collector->_markBitMap, 3289 work_queue(i), 3290 &_collector->_markStack); 3291 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); 3292 } // else nothing to do for this task 3293 } // else nothing to do for this task 3294 } 3295 // We'd be tempted to assert here that since there are no 3296 // more tasks left to claim in this space, the global_finger 3297 // must exceed space->top() and a fortiori space->end(). However, 3298 // that would not quite be correct because the bumping of 3299 // global_finger occurs strictly after the claiming of a task, 3300 // so by the time we reach here the global finger may not yet 3301 // have been bumped up by the thread that claimed the last 3302 // task. 3303 pst->all_tasks_completed(); 3304 } 3305 3306 class ParConcMarkingClosure: public MetadataVisitingOopIterateClosure { 3307 private: 3308 CMSCollector* _collector; 3309 CMSConcMarkingTask* _task; 3310 MemRegion _span; 3311 CMSBitMap* _bit_map; 3312 CMSMarkStack* _overflow_stack; 3313 OopTaskQueue* _work_queue; 3314 protected: 3315 DO_OOP_WORK_DEFN 3316 public: 3317 ParConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, 3318 CMSBitMap* bit_map, CMSMarkStack* overflow_stack): 3319 MetadataVisitingOopIterateClosure(collector->ref_processor()), 3320 _collector(collector), 3321 _task(task), 3322 _span(collector->_span), 3323 _bit_map(bit_map), 3324 _overflow_stack(overflow_stack), 3325 _work_queue(work_queue) 3326 { } 3327 virtual void do_oop(oop* p); 3328 virtual void do_oop(narrowOop* p); 3329 3330 void trim_queue(size_t max); 3331 void handle_stack_overflow(HeapWord* lost); 3332 void do_yield_check() { 3333 if (_task->should_yield()) { 3334 _task->yield(); 3335 } 3336 } 3337 }; 3338 3339 DO_OOP_WORK_IMPL(ParConcMarkingClosure) 3340 3341 // Grey object scanning during work stealing phase -- 3342 // the salient assumption here is that any references 3343 // that are in these stolen objects being scanned must 3344 // already have been initialized (else they would not have 3345 // been published), so we do not need to check for 3346 // uninitialized objects before pushing here. 3347 void ParConcMarkingClosure::do_oop(oop obj) { 3348 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 3349 HeapWord* addr = (HeapWord*)obj; 3350 // Check if oop points into the CMS generation 3351 // and is not marked 3352 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 3353 // a white object ... 3354 // If we manage to "claim" the object, by being the 3355 // first thread to mark it, then we push it on our 3356 // marking stack 3357 if (_bit_map->par_mark(addr)) { // ... now grey 3358 // push on work queue (grey set) 3359 bool simulate_overflow = false; 3360 NOT_PRODUCT( 3361 if (CMSMarkStackOverflowALot && 3362 _collector->simulate_overflow()) { 3363 // simulate a stack overflow 3364 simulate_overflow = true; 3365 } 3366 ) 3367 if (simulate_overflow || 3368 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 3369 // stack overflow 3370 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 3371 // We cannot assert that the overflow stack is full because 3372 // it may have been emptied since. 3373 assert(simulate_overflow || 3374 _work_queue->size() == _work_queue->max_elems(), 3375 "Else push should have succeeded"); 3376 handle_stack_overflow(addr); 3377 } 3378 } // Else, some other thread got there first 3379 do_yield_check(); 3380 } 3381 } 3382 3383 void ParConcMarkingClosure::trim_queue(size_t max) { 3384 while (_work_queue->size() > max) { 3385 oop new_oop; 3386 if (_work_queue->pop_local(new_oop)) { 3387 assert(oopDesc::is_oop(new_oop), "Should be an oop"); 3388 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); 3389 assert(_span.contains((HeapWord*)new_oop), "Not in span"); 3390 new_oop->oop_iterate(this); // do_oop() above 3391 do_yield_check(); 3392 } 3393 } 3394 } 3395 3396 // Upon stack overflow, we discard (part of) the stack, 3397 // remembering the least address amongst those discarded 3398 // in CMSCollector's _restart_address. 3399 void ParConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { 3400 // We need to do this under a mutex to prevent other 3401 // workers from interfering with the work done below. 3402 MutexLockerEx ml(_overflow_stack->par_lock(), 3403 Mutex::_no_safepoint_check_flag); 3404 // Remember the least grey address discarded 3405 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 3406 _collector->lower_restart_addr(ra); 3407 _overflow_stack->reset(); // discard stack contents 3408 _overflow_stack->expand(); // expand the stack if possible 3409 } 3410 3411 3412 void CMSConcMarkingTask::do_work_steal(int i) { 3413 OopTaskQueue* work_q = work_queue(i); 3414 oop obj_to_scan; 3415 CMSBitMap* bm = &(_collector->_markBitMap); 3416 CMSMarkStack* ovflw = &(_collector->_markStack); 3417 ParConcMarkingClosure cl(_collector, this, work_q, bm, ovflw); 3418 while (true) { 3419 cl.trim_queue(0); 3420 assert(work_q->size() == 0, "Should have been emptied above"); 3421 if (get_work_from_overflow_stack(ovflw, work_q)) { 3422 // Can't assert below because the work obtained from the 3423 // overflow stack may already have been stolen from us. 3424 // assert(work_q->size() > 0, "Work from overflow stack"); 3425 continue; 3426 } else if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 3427 assert(oopDesc::is_oop(obj_to_scan), "Should be an oop"); 3428 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); 3429 obj_to_scan->oop_iterate(&cl); 3430 } else if (terminator()->offer_termination(&_term_term)) { 3431 assert(work_q->size() == 0, "Impossible!"); 3432 break; 3433 } else if (yielding() || should_yield()) { 3434 yield(); 3435 } 3436 } 3437 } 3438 3439 // This is run by the CMS (coordinator) thread. 3440 void CMSConcMarkingTask::coordinator_yield() { 3441 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3442 "CMS thread should hold CMS token"); 3443 // First give up the locks, then yield, then re-lock 3444 // We should probably use a constructor/destructor idiom to 3445 // do this unlock/lock or modify the MutexUnlocker class to 3446 // serve our purpose. XXX 3447 assert_lock_strong(_bit_map_lock); 3448 _bit_map_lock->unlock(); 3449 ConcurrentMarkSweepThread::desynchronize(true); 3450 _collector->stopTimer(); 3451 _collector->incrementYields(); 3452 3453 // It is possible for whichever thread initiated the yield request 3454 // not to get a chance to wake up and take the bitmap lock between 3455 // this thread releasing it and reacquiring it. So, while the 3456 // should_yield() flag is on, let's sleep for a bit to give the 3457 // other thread a chance to wake up. The limit imposed on the number 3458 // of iterations is defensive, to avoid any unforseen circumstances 3459 // putting us into an infinite loop. Since it's always been this 3460 // (coordinator_yield()) method that was observed to cause the 3461 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) 3462 // which is by default non-zero. For the other seven methods that 3463 // also perform the yield operation, as are using a different 3464 // parameter (CMSYieldSleepCount) which is by default zero. This way we 3465 // can enable the sleeping for those methods too, if necessary. 3466 // See 6442774. 3467 // 3468 // We really need to reconsider the synchronization between the GC 3469 // thread and the yield-requesting threads in the future and we 3470 // should really use wait/notify, which is the recommended 3471 // way of doing this type of interaction. Additionally, we should 3472 // consolidate the eight methods that do the yield operation and they 3473 // are almost identical into one for better maintainability and 3474 // readability. See 6445193. 3475 // 3476 // Tony 2006.06.29 3477 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && 3478 ConcurrentMarkSweepThread::should_yield() && 3479 !CMSCollector::foregroundGCIsActive(); ++i) { 3480 os::sleep(Thread::current(), 1, false); 3481 } 3482 3483 ConcurrentMarkSweepThread::synchronize(true); 3484 _bit_map_lock->lock_without_safepoint_check(); 3485 _collector->startTimer(); 3486 } 3487 3488 bool CMSCollector::do_marking_mt() { 3489 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); 3490 uint num_workers = AdaptiveSizePolicy::calc_active_conc_workers(conc_workers()->total_workers(), 3491 conc_workers()->active_workers(), 3492 Threads::number_of_non_daemon_threads()); 3493 num_workers = conc_workers()->update_active_workers(num_workers); 3494 log_info(gc,task)("Using %u workers of %u for marking", num_workers, conc_workers()->total_workers()); 3495 3496 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 3497 3498 CMSConcMarkingTask tsk(this, 3499 cms_space, 3500 conc_workers(), 3501 task_queues()); 3502 3503 // Since the actual number of workers we get may be different 3504 // from the number we requested above, do we need to do anything different 3505 // below? In particular, may be we need to subclass the SequantialSubTasksDone 3506 // class?? XXX 3507 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); 3508 3509 // Refs discovery is already non-atomic. 3510 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); 3511 assert(ref_processor()->discovery_is_mt(), "Discovery should be MT"); 3512 conc_workers()->start_task(&tsk); 3513 while (tsk.yielded()) { 3514 tsk.coordinator_yield(); 3515 conc_workers()->continue_task(&tsk); 3516 } 3517 // If the task was aborted, _restart_addr will be non-NULL 3518 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); 3519 while (_restart_addr != NULL) { 3520 // XXX For now we do not make use of ABORTED state and have not 3521 // yet implemented the right abort semantics (even in the original 3522 // single-threaded CMS case). That needs some more investigation 3523 // and is deferred for now; see CR# TBF. 07252005YSR. XXX 3524 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); 3525 // If _restart_addr is non-NULL, a marking stack overflow 3526 // occurred; we need to do a fresh marking iteration from the 3527 // indicated restart address. 3528 if (_foregroundGCIsActive) { 3529 // We may be running into repeated stack overflows, having 3530 // reached the limit of the stack size, while making very 3531 // slow forward progress. It may be best to bail out and 3532 // let the foreground collector do its job. 3533 // Clear _restart_addr, so that foreground GC 3534 // works from scratch. This avoids the headache of 3535 // a "rescan" which would otherwise be needed because 3536 // of the dirty mod union table & card table. 3537 _restart_addr = NULL; 3538 return false; 3539 } 3540 // Adjust the task to restart from _restart_addr 3541 tsk.reset(_restart_addr); 3542 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, 3543 _restart_addr); 3544 _restart_addr = NULL; 3545 // Get the workers going again 3546 conc_workers()->start_task(&tsk); 3547 while (tsk.yielded()) { 3548 tsk.coordinator_yield(); 3549 conc_workers()->continue_task(&tsk); 3550 } 3551 } 3552 assert(tsk.completed(), "Inconsistency"); 3553 assert(tsk.result() == true, "Inconsistency"); 3554 return true; 3555 } 3556 3557 bool CMSCollector::do_marking_st() { 3558 ResourceMark rm; 3559 HandleMark hm; 3560 3561 // Temporarily make refs discovery single threaded (non-MT) 3562 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(ref_processor(), false); 3563 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, 3564 &_markStack, CMSYield); 3565 // the last argument to iterate indicates whether the iteration 3566 // should be incremental with periodic yields. 3567 _markBitMap.iterate(&markFromRootsClosure); 3568 // If _restart_addr is non-NULL, a marking stack overflow 3569 // occurred; we need to do a fresh iteration from the 3570 // indicated restart address. 3571 while (_restart_addr != NULL) { 3572 if (_foregroundGCIsActive) { 3573 // We may be running into repeated stack overflows, having 3574 // reached the limit of the stack size, while making very 3575 // slow forward progress. It may be best to bail out and 3576 // let the foreground collector do its job. 3577 // Clear _restart_addr, so that foreground GC 3578 // works from scratch. This avoids the headache of 3579 // a "rescan" which would otherwise be needed because 3580 // of the dirty mod union table & card table. 3581 _restart_addr = NULL; 3582 return false; // indicating failure to complete marking 3583 } 3584 // Deal with stack overflow: 3585 // we restart marking from _restart_addr 3586 HeapWord* ra = _restart_addr; 3587 markFromRootsClosure.reset(ra); 3588 _restart_addr = NULL; 3589 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); 3590 } 3591 return true; 3592 } 3593 3594 void CMSCollector::preclean() { 3595 check_correct_thread_executing(); 3596 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); 3597 verify_work_stacks_empty(); 3598 verify_overflow_empty(); 3599 _abort_preclean = false; 3600 if (CMSPrecleaningEnabled) { 3601 if (!CMSEdenChunksRecordAlways) { 3602 _eden_chunk_index = 0; 3603 } 3604 size_t used = get_eden_used(); 3605 size_t capacity = get_eden_capacity(); 3606 // Don't start sampling unless we will get sufficiently 3607 // many samples. 3608 if (used < (((capacity / CMSScheduleRemarkSamplingRatio) / 100) 3609 * CMSScheduleRemarkEdenPenetration)) { 3610 _start_sampling = true; 3611 } else { 3612 _start_sampling = false; 3613 } 3614 GCTraceCPUTime tcpu; 3615 CMSPhaseAccounting pa(this, "Concurrent Preclean"); 3616 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); 3617 } 3618 CMSTokenSync x(true); // is cms thread 3619 if (CMSPrecleaningEnabled) { 3620 sample_eden(); 3621 _collectorState = AbortablePreclean; 3622 } else { 3623 _collectorState = FinalMarking; 3624 } 3625 verify_work_stacks_empty(); 3626 verify_overflow_empty(); 3627 } 3628 3629 // Try and schedule the remark such that young gen 3630 // occupancy is CMSScheduleRemarkEdenPenetration %. 3631 void CMSCollector::abortable_preclean() { 3632 check_correct_thread_executing(); 3633 assert(CMSPrecleaningEnabled, "Inconsistent control state"); 3634 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); 3635 3636 // If Eden's current occupancy is below this threshold, 3637 // immediately schedule the remark; else preclean 3638 // past the next scavenge in an effort to 3639 // schedule the pause as described above. By choosing 3640 // CMSScheduleRemarkEdenSizeThreshold >= max eden size 3641 // we will never do an actual abortable preclean cycle. 3642 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { 3643 GCTraceCPUTime tcpu; 3644 CMSPhaseAccounting pa(this, "Concurrent Abortable Preclean"); 3645 // We need more smarts in the abortable preclean 3646 // loop below to deal with cases where allocation 3647 // in young gen is very very slow, and our precleaning 3648 // is running a losing race against a horde of 3649 // mutators intent on flooding us with CMS updates 3650 // (dirty cards). 3651 // One, admittedly dumb, strategy is to give up 3652 // after a certain number of abortable precleaning loops 3653 // or after a certain maximum time. We want to make 3654 // this smarter in the next iteration. 3655 // XXX FIX ME!!! YSR 3656 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; 3657 while (!(should_abort_preclean() || 3658 ConcurrentMarkSweepThread::cmst()->should_terminate())) { 3659 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); 3660 cumworkdone += workdone; 3661 loops++; 3662 // Voluntarily terminate abortable preclean phase if we have 3663 // been at it for too long. 3664 if ((CMSMaxAbortablePrecleanLoops != 0) && 3665 loops >= CMSMaxAbortablePrecleanLoops) { 3666 log_debug(gc)(" CMS: abort preclean due to loops "); 3667 break; 3668 } 3669 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { 3670 log_debug(gc)(" CMS: abort preclean due to time "); 3671 break; 3672 } 3673 // If we are doing little work each iteration, we should 3674 // take a short break. 3675 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { 3676 // Sleep for some time, waiting for work to accumulate 3677 stopTimer(); 3678 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); 3679 startTimer(); 3680 waited++; 3681 } 3682 } 3683 log_trace(gc)(" [" SIZE_FORMAT " iterations, " SIZE_FORMAT " waits, " SIZE_FORMAT " cards)] ", 3684 loops, waited, cumworkdone); 3685 } 3686 CMSTokenSync x(true); // is cms thread 3687 if (_collectorState != Idling) { 3688 assert(_collectorState == AbortablePreclean, 3689 "Spontaneous state transition?"); 3690 _collectorState = FinalMarking; 3691 } // Else, a foreground collection completed this CMS cycle. 3692 return; 3693 } 3694 3695 // Respond to an Eden sampling opportunity 3696 void CMSCollector::sample_eden() { 3697 // Make sure a young gc cannot sneak in between our 3698 // reading and recording of a sample. 3699 assert(Thread::current()->is_ConcurrentGC_thread(), 3700 "Only the cms thread may collect Eden samples"); 3701 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 3702 "Should collect samples while holding CMS token"); 3703 if (!_start_sampling) { 3704 return; 3705 } 3706 // When CMSEdenChunksRecordAlways is true, the eden chunk array 3707 // is populated by the young generation. 3708 if (_eden_chunk_array != NULL && !CMSEdenChunksRecordAlways) { 3709 if (_eden_chunk_index < _eden_chunk_capacity) { 3710 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample 3711 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 3712 "Unexpected state of Eden"); 3713 // We'd like to check that what we just sampled is an oop-start address; 3714 // however, we cannot do that here since the object may not yet have been 3715 // initialized. So we'll instead do the check when we _use_ this sample 3716 // later. 3717 if (_eden_chunk_index == 0 || 3718 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 3719 _eden_chunk_array[_eden_chunk_index-1]) 3720 >= CMSSamplingGrain)) { 3721 _eden_chunk_index++; // commit sample 3722 } 3723 } 3724 } 3725 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { 3726 size_t used = get_eden_used(); 3727 size_t capacity = get_eden_capacity(); 3728 assert(used <= capacity, "Unexpected state of Eden"); 3729 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { 3730 _abort_preclean = true; 3731 } 3732 } 3733 } 3734 3735 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { 3736 assert(_collectorState == Precleaning || 3737 _collectorState == AbortablePreclean, "incorrect state"); 3738 ResourceMark rm; 3739 HandleMark hm; 3740 3741 // Precleaning is currently not MT but the reference processor 3742 // may be set for MT. Disable it temporarily here. 3743 ReferenceProcessor* rp = ref_processor(); 3744 ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false); 3745 3746 // Do one pass of scrubbing the discovered reference lists 3747 // to remove any reference objects with strongly-reachable 3748 // referents. 3749 if (clean_refs) { 3750 CMSPrecleanRefsYieldClosure yield_cl(this); 3751 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 3752 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, 3753 &_markStack, true /* preclean */); 3754 CMSDrainMarkingStackClosure complete_trace(this, 3755 _span, &_markBitMap, &_markStack, 3756 &keep_alive, true /* preclean */); 3757 3758 // We don't want this step to interfere with a young 3759 // collection because we don't want to take CPU 3760 // or memory bandwidth away from the young GC threads 3761 // (which may be as many as there are CPUs). 3762 // Note that we don't need to protect ourselves from 3763 // interference with mutators because they can't 3764 // manipulate the discovered reference lists nor affect 3765 // the computed reachability of the referents, the 3766 // only properties manipulated by the precleaning 3767 // of these reference lists. 3768 stopTimer(); 3769 CMSTokenSyncWithLocks x(true /* is cms thread */, 3770 bitMapLock()); 3771 startTimer(); 3772 sample_eden(); 3773 3774 // The following will yield to allow foreground 3775 // collection to proceed promptly. XXX YSR: 3776 // The code in this method may need further 3777 // tweaking for better performance and some restructuring 3778 // for cleaner interfaces. 3779 GCTimer *gc_timer = NULL; // Currently not tracing concurrent phases 3780 rp->preclean_discovered_references( 3781 rp->is_alive_non_header(), &keep_alive, &complete_trace, &yield_cl, 3782 gc_timer); 3783 } 3784 3785 if (clean_survivor) { // preclean the active survivor space(s) 3786 PushAndMarkClosure pam_cl(this, _span, ref_processor(), 3787 &_markBitMap, &_modUnionTable, 3788 &_markStack, true /* precleaning phase */); 3789 stopTimer(); 3790 CMSTokenSyncWithLocks ts(true /* is cms thread */, 3791 bitMapLock()); 3792 startTimer(); 3793 unsigned int before_count = 3794 CMSHeap::heap()->total_collections(); 3795 SurvivorSpacePrecleanClosure 3796 sss_cl(this, _span, &_markBitMap, &_markStack, 3797 &pam_cl, before_count, CMSYield); 3798 _young_gen->from()->object_iterate_careful(&sss_cl); 3799 _young_gen->to()->object_iterate_careful(&sss_cl); 3800 } 3801 MarkRefsIntoAndScanClosure 3802 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, 3803 &_markStack, this, CMSYield, 3804 true /* precleaning phase */); 3805 // CAUTION: The following closure has persistent state that may need to 3806 // be reset upon a decrease in the sequence of addresses it 3807 // processes. 3808 ScanMarkedObjectsAgainCarefullyClosure 3809 smoac_cl(this, _span, 3810 &_markBitMap, &_markStack, &mrias_cl, CMSYield); 3811 3812 // Preclean dirty cards in ModUnionTable and CardTable using 3813 // appropriate convergence criterion; 3814 // repeat CMSPrecleanIter times unless we find that 3815 // we are losing. 3816 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); 3817 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, 3818 "Bad convergence multiplier"); 3819 assert(CMSPrecleanThreshold >= 100, 3820 "Unreasonably low CMSPrecleanThreshold"); 3821 3822 size_t numIter, cumNumCards, lastNumCards, curNumCards; 3823 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; 3824 numIter < CMSPrecleanIter; 3825 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { 3826 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); 3827 log_trace(gc)(" (modUnionTable: " SIZE_FORMAT " cards)", curNumCards); 3828 // Either there are very few dirty cards, so re-mark 3829 // pause will be small anyway, or our pre-cleaning isn't 3830 // that much faster than the rate at which cards are being 3831 // dirtied, so we might as well stop and re-mark since 3832 // precleaning won't improve our re-mark time by much. 3833 if (curNumCards <= CMSPrecleanThreshold || 3834 (numIter > 0 && 3835 (curNumCards * CMSPrecleanDenominator > 3836 lastNumCards * CMSPrecleanNumerator))) { 3837 numIter++; 3838 cumNumCards += curNumCards; 3839 break; 3840 } 3841 } 3842 3843 preclean_cld(&mrias_cl, _cmsGen->freelistLock()); 3844 3845 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); 3846 cumNumCards += curNumCards; 3847 log_trace(gc)(" (cardTable: " SIZE_FORMAT " cards, re-scanned " SIZE_FORMAT " cards, " SIZE_FORMAT " iterations)", 3848 curNumCards, cumNumCards, numIter); 3849 return cumNumCards; // as a measure of useful work done 3850 } 3851 3852 // PRECLEANING NOTES: 3853 // Precleaning involves: 3854 // . reading the bits of the modUnionTable and clearing the set bits. 3855 // . For the cards corresponding to the set bits, we scan the 3856 // objects on those cards. This means we need the free_list_lock 3857 // so that we can safely iterate over the CMS space when scanning 3858 // for oops. 3859 // . When we scan the objects, we'll be both reading and setting 3860 // marks in the marking bit map, so we'll need the marking bit map. 3861 // . For protecting _collector_state transitions, we take the CGC_lock. 3862 // Note that any races in the reading of of card table entries by the 3863 // CMS thread on the one hand and the clearing of those entries by the 3864 // VM thread or the setting of those entries by the mutator threads on the 3865 // other are quite benign. However, for efficiency it makes sense to keep 3866 // the VM thread from racing with the CMS thread while the latter is 3867 // dirty card info to the modUnionTable. We therefore also use the 3868 // CGC_lock to protect the reading of the card table and the mod union 3869 // table by the CM thread. 3870 // . We run concurrently with mutator updates, so scanning 3871 // needs to be done carefully -- we should not try to scan 3872 // potentially uninitialized objects. 3873 // 3874 // Locking strategy: While holding the CGC_lock, we scan over and 3875 // reset a maximal dirty range of the mod union / card tables, then lock 3876 // the free_list_lock and bitmap lock to do a full marking, then 3877 // release these locks; and repeat the cycle. This allows for a 3878 // certain amount of fairness in the sharing of these locks between 3879 // the CMS collector on the one hand, and the VM thread and the 3880 // mutators on the other. 3881 3882 // NOTE: preclean_mod_union_table() and preclean_card_table() 3883 // further below are largely identical; if you need to modify 3884 // one of these methods, please check the other method too. 3885 3886 size_t CMSCollector::preclean_mod_union_table( 3887 ConcurrentMarkSweepGeneration* old_gen, 3888 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3889 verify_work_stacks_empty(); 3890 verify_overflow_empty(); 3891 3892 // strategy: starting with the first card, accumulate contiguous 3893 // ranges of dirty cards; clear these cards, then scan the region 3894 // covered by these cards. 3895 3896 // Since all of the MUT is committed ahead, we can just use 3897 // that, in case the generations expand while we are precleaning. 3898 // It might also be fine to just use the committed part of the 3899 // generation, but we might potentially miss cards when the 3900 // generation is rapidly expanding while we are in the midst 3901 // of precleaning. 3902 HeapWord* startAddr = old_gen->reserved().start(); 3903 HeapWord* endAddr = old_gen->reserved().end(); 3904 3905 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3906 3907 size_t numDirtyCards, cumNumDirtyCards; 3908 HeapWord *nextAddr, *lastAddr; 3909 for (cumNumDirtyCards = numDirtyCards = 0, 3910 nextAddr = lastAddr = startAddr; 3911 nextAddr < endAddr; 3912 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 3913 3914 ResourceMark rm; 3915 HandleMark hm; 3916 3917 MemRegion dirtyRegion; 3918 { 3919 stopTimer(); 3920 // Potential yield point 3921 CMSTokenSync ts(true); 3922 startTimer(); 3923 sample_eden(); 3924 // Get dirty region starting at nextOffset (inclusive), 3925 // simultaneously clearing it. 3926 dirtyRegion = 3927 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); 3928 assert(dirtyRegion.start() >= nextAddr, 3929 "returned region inconsistent?"); 3930 } 3931 // Remember where the next search should begin. 3932 // The returned region (if non-empty) is a right open interval, 3933 // so lastOffset is obtained from the right end of that 3934 // interval. 3935 lastAddr = dirtyRegion.end(); 3936 // Should do something more transparent and less hacky XXX 3937 numDirtyCards = 3938 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); 3939 3940 // We'll scan the cards in the dirty region (with periodic 3941 // yields for foreground GC as needed). 3942 if (!dirtyRegion.is_empty()) { 3943 assert(numDirtyCards > 0, "consistency check"); 3944 HeapWord* stop_point = NULL; 3945 stopTimer(); 3946 // Potential yield point 3947 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), 3948 bitMapLock()); 3949 startTimer(); 3950 { 3951 verify_work_stacks_empty(); 3952 verify_overflow_empty(); 3953 sample_eden(); 3954 stop_point = 3955 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 3956 } 3957 if (stop_point != NULL) { 3958 // The careful iteration stopped early either because it found an 3959 // uninitialized object, or because we were in the midst of an 3960 // "abortable preclean", which should now be aborted. Redirty 3961 // the bits corresponding to the partially-scanned or unscanned 3962 // cards. We'll either restart at the next block boundary or 3963 // abort the preclean. 3964 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 3965 "Should only be AbortablePreclean."); 3966 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); 3967 if (should_abort_preclean()) { 3968 break; // out of preclean loop 3969 } else { 3970 // Compute the next address at which preclean should pick up; 3971 // might need bitMapLock in order to read P-bits. 3972 lastAddr = next_card_start_after_block(stop_point); 3973 } 3974 } 3975 } else { 3976 assert(lastAddr == endAddr, "consistency check"); 3977 assert(numDirtyCards == 0, "consistency check"); 3978 break; 3979 } 3980 } 3981 verify_work_stacks_empty(); 3982 verify_overflow_empty(); 3983 return cumNumDirtyCards; 3984 } 3985 3986 // NOTE: preclean_mod_union_table() above and preclean_card_table() 3987 // below are largely identical; if you need to modify 3988 // one of these methods, please check the other method too. 3989 3990 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* old_gen, 3991 ScanMarkedObjectsAgainCarefullyClosure* cl) { 3992 // strategy: it's similar to precleamModUnionTable above, in that 3993 // we accumulate contiguous ranges of dirty cards, mark these cards 3994 // precleaned, then scan the region covered by these cards. 3995 HeapWord* endAddr = (HeapWord*)(old_gen->_virtual_space.high()); 3996 HeapWord* startAddr = (HeapWord*)(old_gen->_virtual_space.low()); 3997 3998 cl->setFreelistLock(old_gen->freelistLock()); // needed for yielding 3999 4000 size_t numDirtyCards, cumNumDirtyCards; 4001 HeapWord *lastAddr, *nextAddr; 4002 4003 for (cumNumDirtyCards = numDirtyCards = 0, 4004 nextAddr = lastAddr = startAddr; 4005 nextAddr < endAddr; 4006 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { 4007 4008 ResourceMark rm; 4009 HandleMark hm; 4010 4011 MemRegion dirtyRegion; 4012 { 4013 // See comments in "Precleaning notes" above on why we 4014 // do this locking. XXX Could the locking overheads be 4015 // too high when dirty cards are sparse? [I don't think so.] 4016 stopTimer(); 4017 CMSTokenSync x(true); // is cms thread 4018 startTimer(); 4019 sample_eden(); 4020 // Get and clear dirty region from card table 4021 dirtyRegion = _ct->dirty_card_range_after_reset(MemRegion(nextAddr, endAddr), 4022 true, 4023 CardTable::precleaned_card_val()); 4024 4025 assert(dirtyRegion.start() >= nextAddr, 4026 "returned region inconsistent?"); 4027 } 4028 lastAddr = dirtyRegion.end(); 4029 numDirtyCards = 4030 dirtyRegion.word_size()/CardTable::card_size_in_words; 4031 4032 if (!dirtyRegion.is_empty()) { 4033 stopTimer(); 4034 CMSTokenSyncWithLocks ts(true, old_gen->freelistLock(), bitMapLock()); 4035 startTimer(); 4036 sample_eden(); 4037 verify_work_stacks_empty(); 4038 verify_overflow_empty(); 4039 HeapWord* stop_point = 4040 old_gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); 4041 if (stop_point != NULL) { 4042 assert((_collectorState == AbortablePreclean && should_abort_preclean()), 4043 "Should only be AbortablePreclean."); 4044 _ct->invalidate(MemRegion(stop_point, dirtyRegion.end())); 4045 if (should_abort_preclean()) { 4046 break; // out of preclean loop 4047 } else { 4048 // Compute the next address at which preclean should pick up. 4049 lastAddr = next_card_start_after_block(stop_point); 4050 } 4051 } 4052 } else { 4053 break; 4054 } 4055 } 4056 verify_work_stacks_empty(); 4057 verify_overflow_empty(); 4058 return cumNumDirtyCards; 4059 } 4060 4061 class PrecleanCLDClosure : public CLDClosure { 4062 MetadataVisitingOopsInGenClosure* _cm_closure; 4063 public: 4064 PrecleanCLDClosure(MetadataVisitingOopsInGenClosure* oop_closure) : _cm_closure(oop_closure) {} 4065 void do_cld(ClassLoaderData* cld) { 4066 if (cld->has_accumulated_modified_oops()) { 4067 cld->clear_accumulated_modified_oops(); 4068 4069 _cm_closure->do_cld(cld); 4070 } 4071 } 4072 }; 4073 4074 // The freelist lock is needed to prevent asserts, is it really needed? 4075 void CMSCollector::preclean_cld(MarkRefsIntoAndScanClosure* cl, Mutex* freelistLock) { 4076 // Needed to walk CLDG 4077 MutexLocker ml(ClassLoaderDataGraph_lock); 4078 4079 cl->set_freelistLock(freelistLock); 4080 4081 CMSTokenSyncWithLocks ts(true, freelistLock, bitMapLock()); 4082 4083 // SSS: Add equivalent to ScanMarkedObjectsAgainCarefullyClosure::do_yield_check and should_abort_preclean? 4084 // SSS: We should probably check if precleaning should be aborted, at suitable intervals? 4085 PrecleanCLDClosure preclean_closure(cl); 4086 ClassLoaderDataGraph::cld_do(&preclean_closure); 4087 4088 verify_work_stacks_empty(); 4089 verify_overflow_empty(); 4090 } 4091 4092 void CMSCollector::checkpointRootsFinal() { 4093 assert(_collectorState == FinalMarking, "incorrect state transition?"); 4094 check_correct_thread_executing(); 4095 // world is stopped at this checkpoint 4096 assert(SafepointSynchronize::is_at_safepoint(), 4097 "world should be stopped"); 4098 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 4099 4100 verify_work_stacks_empty(); 4101 verify_overflow_empty(); 4102 4103 log_debug(gc)("YG occupancy: " SIZE_FORMAT " K (" SIZE_FORMAT " K)", 4104 _young_gen->used() / K, _young_gen->capacity() / K); 4105 { 4106 if (CMSScavengeBeforeRemark) { 4107 CMSHeap* heap = CMSHeap::heap(); 4108 // Temporarily set flag to false, GCH->do_collection will 4109 // expect it to be false and set to true 4110 FlagSetting fl(heap->_is_gc_active, false); 4111 4112 heap->do_collection(true, // full (i.e. force, see below) 4113 false, // !clear_all_soft_refs 4114 0, // size 4115 false, // is_tlab 4116 GenCollectedHeap::YoungGen // type 4117 ); 4118 } 4119 FreelistLocker x(this); 4120 MutexLockerEx y(bitMapLock(), 4121 Mutex::_no_safepoint_check_flag); 4122 checkpointRootsFinalWork(); 4123 } 4124 verify_work_stacks_empty(); 4125 verify_overflow_empty(); 4126 } 4127 4128 void CMSCollector::checkpointRootsFinalWork() { 4129 GCTraceTime(Trace, gc, phases) tm("checkpointRootsFinalWork", _gc_timer_cm); 4130 4131 assert(haveFreelistLocks(), "must have free list locks"); 4132 assert_lock_strong(bitMapLock()); 4133 4134 ResourceMark rm; 4135 HandleMark hm; 4136 4137 CMSHeap* heap = CMSHeap::heap(); 4138 4139 if (should_unload_classes()) { 4140 CodeCache::gc_prologue(); 4141 } 4142 assert(haveFreelistLocks(), "must have free list locks"); 4143 assert_lock_strong(bitMapLock()); 4144 4145 // We might assume that we need not fill TLAB's when 4146 // CMSScavengeBeforeRemark is set, because we may have just done 4147 // a scavenge which would have filled all TLAB's -- and besides 4148 // Eden would be empty. This however may not always be the case -- 4149 // for instance although we asked for a scavenge, it may not have 4150 // happened because of a JNI critical section. We probably need 4151 // a policy for deciding whether we can in that case wait until 4152 // the critical section releases and then do the remark following 4153 // the scavenge, and skip it here. In the absence of that policy, 4154 // or of an indication of whether the scavenge did indeed occur, 4155 // we cannot rely on TLAB's having been filled and must do 4156 // so here just in case a scavenge did not happen. 4157 heap->ensure_parsability(false); // fill TLAB's, but no need to retire them 4158 // Update the saved marks which may affect the root scans. 4159 heap->save_marks(); 4160 4161 print_eden_and_survivor_chunk_arrays(); 4162 4163 { 4164 #if COMPILER2_OR_JVMCI 4165 DerivedPointerTableDeactivate dpt_deact; 4166 #endif 4167 4168 // Note on the role of the mod union table: 4169 // Since the marker in "markFromRoots" marks concurrently with 4170 // mutators, it is possible for some reachable objects not to have been 4171 // scanned. For instance, an only reference to an object A was 4172 // placed in object B after the marker scanned B. Unless B is rescanned, 4173 // A would be collected. Such updates to references in marked objects 4174 // are detected via the mod union table which is the set of all cards 4175 // dirtied since the first checkpoint in this GC cycle and prior to 4176 // the most recent young generation GC, minus those cleaned up by the 4177 // concurrent precleaning. 4178 if (CMSParallelRemarkEnabled) { 4179 GCTraceTime(Debug, gc, phases) t("Rescan (parallel)", _gc_timer_cm); 4180 do_remark_parallel(); 4181 } else { 4182 GCTraceTime(Debug, gc, phases) t("Rescan (non-parallel)", _gc_timer_cm); 4183 do_remark_non_parallel(); 4184 } 4185 } 4186 verify_work_stacks_empty(); 4187 verify_overflow_empty(); 4188 4189 { 4190 GCTraceTime(Trace, gc, phases) ts("refProcessingWork", _gc_timer_cm); 4191 refProcessingWork(); 4192 } 4193 verify_work_stacks_empty(); 4194 verify_overflow_empty(); 4195 4196 if (should_unload_classes()) { 4197 CodeCache::gc_epilogue(); 4198 } 4199 JvmtiExport::gc_epilogue(); 4200 4201 // If we encountered any (marking stack / work queue) overflow 4202 // events during the current CMS cycle, take appropriate 4203 // remedial measures, where possible, so as to try and avoid 4204 // recurrence of that condition. 4205 assert(_markStack.isEmpty(), "No grey objects"); 4206 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + 4207 _ser_kac_ovflw + _ser_kac_preclean_ovflw; 4208 if (ser_ovflw > 0) { 4209 log_trace(gc)("Marking stack overflow (benign) (pmc_pc=" SIZE_FORMAT ", pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ", kac_preclean=" SIZE_FORMAT ")", 4210 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, _ser_kac_ovflw, _ser_kac_preclean_ovflw); 4211 _markStack.expand(); 4212 _ser_pmc_remark_ovflw = 0; 4213 _ser_pmc_preclean_ovflw = 0; 4214 _ser_kac_preclean_ovflw = 0; 4215 _ser_kac_ovflw = 0; 4216 } 4217 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { 4218 log_trace(gc)("Work queue overflow (benign) (pmc_rm=" SIZE_FORMAT ", kac=" SIZE_FORMAT ")", 4219 _par_pmc_remark_ovflw, _par_kac_ovflw); 4220 _par_pmc_remark_ovflw = 0; 4221 _par_kac_ovflw = 0; 4222 } 4223 if (_markStack._hit_limit > 0) { 4224 log_trace(gc)(" (benign) Hit max stack size limit (" SIZE_FORMAT ")", 4225 _markStack._hit_limit); 4226 } 4227 if (_markStack._failed_double > 0) { 4228 log_trace(gc)(" (benign) Failed stack doubling (" SIZE_FORMAT "), current capacity " SIZE_FORMAT, 4229 _markStack._failed_double, _markStack.capacity()); 4230 } 4231 _markStack._hit_limit = 0; 4232 _markStack._failed_double = 0; 4233 4234 if ((VerifyAfterGC || VerifyDuringGC) && 4235 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4236 verify_after_remark(); 4237 } 4238 4239 _gc_tracer_cm->report_object_count_after_gc(&_is_alive_closure); 4240 4241 // Change under the freelistLocks. 4242 _collectorState = Sweeping; 4243 // Call isAllClear() under bitMapLock 4244 assert(_modUnionTable.isAllClear(), 4245 "Should be clear by end of the final marking"); 4246 assert(_ct->cld_rem_set()->mod_union_is_clear(), 4247 "Should be clear by end of the final marking"); 4248 } 4249 4250 void CMSParInitialMarkTask::work(uint worker_id) { 4251 elapsedTimer _timer; 4252 ResourceMark rm; 4253 HandleMark hm; 4254 4255 // ---------- scan from roots -------------- 4256 _timer.start(); 4257 CMSHeap* heap = CMSHeap::heap(); 4258 ParMarkRefsIntoClosure par_mri_cl(_collector->_span, &(_collector->_markBitMap)); 4259 4260 // ---------- young gen roots -------------- 4261 { 4262 work_on_young_gen_roots(&par_mri_cl); 4263 _timer.stop(); 4264 log_trace(gc, task)("Finished young gen initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4265 } 4266 4267 // ---------- remaining roots -------------- 4268 _timer.reset(); 4269 _timer.start(); 4270 4271 CLDToOopClosure cld_closure(&par_mri_cl, ClassLoaderData::_claim_strong); 4272 4273 heap->cms_process_roots(_strong_roots_scope, 4274 false, // yg was scanned above 4275 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4276 _collector->should_unload_classes(), 4277 &par_mri_cl, 4278 &cld_closure, 4279 &_par_state_string); 4280 4281 assert(_collector->should_unload_classes() 4282 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4283 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4284 _timer.stop(); 4285 log_trace(gc, task)("Finished remaining root initial mark scan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4286 } 4287 4288 // Parallel remark task 4289 class CMSParRemarkTask: public CMSParMarkTask { 4290 CompactibleFreeListSpace* _cms_space; 4291 4292 // The per-thread work queues, available here for stealing. 4293 OopTaskQueueSet* _task_queues; 4294 ParallelTaskTerminator _term; 4295 StrongRootsScope* _strong_roots_scope; 4296 4297 public: 4298 // A value of 0 passed to n_workers will cause the number of 4299 // workers to be taken from the active workers in the work gang. 4300 CMSParRemarkTask(CMSCollector* collector, 4301 CompactibleFreeListSpace* cms_space, 4302 uint n_workers, WorkGang* workers, 4303 OopTaskQueueSet* task_queues, 4304 StrongRootsScope* strong_roots_scope): 4305 CMSParMarkTask("Rescan roots and grey objects in parallel", 4306 collector, n_workers), 4307 _cms_space(cms_space), 4308 _task_queues(task_queues), 4309 _term(n_workers, task_queues), 4310 _strong_roots_scope(strong_roots_scope) { } 4311 4312 OopTaskQueueSet* task_queues() { return _task_queues; } 4313 4314 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 4315 4316 ParallelTaskTerminator* terminator() { return &_term; } 4317 uint n_workers() { return _n_workers; } 4318 4319 void work(uint worker_id); 4320 4321 private: 4322 // ... of dirty cards in old space 4323 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, 4324 ParMarkRefsIntoAndScanClosure* cl); 4325 4326 // ... work stealing for the above 4327 void do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl); 4328 }; 4329 4330 class RemarkCLDClosure : public CLDClosure { 4331 CLDToOopClosure _cm_closure; 4332 public: 4333 RemarkCLDClosure(OopClosure* oop_closure) : _cm_closure(oop_closure, ClassLoaderData::_claim_strong) {} 4334 void do_cld(ClassLoaderData* cld) { 4335 // Check if we have modified any oops in the CLD during the concurrent marking. 4336 if (cld->has_accumulated_modified_oops()) { 4337 cld->clear_accumulated_modified_oops(); 4338 4339 // We could have transfered the current modified marks to the accumulated marks, 4340 // like we do with the Card Table to Mod Union Table. But it's not really necessary. 4341 } else if (cld->has_modified_oops()) { 4342 // Don't clear anything, this info is needed by the next young collection. 4343 } else { 4344 // No modified oops in the ClassLoaderData. 4345 return; 4346 } 4347 4348 // The klass has modified fields, need to scan the klass. 4349 _cm_closure.do_cld(cld); 4350 } 4351 }; 4352 4353 void CMSParMarkTask::work_on_young_gen_roots(OopsInGenClosure* cl) { 4354 ParNewGeneration* young_gen = _collector->_young_gen; 4355 ContiguousSpace* eden_space = young_gen->eden(); 4356 ContiguousSpace* from_space = young_gen->from(); 4357 ContiguousSpace* to_space = young_gen->to(); 4358 4359 HeapWord** eca = _collector->_eden_chunk_array; 4360 size_t ect = _collector->_eden_chunk_index; 4361 HeapWord** sca = _collector->_survivor_chunk_array; 4362 size_t sct = _collector->_survivor_chunk_index; 4363 4364 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); 4365 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); 4366 4367 do_young_space_rescan(cl, to_space, NULL, 0); 4368 do_young_space_rescan(cl, from_space, sca, sct); 4369 do_young_space_rescan(cl, eden_space, eca, ect); 4370 } 4371 4372 // work_queue(i) is passed to the closure 4373 // ParMarkRefsIntoAndScanClosure. The "i" parameter 4374 // also is passed to do_dirty_card_rescan_tasks() and to 4375 // do_work_steal() to select the i-th task_queue. 4376 4377 void CMSParRemarkTask::work(uint worker_id) { 4378 elapsedTimer _timer; 4379 ResourceMark rm; 4380 HandleMark hm; 4381 4382 // ---------- rescan from roots -------------- 4383 _timer.start(); 4384 CMSHeap* heap = CMSHeap::heap(); 4385 ParMarkRefsIntoAndScanClosure par_mrias_cl(_collector, 4386 _collector->_span, _collector->ref_processor(), 4387 &(_collector->_markBitMap), 4388 work_queue(worker_id)); 4389 4390 // Rescan young gen roots first since these are likely 4391 // coarsely partitioned and may, on that account, constitute 4392 // the critical path; thus, it's best to start off that 4393 // work first. 4394 // ---------- young gen roots -------------- 4395 { 4396 work_on_young_gen_roots(&par_mrias_cl); 4397 _timer.stop(); 4398 log_trace(gc, task)("Finished young gen rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4399 } 4400 4401 // ---------- remaining roots -------------- 4402 _timer.reset(); 4403 _timer.start(); 4404 heap->cms_process_roots(_strong_roots_scope, 4405 false, // yg was scanned above 4406 GenCollectedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), 4407 _collector->should_unload_classes(), 4408 &par_mrias_cl, 4409 NULL, // The dirty klasses will be handled below 4410 &_par_state_string); 4411 4412 assert(_collector->should_unload_classes() 4413 || (_collector->CMSCollector::roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4414 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4415 _timer.stop(); 4416 log_trace(gc, task)("Finished remaining root rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4417 4418 // ---------- unhandled CLD scanning ---------- 4419 if (worker_id == 0) { // Single threaded at the moment. 4420 _timer.reset(); 4421 _timer.start(); 4422 4423 // Scan all new class loader data objects and new dependencies that were 4424 // introduced during concurrent marking. 4425 ResourceMark rm; 4426 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4427 for (int i = 0; i < array->length(); i++) { 4428 Devirtualizer::do_cld(&par_mrias_cl, array->at(i)); 4429 } 4430 4431 // We don't need to keep track of new CLDs anymore. 4432 ClassLoaderDataGraph::remember_new_clds(false); 4433 4434 _timer.stop(); 4435 log_trace(gc, task)("Finished unhandled CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4436 } 4437 4438 // We might have added oops to ClassLoaderData::_handles during the 4439 // concurrent marking phase. These oops do not always point to newly allocated objects 4440 // that are guaranteed to be kept alive. Hence, 4441 // we do have to revisit the _handles block during the remark phase. 4442 4443 // ---------- dirty CLD scanning ---------- 4444 if (worker_id == 0) { // Single threaded at the moment. 4445 _timer.reset(); 4446 _timer.start(); 4447 4448 // Scan all classes that was dirtied during the concurrent marking phase. 4449 RemarkCLDClosure remark_closure(&par_mrias_cl); 4450 ClassLoaderDataGraph::cld_do(&remark_closure); 4451 4452 _timer.stop(); 4453 log_trace(gc, task)("Finished dirty CLD scanning work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4454 } 4455 4456 4457 // ---------- rescan dirty cards ------------ 4458 _timer.reset(); 4459 _timer.start(); 4460 4461 // Do the rescan tasks for each of the two spaces 4462 // (cms_space) in turn. 4463 // "worker_id" is passed to select the task_queue for "worker_id" 4464 do_dirty_card_rescan_tasks(_cms_space, worker_id, &par_mrias_cl); 4465 _timer.stop(); 4466 log_trace(gc, task)("Finished dirty card rescan work in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4467 4468 // ---------- steal work from other threads ... 4469 // ---------- ... and drain overflow list. 4470 _timer.reset(); 4471 _timer.start(); 4472 do_work_steal(worker_id, &par_mrias_cl); 4473 _timer.stop(); 4474 log_trace(gc, task)("Finished work stealing in %dth thread: %3.3f sec", worker_id, _timer.seconds()); 4475 } 4476 4477 void 4478 CMSParMarkTask::do_young_space_rescan( 4479 OopsInGenClosure* cl, ContiguousSpace* space, 4480 HeapWord** chunk_array, size_t chunk_top) { 4481 // Until all tasks completed: 4482 // . claim an unclaimed task 4483 // . compute region boundaries corresponding to task claimed 4484 // using chunk_array 4485 // . par_oop_iterate(cl) over that region 4486 4487 ResourceMark rm; 4488 HandleMark hm; 4489 4490 SequentialSubTasksDone* pst = space->par_seq_tasks(); 4491 4492 uint nth_task = 0; 4493 uint n_tasks = pst->n_tasks(); 4494 4495 if (n_tasks > 0) { 4496 assert(pst->valid(), "Uninitialized use?"); 4497 HeapWord *start, *end; 4498 while (pst->try_claim_task(/* reference */ nth_task)) { 4499 // We claimed task # nth_task; compute its boundaries. 4500 if (chunk_top == 0) { // no samples were taken 4501 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 eden task"); 4502 start = space->bottom(); 4503 end = space->top(); 4504 } else if (nth_task == 0) { 4505 start = space->bottom(); 4506 end = chunk_array[nth_task]; 4507 } else if (nth_task < (uint)chunk_top) { 4508 assert(nth_task >= 1, "Control point invariant"); 4509 start = chunk_array[nth_task - 1]; 4510 end = chunk_array[nth_task]; 4511 } else { 4512 assert(nth_task == (uint)chunk_top, "Control point invariant"); 4513 start = chunk_array[chunk_top - 1]; 4514 end = space->top(); 4515 } 4516 MemRegion mr(start, end); 4517 // Verify that mr is in space 4518 assert(mr.is_empty() || space->used_region().contains(mr), 4519 "Should be in space"); 4520 // Verify that "start" is an object boundary 4521 assert(mr.is_empty() || oopDesc::is_oop(oop(mr.start())), 4522 "Should be an oop"); 4523 space->par_oop_iterate(mr, cl); 4524 } 4525 pst->all_tasks_completed(); 4526 } 4527 } 4528 4529 void 4530 CMSParRemarkTask::do_dirty_card_rescan_tasks( 4531 CompactibleFreeListSpace* sp, int i, 4532 ParMarkRefsIntoAndScanClosure* cl) { 4533 // Until all tasks completed: 4534 // . claim an unclaimed task 4535 // . compute region boundaries corresponding to task claimed 4536 // . transfer dirty bits ct->mut for that region 4537 // . apply rescanclosure to dirty mut bits for that region 4538 4539 ResourceMark rm; 4540 HandleMark hm; 4541 4542 OopTaskQueue* work_q = work_queue(i); 4543 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); 4544 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! 4545 // CAUTION: This closure has state that persists across calls to 4546 // the work method dirty_range_iterate_clear() in that it has 4547 // embedded in it a (subtype of) UpwardsObjectClosure. The 4548 // use of that state in the embedded UpwardsObjectClosure instance 4549 // assumes that the cards are always iterated (even if in parallel 4550 // by several threads) in monotonically increasing order per each 4551 // thread. This is true of the implementation below which picks 4552 // card ranges (chunks) in monotonically increasing order globally 4553 // and, a-fortiori, in monotonically increasing order per thread 4554 // (the latter order being a subsequence of the former). 4555 // If the work code below is ever reorganized into a more chaotic 4556 // work-partitioning form than the current "sequential tasks" 4557 // paradigm, the use of that persistent state will have to be 4558 // revisited and modified appropriately. See also related 4559 // bug 4756801 work on which should examine this code to make 4560 // sure that the changes there do not run counter to the 4561 // assumptions made here and necessary for correctness and 4562 // efficiency. Note also that this code might yield inefficient 4563 // behavior in the case of very large objects that span one or 4564 // more work chunks. Such objects would potentially be scanned 4565 // several times redundantly. Work on 4756801 should try and 4566 // address that performance anomaly if at all possible. XXX 4567 MemRegion full_span = _collector->_span; 4568 CMSBitMap* bm = &(_collector->_markBitMap); // shared 4569 MarkFromDirtyCardsClosure 4570 greyRescanClosure(_collector, full_span, // entire span of interest 4571 sp, bm, work_q, cl); 4572 4573 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); 4574 assert(pst->valid(), "Uninitialized use?"); 4575 uint nth_task = 0; 4576 const int alignment = CardTable::card_size * BitsPerWord; 4577 MemRegion span = sp->used_region(); 4578 HeapWord* start_addr = span.start(); 4579 HeapWord* end_addr = align_up(span.end(), alignment); 4580 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units 4581 assert(is_aligned(start_addr, alignment), "Check alignment"); 4582 assert(is_aligned(chunk_size, alignment), "Check alignment"); 4583 4584 while (pst->try_claim_task(/* reference */ nth_task)) { 4585 // Having claimed the nth_task, compute corresponding mem-region, 4586 // which is a-fortiori aligned correctly (i.e. at a MUT boundary). 4587 // The alignment restriction ensures that we do not need any 4588 // synchronization with other gang-workers while setting or 4589 // clearing bits in thus chunk of the MUT. 4590 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, 4591 start_addr + (nth_task+1)*chunk_size); 4592 // The last chunk's end might be way beyond end of the 4593 // used region. In that case pull back appropriately. 4594 if (this_span.end() > end_addr) { 4595 this_span.set_end(end_addr); 4596 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); 4597 } 4598 // Iterate over the dirty cards covering this chunk, marking them 4599 // precleaned, and setting the corresponding bits in the mod union 4600 // table. Since we have been careful to partition at Card and MUT-word 4601 // boundaries no synchronization is needed between parallel threads. 4602 _collector->_ct->dirty_card_iterate(this_span, 4603 &modUnionClosure); 4604 4605 // Having transferred these marks into the modUnionTable, 4606 // rescan the marked objects on the dirty cards in the modUnionTable. 4607 // Even if this is at a synchronous collection, the initial marking 4608 // may have been done during an asynchronous collection so there 4609 // may be dirty bits in the mod-union table. 4610 _collector->_modUnionTable.dirty_range_iterate_clear( 4611 this_span, &greyRescanClosure); 4612 _collector->_modUnionTable.verifyNoOneBitsInRange( 4613 this_span.start(), 4614 this_span.end()); 4615 } 4616 pst->all_tasks_completed(); // declare that i am done 4617 } 4618 4619 // . see if we can share work_queues with ParNew? XXX 4620 void 4621 CMSParRemarkTask::do_work_steal(int i, ParMarkRefsIntoAndScanClosure* cl) { 4622 OopTaskQueue* work_q = work_queue(i); 4623 NOT_PRODUCT(int num_steals = 0;) 4624 oop obj_to_scan; 4625 CMSBitMap* bm = &(_collector->_markBitMap); 4626 4627 while (true) { 4628 // Completely finish any left over work from (an) earlier round(s) 4629 cl->trim_queue(0); 4630 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 4631 (size_t)ParGCDesiredObjsFromOverflowList); 4632 // Now check if there's any work in the overflow list 4633 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 4634 // only affects the number of attempts made to get work from the 4635 // overflow list and does not affect the number of workers. Just 4636 // pass ParallelGCThreads so this behavior is unchanged. 4637 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 4638 work_q, 4639 ParallelGCThreads)) { 4640 // found something in global overflow list; 4641 // not yet ready to go stealing work from others. 4642 // We'd like to assert(work_q->size() != 0, ...) 4643 // because we just took work from the overflow list, 4644 // but of course we can't since all of that could have 4645 // been already stolen from us. 4646 // "He giveth and He taketh away." 4647 continue; 4648 } 4649 // Verify that we have no work before we resort to stealing 4650 assert(work_q->size() == 0, "Have work, shouldn't steal"); 4651 // Try to steal from other queues that have work 4652 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 4653 NOT_PRODUCT(num_steals++;) 4654 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 4655 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 4656 // Do scanning work 4657 obj_to_scan->oop_iterate(cl); 4658 // Loop around, finish this work, and try to steal some more 4659 } else if (terminator()->offer_termination()) { 4660 break; // nirvana from the infinite cycle 4661 } 4662 } 4663 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 4664 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), 4665 "Else our work is not yet done"); 4666 } 4667 4668 // Record object boundaries in _eden_chunk_array by sampling the eden 4669 // top in the slow-path eden object allocation code path and record 4670 // the boundaries, if CMSEdenChunksRecordAlways is true. If 4671 // CMSEdenChunksRecordAlways is false, we use the other asynchronous 4672 // sampling in sample_eden() that activates during the part of the 4673 // preclean phase. 4674 void CMSCollector::sample_eden_chunk() { 4675 if (CMSEdenChunksRecordAlways && _eden_chunk_array != NULL) { 4676 if (_eden_chunk_lock->try_lock()) { 4677 // Record a sample. This is the critical section. The contents 4678 // of the _eden_chunk_array have to be non-decreasing in the 4679 // address order. 4680 _eden_chunk_array[_eden_chunk_index] = *_top_addr; 4681 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, 4682 "Unexpected state of Eden"); 4683 if (_eden_chunk_index == 0 || 4684 ((_eden_chunk_array[_eden_chunk_index] > _eden_chunk_array[_eden_chunk_index-1]) && 4685 (pointer_delta(_eden_chunk_array[_eden_chunk_index], 4686 _eden_chunk_array[_eden_chunk_index-1]) >= CMSSamplingGrain))) { 4687 _eden_chunk_index++; // commit sample 4688 } 4689 _eden_chunk_lock->unlock(); 4690 } 4691 } 4692 } 4693 4694 // Return a thread-local PLAB recording array, as appropriate. 4695 void* CMSCollector::get_data_recorder(int thr_num) { 4696 if (_survivor_plab_array != NULL && 4697 (CMSPLABRecordAlways || 4698 (_collectorState > Marking && _collectorState < FinalMarking))) { 4699 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); 4700 ChunkArray* ca = &_survivor_plab_array[thr_num]; 4701 ca->reset(); // clear it so that fresh data is recorded 4702 return (void*) ca; 4703 } else { 4704 return NULL; 4705 } 4706 } 4707 4708 // Reset all the thread-local PLAB recording arrays 4709 void CMSCollector::reset_survivor_plab_arrays() { 4710 for (uint i = 0; i < ParallelGCThreads; i++) { 4711 _survivor_plab_array[i].reset(); 4712 } 4713 } 4714 4715 // Merge the per-thread plab arrays into the global survivor chunk 4716 // array which will provide the partitioning of the survivor space 4717 // for CMS initial scan and rescan. 4718 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, 4719 int no_of_gc_threads) { 4720 assert(_survivor_plab_array != NULL, "Error"); 4721 assert(_survivor_chunk_array != NULL, "Error"); 4722 assert(_collectorState == FinalMarking || 4723 (CMSParallelInitialMarkEnabled && _collectorState == InitialMarking), "Error"); 4724 for (int j = 0; j < no_of_gc_threads; j++) { 4725 _cursor[j] = 0; 4726 } 4727 HeapWord* top = surv->top(); 4728 size_t i; 4729 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries 4730 HeapWord* min_val = top; // Higher than any PLAB address 4731 uint min_tid = 0; // position of min_val this round 4732 for (int j = 0; j < no_of_gc_threads; j++) { 4733 ChunkArray* cur_sca = &_survivor_plab_array[j]; 4734 if (_cursor[j] == cur_sca->end()) { 4735 continue; 4736 } 4737 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); 4738 HeapWord* cur_val = cur_sca->nth(_cursor[j]); 4739 assert(surv->used_region().contains(cur_val), "Out of bounds value"); 4740 if (cur_val < min_val) { 4741 min_tid = j; 4742 min_val = cur_val; 4743 } else { 4744 assert(cur_val < top, "All recorded addresses should be less"); 4745 } 4746 } 4747 // At this point min_val and min_tid are respectively 4748 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) 4749 // and the thread (j) that witnesses that address. 4750 // We record this address in the _survivor_chunk_array[i] 4751 // and increment _cursor[min_tid] prior to the next round i. 4752 if (min_val == top) { 4753 break; 4754 } 4755 _survivor_chunk_array[i] = min_val; 4756 _cursor[min_tid]++; 4757 } 4758 // We are all done; record the size of the _survivor_chunk_array 4759 _survivor_chunk_index = i; // exclusive: [0, i) 4760 log_trace(gc, survivor)(" (Survivor:" SIZE_FORMAT "chunks) ", i); 4761 // Verify that we used up all the recorded entries 4762 #ifdef ASSERT 4763 size_t total = 0; 4764 for (int j = 0; j < no_of_gc_threads; j++) { 4765 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); 4766 total += _cursor[j]; 4767 } 4768 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); 4769 // Check that the merged array is in sorted order 4770 if (total > 0) { 4771 for (size_t i = 0; i < total - 1; i++) { 4772 log_develop_trace(gc, survivor)(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", 4773 i, p2i(_survivor_chunk_array[i])); 4774 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], 4775 "Not sorted"); 4776 } 4777 } 4778 #endif // ASSERT 4779 } 4780 4781 // Set up the space's par_seq_tasks structure for work claiming 4782 // for parallel initial scan and rescan of young gen. 4783 // See ParRescanTask where this is currently used. 4784 void 4785 CMSCollector:: 4786 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { 4787 assert(n_threads > 0, "Unexpected n_threads argument"); 4788 4789 // Eden space 4790 if (!_young_gen->eden()->is_empty()) { 4791 SequentialSubTasksDone* pst = _young_gen->eden()->par_seq_tasks(); 4792 assert(!pst->valid(), "Clobbering existing data?"); 4793 // Each valid entry in [0, _eden_chunk_index) represents a task. 4794 size_t n_tasks = _eden_chunk_index + 1; 4795 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); 4796 // Sets the condition for completion of the subtask (how many threads 4797 // need to finish in order to be done). 4798 pst->set_n_threads(n_threads); 4799 pst->set_n_tasks((int)n_tasks); 4800 } 4801 4802 // Merge the survivor plab arrays into _survivor_chunk_array 4803 if (_survivor_plab_array != NULL) { 4804 merge_survivor_plab_arrays(_young_gen->from(), n_threads); 4805 } else { 4806 assert(_survivor_chunk_index == 0, "Error"); 4807 } 4808 4809 // To space 4810 { 4811 SequentialSubTasksDone* pst = _young_gen->to()->par_seq_tasks(); 4812 assert(!pst->valid(), "Clobbering existing data?"); 4813 // Sets the condition for completion of the subtask (how many threads 4814 // need to finish in order to be done). 4815 pst->set_n_threads(n_threads); 4816 pst->set_n_tasks(1); 4817 assert(pst->valid(), "Error"); 4818 } 4819 4820 // From space 4821 { 4822 SequentialSubTasksDone* pst = _young_gen->from()->par_seq_tasks(); 4823 assert(!pst->valid(), "Clobbering existing data?"); 4824 size_t n_tasks = _survivor_chunk_index + 1; 4825 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); 4826 // Sets the condition for completion of the subtask (how many threads 4827 // need to finish in order to be done). 4828 pst->set_n_threads(n_threads); 4829 pst->set_n_tasks((int)n_tasks); 4830 assert(pst->valid(), "Error"); 4831 } 4832 } 4833 4834 // Parallel version of remark 4835 void CMSCollector::do_remark_parallel() { 4836 CMSHeap* heap = CMSHeap::heap(); 4837 WorkGang* workers = heap->workers(); 4838 assert(workers != NULL, "Need parallel worker threads."); 4839 // Choose to use the number of GC workers most recently set 4840 // into "active_workers". 4841 uint n_workers = workers->active_workers(); 4842 4843 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); 4844 4845 StrongRootsScope srs(n_workers); 4846 4847 CMSParRemarkTask tsk(this, cms_space, n_workers, workers, task_queues(), &srs); 4848 4849 // We won't be iterating over the cards in the card table updating 4850 // the younger_gen cards, so we shouldn't call the following else 4851 // the verification code as well as subsequent younger_refs_iterate 4852 // code would get confused. XXX 4853 // heap->rem_set()->prepare_for_younger_refs_iterate(true); // parallel 4854 4855 // The young gen rescan work will not be done as part of 4856 // process_roots (which currently doesn't know how to 4857 // parallelize such a scan), but rather will be broken up into 4858 // a set of parallel tasks (via the sampling that the [abortable] 4859 // preclean phase did of eden, plus the [two] tasks of 4860 // scanning the [two] survivor spaces. Further fine-grain 4861 // parallelization of the scanning of the survivor spaces 4862 // themselves, and of precleaning of the young gen itself 4863 // is deferred to the future. 4864 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); 4865 4866 // The dirty card rescan work is broken up into a "sequence" 4867 // of parallel tasks (per constituent space) that are dynamically 4868 // claimed by the parallel threads. 4869 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); 4870 4871 // It turns out that even when we're using 1 thread, doing the work in a 4872 // separate thread causes wide variance in run times. We can't help this 4873 // in the multi-threaded case, but we special-case n=1 here to get 4874 // repeatable measurements of the 1-thread overhead of the parallel code. 4875 if (n_workers > 1) { 4876 // Make refs discovery MT-safe, if it isn't already: it may not 4877 // necessarily be so, since it's possible that we are doing 4878 // ST marking. 4879 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), true); 4880 workers->run_task(&tsk); 4881 } else { 4882 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4883 tsk.work(0); 4884 } 4885 4886 // restore, single-threaded for now, any preserved marks 4887 // as a result of work_q overflow 4888 restore_preserved_marks_if_any(); 4889 } 4890 4891 // Non-parallel version of remark 4892 void CMSCollector::do_remark_non_parallel() { 4893 ResourceMark rm; 4894 HandleMark hm; 4895 CMSHeap* heap = CMSHeap::heap(); 4896 ReferenceProcessorMTDiscoveryMutator mt(ref_processor(), false); 4897 4898 MarkRefsIntoAndScanClosure 4899 mrias_cl(_span, ref_processor(), &_markBitMap, NULL /* not precleaning */, 4900 &_markStack, this, 4901 false /* should_yield */, false /* not precleaning */); 4902 MarkFromDirtyCardsClosure 4903 markFromDirtyCardsClosure(this, _span, 4904 NULL, // space is set further below 4905 &_markBitMap, &_markStack, &mrias_cl); 4906 { 4907 GCTraceTime(Trace, gc, phases) t("Grey Object Rescan", _gc_timer_cm); 4908 // Iterate over the dirty cards, setting the corresponding bits in the 4909 // mod union table. 4910 { 4911 ModUnionClosure modUnionClosure(&_modUnionTable); 4912 _ct->dirty_card_iterate(_cmsGen->used_region(), 4913 &modUnionClosure); 4914 } 4915 // Having transferred these marks into the modUnionTable, we just need 4916 // to rescan the marked objects on the dirty cards in the modUnionTable. 4917 // The initial marking may have been done during an asynchronous 4918 // collection so there may be dirty bits in the mod-union table. 4919 const int alignment = CardTable::card_size * BitsPerWord; 4920 { 4921 // ... First handle dirty cards in CMS gen 4922 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); 4923 MemRegion ur = _cmsGen->used_region(); 4924 HeapWord* lb = ur.start(); 4925 HeapWord* ub = align_up(ur.end(), alignment); 4926 MemRegion cms_span(lb, ub); 4927 _modUnionTable.dirty_range_iterate_clear(cms_span, 4928 &markFromDirtyCardsClosure); 4929 verify_work_stacks_empty(); 4930 log_trace(gc)(" (re-scanned " SIZE_FORMAT " dirty cards in cms gen) ", markFromDirtyCardsClosure.num_dirty_cards()); 4931 } 4932 } 4933 if (VerifyDuringGC && 4934 CMSHeap::heap()->total_collections() >= VerifyGCStartAt) { 4935 HandleMark hm; // Discard invalid handles created during verification 4936 Universe::verify(); 4937 } 4938 { 4939 GCTraceTime(Trace, gc, phases) t("Root Rescan", _gc_timer_cm); 4940 4941 verify_work_stacks_empty(); 4942 4943 heap->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. 4944 StrongRootsScope srs(1); 4945 4946 heap->cms_process_roots(&srs, 4947 true, // young gen as roots 4948 GenCollectedHeap::ScanningOption(roots_scanning_options()), 4949 should_unload_classes(), 4950 &mrias_cl, 4951 NULL); // The dirty klasses will be handled below 4952 4953 assert(should_unload_classes() 4954 || (roots_scanning_options() & GenCollectedHeap::SO_AllCodeCache), 4955 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); 4956 } 4957 4958 { 4959 GCTraceTime(Trace, gc, phases) t("Visit Unhandled CLDs", _gc_timer_cm); 4960 4961 verify_work_stacks_empty(); 4962 4963 // Scan all class loader data objects that might have been introduced 4964 // during concurrent marking. 4965 ResourceMark rm; 4966 GrowableArray<ClassLoaderData*>* array = ClassLoaderDataGraph::new_clds(); 4967 for (int i = 0; i < array->length(); i++) { 4968 Devirtualizer::do_cld(&mrias_cl, array->at(i)); 4969 } 4970 4971 // We don't need to keep track of new CLDs anymore. 4972 ClassLoaderDataGraph::remember_new_clds(false); 4973 4974 verify_work_stacks_empty(); 4975 } 4976 4977 // We might have added oops to ClassLoaderData::_handles during the 4978 // concurrent marking phase. These oops do not point to newly allocated objects 4979 // that are guaranteed to be kept alive. Hence, 4980 // we do have to revisit the _handles block during the remark phase. 4981 { 4982 GCTraceTime(Trace, gc, phases) t("Dirty CLD Scan", _gc_timer_cm); 4983 4984 verify_work_stacks_empty(); 4985 4986 RemarkCLDClosure remark_closure(&mrias_cl); 4987 ClassLoaderDataGraph::cld_do(&remark_closure); 4988 4989 verify_work_stacks_empty(); 4990 } 4991 4992 verify_work_stacks_empty(); 4993 // Restore evacuated mark words, if any, used for overflow list links 4994 restore_preserved_marks_if_any(); 4995 4996 verify_overflow_empty(); 4997 } 4998 4999 //////////////////////////////////////////////////////// 5000 // Parallel Reference Processing Task Proxy Class 5001 //////////////////////////////////////////////////////// 5002 class AbstractGangTaskWOopQueues : public AbstractGangTask { 5003 OopTaskQueueSet* _queues; 5004 ParallelTaskTerminator _terminator; 5005 public: 5006 AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues, uint n_threads) : 5007 AbstractGangTask(name), _queues(queues), _terminator(n_threads, _queues) {} 5008 ParallelTaskTerminator* terminator() { return &_terminator; } 5009 OopTaskQueueSet* queues() { return _queues; } 5010 }; 5011 5012 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { 5013 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5014 CMSCollector* _collector; 5015 CMSBitMap* _mark_bit_map; 5016 const MemRegion _span; 5017 ProcessTask& _task; 5018 5019 public: 5020 CMSRefProcTaskProxy(ProcessTask& task, 5021 CMSCollector* collector, 5022 const MemRegion& span, 5023 CMSBitMap* mark_bit_map, 5024 AbstractWorkGang* workers, 5025 OopTaskQueueSet* task_queues): 5026 AbstractGangTaskWOopQueues("Process referents by policy in parallel", 5027 task_queues, 5028 workers->active_workers()), 5029 _collector(collector), 5030 _mark_bit_map(mark_bit_map), 5031 _span(span), 5032 _task(task) 5033 { 5034 assert(_collector->_span.equals(_span) && !_span.is_empty(), 5035 "Inconsistency in _span"); 5036 } 5037 5038 OopTaskQueueSet* task_queues() { return queues(); } 5039 5040 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } 5041 5042 void do_work_steal(int i, 5043 CMSParDrainMarkingStackClosure* drain, 5044 CMSParKeepAliveClosure* keep_alive); 5045 5046 virtual void work(uint worker_id); 5047 }; 5048 5049 void CMSRefProcTaskProxy::work(uint worker_id) { 5050 ResourceMark rm; 5051 HandleMark hm; 5052 assert(_collector->_span.equals(_span), "Inconsistency in _span"); 5053 CMSParKeepAliveClosure par_keep_alive(_collector, _span, 5054 _mark_bit_map, 5055 work_queue(worker_id)); 5056 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, 5057 _mark_bit_map, 5058 work_queue(worker_id)); 5059 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); 5060 _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack); 5061 if (_task.marks_oops_alive()) { 5062 do_work_steal(worker_id, &par_drain_stack, &par_keep_alive); 5063 } 5064 assert(work_queue(worker_id)->size() == 0, "work_queue should be empty"); 5065 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); 5066 } 5067 5068 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, 5069 MemRegion span, CMSBitMap* bit_map, OopTaskQueue* work_queue): 5070 _span(span), 5071 _work_queue(work_queue), 5072 _bit_map(bit_map), 5073 _mark_and_push(collector, span, bit_map, work_queue), 5074 _low_water_mark(MIN2((work_queue->max_elems()/4), 5075 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))) 5076 { } 5077 5078 // . see if we can share work_queues with ParNew? XXX 5079 void CMSRefProcTaskProxy::do_work_steal(int i, 5080 CMSParDrainMarkingStackClosure* drain, 5081 CMSParKeepAliveClosure* keep_alive) { 5082 OopTaskQueue* work_q = work_queue(i); 5083 NOT_PRODUCT(int num_steals = 0;) 5084 oop obj_to_scan; 5085 5086 while (true) { 5087 // Completely finish any left over work from (an) earlier round(s) 5088 drain->trim_queue(0); 5089 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, 5090 (size_t)ParGCDesiredObjsFromOverflowList); 5091 // Now check if there's any work in the overflow list 5092 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, 5093 // only affects the number of attempts made to get work from the 5094 // overflow list and does not affect the number of workers. Just 5095 // pass ParallelGCThreads so this behavior is unchanged. 5096 if (_collector->par_take_from_overflow_list(num_from_overflow_list, 5097 work_q, 5098 ParallelGCThreads)) { 5099 // Found something in global overflow list; 5100 // not yet ready to go stealing work from others. 5101 // We'd like to assert(work_q->size() != 0, ...) 5102 // because we just took work from the overflow list, 5103 // but of course we can't, since all of that might have 5104 // been already stolen from us. 5105 continue; 5106 } 5107 // Verify that we have no work before we resort to stealing 5108 assert(work_q->size() == 0, "Have work, shouldn't steal"); 5109 // Try to steal from other queues that have work 5110 if (task_queues()->steal(i, /* reference */ obj_to_scan)) { 5111 NOT_PRODUCT(num_steals++;) 5112 assert(oopDesc::is_oop(obj_to_scan), "Oops, not an oop!"); 5113 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); 5114 // Do scanning work 5115 obj_to_scan->oop_iterate(keep_alive); 5116 // Loop around, finish this work, and try to steal some more 5117 } else if (terminator()->offer_termination()) { 5118 break; // nirvana from the infinite cycle 5119 } 5120 } 5121 log_develop_trace(gc, task)("\t(%d: stole %d oops)", i, num_steals); 5122 } 5123 5124 void CMSRefProcTaskExecutor::execute(ProcessTask& task, uint ergo_workers) { 5125 CMSHeap* heap = CMSHeap::heap(); 5126 WorkGang* workers = heap->workers(); 5127 assert(workers != NULL, "Need parallel worker threads."); 5128 assert(workers->active_workers() == ergo_workers, 5129 "Ergonomically chosen workers (%u) must be equal to active workers (%u)", 5130 ergo_workers, workers->active_workers()); 5131 CMSRefProcTaskProxy rp_task(task, &_collector, 5132 _collector.ref_processor_span(), 5133 _collector.markBitMap(), 5134 workers, _collector.task_queues()); 5135 workers->run_task(&rp_task, workers->active_workers()); 5136 } 5137 5138 void CMSCollector::refProcessingWork() { 5139 ResourceMark rm; 5140 HandleMark hm; 5141 5142 ReferenceProcessor* rp = ref_processor(); 5143 assert(_span_based_discoverer.span().equals(_span), "Spans should be equal"); 5144 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); 5145 // Process weak references. 5146 rp->setup_policy(false); 5147 verify_work_stacks_empty(); 5148 5149 ReferenceProcessorPhaseTimes pt(_gc_timer_cm, rp->max_num_queues()); 5150 { 5151 GCTraceTime(Debug, gc, phases) t("Reference Processing", _gc_timer_cm); 5152 5153 // Setup keep_alive and complete closures. 5154 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, 5155 &_markStack, false /* !preclean */); 5156 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, 5157 _span, &_markBitMap, &_markStack, 5158 &cmsKeepAliveClosure, false /* !preclean */); 5159 5160 ReferenceProcessorStats stats; 5161 if (rp->processing_is_mt()) { 5162 // Set the degree of MT here. If the discovery is done MT, there 5163 // may have been a different number of threads doing the discovery 5164 // and a different number of discovered lists may have Ref objects. 5165 // That is OK as long as the Reference lists are balanced (see 5166 // balance_all_queues() and balance_queues()). 5167 CMSHeap* heap = CMSHeap::heap(); 5168 uint active_workers = ParallelGCThreads; 5169 WorkGang* workers = heap->workers(); 5170 if (workers != NULL) { 5171 active_workers = workers->active_workers(); 5172 // The expectation is that active_workers will have already 5173 // been set to a reasonable value. If it has not been set, 5174 // investigate. 5175 assert(active_workers > 0, "Should have been set during scavenge"); 5176 } 5177 rp->set_active_mt_degree(active_workers); 5178 CMSRefProcTaskExecutor task_executor(*this); 5179 stats = rp->process_discovered_references(&_is_alive_closure, 5180 &cmsKeepAliveClosure, 5181 &cmsDrainMarkingStackClosure, 5182 &task_executor, 5183 &pt); 5184 } else { 5185 stats = rp->process_discovered_references(&_is_alive_closure, 5186 &cmsKeepAliveClosure, 5187 &cmsDrainMarkingStackClosure, 5188 NULL, 5189 &pt); 5190 } 5191 _gc_tracer_cm->report_gc_reference_stats(stats); 5192 pt.print_all_references(); 5193 } 5194 5195 // This is the point where the entire marking should have completed. 5196 verify_work_stacks_empty(); 5197 5198 { 5199 GCTraceTime(Debug, gc, phases) t("Weak Processing", _gc_timer_cm); 5200 WeakProcessor::weak_oops_do(&_is_alive_closure, &do_nothing_cl); 5201 } 5202 5203 if (should_unload_classes()) { 5204 { 5205 GCTraceTime(Debug, gc, phases) t("Class Unloading", _gc_timer_cm); 5206 5207 // Unload classes and purge the SystemDictionary. 5208 bool purged_class = SystemDictionary::do_unloading(_gc_timer_cm); 5209 5210 // Unload nmethods. 5211 CodeCache::do_unloading(&_is_alive_closure, purged_class); 5212 5213 // Prune dead klasses from subklass/sibling/implementor lists. 5214 Klass::clean_weak_klass_links(purged_class); 5215 } 5216 5217 { 5218 GCTraceTime(Debug, gc, phases) t("Scrub Symbol Table", _gc_timer_cm); 5219 // Clean up unreferenced symbols in symbol table. 5220 SymbolTable::unlink(); 5221 } 5222 5223 { 5224 GCTraceTime(Debug, gc, phases) t("Scrub String Table", _gc_timer_cm); 5225 // Delete entries for dead interned strings. 5226 StringTable::unlink(&_is_alive_closure); 5227 } 5228 } 5229 5230 // Restore any preserved marks as a result of mark stack or 5231 // work queue overflow 5232 restore_preserved_marks_if_any(); // done single-threaded for now 5233 5234 rp->set_enqueuing_is_done(true); 5235 rp->verify_no_references_recorded(); 5236 } 5237 5238 #ifndef PRODUCT 5239 void CMSCollector::check_correct_thread_executing() { 5240 Thread* t = Thread::current(); 5241 // Only the VM thread or the CMS thread should be here. 5242 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), 5243 "Unexpected thread type"); 5244 // If this is the vm thread, the foreground process 5245 // should not be waiting. Note that _foregroundGCIsActive is 5246 // true while the foreground collector is waiting. 5247 if (_foregroundGCShouldWait) { 5248 // We cannot be the VM thread 5249 assert(t->is_ConcurrentGC_thread(), 5250 "Should be CMS thread"); 5251 } else { 5252 // We can be the CMS thread only if we are in a stop-world 5253 // phase of CMS collection. 5254 if (t->is_ConcurrentGC_thread()) { 5255 assert(_collectorState == InitialMarking || 5256 _collectorState == FinalMarking, 5257 "Should be a stop-world phase"); 5258 // The CMS thread should be holding the CMS_token. 5259 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5260 "Potential interference with concurrently " 5261 "executing VM thread"); 5262 } 5263 } 5264 } 5265 #endif 5266 5267 void CMSCollector::sweep() { 5268 assert(_collectorState == Sweeping, "just checking"); 5269 check_correct_thread_executing(); 5270 verify_work_stacks_empty(); 5271 verify_overflow_empty(); 5272 increment_sweep_count(); 5273 TraceCMSMemoryManagerStats tms(_collectorState, CMSHeap::heap()->gc_cause()); 5274 5275 _inter_sweep_timer.stop(); 5276 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); 5277 5278 assert(!_intra_sweep_timer.is_active(), "Should not be active"); 5279 _intra_sweep_timer.reset(); 5280 _intra_sweep_timer.start(); 5281 { 5282 GCTraceCPUTime tcpu; 5283 CMSPhaseAccounting pa(this, "Concurrent Sweep"); 5284 // First sweep the old gen 5285 { 5286 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), 5287 bitMapLock()); 5288 sweepWork(_cmsGen); 5289 } 5290 5291 // Update Universe::_heap_*_at_gc figures. 5292 // We need all the free list locks to make the abstract state 5293 // transition from Sweeping to Resetting. See detailed note 5294 // further below. 5295 { 5296 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock()); 5297 // Update heap occupancy information which is used as 5298 // input to soft ref clearing policy at the next gc. 5299 Universe::update_heap_info_at_gc(); 5300 _collectorState = Resizing; 5301 } 5302 } 5303 verify_work_stacks_empty(); 5304 verify_overflow_empty(); 5305 5306 if (should_unload_classes()) { 5307 // Delay purge to the beginning of the next safepoint. Metaspace::contains 5308 // requires that the virtual spaces are stable and not deleted. 5309 ClassLoaderDataGraph::set_should_purge(true); 5310 } 5311 5312 _intra_sweep_timer.stop(); 5313 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); 5314 5315 _inter_sweep_timer.reset(); 5316 _inter_sweep_timer.start(); 5317 5318 // We need to use a monotonically non-decreasing time in ms 5319 // or we will see time-warp warnings and os::javaTimeMillis() 5320 // does not guarantee monotonicity. 5321 jlong now = os::javaTimeNanos() / NANOSECS_PER_MILLISEC; 5322 update_time_of_last_gc(now); 5323 5324 // NOTE on abstract state transitions: 5325 // Mutators allocate-live and/or mark the mod-union table dirty 5326 // based on the state of the collection. The former is done in 5327 // the interval [Marking, Sweeping] and the latter in the interval 5328 // [Marking, Sweeping). Thus the transitions into the Marking state 5329 // and out of the Sweeping state must be synchronously visible 5330 // globally to the mutators. 5331 // The transition into the Marking state happens with the world 5332 // stopped so the mutators will globally see it. Sweeping is 5333 // done asynchronously by the background collector so the transition 5334 // from the Sweeping state to the Resizing state must be done 5335 // under the freelistLock (as is the check for whether to 5336 // allocate-live and whether to dirty the mod-union table). 5337 assert(_collectorState == Resizing, "Change of collector state to" 5338 " Resizing must be done under the freelistLocks (plural)"); 5339 5340 // Now that sweeping has been completed, we clear 5341 // the incremental_collection_failed flag, 5342 // thus inviting a younger gen collection to promote into 5343 // this generation. If such a promotion may still fail, 5344 // the flag will be set again when a young collection is 5345 // attempted. 5346 CMSHeap* heap = CMSHeap::heap(); 5347 heap->clear_incremental_collection_failed(); // Worth retrying as fresh space may have been freed up 5348 heap->update_full_collections_completed(_collection_count_start); 5349 } 5350 5351 // FIX ME!!! Looks like this belongs in CFLSpace, with 5352 // CMSGen merely delegating to it. 5353 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { 5354 double nearLargestPercent = FLSLargestBlockCoalesceProximity; 5355 HeapWord* minAddr = _cmsSpace->bottom(); 5356 HeapWord* largestAddr = 5357 (HeapWord*) _cmsSpace->dictionary()->find_largest_dict(); 5358 if (largestAddr == NULL) { 5359 // The dictionary appears to be empty. In this case 5360 // try to coalesce at the end of the heap. 5361 largestAddr = _cmsSpace->end(); 5362 } 5363 size_t largestOffset = pointer_delta(largestAddr, minAddr); 5364 size_t nearLargestOffset = 5365 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; 5366 log_debug(gc, freelist)("CMS: Large Block: " PTR_FORMAT "; Proximity: " PTR_FORMAT " -> " PTR_FORMAT, 5367 p2i(largestAddr), p2i(_cmsSpace->nearLargestChunk()), p2i(minAddr + nearLargestOffset)); 5368 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); 5369 } 5370 5371 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { 5372 return addr >= _cmsSpace->nearLargestChunk(); 5373 } 5374 5375 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { 5376 return _cmsSpace->find_chunk_at_end(); 5377 } 5378 5379 void ConcurrentMarkSweepGeneration::update_gc_stats(Generation* current_generation, 5380 bool full) { 5381 // If the young generation has been collected, gather any statistics 5382 // that are of interest at this point. 5383 bool current_is_young = CMSHeap::heap()->is_young_gen(current_generation); 5384 if (!full && current_is_young) { 5385 // Gather statistics on the young generation collection. 5386 collector()->stats().record_gc0_end(used()); 5387 } 5388 } 5389 5390 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* old_gen) { 5391 // We iterate over the space(s) underlying this generation, 5392 // checking the mark bit map to see if the bits corresponding 5393 // to specific blocks are marked or not. Blocks that are 5394 // marked are live and are not swept up. All remaining blocks 5395 // are swept up, with coalescing on-the-fly as we sweep up 5396 // contiguous free and/or garbage blocks: 5397 // We need to ensure that the sweeper synchronizes with allocators 5398 // and stop-the-world collectors. In particular, the following 5399 // locks are used: 5400 // . CMS token: if this is held, a stop the world collection cannot occur 5401 // . freelistLock: if this is held no allocation can occur from this 5402 // generation by another thread 5403 // . bitMapLock: if this is held, no other thread can access or update 5404 // 5405 5406 // Note that we need to hold the freelistLock if we use 5407 // block iterate below; else the iterator might go awry if 5408 // a mutator (or promotion) causes block contents to change 5409 // (for instance if the allocator divvies up a block). 5410 // If we hold the free list lock, for all practical purposes 5411 // young generation GC's can't occur (they'll usually need to 5412 // promote), so we might as well prevent all young generation 5413 // GC's while we do a sweeping step. For the same reason, we might 5414 // as well take the bit map lock for the entire duration 5415 5416 // check that we hold the requisite locks 5417 assert(have_cms_token(), "Should hold cms token"); 5418 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), "Should possess CMS token to sweep"); 5419 assert_lock_strong(old_gen->freelistLock()); 5420 assert_lock_strong(bitMapLock()); 5421 5422 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); 5423 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); 5424 old_gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), 5425 _inter_sweep_estimate.padded_average(), 5426 _intra_sweep_estimate.padded_average()); 5427 old_gen->setNearLargestChunk(); 5428 5429 { 5430 SweepClosure sweepClosure(this, old_gen, &_markBitMap, CMSYield); 5431 old_gen->cmsSpace()->blk_iterate_careful(&sweepClosure); 5432 // We need to free-up/coalesce garbage/blocks from a 5433 // co-terminal free run. This is done in the SweepClosure 5434 // destructor; so, do not remove this scope, else the 5435 // end-of-sweep-census below will be off by a little bit. 5436 } 5437 old_gen->cmsSpace()->sweep_completed(); 5438 old_gen->cmsSpace()->endSweepFLCensus(sweep_count()); 5439 if (should_unload_classes()) { // unloaded classes this cycle, 5440 _concurrent_cycles_since_last_unload = 0; // ... reset count 5441 } else { // did not unload classes, 5442 _concurrent_cycles_since_last_unload++; // ... increment count 5443 } 5444 } 5445 5446 // Reset CMS data structures (for now just the marking bit map) 5447 // preparatory for the next cycle. 5448 void CMSCollector::reset_concurrent() { 5449 CMSTokenSyncWithLocks ts(true, bitMapLock()); 5450 5451 // If the state is not "Resetting", the foreground thread 5452 // has done a collection and the resetting. 5453 if (_collectorState != Resetting) { 5454 assert(_collectorState == Idling, "The state should only change" 5455 " because the foreground collector has finished the collection"); 5456 return; 5457 } 5458 5459 { 5460 // Clear the mark bitmap (no grey objects to start with) 5461 // for the next cycle. 5462 GCTraceCPUTime tcpu; 5463 CMSPhaseAccounting cmspa(this, "Concurrent Reset"); 5464 5465 HeapWord* curAddr = _markBitMap.startWord(); 5466 while (curAddr < _markBitMap.endWord()) { 5467 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); 5468 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); 5469 _markBitMap.clear_large_range(chunk); 5470 if (ConcurrentMarkSweepThread::should_yield() && 5471 !foregroundGCIsActive() && 5472 CMSYield) { 5473 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5474 "CMS thread should hold CMS token"); 5475 assert_lock_strong(bitMapLock()); 5476 bitMapLock()->unlock(); 5477 ConcurrentMarkSweepThread::desynchronize(true); 5478 stopTimer(); 5479 incrementYields(); 5480 5481 // See the comment in coordinator_yield() 5482 for (unsigned i = 0; i < CMSYieldSleepCount && 5483 ConcurrentMarkSweepThread::should_yield() && 5484 !CMSCollector::foregroundGCIsActive(); ++i) { 5485 os::sleep(Thread::current(), 1, false); 5486 } 5487 5488 ConcurrentMarkSweepThread::synchronize(true); 5489 bitMapLock()->lock_without_safepoint_check(); 5490 startTimer(); 5491 } 5492 curAddr = chunk.end(); 5493 } 5494 // A successful mostly concurrent collection has been done. 5495 // Because only the full (i.e., concurrent mode failure) collections 5496 // are being measured for gc overhead limits, clean the "near" flag 5497 // and count. 5498 size_policy()->reset_gc_overhead_limit_count(); 5499 _collectorState = Idling; 5500 } 5501 5502 register_gc_end(); 5503 } 5504 5505 // Same as above but for STW paths 5506 void CMSCollector::reset_stw() { 5507 // already have the lock 5508 assert(_collectorState == Resetting, "just checking"); 5509 assert_lock_strong(bitMapLock()); 5510 GCIdMark gc_id_mark(_cmsThread->gc_id()); 5511 _markBitMap.clear_all(); 5512 _collectorState = Idling; 5513 register_gc_end(); 5514 } 5515 5516 void CMSCollector::do_CMS_operation(CMS_op_type op, GCCause::Cause gc_cause) { 5517 GCTraceCPUTime tcpu; 5518 TraceCollectorStats tcs_cgc(cgc_counters()); 5519 5520 switch (op) { 5521 case CMS_op_checkpointRootsInitial: { 5522 GCTraceTime(Info, gc) t("Pause Initial Mark", NULL, GCCause::_no_gc, true); 5523 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5524 checkpointRootsInitial(); 5525 break; 5526 } 5527 case CMS_op_checkpointRootsFinal: { 5528 GCTraceTime(Info, gc) t("Pause Remark", NULL, GCCause::_no_gc, true); 5529 SvcGCMarker sgcm(SvcGCMarker::CONCURRENT); 5530 checkpointRootsFinal(); 5531 break; 5532 } 5533 default: 5534 fatal("No such CMS_op"); 5535 } 5536 } 5537 5538 #ifndef PRODUCT 5539 size_t const CMSCollector::skip_header_HeapWords() { 5540 return FreeChunk::header_size(); 5541 } 5542 5543 // Try and collect here conditions that should hold when 5544 // CMS thread is exiting. The idea is that the foreground GC 5545 // thread should not be blocked if it wants to terminate 5546 // the CMS thread and yet continue to run the VM for a while 5547 // after that. 5548 void CMSCollector::verify_ok_to_terminate() const { 5549 assert(Thread::current()->is_ConcurrentGC_thread(), 5550 "should be called by CMS thread"); 5551 assert(!_foregroundGCShouldWait, "should be false"); 5552 // We could check here that all the various low-level locks 5553 // are not held by the CMS thread, but that is overkill; see 5554 // also CMSThread::verify_ok_to_terminate() where the CGC_lock 5555 // is checked. 5556 } 5557 #endif 5558 5559 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { 5560 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), 5561 "missing Printezis mark?"); 5562 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5563 size_t size = pointer_delta(nextOneAddr + 1, addr); 5564 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5565 "alignment problem"); 5566 assert(size >= 3, "Necessary for Printezis marks to work"); 5567 return size; 5568 } 5569 5570 // A variant of the above (block_size_using_printezis_bits()) except 5571 // that we return 0 if the P-bits are not yet set. 5572 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { 5573 if (_markBitMap.isMarked(addr + 1)) { 5574 assert(_markBitMap.isMarked(addr), "P-bit can be set only for marked objects"); 5575 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); 5576 size_t size = pointer_delta(nextOneAddr + 1, addr); 5577 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 5578 "alignment problem"); 5579 assert(size >= 3, "Necessary for Printezis marks to work"); 5580 return size; 5581 } 5582 return 0; 5583 } 5584 5585 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { 5586 size_t sz = 0; 5587 oop p = (oop)addr; 5588 if (p->klass_or_null_acquire() != NULL) { 5589 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); 5590 } else { 5591 sz = block_size_using_printezis_bits(addr); 5592 } 5593 assert(sz > 0, "size must be nonzero"); 5594 HeapWord* next_block = addr + sz; 5595 HeapWord* next_card = align_up(next_block, CardTable::card_size); 5596 assert(align_down((uintptr_t)addr, CardTable::card_size) < 5597 align_down((uintptr_t)next_card, CardTable::card_size), 5598 "must be different cards"); 5599 return next_card; 5600 } 5601 5602 5603 // CMS Bit Map Wrapper ///////////////////////////////////////// 5604 5605 // Construct a CMS bit map infrastructure, but don't create the 5606 // bit vector itself. That is done by a separate call CMSBitMap::allocate() 5607 // further below. 5608 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): 5609 _shifter(shifter), 5610 _bm(), 5611 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true, 5612 Monitor::_safepoint_check_sometimes) : NULL) 5613 { 5614 _bmStartWord = 0; 5615 _bmWordSize = 0; 5616 } 5617 5618 bool CMSBitMap::allocate(MemRegion mr) { 5619 _bmStartWord = mr.start(); 5620 _bmWordSize = mr.word_size(); 5621 ReservedSpace brs(ReservedSpace::allocation_align_size_up( 5622 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); 5623 if (!brs.is_reserved()) { 5624 log_warning(gc)("CMS bit map allocation failure"); 5625 return false; 5626 } 5627 // For now we'll just commit all of the bit map up front. 5628 // Later on we'll try to be more parsimonious with swap. 5629 if (!_virtual_space.initialize(brs, brs.size())) { 5630 log_warning(gc)("CMS bit map backing store failure"); 5631 return false; 5632 } 5633 assert(_virtual_space.committed_size() == brs.size(), 5634 "didn't reserve backing store for all of CMS bit map?"); 5635 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= 5636 _bmWordSize, "inconsistency in bit map sizing"); 5637 _bm = BitMapView((BitMap::bm_word_t*)_virtual_space.low(), _bmWordSize >> _shifter); 5638 5639 // bm.clear(); // can we rely on getting zero'd memory? verify below 5640 assert(isAllClear(), 5641 "Expected zero'd memory from ReservedSpace constructor"); 5642 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), 5643 "consistency check"); 5644 return true; 5645 } 5646 5647 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { 5648 HeapWord *next_addr, *end_addr, *last_addr; 5649 assert_locked(); 5650 assert(covers(mr), "out-of-range error"); 5651 // XXX assert that start and end are appropriately aligned 5652 for (next_addr = mr.start(), end_addr = mr.end(); 5653 next_addr < end_addr; next_addr = last_addr) { 5654 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); 5655 last_addr = dirty_region.end(); 5656 if (!dirty_region.is_empty()) { 5657 cl->do_MemRegion(dirty_region); 5658 } else { 5659 assert(last_addr == end_addr, "program logic"); 5660 return; 5661 } 5662 } 5663 } 5664 5665 void CMSBitMap::print_on_error(outputStream* st, const char* prefix) const { 5666 _bm.print_on_error(st, prefix); 5667 } 5668 5669 #ifndef PRODUCT 5670 void CMSBitMap::assert_locked() const { 5671 CMSLockVerifier::assert_locked(lock()); 5672 } 5673 5674 bool CMSBitMap::covers(MemRegion mr) const { 5675 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); 5676 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), 5677 "size inconsistency"); 5678 return (mr.start() >= _bmStartWord) && 5679 (mr.end() <= endWord()); 5680 } 5681 5682 bool CMSBitMap::covers(HeapWord* start, size_t size) const { 5683 return (start >= _bmStartWord && (start + size) <= endWord()); 5684 } 5685 5686 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { 5687 // verify that there are no 1 bits in the interval [left, right) 5688 FalseBitMapClosure falseBitMapClosure; 5689 iterate(&falseBitMapClosure, left, right); 5690 } 5691 5692 void CMSBitMap::region_invariant(MemRegion mr) 5693 { 5694 assert_locked(); 5695 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); 5696 assert(!mr.is_empty(), "unexpected empty region"); 5697 assert(covers(mr), "mr should be covered by bit map"); 5698 // convert address range into offset range 5699 size_t start_ofs = heapWordToOffset(mr.start()); 5700 // Make sure that end() is appropriately aligned 5701 assert(mr.end() == align_up(mr.end(), (1 << (_shifter+LogHeapWordSize))), 5702 "Misaligned mr.end()"); 5703 size_t end_ofs = heapWordToOffset(mr.end()); 5704 assert(end_ofs > start_ofs, "Should mark at least one bit"); 5705 } 5706 5707 #endif 5708 5709 bool CMSMarkStack::allocate(size_t size) { 5710 // allocate a stack of the requisite depth 5711 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5712 size * sizeof(oop))); 5713 if (!rs.is_reserved()) { 5714 log_warning(gc)("CMSMarkStack allocation failure"); 5715 return false; 5716 } 5717 if (!_virtual_space.initialize(rs, rs.size())) { 5718 log_warning(gc)("CMSMarkStack backing store failure"); 5719 return false; 5720 } 5721 assert(_virtual_space.committed_size() == rs.size(), 5722 "didn't reserve backing store for all of CMS stack?"); 5723 _base = (oop*)(_virtual_space.low()); 5724 _index = 0; 5725 _capacity = size; 5726 NOT_PRODUCT(_max_depth = 0); 5727 return true; 5728 } 5729 5730 // XXX FIX ME !!! In the MT case we come in here holding a 5731 // leaf lock. For printing we need to take a further lock 5732 // which has lower rank. We need to recalibrate the two 5733 // lock-ranks involved in order to be able to print the 5734 // messages below. (Or defer the printing to the caller. 5735 // For now we take the expedient path of just disabling the 5736 // messages for the problematic case.) 5737 void CMSMarkStack::expand() { 5738 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); 5739 if (_capacity == MarkStackSizeMax) { 5740 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled) { 5741 // We print a warning message only once per CMS cycle. 5742 log_debug(gc)(" (benign) Hit CMSMarkStack max size limit"); 5743 } 5744 return; 5745 } 5746 // Double capacity if possible 5747 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); 5748 // Do not give up existing stack until we have managed to 5749 // get the double capacity that we desired. 5750 ReservedSpace rs(ReservedSpace::allocation_align_size_up( 5751 new_capacity * sizeof(oop))); 5752 if (rs.is_reserved()) { 5753 // Release the backing store associated with old stack 5754 _virtual_space.release(); 5755 // Reinitialize virtual space for new stack 5756 if (!_virtual_space.initialize(rs, rs.size())) { 5757 fatal("Not enough swap for expanded marking stack"); 5758 } 5759 _base = (oop*)(_virtual_space.low()); 5760 _index = 0; 5761 _capacity = new_capacity; 5762 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled) { 5763 // Failed to double capacity, continue; 5764 // we print a detail message only once per CMS cycle. 5765 log_debug(gc)(" (benign) Failed to expand marking stack from " SIZE_FORMAT "K to " SIZE_FORMAT "K", 5766 _capacity / K, new_capacity / K); 5767 } 5768 } 5769 5770 5771 // Closures 5772 // XXX: there seems to be a lot of code duplication here; 5773 // should refactor and consolidate common code. 5774 5775 // This closure is used to mark refs into the CMS generation in 5776 // the CMS bit map. Called at the first checkpoint. This closure 5777 // assumes that we do not need to re-mark dirty cards; if the CMS 5778 // generation on which this is used is not an oldest 5779 // generation then this will lose younger_gen cards! 5780 5781 MarkRefsIntoClosure::MarkRefsIntoClosure( 5782 MemRegion span, CMSBitMap* bitMap): 5783 _span(span), 5784 _bitMap(bitMap) 5785 { 5786 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5787 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5788 } 5789 5790 void MarkRefsIntoClosure::do_oop(oop obj) { 5791 // if p points into _span, then mark corresponding bit in _markBitMap 5792 assert(oopDesc::is_oop(obj), "expected an oop"); 5793 HeapWord* addr = (HeapWord*)obj; 5794 if (_span.contains(addr)) { 5795 // this should be made more efficient 5796 _bitMap->mark(addr); 5797 } 5798 } 5799 5800 ParMarkRefsIntoClosure::ParMarkRefsIntoClosure( 5801 MemRegion span, CMSBitMap* bitMap): 5802 _span(span), 5803 _bitMap(bitMap) 5804 { 5805 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5806 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); 5807 } 5808 5809 void ParMarkRefsIntoClosure::do_oop(oop obj) { 5810 // if p points into _span, then mark corresponding bit in _markBitMap 5811 assert(oopDesc::is_oop(obj), "expected an oop"); 5812 HeapWord* addr = (HeapWord*)obj; 5813 if (_span.contains(addr)) { 5814 // this should be made more efficient 5815 _bitMap->par_mark(addr); 5816 } 5817 } 5818 5819 // A variant of the above, used for CMS marking verification. 5820 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( 5821 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): 5822 _span(span), 5823 _verification_bm(verification_bm), 5824 _cms_bm(cms_bm) 5825 { 5826 assert(ref_discoverer() == NULL, "deliberately left NULL"); 5827 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); 5828 } 5829 5830 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { 5831 // if p points into _span, then mark corresponding bit in _markBitMap 5832 assert(oopDesc::is_oop(obj), "expected an oop"); 5833 HeapWord* addr = (HeapWord*)obj; 5834 if (_span.contains(addr)) { 5835 _verification_bm->mark(addr); 5836 if (!_cms_bm->isMarked(addr)) { 5837 Log(gc, verify) log; 5838 ResourceMark rm; 5839 LogStream ls(log.error()); 5840 oop(addr)->print_on(&ls); 5841 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 5842 fatal("... aborting"); 5843 } 5844 } 5845 } 5846 5847 ////////////////////////////////////////////////// 5848 // MarkRefsIntoAndScanClosure 5849 ////////////////////////////////////////////////// 5850 5851 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, 5852 ReferenceDiscoverer* rd, 5853 CMSBitMap* bit_map, 5854 CMSBitMap* mod_union_table, 5855 CMSMarkStack* mark_stack, 5856 CMSCollector* collector, 5857 bool should_yield, 5858 bool concurrent_precleaning): 5859 _span(span), 5860 _bit_map(bit_map), 5861 _mark_stack(mark_stack), 5862 _pushAndMarkClosure(collector, span, rd, bit_map, mod_union_table, 5863 mark_stack, concurrent_precleaning), 5864 _collector(collector), 5865 _freelistLock(NULL), 5866 _yield(should_yield), 5867 _concurrent_precleaning(concurrent_precleaning) 5868 { 5869 // FIXME: Should initialize in base class constructor. 5870 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5871 set_ref_discoverer_internal(rd); 5872 } 5873 5874 // This closure is used to mark refs into the CMS generation at the 5875 // second (final) checkpoint, and to scan and transitively follow 5876 // the unmarked oops. It is also used during the concurrent precleaning 5877 // phase while scanning objects on dirty cards in the CMS generation. 5878 // The marks are made in the marking bit map and the marking stack is 5879 // used for keeping the (newly) grey objects during the scan. 5880 // The parallel version (Par_...) appears further below. 5881 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { 5882 if (obj != NULL) { 5883 assert(oopDesc::is_oop(obj), "expected an oop"); 5884 HeapWord* addr = (HeapWord*)obj; 5885 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 5886 assert(_collector->overflow_list_is_empty(), 5887 "overflow list should be empty"); 5888 if (_span.contains(addr) && 5889 !_bit_map->isMarked(addr)) { 5890 // mark bit map (object is now grey) 5891 _bit_map->mark(addr); 5892 // push on marking stack (stack should be empty), and drain the 5893 // stack by applying this closure to the oops in the oops popped 5894 // from the stack (i.e. blacken the grey objects) 5895 bool res = _mark_stack->push(obj); 5896 assert(res, "Should have space to push on empty stack"); 5897 do { 5898 oop new_oop = _mark_stack->pop(); 5899 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 5900 assert(_bit_map->isMarked((HeapWord*)new_oop), 5901 "only grey objects on this stack"); 5902 // iterate over the oops in this oop, marking and pushing 5903 // the ones in CMS heap (i.e. in _span). 5904 new_oop->oop_iterate(&_pushAndMarkClosure); 5905 // check if it's time to yield 5906 do_yield_check(); 5907 } while (!_mark_stack->isEmpty() || 5908 (!_concurrent_precleaning && take_from_overflow_list())); 5909 // if marking stack is empty, and we are not doing this 5910 // during precleaning, then check the overflow list 5911 } 5912 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 5913 assert(_collector->overflow_list_is_empty(), 5914 "overflow list was drained above"); 5915 5916 assert(_collector->no_preserved_marks(), 5917 "All preserved marks should have been restored above"); 5918 } 5919 } 5920 5921 void MarkRefsIntoAndScanClosure::do_yield_work() { 5922 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 5923 "CMS thread should hold CMS token"); 5924 assert_lock_strong(_freelistLock); 5925 assert_lock_strong(_bit_map->lock()); 5926 // relinquish the free_list_lock and bitMaplock() 5927 _bit_map->lock()->unlock(); 5928 _freelistLock->unlock(); 5929 ConcurrentMarkSweepThread::desynchronize(true); 5930 _collector->stopTimer(); 5931 _collector->incrementYields(); 5932 5933 // See the comment in coordinator_yield() 5934 for (unsigned i = 0; 5935 i < CMSYieldSleepCount && 5936 ConcurrentMarkSweepThread::should_yield() && 5937 !CMSCollector::foregroundGCIsActive(); 5938 ++i) { 5939 os::sleep(Thread::current(), 1, false); 5940 } 5941 5942 ConcurrentMarkSweepThread::synchronize(true); 5943 _freelistLock->lock_without_safepoint_check(); 5944 _bit_map->lock()->lock_without_safepoint_check(); 5945 _collector->startTimer(); 5946 } 5947 5948 /////////////////////////////////////////////////////////// 5949 // ParMarkRefsIntoAndScanClosure: a parallel version of 5950 // MarkRefsIntoAndScanClosure 5951 /////////////////////////////////////////////////////////// 5952 ParMarkRefsIntoAndScanClosure::ParMarkRefsIntoAndScanClosure( 5953 CMSCollector* collector, MemRegion span, ReferenceDiscoverer* rd, 5954 CMSBitMap* bit_map, OopTaskQueue* work_queue): 5955 _span(span), 5956 _bit_map(bit_map), 5957 _work_queue(work_queue), 5958 _low_water_mark(MIN2((work_queue->max_elems()/4), 5959 ((uint)CMSWorkQueueDrainThreshold * ParallelGCThreads))), 5960 _parPushAndMarkClosure(collector, span, rd, bit_map, work_queue) 5961 { 5962 // FIXME: Should initialize in base class constructor. 5963 assert(rd != NULL, "ref_discoverer shouldn't be NULL"); 5964 set_ref_discoverer_internal(rd); 5965 } 5966 5967 // This closure is used to mark refs into the CMS generation at the 5968 // second (final) checkpoint, and to scan and transitively follow 5969 // the unmarked oops. The marks are made in the marking bit map and 5970 // the work_queue is used for keeping the (newly) grey objects during 5971 // the scan phase whence they are also available for stealing by parallel 5972 // threads. Since the marking bit map is shared, updates are 5973 // synchronized (via CAS). 5974 void ParMarkRefsIntoAndScanClosure::do_oop(oop obj) { 5975 if (obj != NULL) { 5976 // Ignore mark word because this could be an already marked oop 5977 // that may be chained at the end of the overflow list. 5978 assert(oopDesc::is_oop(obj, true), "expected an oop"); 5979 HeapWord* addr = (HeapWord*)obj; 5980 if (_span.contains(addr) && 5981 !_bit_map->isMarked(addr)) { 5982 // mark bit map (object will become grey): 5983 // It is possible for several threads to be 5984 // trying to "claim" this object concurrently; 5985 // the unique thread that succeeds in marking the 5986 // object first will do the subsequent push on 5987 // to the work queue (or overflow list). 5988 if (_bit_map->par_mark(addr)) { 5989 // push on work_queue (which may not be empty), and trim the 5990 // queue to an appropriate length by applying this closure to 5991 // the oops in the oops popped from the stack (i.e. blacken the 5992 // grey objects) 5993 bool res = _work_queue->push(obj); 5994 assert(res, "Low water mark should be less than capacity?"); 5995 trim_queue(_low_water_mark); 5996 } // Else, another thread claimed the object 5997 } 5998 } 5999 } 6000 6001 // This closure is used to rescan the marked objects on the dirty cards 6002 // in the mod union table and the card table proper. 6003 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( 6004 oop p, MemRegion mr) { 6005 6006 size_t size = 0; 6007 HeapWord* addr = (HeapWord*)p; 6008 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6009 assert(_span.contains(addr), "we are scanning the CMS generation"); 6010 // check if it's time to yield 6011 if (do_yield_check()) { 6012 // We yielded for some foreground stop-world work, 6013 // and we have been asked to abort this ongoing preclean cycle. 6014 return 0; 6015 } 6016 if (_bitMap->isMarked(addr)) { 6017 // it's marked; is it potentially uninitialized? 6018 if (p->klass_or_null_acquire() != NULL) { 6019 // an initialized object; ignore mark word in verification below 6020 // since we are running concurrent with mutators 6021 assert(oopDesc::is_oop(p, true), "should be an oop"); 6022 if (p->is_objArray()) { 6023 // objArrays are precisely marked; restrict scanning 6024 // to dirty cards only. 6025 size = CompactibleFreeListSpace::adjustObjectSize( 6026 p->oop_iterate_size(_scanningClosure, mr)); 6027 } else { 6028 // A non-array may have been imprecisely marked; we need 6029 // to scan object in its entirety. 6030 size = CompactibleFreeListSpace::adjustObjectSize( 6031 p->oop_iterate_size(_scanningClosure)); 6032 } 6033 #ifdef ASSERT 6034 size_t direct_size = 6035 CompactibleFreeListSpace::adjustObjectSize(p->size()); 6036 assert(size == direct_size, "Inconsistency in size"); 6037 assert(size >= 3, "Necessary for Printezis marks to work"); 6038 HeapWord* start_pbit = addr + 1; 6039 HeapWord* end_pbit = addr + size - 1; 6040 assert(_bitMap->isMarked(start_pbit) == _bitMap->isMarked(end_pbit), 6041 "inconsistent Printezis mark"); 6042 // Verify inner mark bits (between Printezis bits) are clear, 6043 // but don't repeat if there are multiple dirty regions for 6044 // the same object, to avoid potential O(N^2) performance. 6045 if (addr != _last_scanned_object) { 6046 _bitMap->verifyNoOneBitsInRange(start_pbit + 1, end_pbit); 6047 _last_scanned_object = addr; 6048 } 6049 #endif // ASSERT 6050 } else { 6051 // An uninitialized object. 6052 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); 6053 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 6054 size = pointer_delta(nextOneAddr + 1, addr); 6055 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 6056 "alignment problem"); 6057 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() 6058 // will dirty the card when the klass pointer is installed in the 6059 // object (signaling the completion of initialization). 6060 } 6061 } else { 6062 // Either a not yet marked object or an uninitialized object 6063 if (p->klass_or_null_acquire() == NULL) { 6064 // An uninitialized object, skip to the next card, since 6065 // we may not be able to read its P-bits yet. 6066 assert(size == 0, "Initial value"); 6067 } else { 6068 // An object not (yet) reached by marking: we merely need to 6069 // compute its size so as to go look at the next block. 6070 assert(oopDesc::is_oop(p, true), "should be an oop"); 6071 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); 6072 } 6073 } 6074 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6075 return size; 6076 } 6077 6078 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { 6079 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6080 "CMS thread should hold CMS token"); 6081 assert_lock_strong(_freelistLock); 6082 assert_lock_strong(_bitMap->lock()); 6083 // relinquish the free_list_lock and bitMaplock() 6084 _bitMap->lock()->unlock(); 6085 _freelistLock->unlock(); 6086 ConcurrentMarkSweepThread::desynchronize(true); 6087 _collector->stopTimer(); 6088 _collector->incrementYields(); 6089 6090 // See the comment in coordinator_yield() 6091 for (unsigned i = 0; i < CMSYieldSleepCount && 6092 ConcurrentMarkSweepThread::should_yield() && 6093 !CMSCollector::foregroundGCIsActive(); ++i) { 6094 os::sleep(Thread::current(), 1, false); 6095 } 6096 6097 ConcurrentMarkSweepThread::synchronize(true); 6098 _freelistLock->lock_without_safepoint_check(); 6099 _bitMap->lock()->lock_without_safepoint_check(); 6100 _collector->startTimer(); 6101 } 6102 6103 6104 ////////////////////////////////////////////////////////////////// 6105 // SurvivorSpacePrecleanClosure 6106 ////////////////////////////////////////////////////////////////// 6107 // This (single-threaded) closure is used to preclean the oops in 6108 // the survivor spaces. 6109 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { 6110 6111 HeapWord* addr = (HeapWord*)p; 6112 DEBUG_ONLY(_collector->verify_work_stacks_empty();) 6113 assert(!_span.contains(addr), "we are scanning the survivor spaces"); 6114 assert(p->klass_or_null() != NULL, "object should be initialized"); 6115 // an initialized object; ignore mark word in verification below 6116 // since we are running concurrent with mutators 6117 assert(oopDesc::is_oop(p, true), "should be an oop"); 6118 // Note that we do not yield while we iterate over 6119 // the interior oops of p, pushing the relevant ones 6120 // on our marking stack. 6121 size_t size = p->oop_iterate_size(_scanning_closure); 6122 do_yield_check(); 6123 // Observe that below, we do not abandon the preclean 6124 // phase as soon as we should; rather we empty the 6125 // marking stack before returning. This is to satisfy 6126 // some existing assertions. In general, it may be a 6127 // good idea to abort immediately and complete the marking 6128 // from the grey objects at a later time. 6129 while (!_mark_stack->isEmpty()) { 6130 oop new_oop = _mark_stack->pop(); 6131 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 6132 assert(_bit_map->isMarked((HeapWord*)new_oop), 6133 "only grey objects on this stack"); 6134 // iterate over the oops in this oop, marking and pushing 6135 // the ones in CMS heap (i.e. in _span). 6136 new_oop->oop_iterate(_scanning_closure); 6137 // check if it's time to yield 6138 do_yield_check(); 6139 } 6140 unsigned int after_count = 6141 CMSHeap::heap()->total_collections(); 6142 bool abort = (_before_count != after_count) || 6143 _collector->should_abort_preclean(); 6144 return abort ? 0 : size; 6145 } 6146 6147 void SurvivorSpacePrecleanClosure::do_yield_work() { 6148 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6149 "CMS thread should hold CMS token"); 6150 assert_lock_strong(_bit_map->lock()); 6151 // Relinquish the bit map lock 6152 _bit_map->lock()->unlock(); 6153 ConcurrentMarkSweepThread::desynchronize(true); 6154 _collector->stopTimer(); 6155 _collector->incrementYields(); 6156 6157 // See the comment in coordinator_yield() 6158 for (unsigned i = 0; i < CMSYieldSleepCount && 6159 ConcurrentMarkSweepThread::should_yield() && 6160 !CMSCollector::foregroundGCIsActive(); ++i) { 6161 os::sleep(Thread::current(), 1, false); 6162 } 6163 6164 ConcurrentMarkSweepThread::synchronize(true); 6165 _bit_map->lock()->lock_without_safepoint_check(); 6166 _collector->startTimer(); 6167 } 6168 6169 // This closure is used to rescan the marked objects on the dirty cards 6170 // in the mod union table and the card table proper. In the parallel 6171 // case, although the bitMap is shared, we do a single read so the 6172 // isMarked() query is "safe". 6173 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { 6174 // Ignore mark word because we are running concurrent with mutators 6175 assert(oopDesc::is_oop_or_null(p, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(p)); 6176 HeapWord* addr = (HeapWord*)p; 6177 assert(_span.contains(addr), "we are scanning the CMS generation"); 6178 bool is_obj_array = false; 6179 #ifdef ASSERT 6180 if (!_parallel) { 6181 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); 6182 assert(_collector->overflow_list_is_empty(), 6183 "overflow list should be empty"); 6184 6185 } 6186 #endif // ASSERT 6187 if (_bit_map->isMarked(addr)) { 6188 // Obj arrays are precisely marked, non-arrays are not; 6189 // so we scan objArrays precisely and non-arrays in their 6190 // entirety. 6191 if (p->is_objArray()) { 6192 is_obj_array = true; 6193 if (_parallel) { 6194 p->oop_iterate(_par_scan_closure, mr); 6195 } else { 6196 p->oop_iterate(_scan_closure, mr); 6197 } 6198 } else { 6199 if (_parallel) { 6200 p->oop_iterate(_par_scan_closure); 6201 } else { 6202 p->oop_iterate(_scan_closure); 6203 } 6204 } 6205 } 6206 #ifdef ASSERT 6207 if (!_parallel) { 6208 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); 6209 assert(_collector->overflow_list_is_empty(), 6210 "overflow list should be empty"); 6211 6212 } 6213 #endif // ASSERT 6214 return is_obj_array; 6215 } 6216 6217 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, 6218 MemRegion span, 6219 CMSBitMap* bitMap, CMSMarkStack* markStack, 6220 bool should_yield, bool verifying): 6221 _collector(collector), 6222 _span(span), 6223 _bitMap(bitMap), 6224 _mut(&collector->_modUnionTable), 6225 _markStack(markStack), 6226 _yield(should_yield), 6227 _skipBits(0) 6228 { 6229 assert(_markStack->isEmpty(), "stack should be empty"); 6230 _finger = _bitMap->startWord(); 6231 _threshold = _finger; 6232 assert(_collector->_restart_addr == NULL, "Sanity check"); 6233 assert(_span.contains(_finger), "Out of bounds _finger?"); 6234 DEBUG_ONLY(_verifying = verifying;) 6235 } 6236 6237 void MarkFromRootsClosure::reset(HeapWord* addr) { 6238 assert(_markStack->isEmpty(), "would cause duplicates on stack"); 6239 assert(_span.contains(addr), "Out of bounds _finger?"); 6240 _finger = addr; 6241 _threshold = align_up(_finger, CardTable::card_size); 6242 } 6243 6244 // Should revisit to see if this should be restructured for 6245 // greater efficiency. 6246 bool MarkFromRootsClosure::do_bit(size_t offset) { 6247 if (_skipBits > 0) { 6248 _skipBits--; 6249 return true; 6250 } 6251 // convert offset into a HeapWord* 6252 HeapWord* addr = _bitMap->startWord() + offset; 6253 assert(_bitMap->endWord() && addr < _bitMap->endWord(), 6254 "address out of range"); 6255 assert(_bitMap->isMarked(addr), "tautology"); 6256 if (_bitMap->isMarked(addr+1)) { 6257 // this is an allocated but not yet initialized object 6258 assert(_skipBits == 0, "tautology"); 6259 _skipBits = 2; // skip next two marked bits ("Printezis-marks") 6260 oop p = oop(addr); 6261 if (p->klass_or_null_acquire() == NULL) { 6262 DEBUG_ONLY(if (!_verifying) {) 6263 // We re-dirty the cards on which this object lies and increase 6264 // the _threshold so that we'll come back to scan this object 6265 // during the preclean or remark phase. (CMSCleanOnEnter) 6266 if (CMSCleanOnEnter) { 6267 size_t sz = _collector->block_size_using_printezis_bits(addr); 6268 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6269 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6270 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6271 // Bump _threshold to end_card_addr; note that 6272 // _threshold cannot possibly exceed end_card_addr, anyhow. 6273 // This prevents future clearing of the card as the scan proceeds 6274 // to the right. 6275 assert(_threshold <= end_card_addr, 6276 "Because we are just scanning into this object"); 6277 if (_threshold < end_card_addr) { 6278 _threshold = end_card_addr; 6279 } 6280 if (p->klass_or_null_acquire() != NULL) { 6281 // Redirty the range of cards... 6282 _mut->mark_range(redirty_range); 6283 } // ...else the setting of klass will dirty the card anyway. 6284 } 6285 DEBUG_ONLY(}) 6286 return true; 6287 } 6288 } 6289 scanOopsInOop(addr); 6290 return true; 6291 } 6292 6293 // We take a break if we've been at this for a while, 6294 // so as to avoid monopolizing the locks involved. 6295 void MarkFromRootsClosure::do_yield_work() { 6296 // First give up the locks, then yield, then re-lock 6297 // We should probably use a constructor/destructor idiom to 6298 // do this unlock/lock or modify the MutexUnlocker class to 6299 // serve our purpose. XXX 6300 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6301 "CMS thread should hold CMS token"); 6302 assert_lock_strong(_bitMap->lock()); 6303 _bitMap->lock()->unlock(); 6304 ConcurrentMarkSweepThread::desynchronize(true); 6305 _collector->stopTimer(); 6306 _collector->incrementYields(); 6307 6308 // See the comment in coordinator_yield() 6309 for (unsigned i = 0; i < CMSYieldSleepCount && 6310 ConcurrentMarkSweepThread::should_yield() && 6311 !CMSCollector::foregroundGCIsActive(); ++i) { 6312 os::sleep(Thread::current(), 1, false); 6313 } 6314 6315 ConcurrentMarkSweepThread::synchronize(true); 6316 _bitMap->lock()->lock_without_safepoint_check(); 6317 _collector->startTimer(); 6318 } 6319 6320 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { 6321 assert(_bitMap->isMarked(ptr), "expected bit to be set"); 6322 assert(_markStack->isEmpty(), 6323 "should drain stack to limit stack usage"); 6324 // convert ptr to an oop preparatory to scanning 6325 oop obj = oop(ptr); 6326 // Ignore mark word in verification below, since we 6327 // may be running concurrent with mutators. 6328 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6329 assert(_finger <= ptr, "_finger runneth ahead"); 6330 // advance the finger to right end of this object 6331 _finger = ptr + obj->size(); 6332 assert(_finger > ptr, "we just incremented it above"); 6333 // On large heaps, it may take us some time to get through 6334 // the marking phase. During 6335 // this time it's possible that a lot of mutations have 6336 // accumulated in the card table and the mod union table -- 6337 // these mutation records are redundant until we have 6338 // actually traced into the corresponding card. 6339 // Here, we check whether advancing the finger would make 6340 // us cross into a new card, and if so clear corresponding 6341 // cards in the MUT (preclean them in the card-table in the 6342 // future). 6343 6344 DEBUG_ONLY(if (!_verifying) {) 6345 // The clean-on-enter optimization is disabled by default, 6346 // until we fix 6178663. 6347 if (CMSCleanOnEnter && (_finger > _threshold)) { 6348 // [_threshold, _finger) represents the interval 6349 // of cards to be cleared in MUT (or precleaned in card table). 6350 // The set of cards to be cleared is all those that overlap 6351 // with the interval [_threshold, _finger); note that 6352 // _threshold is always kept card-aligned but _finger isn't 6353 // always card-aligned. 6354 HeapWord* old_threshold = _threshold; 6355 assert(is_aligned(old_threshold, CardTable::card_size), 6356 "_threshold should always be card-aligned"); 6357 _threshold = align_up(_finger, CardTable::card_size); 6358 MemRegion mr(old_threshold, _threshold); 6359 assert(!mr.is_empty(), "Control point invariant"); 6360 assert(_span.contains(mr), "Should clear within span"); 6361 _mut->clear_range(mr); 6362 } 6363 DEBUG_ONLY(}) 6364 // Note: the finger doesn't advance while we drain 6365 // the stack below. 6366 PushOrMarkClosure pushOrMarkClosure(_collector, 6367 _span, _bitMap, _markStack, 6368 _finger, this); 6369 bool res = _markStack->push(obj); 6370 assert(res, "Empty non-zero size stack should have space for single push"); 6371 while (!_markStack->isEmpty()) { 6372 oop new_oop = _markStack->pop(); 6373 // Skip verifying header mark word below because we are 6374 // running concurrent with mutators. 6375 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6376 // now scan this oop's oops 6377 new_oop->oop_iterate(&pushOrMarkClosure); 6378 do_yield_check(); 6379 } 6380 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); 6381 } 6382 6383 ParMarkFromRootsClosure::ParMarkFromRootsClosure(CMSConcMarkingTask* task, 6384 CMSCollector* collector, MemRegion span, 6385 CMSBitMap* bit_map, 6386 OopTaskQueue* work_queue, 6387 CMSMarkStack* overflow_stack): 6388 _collector(collector), 6389 _whole_span(collector->_span), 6390 _span(span), 6391 _bit_map(bit_map), 6392 _mut(&collector->_modUnionTable), 6393 _work_queue(work_queue), 6394 _overflow_stack(overflow_stack), 6395 _skip_bits(0), 6396 _task(task) 6397 { 6398 assert(_work_queue->size() == 0, "work_queue should be empty"); 6399 _finger = span.start(); 6400 _threshold = _finger; // XXX Defer clear-on-enter optimization for now 6401 assert(_span.contains(_finger), "Out of bounds _finger?"); 6402 } 6403 6404 // Should revisit to see if this should be restructured for 6405 // greater efficiency. 6406 bool ParMarkFromRootsClosure::do_bit(size_t offset) { 6407 if (_skip_bits > 0) { 6408 _skip_bits--; 6409 return true; 6410 } 6411 // convert offset into a HeapWord* 6412 HeapWord* addr = _bit_map->startWord() + offset; 6413 assert(_bit_map->endWord() && addr < _bit_map->endWord(), 6414 "address out of range"); 6415 assert(_bit_map->isMarked(addr), "tautology"); 6416 if (_bit_map->isMarked(addr+1)) { 6417 // this is an allocated object that might not yet be initialized 6418 assert(_skip_bits == 0, "tautology"); 6419 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") 6420 oop p = oop(addr); 6421 if (p->klass_or_null_acquire() == NULL) { 6422 // in the case of Clean-on-Enter optimization, redirty card 6423 // and avoid clearing card by increasing the threshold. 6424 return true; 6425 } 6426 } 6427 scan_oops_in_oop(addr); 6428 return true; 6429 } 6430 6431 void ParMarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { 6432 assert(_bit_map->isMarked(ptr), "expected bit to be set"); 6433 // Should we assert that our work queue is empty or 6434 // below some drain limit? 6435 assert(_work_queue->size() == 0, 6436 "should drain stack to limit stack usage"); 6437 // convert ptr to an oop preparatory to scanning 6438 oop obj = oop(ptr); 6439 // Ignore mark word in verification below, since we 6440 // may be running concurrent with mutators. 6441 assert(oopDesc::is_oop(obj, true), "should be an oop"); 6442 assert(_finger <= ptr, "_finger runneth ahead"); 6443 // advance the finger to right end of this object 6444 _finger = ptr + obj->size(); 6445 assert(_finger > ptr, "we just incremented it above"); 6446 // On large heaps, it may take us some time to get through 6447 // the marking phase. During 6448 // this time it's possible that a lot of mutations have 6449 // accumulated in the card table and the mod union table -- 6450 // these mutation records are redundant until we have 6451 // actually traced into the corresponding card. 6452 // Here, we check whether advancing the finger would make 6453 // us cross into a new card, and if so clear corresponding 6454 // cards in the MUT (preclean them in the card-table in the 6455 // future). 6456 6457 // The clean-on-enter optimization is disabled by default, 6458 // until we fix 6178663. 6459 if (CMSCleanOnEnter && (_finger > _threshold)) { 6460 // [_threshold, _finger) represents the interval 6461 // of cards to be cleared in MUT (or precleaned in card table). 6462 // The set of cards to be cleared is all those that overlap 6463 // with the interval [_threshold, _finger); note that 6464 // _threshold is always kept card-aligned but _finger isn't 6465 // always card-aligned. 6466 HeapWord* old_threshold = _threshold; 6467 assert(is_aligned(old_threshold, CardTable::card_size), 6468 "_threshold should always be card-aligned"); 6469 _threshold = align_up(_finger, CardTable::card_size); 6470 MemRegion mr(old_threshold, _threshold); 6471 assert(!mr.is_empty(), "Control point invariant"); 6472 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? 6473 _mut->clear_range(mr); 6474 } 6475 6476 // Note: the local finger doesn't advance while we drain 6477 // the stack below, but the global finger sure can and will. 6478 HeapWord* volatile* gfa = _task->global_finger_addr(); 6479 ParPushOrMarkClosure pushOrMarkClosure(_collector, 6480 _span, _bit_map, 6481 _work_queue, 6482 _overflow_stack, 6483 _finger, 6484 gfa, this); 6485 bool res = _work_queue->push(obj); // overflow could occur here 6486 assert(res, "Will hold once we use workqueues"); 6487 while (true) { 6488 oop new_oop; 6489 if (!_work_queue->pop_local(new_oop)) { 6490 // We emptied our work_queue; check if there's stuff that can 6491 // be gotten from the overflow stack. 6492 if (CMSConcMarkingTask::get_work_from_overflow_stack( 6493 _overflow_stack, _work_queue)) { 6494 do_yield_check(); 6495 continue; 6496 } else { // done 6497 break; 6498 } 6499 } 6500 // Skip verifying header mark word below because we are 6501 // running concurrent with mutators. 6502 assert(oopDesc::is_oop(new_oop, true), "Oops! expected to pop an oop"); 6503 // now scan this oop's oops 6504 new_oop->oop_iterate(&pushOrMarkClosure); 6505 do_yield_check(); 6506 } 6507 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); 6508 } 6509 6510 // Yield in response to a request from VM Thread or 6511 // from mutators. 6512 void ParMarkFromRootsClosure::do_yield_work() { 6513 assert(_task != NULL, "sanity"); 6514 _task->yield(); 6515 } 6516 6517 // A variant of the above used for verifying CMS marking work. 6518 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, 6519 MemRegion span, 6520 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6521 CMSMarkStack* mark_stack): 6522 _collector(collector), 6523 _span(span), 6524 _verification_bm(verification_bm), 6525 _cms_bm(cms_bm), 6526 _mark_stack(mark_stack), 6527 _pam_verify_closure(collector, span, verification_bm, cms_bm, 6528 mark_stack) 6529 { 6530 assert(_mark_stack->isEmpty(), "stack should be empty"); 6531 _finger = _verification_bm->startWord(); 6532 assert(_collector->_restart_addr == NULL, "Sanity check"); 6533 assert(_span.contains(_finger), "Out of bounds _finger?"); 6534 } 6535 6536 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { 6537 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); 6538 assert(_span.contains(addr), "Out of bounds _finger?"); 6539 _finger = addr; 6540 } 6541 6542 // Should revisit to see if this should be restructured for 6543 // greater efficiency. 6544 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { 6545 // convert offset into a HeapWord* 6546 HeapWord* addr = _verification_bm->startWord() + offset; 6547 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), 6548 "address out of range"); 6549 assert(_verification_bm->isMarked(addr), "tautology"); 6550 assert(_cms_bm->isMarked(addr), "tautology"); 6551 6552 assert(_mark_stack->isEmpty(), 6553 "should drain stack to limit stack usage"); 6554 // convert addr to an oop preparatory to scanning 6555 oop obj = oop(addr); 6556 assert(oopDesc::is_oop(obj), "should be an oop"); 6557 assert(_finger <= addr, "_finger runneth ahead"); 6558 // advance the finger to right end of this object 6559 _finger = addr + obj->size(); 6560 assert(_finger > addr, "we just incremented it above"); 6561 // Note: the finger doesn't advance while we drain 6562 // the stack below. 6563 bool res = _mark_stack->push(obj); 6564 assert(res, "Empty non-zero size stack should have space for single push"); 6565 while (!_mark_stack->isEmpty()) { 6566 oop new_oop = _mark_stack->pop(); 6567 assert(oopDesc::is_oop(new_oop), "Oops! expected to pop an oop"); 6568 // now scan this oop's oops 6569 new_oop->oop_iterate(&_pam_verify_closure); 6570 } 6571 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); 6572 return true; 6573 } 6574 6575 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( 6576 CMSCollector* collector, MemRegion span, 6577 CMSBitMap* verification_bm, CMSBitMap* cms_bm, 6578 CMSMarkStack* mark_stack): 6579 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6580 _collector(collector), 6581 _span(span), 6582 _verification_bm(verification_bm), 6583 _cms_bm(cms_bm), 6584 _mark_stack(mark_stack) 6585 { } 6586 6587 template <class T> void PushAndMarkVerifyClosure::do_oop_work(T *p) { 6588 oop obj = RawAccess<>::oop_load(p); 6589 do_oop(obj); 6590 } 6591 6592 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6593 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } 6594 6595 // Upon stack overflow, we discard (part of) the stack, 6596 // remembering the least address amongst those discarded 6597 // in CMSCollector's _restart_address. 6598 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { 6599 // Remember the least grey address discarded 6600 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); 6601 _collector->lower_restart_addr(ra); 6602 _mark_stack->reset(); // discard stack contents 6603 _mark_stack->expand(); // expand the stack if possible 6604 } 6605 6606 void PushAndMarkVerifyClosure::do_oop(oop obj) { 6607 assert(oopDesc::is_oop_or_null(obj), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6608 HeapWord* addr = (HeapWord*)obj; 6609 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { 6610 // Oop lies in _span and isn't yet grey or black 6611 _verification_bm->mark(addr); // now grey 6612 if (!_cms_bm->isMarked(addr)) { 6613 Log(gc, verify) log; 6614 ResourceMark rm; 6615 LogStream ls(log.error()); 6616 oop(addr)->print_on(&ls); 6617 log.error(" (" INTPTR_FORMAT " should have been marked)", p2i(addr)); 6618 fatal("... aborting"); 6619 } 6620 6621 if (!_mark_stack->push(obj)) { // stack overflow 6622 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _mark_stack->capacity()); 6623 assert(_mark_stack->isFull(), "Else push should have succeeded"); 6624 handle_stack_overflow(addr); 6625 } 6626 // anything including and to the right of _finger 6627 // will be scanned as we iterate over the remainder of the 6628 // bit map 6629 } 6630 } 6631 6632 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, 6633 MemRegion span, 6634 CMSBitMap* bitMap, CMSMarkStack* markStack, 6635 HeapWord* finger, MarkFromRootsClosure* parent) : 6636 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6637 _collector(collector), 6638 _span(span), 6639 _bitMap(bitMap), 6640 _markStack(markStack), 6641 _finger(finger), 6642 _parent(parent) 6643 { } 6644 6645 ParPushOrMarkClosure::ParPushOrMarkClosure(CMSCollector* collector, 6646 MemRegion span, 6647 CMSBitMap* bit_map, 6648 OopTaskQueue* work_queue, 6649 CMSMarkStack* overflow_stack, 6650 HeapWord* finger, 6651 HeapWord* volatile* global_finger_addr, 6652 ParMarkFromRootsClosure* parent) : 6653 MetadataVisitingOopIterateClosure(collector->ref_processor()), 6654 _collector(collector), 6655 _whole_span(collector->_span), 6656 _span(span), 6657 _bit_map(bit_map), 6658 _work_queue(work_queue), 6659 _overflow_stack(overflow_stack), 6660 _finger(finger), 6661 _global_finger_addr(global_finger_addr), 6662 _parent(parent) 6663 { } 6664 6665 // Assumes thread-safe access by callers, who are 6666 // responsible for mutual exclusion. 6667 void CMSCollector::lower_restart_addr(HeapWord* low) { 6668 assert(_span.contains(low), "Out of bounds addr"); 6669 if (_restart_addr == NULL) { 6670 _restart_addr = low; 6671 } else { 6672 _restart_addr = MIN2(_restart_addr, low); 6673 } 6674 } 6675 6676 // Upon stack overflow, we discard (part of) the stack, 6677 // remembering the least address amongst those discarded 6678 // in CMSCollector's _restart_address. 6679 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6680 // Remember the least grey address discarded 6681 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); 6682 _collector->lower_restart_addr(ra); 6683 _markStack->reset(); // discard stack contents 6684 _markStack->expand(); // expand the stack if possible 6685 } 6686 6687 // Upon stack overflow, we discard (part of) the stack, 6688 // remembering the least address amongst those discarded 6689 // in CMSCollector's _restart_address. 6690 void ParPushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { 6691 // We need to do this under a mutex to prevent other 6692 // workers from interfering with the work done below. 6693 MutexLockerEx ml(_overflow_stack->par_lock(), 6694 Mutex::_no_safepoint_check_flag); 6695 // Remember the least grey address discarded 6696 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); 6697 _collector->lower_restart_addr(ra); 6698 _overflow_stack->reset(); // discard stack contents 6699 _overflow_stack->expand(); // expand the stack if possible 6700 } 6701 6702 void PushOrMarkClosure::do_oop(oop obj) { 6703 // Ignore mark word because we are running concurrent with mutators. 6704 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6705 HeapWord* addr = (HeapWord*)obj; 6706 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { 6707 // Oop lies in _span and isn't yet grey or black 6708 _bitMap->mark(addr); // now grey 6709 if (addr < _finger) { 6710 // the bit map iteration has already either passed, or 6711 // sampled, this bit in the bit map; we'll need to 6712 // use the marking stack to scan this oop's oops. 6713 bool simulate_overflow = false; 6714 NOT_PRODUCT( 6715 if (CMSMarkStackOverflowALot && 6716 _collector->simulate_overflow()) { 6717 // simulate a stack overflow 6718 simulate_overflow = true; 6719 } 6720 ) 6721 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow 6722 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _markStack->capacity()); 6723 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); 6724 handle_stack_overflow(addr); 6725 } 6726 } 6727 // anything including and to the right of _finger 6728 // will be scanned as we iterate over the remainder of the 6729 // bit map 6730 do_yield_check(); 6731 } 6732 } 6733 6734 void ParPushOrMarkClosure::do_oop(oop obj) { 6735 // Ignore mark word because we are running concurrent with mutators. 6736 assert(oopDesc::is_oop_or_null(obj, true), "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6737 HeapWord* addr = (HeapWord*)obj; 6738 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { 6739 // Oop lies in _span and isn't yet grey or black 6740 // We read the global_finger (volatile read) strictly after marking oop 6741 bool res = _bit_map->par_mark(addr); // now grey 6742 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; 6743 // Should we push this marked oop on our stack? 6744 // -- if someone else marked it, nothing to do 6745 // -- if target oop is above global finger nothing to do 6746 // -- if target oop is in chunk and above local finger 6747 // then nothing to do 6748 // -- else push on work queue 6749 if ( !res // someone else marked it, they will deal with it 6750 || (addr >= *gfa) // will be scanned in a later task 6751 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk 6752 return; 6753 } 6754 // the bit map iteration has already either passed, or 6755 // sampled, this bit in the bit map; we'll need to 6756 // use the marking stack to scan this oop's oops. 6757 bool simulate_overflow = false; 6758 NOT_PRODUCT( 6759 if (CMSMarkStackOverflowALot && 6760 _collector->simulate_overflow()) { 6761 // simulate a stack overflow 6762 simulate_overflow = true; 6763 } 6764 ) 6765 if (simulate_overflow || 6766 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { 6767 // stack overflow 6768 log_trace(gc)("CMS marking stack overflow (benign) at " SIZE_FORMAT, _overflow_stack->capacity()); 6769 // We cannot assert that the overflow stack is full because 6770 // it may have been emptied since. 6771 assert(simulate_overflow || 6772 _work_queue->size() == _work_queue->max_elems(), 6773 "Else push should have succeeded"); 6774 handle_stack_overflow(addr); 6775 } 6776 do_yield_check(); 6777 } 6778 } 6779 6780 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, 6781 MemRegion span, 6782 ReferenceDiscoverer* rd, 6783 CMSBitMap* bit_map, 6784 CMSBitMap* mod_union_table, 6785 CMSMarkStack* mark_stack, 6786 bool concurrent_precleaning): 6787 MetadataVisitingOopIterateClosure(rd), 6788 _collector(collector), 6789 _span(span), 6790 _bit_map(bit_map), 6791 _mod_union_table(mod_union_table), 6792 _mark_stack(mark_stack), 6793 _concurrent_precleaning(concurrent_precleaning) 6794 { 6795 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6796 } 6797 6798 // Grey object rescan during pre-cleaning and second checkpoint phases -- 6799 // the non-parallel version (the parallel version appears further below.) 6800 void PushAndMarkClosure::do_oop(oop obj) { 6801 // Ignore mark word verification. If during concurrent precleaning, 6802 // the object monitor may be locked. If during the checkpoint 6803 // phases, the object may already have been reached by a different 6804 // path and may be at the end of the global overflow list (so 6805 // the mark word may be NULL). 6806 assert(oopDesc::is_oop_or_null(obj, true /* ignore mark word */), 6807 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6808 HeapWord* addr = (HeapWord*)obj; 6809 // Check if oop points into the CMS generation 6810 // and is not marked 6811 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6812 // a white object ... 6813 _bit_map->mark(addr); // ... now grey 6814 // push on the marking stack (grey set) 6815 bool simulate_overflow = false; 6816 NOT_PRODUCT( 6817 if (CMSMarkStackOverflowALot && 6818 _collector->simulate_overflow()) { 6819 // simulate a stack overflow 6820 simulate_overflow = true; 6821 } 6822 ) 6823 if (simulate_overflow || !_mark_stack->push(obj)) { 6824 if (_concurrent_precleaning) { 6825 // During precleaning we can just dirty the appropriate card(s) 6826 // in the mod union table, thus ensuring that the object remains 6827 // in the grey set and continue. In the case of object arrays 6828 // we need to dirty all of the cards that the object spans, 6829 // since the rescan of object arrays will be limited to the 6830 // dirty cards. 6831 // Note that no one can be interfering with us in this action 6832 // of dirtying the mod union table, so no locking or atomics 6833 // are required. 6834 if (obj->is_objArray()) { 6835 size_t sz = obj->size(); 6836 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 6837 MemRegion redirty_range = MemRegion(addr, end_card_addr); 6838 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 6839 _mod_union_table->mark_range(redirty_range); 6840 } else { 6841 _mod_union_table->mark(addr); 6842 } 6843 _collector->_ser_pmc_preclean_ovflw++; 6844 } else { 6845 // During the remark phase, we need to remember this oop 6846 // in the overflow list. 6847 _collector->push_on_overflow_list(obj); 6848 _collector->_ser_pmc_remark_ovflw++; 6849 } 6850 } 6851 } 6852 } 6853 6854 ParPushAndMarkClosure::ParPushAndMarkClosure(CMSCollector* collector, 6855 MemRegion span, 6856 ReferenceDiscoverer* rd, 6857 CMSBitMap* bit_map, 6858 OopTaskQueue* work_queue): 6859 MetadataVisitingOopIterateClosure(rd), 6860 _collector(collector), 6861 _span(span), 6862 _bit_map(bit_map), 6863 _work_queue(work_queue) 6864 { 6865 assert(ref_discoverer() != NULL, "ref_discoverer shouldn't be NULL"); 6866 } 6867 6868 // Grey object rescan during second checkpoint phase -- 6869 // the parallel version. 6870 void ParPushAndMarkClosure::do_oop(oop obj) { 6871 // In the assert below, we ignore the mark word because 6872 // this oop may point to an already visited object that is 6873 // on the overflow stack (in which case the mark word has 6874 // been hijacked for chaining into the overflow stack -- 6875 // if this is the last object in the overflow stack then 6876 // its mark word will be NULL). Because this object may 6877 // have been subsequently popped off the global overflow 6878 // stack, and the mark word possibly restored to the prototypical 6879 // value, by the time we get to examined this failing assert in 6880 // the debugger, is_oop_or_null(false) may subsequently start 6881 // to hold. 6882 assert(oopDesc::is_oop_or_null(obj, true), 6883 "Expected an oop or NULL at " PTR_FORMAT, p2i(obj)); 6884 HeapWord* addr = (HeapWord*)obj; 6885 // Check if oop points into the CMS generation 6886 // and is not marked 6887 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { 6888 // a white object ... 6889 // If we manage to "claim" the object, by being the 6890 // first thread to mark it, then we push it on our 6891 // marking stack 6892 if (_bit_map->par_mark(addr)) { // ... now grey 6893 // push on work queue (grey set) 6894 bool simulate_overflow = false; 6895 NOT_PRODUCT( 6896 if (CMSMarkStackOverflowALot && 6897 _collector->par_simulate_overflow()) { 6898 // simulate a stack overflow 6899 simulate_overflow = true; 6900 } 6901 ) 6902 if (simulate_overflow || !_work_queue->push(obj)) { 6903 _collector->par_push_on_overflow_list(obj); 6904 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS 6905 } 6906 } // Else, some other thread got there first 6907 } 6908 } 6909 6910 void CMSPrecleanRefsYieldClosure::do_yield_work() { 6911 Mutex* bml = _collector->bitMapLock(); 6912 assert_lock_strong(bml); 6913 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 6914 "CMS thread should hold CMS token"); 6915 6916 bml->unlock(); 6917 ConcurrentMarkSweepThread::desynchronize(true); 6918 6919 _collector->stopTimer(); 6920 _collector->incrementYields(); 6921 6922 // See the comment in coordinator_yield() 6923 for (unsigned i = 0; i < CMSYieldSleepCount && 6924 ConcurrentMarkSweepThread::should_yield() && 6925 !CMSCollector::foregroundGCIsActive(); ++i) { 6926 os::sleep(Thread::current(), 1, false); 6927 } 6928 6929 ConcurrentMarkSweepThread::synchronize(true); 6930 bml->lock(); 6931 6932 _collector->startTimer(); 6933 } 6934 6935 bool CMSPrecleanRefsYieldClosure::should_return() { 6936 if (ConcurrentMarkSweepThread::should_yield()) { 6937 do_yield_work(); 6938 } 6939 return _collector->foregroundGCIsActive(); 6940 } 6941 6942 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { 6943 assert(((size_t)mr.start())%CardTable::card_size_in_words == 0, 6944 "mr should be aligned to start at a card boundary"); 6945 // We'd like to assert: 6946 // assert(mr.word_size()%CardTable::card_size_in_words == 0, 6947 // "mr should be a range of cards"); 6948 // However, that would be too strong in one case -- the last 6949 // partition ends at _unallocated_block which, in general, can be 6950 // an arbitrary boundary, not necessarily card aligned. 6951 _num_dirty_cards += mr.word_size()/CardTable::card_size_in_words; 6952 _space->object_iterate_mem(mr, &_scan_cl); 6953 } 6954 6955 SweepClosure::SweepClosure(CMSCollector* collector, 6956 ConcurrentMarkSweepGeneration* g, 6957 CMSBitMap* bitMap, bool should_yield) : 6958 _collector(collector), 6959 _g(g), 6960 _sp(g->cmsSpace()), 6961 _limit(_sp->sweep_limit()), 6962 _freelistLock(_sp->freelistLock()), 6963 _bitMap(bitMap), 6964 _inFreeRange(false), // No free range at beginning of sweep 6965 _freeRangeInFreeLists(false), // No free range at beginning of sweep 6966 _lastFreeRangeCoalesced(false), 6967 _yield(should_yield), 6968 _freeFinger(g->used_region().start()) 6969 { 6970 NOT_PRODUCT( 6971 _numObjectsFreed = 0; 6972 _numWordsFreed = 0; 6973 _numObjectsLive = 0; 6974 _numWordsLive = 0; 6975 _numObjectsAlreadyFree = 0; 6976 _numWordsAlreadyFree = 0; 6977 _last_fc = NULL; 6978 6979 _sp->initializeIndexedFreeListArrayReturnedBytes(); 6980 _sp->dictionary()->initialize_dict_returned_bytes(); 6981 ) 6982 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 6983 "sweep _limit out of bounds"); 6984 log_develop_trace(gc, sweep)("===================="); 6985 log_develop_trace(gc, sweep)("Starting new sweep with limit " PTR_FORMAT, p2i(_limit)); 6986 } 6987 6988 void SweepClosure::print_on(outputStream* st) const { 6989 st->print_cr("_sp = [" PTR_FORMAT "," PTR_FORMAT ")", 6990 p2i(_sp->bottom()), p2i(_sp->end())); 6991 st->print_cr("_limit = " PTR_FORMAT, p2i(_limit)); 6992 st->print_cr("_freeFinger = " PTR_FORMAT, p2i(_freeFinger)); 6993 NOT_PRODUCT(st->print_cr("_last_fc = " PTR_FORMAT, p2i(_last_fc));) 6994 st->print_cr("_inFreeRange = %d, _freeRangeInFreeLists = %d, _lastFreeRangeCoalesced = %d", 6995 _inFreeRange, _freeRangeInFreeLists, _lastFreeRangeCoalesced); 6996 } 6997 6998 #ifndef PRODUCT 6999 // Assertion checking only: no useful work in product mode -- 7000 // however, if any of the flags below become product flags, 7001 // you may need to review this code to see if it needs to be 7002 // enabled in product mode. 7003 SweepClosure::~SweepClosure() { 7004 assert_lock_strong(_freelistLock); 7005 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7006 "sweep _limit out of bounds"); 7007 if (inFreeRange()) { 7008 Log(gc, sweep) log; 7009 log.error("inFreeRange() should have been reset; dumping state of SweepClosure"); 7010 ResourceMark rm; 7011 LogStream ls(log.error()); 7012 print_on(&ls); 7013 ShouldNotReachHere(); 7014 } 7015 7016 if (log_is_enabled(Debug, gc, sweep)) { 7017 log_debug(gc, sweep)("Collected " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7018 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); 7019 log_debug(gc, sweep)("Live " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes Already free " SIZE_FORMAT " objects, " SIZE_FORMAT " bytes", 7020 _numObjectsLive, _numWordsLive*sizeof(HeapWord), _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); 7021 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * sizeof(HeapWord); 7022 log_debug(gc, sweep)("Total sweep: " SIZE_FORMAT " bytes", totalBytes); 7023 } 7024 7025 if (log_is_enabled(Trace, gc, sweep) && CMSVerifyReturnedBytes) { 7026 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); 7027 size_t dict_returned_bytes = _sp->dictionary()->sum_dict_returned_bytes(); 7028 size_t returned_bytes = indexListReturnedBytes + dict_returned_bytes; 7029 log_trace(gc, sweep)("Returned " SIZE_FORMAT " bytes Indexed List Returned " SIZE_FORMAT " bytes Dictionary Returned " SIZE_FORMAT " bytes", 7030 returned_bytes, indexListReturnedBytes, dict_returned_bytes); 7031 } 7032 log_develop_trace(gc, sweep)("end of sweep with _limit = " PTR_FORMAT, p2i(_limit)); 7033 log_develop_trace(gc, sweep)("================"); 7034 } 7035 #endif // PRODUCT 7036 7037 void SweepClosure::initialize_free_range(HeapWord* freeFinger, 7038 bool freeRangeInFreeLists) { 7039 log_develop_trace(gc, sweep)("---- Start free range at " PTR_FORMAT " with free block (%d)", 7040 p2i(freeFinger), freeRangeInFreeLists); 7041 assert(!inFreeRange(), "Trampling existing free range"); 7042 set_inFreeRange(true); 7043 set_lastFreeRangeCoalesced(false); 7044 7045 set_freeFinger(freeFinger); 7046 set_freeRangeInFreeLists(freeRangeInFreeLists); 7047 if (CMSTestInFreeList) { 7048 if (freeRangeInFreeLists) { 7049 FreeChunk* fc = (FreeChunk*) freeFinger; 7050 assert(fc->is_free(), "A chunk on the free list should be free."); 7051 assert(fc->size() > 0, "Free range should have a size"); 7052 assert(_sp->verify_chunk_in_free_list(fc), "Chunk is not in free lists"); 7053 } 7054 } 7055 } 7056 7057 // Note that the sweeper runs concurrently with mutators. Thus, 7058 // it is possible for direct allocation in this generation to happen 7059 // in the middle of the sweep. Note that the sweeper also coalesces 7060 // contiguous free blocks. Thus, unless the sweeper and the allocator 7061 // synchronize appropriately freshly allocated blocks may get swept up. 7062 // This is accomplished by the sweeper locking the free lists while 7063 // it is sweeping. Thus blocks that are determined to be free are 7064 // indeed free. There is however one additional complication: 7065 // blocks that have been allocated since the final checkpoint and 7066 // mark, will not have been marked and so would be treated as 7067 // unreachable and swept up. To prevent this, the allocator marks 7068 // the bit map when allocating during the sweep phase. This leads, 7069 // however, to a further complication -- objects may have been allocated 7070 // but not yet initialized -- in the sense that the header isn't yet 7071 // installed. The sweeper can not then determine the size of the block 7072 // in order to skip over it. To deal with this case, we use a technique 7073 // (due to Printezis) to encode such uninitialized block sizes in the 7074 // bit map. Since the bit map uses a bit per every HeapWord, but the 7075 // CMS generation has a minimum object size of 3 HeapWords, it follows 7076 // that "normal marks" won't be adjacent in the bit map (there will 7077 // always be at least two 0 bits between successive 1 bits). We make use 7078 // of these "unused" bits to represent uninitialized blocks -- the bit 7079 // corresponding to the start of the uninitialized object and the next 7080 // bit are both set. Finally, a 1 bit marks the end of the object that 7081 // started with the two consecutive 1 bits to indicate its potentially 7082 // uninitialized state. 7083 7084 size_t SweepClosure::do_blk_careful(HeapWord* addr) { 7085 FreeChunk* fc = (FreeChunk*)addr; 7086 size_t res; 7087 7088 // Check if we are done sweeping. Below we check "addr >= _limit" rather 7089 // than "addr == _limit" because although _limit was a block boundary when 7090 // we started the sweep, it may no longer be one because heap expansion 7091 // may have caused us to coalesce the block ending at the address _limit 7092 // with a newly expanded chunk (this happens when _limit was set to the 7093 // previous _end of the space), so we may have stepped past _limit: 7094 // see the following Zeno-like trail of CRs 6977970, 7008136, 7042740. 7095 if (addr >= _limit) { // we have swept up to or past the limit: finish up 7096 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), 7097 "sweep _limit out of bounds"); 7098 assert(addr < _sp->end(), "addr out of bounds"); 7099 // Flush any free range we might be holding as a single 7100 // coalesced chunk to the appropriate free list. 7101 if (inFreeRange()) { 7102 assert(freeFinger() >= _sp->bottom() && freeFinger() < _limit, 7103 "freeFinger() " PTR_FORMAT " is out of bounds", p2i(freeFinger())); 7104 flush_cur_free_chunk(freeFinger(), 7105 pointer_delta(addr, freeFinger())); 7106 log_develop_trace(gc, sweep)("Sweep: last chunk: put_free_blk " PTR_FORMAT " (" SIZE_FORMAT ") [coalesced:%d]", 7107 p2i(freeFinger()), pointer_delta(addr, freeFinger()), 7108 lastFreeRangeCoalesced() ? 1 : 0); 7109 } 7110 7111 // help the iterator loop finish 7112 return pointer_delta(_sp->end(), addr); 7113 } 7114 7115 assert(addr < _limit, "sweep invariant"); 7116 // check if we should yield 7117 do_yield_check(addr); 7118 if (fc->is_free()) { 7119 // Chunk that is already free 7120 res = fc->size(); 7121 do_already_free_chunk(fc); 7122 debug_only(_sp->verifyFreeLists()); 7123 // If we flush the chunk at hand in lookahead_and_flush() 7124 // and it's coalesced with a preceding chunk, then the 7125 // process of "mangling" the payload of the coalesced block 7126 // will cause erasure of the size information from the 7127 // (erstwhile) header of all the coalesced blocks but the 7128 // first, so the first disjunct in the assert will not hold 7129 // in that specific case (in which case the second disjunct 7130 // will hold). 7131 assert(res == fc->size() || ((HeapWord*)fc) + res >= _limit, 7132 "Otherwise the size info doesn't change at this step"); 7133 NOT_PRODUCT( 7134 _numObjectsAlreadyFree++; 7135 _numWordsAlreadyFree += res; 7136 ) 7137 NOT_PRODUCT(_last_fc = fc;) 7138 } else if (!_bitMap->isMarked(addr)) { 7139 // Chunk is fresh garbage 7140 res = do_garbage_chunk(fc); 7141 debug_only(_sp->verifyFreeLists()); 7142 NOT_PRODUCT( 7143 _numObjectsFreed++; 7144 _numWordsFreed += res; 7145 ) 7146 } else { 7147 // Chunk that is alive. 7148 res = do_live_chunk(fc); 7149 debug_only(_sp->verifyFreeLists()); 7150 NOT_PRODUCT( 7151 _numObjectsLive++; 7152 _numWordsLive += res; 7153 ) 7154 } 7155 return res; 7156 } 7157 7158 // For the smart allocation, record following 7159 // split deaths - a free chunk is removed from its free list because 7160 // it is being split into two or more chunks. 7161 // split birth - a free chunk is being added to its free list because 7162 // a larger free chunk has been split and resulted in this free chunk. 7163 // coal death - a free chunk is being removed from its free list because 7164 // it is being coalesced into a large free chunk. 7165 // coal birth - a free chunk is being added to its free list because 7166 // it was created when two or more free chunks where coalesced into 7167 // this free chunk. 7168 // 7169 // These statistics are used to determine the desired number of free 7170 // chunks of a given size. The desired number is chosen to be relative 7171 // to the end of a CMS sweep. The desired number at the end of a sweep 7172 // is the 7173 // count-at-end-of-previous-sweep (an amount that was enough) 7174 // - count-at-beginning-of-current-sweep (the excess) 7175 // + split-births (gains in this size during interval) 7176 // - split-deaths (demands on this size during interval) 7177 // where the interval is from the end of one sweep to the end of the 7178 // next. 7179 // 7180 // When sweeping the sweeper maintains an accumulated chunk which is 7181 // the chunk that is made up of chunks that have been coalesced. That 7182 // will be termed the left-hand chunk. A new chunk of garbage that 7183 // is being considered for coalescing will be referred to as the 7184 // right-hand chunk. 7185 // 7186 // When making a decision on whether to coalesce a right-hand chunk with 7187 // the current left-hand chunk, the current count vs. the desired count 7188 // of the left-hand chunk is considered. Also if the right-hand chunk 7189 // is near the large chunk at the end of the heap (see 7190 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the 7191 // left-hand chunk is coalesced. 7192 // 7193 // When making a decision about whether to split a chunk, the desired count 7194 // vs. the current count of the candidate to be split is also considered. 7195 // If the candidate is underpopulated (currently fewer chunks than desired) 7196 // a chunk of an overpopulated (currently more chunks than desired) size may 7197 // be chosen. The "hint" associated with a free list, if non-null, points 7198 // to a free list which may be overpopulated. 7199 // 7200 7201 void SweepClosure::do_already_free_chunk(FreeChunk* fc) { 7202 const size_t size = fc->size(); 7203 // Chunks that cannot be coalesced are not in the 7204 // free lists. 7205 if (CMSTestInFreeList && !fc->cantCoalesce()) { 7206 assert(_sp->verify_chunk_in_free_list(fc), 7207 "free chunk should be in free lists"); 7208 } 7209 // a chunk that is already free, should not have been 7210 // marked in the bit map 7211 HeapWord* const addr = (HeapWord*) fc; 7212 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); 7213 // Verify that the bit map has no bits marked between 7214 // addr and purported end of this block. 7215 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7216 7217 // Some chunks cannot be coalesced under any circumstances. 7218 // See the definition of cantCoalesce(). 7219 if (!fc->cantCoalesce()) { 7220 // This chunk can potentially be coalesced. 7221 // All the work is done in 7222 do_post_free_or_garbage_chunk(fc, size); 7223 // Note that if the chunk is not coalescable (the else arm 7224 // below), we unconditionally flush, without needing to do 7225 // a "lookahead," as we do below. 7226 if (inFreeRange()) lookahead_and_flush(fc, size); 7227 } else { 7228 // Code path common to both original and adaptive free lists. 7229 7230 // cant coalesce with previous block; this should be treated 7231 // as the end of a free run if any 7232 if (inFreeRange()) { 7233 // we kicked some butt; time to pick up the garbage 7234 assert(freeFinger() < addr, "freeFinger points too high"); 7235 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7236 } 7237 // else, nothing to do, just continue 7238 } 7239 } 7240 7241 size_t SweepClosure::do_garbage_chunk(FreeChunk* fc) { 7242 // This is a chunk of garbage. It is not in any free list. 7243 // Add it to a free list or let it possibly be coalesced into 7244 // a larger chunk. 7245 HeapWord* const addr = (HeapWord*) fc; 7246 const size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7247 7248 // Verify that the bit map has no bits marked between 7249 // addr and purported end of just dead object. 7250 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); 7251 do_post_free_or_garbage_chunk(fc, size); 7252 7253 assert(_limit >= addr + size, 7254 "A freshly garbage chunk can't possibly straddle over _limit"); 7255 if (inFreeRange()) lookahead_and_flush(fc, size); 7256 return size; 7257 } 7258 7259 size_t SweepClosure::do_live_chunk(FreeChunk* fc) { 7260 HeapWord* addr = (HeapWord*) fc; 7261 // The sweeper has just found a live object. Return any accumulated 7262 // left hand chunk to the free lists. 7263 if (inFreeRange()) { 7264 assert(freeFinger() < addr, "freeFinger points too high"); 7265 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7266 } 7267 7268 // This object is live: we'd normally expect this to be 7269 // an oop, and like to assert the following: 7270 // assert(oopDesc::is_oop(oop(addr)), "live block should be an oop"); 7271 // However, as we commented above, this may be an object whose 7272 // header hasn't yet been initialized. 7273 size_t size; 7274 assert(_bitMap->isMarked(addr), "Tautology for this control point"); 7275 if (_bitMap->isMarked(addr + 1)) { 7276 // Determine the size from the bit map, rather than trying to 7277 // compute it from the object header. 7278 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); 7279 size = pointer_delta(nextOneAddr + 1, addr); 7280 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), 7281 "alignment problem"); 7282 7283 #ifdef ASSERT 7284 if (oop(addr)->klass_or_null_acquire() != NULL) { 7285 // Ignore mark word because we are running concurrent with mutators 7286 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7287 assert(size == 7288 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), 7289 "P-mark and computed size do not agree"); 7290 } 7291 #endif 7292 7293 } else { 7294 // This should be an initialized object that's alive. 7295 assert(oop(addr)->klass_or_null_acquire() != NULL, 7296 "Should be an initialized object"); 7297 // Ignore mark word because we are running concurrent with mutators 7298 assert(oopDesc::is_oop(oop(addr), true), "live block should be an oop"); 7299 // Verify that the bit map has no bits marked between 7300 // addr and purported end of this block. 7301 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); 7302 assert(size >= 3, "Necessary for Printezis marks to work"); 7303 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); 7304 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) 7305 } 7306 return size; 7307 } 7308 7309 void SweepClosure::do_post_free_or_garbage_chunk(FreeChunk* fc, 7310 size_t chunkSize) { 7311 // do_post_free_or_garbage_chunk() should only be called in the case 7312 // of the adaptive free list allocator. 7313 const bool fcInFreeLists = fc->is_free(); 7314 assert((HeapWord*)fc <= _limit, "sweep invariant"); 7315 if (CMSTestInFreeList && fcInFreeLists) { 7316 assert(_sp->verify_chunk_in_free_list(fc), "free chunk is not in free lists"); 7317 } 7318 7319 log_develop_trace(gc, sweep)(" -- pick up another chunk at " PTR_FORMAT " (" SIZE_FORMAT ")", p2i(fc), chunkSize); 7320 7321 HeapWord* const fc_addr = (HeapWord*) fc; 7322 7323 bool coalesce = false; 7324 const size_t left = pointer_delta(fc_addr, freeFinger()); 7325 const size_t right = chunkSize; 7326 switch (FLSCoalescePolicy) { 7327 // numeric value forms a coalition aggressiveness metric 7328 case 0: { // never coalesce 7329 coalesce = false; 7330 break; 7331 } 7332 case 1: { // coalesce if left & right chunks on overpopulated lists 7333 coalesce = _sp->coalOverPopulated(left) && 7334 _sp->coalOverPopulated(right); 7335 break; 7336 } 7337 case 2: { // coalesce if left chunk on overpopulated list (default) 7338 coalesce = _sp->coalOverPopulated(left); 7339 break; 7340 } 7341 case 3: { // coalesce if left OR right chunk on overpopulated list 7342 coalesce = _sp->coalOverPopulated(left) || 7343 _sp->coalOverPopulated(right); 7344 break; 7345 } 7346 case 4: { // always coalesce 7347 coalesce = true; 7348 break; 7349 } 7350 default: 7351 ShouldNotReachHere(); 7352 } 7353 7354 // Should the current free range be coalesced? 7355 // If the chunk is in a free range and either we decided to coalesce above 7356 // or the chunk is near the large block at the end of the heap 7357 // (isNearLargestChunk() returns true), then coalesce this chunk. 7358 const bool doCoalesce = inFreeRange() 7359 && (coalesce || _g->isNearLargestChunk(fc_addr)); 7360 if (doCoalesce) { 7361 // Coalesce the current free range on the left with the new 7362 // chunk on the right. If either is on a free list, 7363 // it must be removed from the list and stashed in the closure. 7364 if (freeRangeInFreeLists()) { 7365 FreeChunk* const ffc = (FreeChunk*)freeFinger(); 7366 assert(ffc->size() == pointer_delta(fc_addr, freeFinger()), 7367 "Size of free range is inconsistent with chunk size."); 7368 if (CMSTestInFreeList) { 7369 assert(_sp->verify_chunk_in_free_list(ffc), 7370 "Chunk is not in free lists"); 7371 } 7372 _sp->coalDeath(ffc->size()); 7373 _sp->removeFreeChunkFromFreeLists(ffc); 7374 set_freeRangeInFreeLists(false); 7375 } 7376 if (fcInFreeLists) { 7377 _sp->coalDeath(chunkSize); 7378 assert(fc->size() == chunkSize, 7379 "The chunk has the wrong size or is not in the free lists"); 7380 _sp->removeFreeChunkFromFreeLists(fc); 7381 } 7382 set_lastFreeRangeCoalesced(true); 7383 print_free_block_coalesced(fc); 7384 } else { // not in a free range and/or should not coalesce 7385 // Return the current free range and start a new one. 7386 if (inFreeRange()) { 7387 // In a free range but cannot coalesce with the right hand chunk. 7388 // Put the current free range into the free lists. 7389 flush_cur_free_chunk(freeFinger(), 7390 pointer_delta(fc_addr, freeFinger())); 7391 } 7392 // Set up for new free range. Pass along whether the right hand 7393 // chunk is in the free lists. 7394 initialize_free_range((HeapWord*)fc, fcInFreeLists); 7395 } 7396 } 7397 7398 // Lookahead flush: 7399 // If we are tracking a free range, and this is the last chunk that 7400 // we'll look at because its end crosses past _limit, we'll preemptively 7401 // flush it along with any free range we may be holding on to. Note that 7402 // this can be the case only for an already free or freshly garbage 7403 // chunk. If this block is an object, it can never straddle 7404 // over _limit. The "straddling" occurs when _limit is set at 7405 // the previous end of the space when this cycle started, and 7406 // a subsequent heap expansion caused the previously co-terminal 7407 // free block to be coalesced with the newly expanded portion, 7408 // thus rendering _limit a non-block-boundary making it dangerous 7409 // for the sweeper to step over and examine. 7410 void SweepClosure::lookahead_and_flush(FreeChunk* fc, size_t chunk_size) { 7411 assert(inFreeRange(), "Should only be called if currently in a free range."); 7412 HeapWord* const eob = ((HeapWord*)fc) + chunk_size; 7413 assert(_sp->used_region().contains(eob - 1), 7414 "eob = " PTR_FORMAT " eob-1 = " PTR_FORMAT " _limit = " PTR_FORMAT 7415 " out of bounds wrt _sp = [" PTR_FORMAT "," PTR_FORMAT ")" 7416 " when examining fc = " PTR_FORMAT "(" SIZE_FORMAT ")", 7417 p2i(eob), p2i(eob-1), p2i(_limit), p2i(_sp->bottom()), p2i(_sp->end()), p2i(fc), chunk_size); 7418 if (eob >= _limit) { 7419 assert(eob == _limit || fc->is_free(), "Only a free chunk should allow us to cross over the limit"); 7420 log_develop_trace(gc, sweep)("_limit " PTR_FORMAT " reached or crossed by block " 7421 "[" PTR_FORMAT "," PTR_FORMAT ") in space " 7422 "[" PTR_FORMAT "," PTR_FORMAT ")", 7423 p2i(_limit), p2i(fc), p2i(eob), p2i(_sp->bottom()), p2i(_sp->end())); 7424 // Return the storage we are tracking back into the free lists. 7425 log_develop_trace(gc, sweep)("Flushing ... "); 7426 assert(freeFinger() < eob, "Error"); 7427 flush_cur_free_chunk( freeFinger(), pointer_delta(eob, freeFinger())); 7428 } 7429 } 7430 7431 void SweepClosure::flush_cur_free_chunk(HeapWord* chunk, size_t size) { 7432 assert(inFreeRange(), "Should only be called if currently in a free range."); 7433 assert(size > 0, 7434 "A zero sized chunk cannot be added to the free lists."); 7435 if (!freeRangeInFreeLists()) { 7436 if (CMSTestInFreeList) { 7437 FreeChunk* fc = (FreeChunk*) chunk; 7438 fc->set_size(size); 7439 assert(!_sp->verify_chunk_in_free_list(fc), 7440 "chunk should not be in free lists yet"); 7441 } 7442 log_develop_trace(gc, sweep)(" -- add free block " PTR_FORMAT " (" SIZE_FORMAT ") to free lists", p2i(chunk), size); 7443 // A new free range is going to be starting. The current 7444 // free range has not been added to the free lists yet or 7445 // was removed so add it back. 7446 // If the current free range was coalesced, then the death 7447 // of the free range was recorded. Record a birth now. 7448 if (lastFreeRangeCoalesced()) { 7449 _sp->coalBirth(size); 7450 } 7451 _sp->addChunkAndRepairOffsetTable(chunk, size, 7452 lastFreeRangeCoalesced()); 7453 } else { 7454 log_develop_trace(gc, sweep)("Already in free list: nothing to flush"); 7455 } 7456 set_inFreeRange(false); 7457 set_freeRangeInFreeLists(false); 7458 } 7459 7460 // We take a break if we've been at this for a while, 7461 // so as to avoid monopolizing the locks involved. 7462 void SweepClosure::do_yield_work(HeapWord* addr) { 7463 // Return current free chunk being used for coalescing (if any) 7464 // to the appropriate freelist. After yielding, the next 7465 // free block encountered will start a coalescing range of 7466 // free blocks. If the next free block is adjacent to the 7467 // chunk just flushed, they will need to wait for the next 7468 // sweep to be coalesced. 7469 if (inFreeRange()) { 7470 flush_cur_free_chunk(freeFinger(), pointer_delta(addr, freeFinger())); 7471 } 7472 7473 // First give up the locks, then yield, then re-lock. 7474 // We should probably use a constructor/destructor idiom to 7475 // do this unlock/lock or modify the MutexUnlocker class to 7476 // serve our purpose. XXX 7477 assert_lock_strong(_bitMap->lock()); 7478 assert_lock_strong(_freelistLock); 7479 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), 7480 "CMS thread should hold CMS token"); 7481 _bitMap->lock()->unlock(); 7482 _freelistLock->unlock(); 7483 ConcurrentMarkSweepThread::desynchronize(true); 7484 _collector->stopTimer(); 7485 _collector->incrementYields(); 7486 7487 // See the comment in coordinator_yield() 7488 for (unsigned i = 0; i < CMSYieldSleepCount && 7489 ConcurrentMarkSweepThread::should_yield() && 7490 !CMSCollector::foregroundGCIsActive(); ++i) { 7491 os::sleep(Thread::current(), 1, false); 7492 } 7493 7494 ConcurrentMarkSweepThread::synchronize(true); 7495 _freelistLock->lock(); 7496 _bitMap->lock()->lock_without_safepoint_check(); 7497 _collector->startTimer(); 7498 } 7499 7500 #ifndef PRODUCT 7501 // This is actually very useful in a product build if it can 7502 // be called from the debugger. Compile it into the product 7503 // as needed. 7504 bool debug_verify_chunk_in_free_list(FreeChunk* fc) { 7505 return debug_cms_space->verify_chunk_in_free_list(fc); 7506 } 7507 #endif 7508 7509 void SweepClosure::print_free_block_coalesced(FreeChunk* fc) const { 7510 log_develop_trace(gc, sweep)("Sweep:coal_free_blk " PTR_FORMAT " (" SIZE_FORMAT ")", 7511 p2i(fc), fc->size()); 7512 } 7513 7514 // CMSIsAliveClosure 7515 bool CMSIsAliveClosure::do_object_b(oop obj) { 7516 HeapWord* addr = (HeapWord*)obj; 7517 return addr != NULL && 7518 (!_span.contains(addr) || _bit_map->isMarked(addr)); 7519 } 7520 7521 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, 7522 MemRegion span, 7523 CMSBitMap* bit_map, CMSMarkStack* mark_stack, 7524 bool cpc): 7525 _collector(collector), 7526 _span(span), 7527 _mark_stack(mark_stack), 7528 _bit_map(bit_map), 7529 _concurrent_precleaning(cpc) { 7530 assert(!_span.is_empty(), "Empty span could spell trouble"); 7531 } 7532 7533 7534 // CMSKeepAliveClosure: the serial version 7535 void CMSKeepAliveClosure::do_oop(oop obj) { 7536 HeapWord* addr = (HeapWord*)obj; 7537 if (_span.contains(addr) && 7538 !_bit_map->isMarked(addr)) { 7539 _bit_map->mark(addr); 7540 bool simulate_overflow = false; 7541 NOT_PRODUCT( 7542 if (CMSMarkStackOverflowALot && 7543 _collector->simulate_overflow()) { 7544 // simulate a stack overflow 7545 simulate_overflow = true; 7546 } 7547 ) 7548 if (simulate_overflow || !_mark_stack->push(obj)) { 7549 if (_concurrent_precleaning) { 7550 // We dirty the overflown object and let the remark 7551 // phase deal with it. 7552 assert(_collector->overflow_list_is_empty(), "Error"); 7553 // In the case of object arrays, we need to dirty all of 7554 // the cards that the object spans. No locking or atomics 7555 // are needed since no one else can be mutating the mod union 7556 // table. 7557 if (obj->is_objArray()) { 7558 size_t sz = obj->size(); 7559 HeapWord* end_card_addr = align_up(addr + sz, CardTable::card_size); 7560 MemRegion redirty_range = MemRegion(addr, end_card_addr); 7561 assert(!redirty_range.is_empty(), "Arithmetical tautology"); 7562 _collector->_modUnionTable.mark_range(redirty_range); 7563 } else { 7564 _collector->_modUnionTable.mark(addr); 7565 } 7566 _collector->_ser_kac_preclean_ovflw++; 7567 } else { 7568 _collector->push_on_overflow_list(obj); 7569 _collector->_ser_kac_ovflw++; 7570 } 7571 } 7572 } 7573 } 7574 7575 // CMSParKeepAliveClosure: a parallel version of the above. 7576 // The work queues are private to each closure (thread), 7577 // but (may be) available for stealing by other threads. 7578 void CMSParKeepAliveClosure::do_oop(oop obj) { 7579 HeapWord* addr = (HeapWord*)obj; 7580 if (_span.contains(addr) && 7581 !_bit_map->isMarked(addr)) { 7582 // In general, during recursive tracing, several threads 7583 // may be concurrently getting here; the first one to 7584 // "tag" it, claims it. 7585 if (_bit_map->par_mark(addr)) { 7586 bool res = _work_queue->push(obj); 7587 assert(res, "Low water mark should be much less than capacity"); 7588 // Do a recursive trim in the hope that this will keep 7589 // stack usage lower, but leave some oops for potential stealers 7590 trim_queue(_low_water_mark); 7591 } // Else, another thread got there first 7592 } 7593 } 7594 7595 void CMSParKeepAliveClosure::trim_queue(uint max) { 7596 while (_work_queue->size() > max) { 7597 oop new_oop; 7598 if (_work_queue->pop_local(new_oop)) { 7599 assert(new_oop != NULL && oopDesc::is_oop(new_oop), "Expected an oop"); 7600 assert(_bit_map->isMarked((HeapWord*)new_oop), 7601 "no white objects on this stack!"); 7602 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7603 // iterate over the oops in this oop, marking and pushing 7604 // the ones in CMS heap (i.e. in _span). 7605 new_oop->oop_iterate(&_mark_and_push); 7606 } 7607 } 7608 } 7609 7610 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( 7611 CMSCollector* collector, 7612 MemRegion span, CMSBitMap* bit_map, 7613 OopTaskQueue* work_queue): 7614 _collector(collector), 7615 _span(span), 7616 _work_queue(work_queue), 7617 _bit_map(bit_map) { } 7618 7619 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { 7620 HeapWord* addr = (HeapWord*)obj; 7621 if (_span.contains(addr) && 7622 !_bit_map->isMarked(addr)) { 7623 if (_bit_map->par_mark(addr)) { 7624 bool simulate_overflow = false; 7625 NOT_PRODUCT( 7626 if (CMSMarkStackOverflowALot && 7627 _collector->par_simulate_overflow()) { 7628 // simulate a stack overflow 7629 simulate_overflow = true; 7630 } 7631 ) 7632 if (simulate_overflow || !_work_queue->push(obj)) { 7633 _collector->par_push_on_overflow_list(obj); 7634 _collector->_par_kac_ovflw++; 7635 } 7636 } // Else another thread got there already 7637 } 7638 } 7639 7640 ////////////////////////////////////////////////////////////////// 7641 // CMSExpansionCause ///////////////////////////// 7642 ////////////////////////////////////////////////////////////////// 7643 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { 7644 switch (cause) { 7645 case _no_expansion: 7646 return "No expansion"; 7647 case _satisfy_free_ratio: 7648 return "Free ratio"; 7649 case _satisfy_promotion: 7650 return "Satisfy promotion"; 7651 case _satisfy_allocation: 7652 return "allocation"; 7653 case _allocate_par_lab: 7654 return "Par LAB"; 7655 case _allocate_par_spooling_space: 7656 return "Par Spooling Space"; 7657 case _adaptive_size_policy: 7658 return "Ergonomics"; 7659 default: 7660 return "unknown"; 7661 } 7662 } 7663 7664 void CMSDrainMarkingStackClosure::do_void() { 7665 // the max number to take from overflow list at a time 7666 const size_t num = _mark_stack->capacity()/4; 7667 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), 7668 "Overflow list should be NULL during concurrent phases"); 7669 while (!_mark_stack->isEmpty() || 7670 // if stack is empty, check the overflow list 7671 _collector->take_from_overflow_list(num, _mark_stack)) { 7672 oop obj = _mark_stack->pop(); 7673 HeapWord* addr = (HeapWord*)obj; 7674 assert(_span.contains(addr), "Should be within span"); 7675 assert(_bit_map->isMarked(addr), "Should be marked"); 7676 assert(oopDesc::is_oop(obj), "Should be an oop"); 7677 obj->oop_iterate(_keep_alive); 7678 } 7679 } 7680 7681 void CMSParDrainMarkingStackClosure::do_void() { 7682 // drain queue 7683 trim_queue(0); 7684 } 7685 7686 // Trim our work_queue so its length is below max at return 7687 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { 7688 while (_work_queue->size() > max) { 7689 oop new_oop; 7690 if (_work_queue->pop_local(new_oop)) { 7691 assert(oopDesc::is_oop(new_oop), "Expected an oop"); 7692 assert(_bit_map->isMarked((HeapWord*)new_oop), 7693 "no white objects on this stack!"); 7694 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); 7695 // iterate over the oops in this oop, marking and pushing 7696 // the ones in CMS heap (i.e. in _span). 7697 new_oop->oop_iterate(&_mark_and_push); 7698 } 7699 } 7700 } 7701 7702 //////////////////////////////////////////////////////////////////// 7703 // Support for Marking Stack Overflow list handling and related code 7704 //////////////////////////////////////////////////////////////////// 7705 // Much of the following code is similar in shape and spirit to the 7706 // code used in ParNewGC. We should try and share that code 7707 // as much as possible in the future. 7708 7709 #ifndef PRODUCT 7710 // Debugging support for CMSStackOverflowALot 7711 7712 // It's OK to call this multi-threaded; the worst thing 7713 // that can happen is that we'll get a bunch of closely 7714 // spaced simulated overflows, but that's OK, in fact 7715 // probably good as it would exercise the overflow code 7716 // under contention. 7717 bool CMSCollector::simulate_overflow() { 7718 if (_overflow_counter-- <= 0) { // just being defensive 7719 _overflow_counter = CMSMarkStackOverflowInterval; 7720 return true; 7721 } else { 7722 return false; 7723 } 7724 } 7725 7726 bool CMSCollector::par_simulate_overflow() { 7727 return simulate_overflow(); 7728 } 7729 #endif 7730 7731 // Single-threaded 7732 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { 7733 assert(stack->isEmpty(), "Expected precondition"); 7734 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); 7735 size_t i = num; 7736 oop cur = _overflow_list; 7737 const markOop proto = markOopDesc::prototype(); 7738 NOT_PRODUCT(ssize_t n = 0;) 7739 for (oop next; i > 0 && cur != NULL; cur = next, i--) { 7740 next = oop(cur->mark_raw()); 7741 cur->set_mark_raw(proto); // until proven otherwise 7742 assert(oopDesc::is_oop(cur), "Should be an oop"); 7743 bool res = stack->push(cur); 7744 assert(res, "Bit off more than can chew?"); 7745 NOT_PRODUCT(n++;) 7746 } 7747 _overflow_list = cur; 7748 #ifndef PRODUCT 7749 assert(_num_par_pushes >= n, "Too many pops?"); 7750 _num_par_pushes -=n; 7751 #endif 7752 return !stack->isEmpty(); 7753 } 7754 7755 #define BUSY (cast_to_oop<intptr_t>(0x1aff1aff)) 7756 // (MT-safe) Get a prefix of at most "num" from the list. 7757 // The overflow list is chained through the mark word of 7758 // each object in the list. We fetch the entire list, 7759 // break off a prefix of the right size and return the 7760 // remainder. If other threads try to take objects from 7761 // the overflow list at that time, they will wait for 7762 // some time to see if data becomes available. If (and 7763 // only if) another thread places one or more object(s) 7764 // on the global list before we have returned the suffix 7765 // to the global list, we will walk down our local list 7766 // to find its end and append the global list to 7767 // our suffix before returning it. This suffix walk can 7768 // prove to be expensive (quadratic in the amount of traffic) 7769 // when there are many objects in the overflow list and 7770 // there is much producer-consumer contention on the list. 7771 // *NOTE*: The overflow list manipulation code here and 7772 // in ParNewGeneration:: are very similar in shape, 7773 // except that in the ParNew case we use the old (from/eden) 7774 // copy of the object to thread the list via its klass word. 7775 // Because of the common code, if you make any changes in 7776 // the code below, please check the ParNew version to see if 7777 // similar changes might be needed. 7778 // CR 6797058 has been filed to consolidate the common code. 7779 bool CMSCollector::par_take_from_overflow_list(size_t num, 7780 OopTaskQueue* work_q, 7781 int no_of_gc_threads) { 7782 assert(work_q->size() == 0, "First empty local work queue"); 7783 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); 7784 if (_overflow_list == NULL) { 7785 return false; 7786 } 7787 // Grab the entire list; we'll put back a suffix 7788 oop prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7789 Thread* tid = Thread::current(); 7790 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was 7791 // set to ParallelGCThreads. 7792 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; 7793 size_t sleep_time_millis = MAX2((size_t)1, num/100); 7794 // If the list is busy, we spin for a short while, 7795 // sleeping between attempts to get the list. 7796 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { 7797 os::sleep(tid, sleep_time_millis, false); 7798 if (_overflow_list == NULL) { 7799 // Nothing left to take 7800 return false; 7801 } else if (_overflow_list != BUSY) { 7802 // Try and grab the prefix 7803 prefix = cast_to_oop(Atomic::xchg((oopDesc*)BUSY, &_overflow_list)); 7804 } 7805 } 7806 // If the list was found to be empty, or we spun long 7807 // enough, we give up and return empty-handed. If we leave 7808 // the list in the BUSY state below, it must be the case that 7809 // some other thread holds the overflow list and will set it 7810 // to a non-BUSY state in the future. 7811 if (prefix == NULL || prefix == BUSY) { 7812 // Nothing to take or waited long enough 7813 if (prefix == NULL) { 7814 // Write back the NULL in case we overwrote it with BUSY above 7815 // and it is still the same value. 7816 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7817 } 7818 return false; 7819 } 7820 assert(prefix != NULL && prefix != BUSY, "Error"); 7821 size_t i = num; 7822 oop cur = prefix; 7823 // Walk down the first "num" objects, unless we reach the end. 7824 for (; i > 1 && cur->mark_raw() != NULL; cur = oop(cur->mark_raw()), i--); 7825 if (cur->mark_raw() == NULL) { 7826 // We have "num" or fewer elements in the list, so there 7827 // is nothing to return to the global list. 7828 // Write back the NULL in lieu of the BUSY we wrote 7829 // above, if it is still the same value. 7830 if (_overflow_list == BUSY) { 7831 Atomic::cmpxchg((oopDesc*)NULL, &_overflow_list, (oopDesc*)BUSY); 7832 } 7833 } else { 7834 // Chop off the suffix and return it to the global list. 7835 assert(cur->mark_raw() != BUSY, "Error"); 7836 oop suffix_head = cur->mark_raw(); // suffix will be put back on global list 7837 cur->set_mark_raw(NULL); // break off suffix 7838 // It's possible that the list is still in the empty(busy) state 7839 // we left it in a short while ago; in that case we may be 7840 // able to place back the suffix without incurring the cost 7841 // of a walk down the list. 7842 oop observed_overflow_list = _overflow_list; 7843 oop cur_overflow_list = observed_overflow_list; 7844 bool attached = false; 7845 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { 7846 observed_overflow_list = 7847 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7848 if (cur_overflow_list == observed_overflow_list) { 7849 attached = true; 7850 break; 7851 } else cur_overflow_list = observed_overflow_list; 7852 } 7853 if (!attached) { 7854 // Too bad, someone else sneaked in (at least) an element; we'll need 7855 // to do a splice. Find tail of suffix so we can prepend suffix to global 7856 // list. 7857 for (cur = suffix_head; cur->mark_raw() != NULL; cur = (oop)(cur->mark_raw())); 7858 oop suffix_tail = cur; 7859 assert(suffix_tail != NULL && suffix_tail->mark_raw() == NULL, 7860 "Tautology"); 7861 observed_overflow_list = _overflow_list; 7862 do { 7863 cur_overflow_list = observed_overflow_list; 7864 if (cur_overflow_list != BUSY) { 7865 // Do the splice ... 7866 suffix_tail->set_mark_raw(markOop(cur_overflow_list)); 7867 } else { // cur_overflow_list == BUSY 7868 suffix_tail->set_mark_raw(NULL); 7869 } 7870 // ... and try to place spliced list back on overflow_list ... 7871 observed_overflow_list = 7872 Atomic::cmpxchg((oopDesc*)suffix_head, &_overflow_list, (oopDesc*)cur_overflow_list); 7873 } while (cur_overflow_list != observed_overflow_list); 7874 // ... until we have succeeded in doing so. 7875 } 7876 } 7877 7878 // Push the prefix elements on work_q 7879 assert(prefix != NULL, "control point invariant"); 7880 const markOop proto = markOopDesc::prototype(); 7881 oop next; 7882 NOT_PRODUCT(ssize_t n = 0;) 7883 for (cur = prefix; cur != NULL; cur = next) { 7884 next = oop(cur->mark_raw()); 7885 cur->set_mark_raw(proto); // until proven otherwise 7886 assert(oopDesc::is_oop(cur), "Should be an oop"); 7887 bool res = work_q->push(cur); 7888 assert(res, "Bit off more than we can chew?"); 7889 NOT_PRODUCT(n++;) 7890 } 7891 #ifndef PRODUCT 7892 assert(_num_par_pushes >= n, "Too many pops?"); 7893 Atomic::sub(n, &_num_par_pushes); 7894 #endif 7895 return true; 7896 } 7897 7898 // Single-threaded 7899 void CMSCollector::push_on_overflow_list(oop p) { 7900 NOT_PRODUCT(_num_par_pushes++;) 7901 assert(oopDesc::is_oop(p), "Not an oop"); 7902 preserve_mark_if_necessary(p); 7903 p->set_mark_raw((markOop)_overflow_list); 7904 _overflow_list = p; 7905 } 7906 7907 // Multi-threaded; use CAS to prepend to overflow list 7908 void CMSCollector::par_push_on_overflow_list(oop p) { 7909 NOT_PRODUCT(Atomic::inc(&_num_par_pushes);) 7910 assert(oopDesc::is_oop(p), "Not an oop"); 7911 par_preserve_mark_if_necessary(p); 7912 oop observed_overflow_list = _overflow_list; 7913 oop cur_overflow_list; 7914 do { 7915 cur_overflow_list = observed_overflow_list; 7916 if (cur_overflow_list != BUSY) { 7917 p->set_mark_raw(markOop(cur_overflow_list)); 7918 } else { 7919 p->set_mark_raw(NULL); 7920 } 7921 observed_overflow_list = 7922 Atomic::cmpxchg((oopDesc*)p, &_overflow_list, (oopDesc*)cur_overflow_list); 7923 } while (cur_overflow_list != observed_overflow_list); 7924 } 7925 #undef BUSY 7926 7927 // Single threaded 7928 // General Note on GrowableArray: pushes may silently fail 7929 // because we are (temporarily) out of C-heap for expanding 7930 // the stack. The problem is quite ubiquitous and affects 7931 // a lot of code in the JVM. The prudent thing for GrowableArray 7932 // to do (for now) is to exit with an error. However, that may 7933 // be too draconian in some cases because the caller may be 7934 // able to recover without much harm. For such cases, we 7935 // should probably introduce a "soft_push" method which returns 7936 // an indication of success or failure with the assumption that 7937 // the caller may be able to recover from a failure; code in 7938 // the VM can then be changed, incrementally, to deal with such 7939 // failures where possible, thus, incrementally hardening the VM 7940 // in such low resource situations. 7941 void CMSCollector::preserve_mark_work(oop p, markOop m) { 7942 _preserved_oop_stack.push(p); 7943 _preserved_mark_stack.push(m); 7944 assert(m == p->mark_raw(), "Mark word changed"); 7945 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7946 "bijection"); 7947 } 7948 7949 // Single threaded 7950 void CMSCollector::preserve_mark_if_necessary(oop p) { 7951 markOop m = p->mark_raw(); 7952 if (m->must_be_preserved(p)) { 7953 preserve_mark_work(p, m); 7954 } 7955 } 7956 7957 void CMSCollector::par_preserve_mark_if_necessary(oop p) { 7958 markOop m = p->mark_raw(); 7959 if (m->must_be_preserved(p)) { 7960 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 7961 // Even though we read the mark word without holding 7962 // the lock, we are assured that it will not change 7963 // because we "own" this oop, so no other thread can 7964 // be trying to push it on the overflow list; see 7965 // the assertion in preserve_mark_work() that checks 7966 // that m == p->mark_raw(). 7967 preserve_mark_work(p, m); 7968 } 7969 } 7970 7971 // We should be able to do this multi-threaded, 7972 // a chunk of stack being a task (this is 7973 // correct because each oop only ever appears 7974 // once in the overflow list. However, it's 7975 // not very easy to completely overlap this with 7976 // other operations, so will generally not be done 7977 // until all work's been completed. Because we 7978 // expect the preserved oop stack (set) to be small, 7979 // it's probably fine to do this single-threaded. 7980 // We can explore cleverer concurrent/overlapped/parallel 7981 // processing of preserved marks if we feel the 7982 // need for this in the future. Stack overflow should 7983 // be so rare in practice and, when it happens, its 7984 // effect on performance so great that this will 7985 // likely just be in the noise anyway. 7986 void CMSCollector::restore_preserved_marks_if_any() { 7987 assert(SafepointSynchronize::is_at_safepoint(), 7988 "world should be stopped"); 7989 assert(Thread::current()->is_ConcurrentGC_thread() || 7990 Thread::current()->is_VM_thread(), 7991 "should be single-threaded"); 7992 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), 7993 "bijection"); 7994 7995 while (!_preserved_oop_stack.is_empty()) { 7996 oop p = _preserved_oop_stack.pop(); 7997 assert(oopDesc::is_oop(p), "Should be an oop"); 7998 assert(_span.contains(p), "oop should be in _span"); 7999 assert(p->mark_raw() == markOopDesc::prototype(), 8000 "Set when taken from overflow list"); 8001 markOop m = _preserved_mark_stack.pop(); 8002 p->set_mark_raw(m); 8003 } 8004 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), 8005 "stacks were cleared above"); 8006 } 8007 8008 #ifndef PRODUCT 8009 bool CMSCollector::no_preserved_marks() const { 8010 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); 8011 } 8012 #endif 8013 8014 // Transfer some number of overflown objects to usual marking 8015 // stack. Return true if some objects were transferred. 8016 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { 8017 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, 8018 (size_t)ParGCDesiredObjsFromOverflowList); 8019 8020 bool res = _collector->take_from_overflow_list(num, _mark_stack); 8021 assert(_collector->overflow_list_is_empty() || res, 8022 "If list is not empty, we should have taken something"); 8023 assert(!res || !_mark_stack->isEmpty(), 8024 "If we took something, it should now be on our stack"); 8025 return res; 8026 } 8027 8028 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { 8029 size_t res = _sp->block_size_no_stall(addr, _collector); 8030 if (_sp->block_is_obj(addr)) { 8031 if (_live_bit_map->isMarked(addr)) { 8032 // It can't have been dead in a previous cycle 8033 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); 8034 } else { 8035 _dead_bit_map->mark(addr); // mark the dead object 8036 } 8037 } 8038 // Could be 0, if the block size could not be computed without stalling. 8039 return res; 8040 } 8041 8042 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase, GCCause::Cause cause): TraceMemoryManagerStats() { 8043 GCMemoryManager* manager = CMSHeap::heap()->old_manager(); 8044 switch (phase) { 8045 case CMSCollector::InitialMarking: 8046 initialize(manager /* GC manager */ , 8047 cause /* cause of the GC */, 8048 true /* allMemoryPoolsAffected */, 8049 true /* recordGCBeginTime */, 8050 true /* recordPreGCUsage */, 8051 false /* recordPeakUsage */, 8052 false /* recordPostGCusage */, 8053 true /* recordAccumulatedGCTime */, 8054 false /* recordGCEndTime */, 8055 false /* countCollection */ ); 8056 break; 8057 8058 case CMSCollector::FinalMarking: 8059 initialize(manager /* GC manager */ , 8060 cause /* cause of the GC */, 8061 true /* allMemoryPoolsAffected */, 8062 false /* recordGCBeginTime */, 8063 false /* recordPreGCUsage */, 8064 false /* recordPeakUsage */, 8065 false /* recordPostGCusage */, 8066 true /* recordAccumulatedGCTime */, 8067 false /* recordGCEndTime */, 8068 false /* countCollection */ ); 8069 break; 8070 8071 case CMSCollector::Sweeping: 8072 initialize(manager /* GC manager */ , 8073 cause /* cause of the GC */, 8074 true /* allMemoryPoolsAffected */, 8075 false /* recordGCBeginTime */, 8076 false /* recordPreGCUsage */, 8077 true /* recordPeakUsage */, 8078 true /* recordPostGCusage */, 8079 false /* recordAccumulatedGCTime */, 8080 true /* recordGCEndTime */, 8081 true /* countCollection */ ); 8082 break; 8083 8084 default: 8085 ShouldNotReachHere(); 8086 } 8087 }